Add files using upload-large-folder tool

example/timerpo_trainer/run_anomseer.sh

@@ -0,0 +1,63 @@
+#!/bin/bash
+
+MODEL_PATH=Qwen/Qwen2.5-VL-3B-Instruct
+# MODEL_PATH=Qwen/Qwen2.5-VL-7B-Instruct
+EVAL=False
+
+TRAIN_FILE=${TRAIN_FILE:-./data/anol_processed_mllm_data/train_full.parquet}
+VAL_FILE=${VAL_FILE:-./data/anol_processed_mllm_data/test_full.parquet}
+
+if [ "$EVAL" = "True" ]; then
+    VAL_ONLY=True
+else
+    VAL_ONLY=False
+fi
+
+python3 -m verl.trainer.main_ppo \
+    algorithm.adv_estimator=grpo \
+    data.train_files=$TRAIN_FILE \
+    data.val_files=$VAL_FILE \
+    data.train_batch_size=128 \
+    data.max_prompt_length=1024 \
+    data.max_response_length=1024 \
+    data.filter_overlong_prompts=True \
+    data.truncation='error' \
+    data.image_key=images \
+    actor_rollout_ref.model.path=$MODEL_PATH \
+    actor_rollout_ref.actor.optim.lr=1e-6 \
+    actor_rollout_ref.model.use_remove_padding=True \
+    actor_rollout_ref.actor.ppo_mini_batch_size=128 \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=8 \
+    actor_rollout_ref.actor.use_kl_loss=True \
+    actor_rollout_ref.actor.kl_loss_coef=0.001 \
+    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
+    actor_rollout_ref.model.enable_gradient_checkpointing=True \
+    actor_rollout_ref.actor.fsdp_config.param_offload=False \
+    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
+    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \
+    actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+    actor_rollout_ref.rollout.name=vllm \
+    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
+    actor_rollout_ref.rollout.enable_chunked_prefill=False \
+    actor_rollout_ref.rollout.enforce_eager=False \
+    actor_rollout_ref.rollout.free_cache_engine=False \
+    actor_rollout_ref.rollout.n=5 \
+    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \
+    actor_rollout_ref.ref.fsdp_config.param_offload=True \
+    algorithm.kl_ctrl.kl_coef=0.001 \
+    trainer.critic_warmup=0 \
+    trainer.logger=['console','wandb'] \
+    trainer.project_name='anomseer' \
+    trainer.experiment_name='anomseer_timerpo' \
+    trainer.n_gpus_per_node=4 \
+    trainer.nnodes=1 \
+    trainer.save_freq=500 \
+    trainer.test_freq=10 \
+    trainer.val_only=$VAL_ONLY \
+    trainer.val_before_train=True \
+    trainer.total_epochs=10 \
+    ts.use_sem_orth=True \
+    ts.adv_mix=0.3 \
+    ts.similarity_method=ot \
+    ts.ot_eps=0.08 \
+    ts.ot_n_iter=50 $@

example/timerpo_trainer/run_anomseer_rats.sh

@@ -0,0 +1,73 @@
+#!/bin/bash
+# AnomSeer (TimerPO) on the Time-RA RATs-Uni univariate dataset.
+# Task: 15-class anomaly classification + reasoning (no interval localization).
+#
+# Prepare data first:
+#   python multimodal_data_processing/rats_uni.py \
+#       --json_path /path/to/RATs40K/RATs-Uni-TSImage_Reason.json \
+#       --out_dir   ./data/rats_uni_processed
+#
+# Train:           bash example/timerpo_trainer/run_anomseer_rats.sh
+# Eval (val only): EVAL=True bash example/timerpo_trainer/run_anomseer_rats.sh
+
+MODEL_PATH=${MODEL_PATH:-Qwen/Qwen2.5-VL-3B-Instruct}
+# MODEL_PATH=Qwen/Qwen2.5-VL-7B-Instruct
+EVAL=${EVAL:-False}
+
+TRAIN_FILE=${TRAIN_FILE:-./data/rats_uni_processed/train_full.parquet}
+VAL_FILE=${VAL_FILE:-./data/rats_uni_processed/test_full.parquet}
+
+if [ "$EVAL" = "True" ]; then
+    VAL_ONLY=True
+else
+    VAL_ONLY=False
+fi
+
+python3 -m verl.trainer.main_ppo \
+    algorithm.adv_estimator=grpo \
+    data.train_files=$TRAIN_FILE \
+    data.val_files=$VAL_FILE \
+    data.train_batch_size=128 \
+    data.max_prompt_length=1024 \
+    data.max_response_length=1024 \
+    data.filter_overlong_prompts=True \
+    data.truncation='error' \
+    data.image_key=images \
+    actor_rollout_ref.model.path=$MODEL_PATH \
+    actor_rollout_ref.actor.optim.lr=1e-6 \
+    actor_rollout_ref.model.use_remove_padding=True \
+    actor_rollout_ref.actor.ppo_mini_batch_size=128 \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=8 \
+    actor_rollout_ref.actor.use_kl_loss=True \
+    actor_rollout_ref.actor.kl_loss_coef=0.001 \
+    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
+    actor_rollout_ref.model.enable_gradient_checkpointing=True \
+    actor_rollout_ref.actor.fsdp_config.param_offload=False \
+    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
+    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \
+    actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+    actor_rollout_ref.rollout.name=vllm \
+    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
+    actor_rollout_ref.rollout.enable_chunked_prefill=False \
+    actor_rollout_ref.rollout.enforce_eager=False \
+    actor_rollout_ref.rollout.free_cache_engine=False \
+    actor_rollout_ref.rollout.n=5 \
+    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \
+    actor_rollout_ref.ref.fsdp_config.param_offload=True \
+    algorithm.kl_ctrl.kl_coef=0.001 \
+    trainer.critic_warmup=0 \
+    trainer.logger=['console','wandb'] \
+    trainer.project_name='anomseer' \
+    trainer.experiment_name='anomseer_rats_uni' \
+    trainer.n_gpus_per_node=4 \
+    trainer.nnodes=1 \
+    trainer.save_freq=500 \
+    trainer.test_freq=10 \
+    trainer.val_only=$VAL_ONLY \
+    trainer.val_before_train=True \
+    trainer.total_epochs=10 \
+    ts.use_sem_orth=True \
+    ts.adv_mix=0.3 \
+    ts.similarity_method=ot \
+    ts.ot_eps=0.08 \
+    ts.ot_n_iter=50 $@

example/timerpo_trainer/run_rats_2gpu.sh

@@ -0,0 +1,263 @@
+#!/bin/bash
+# =============================================================================
+# AnomSeer / TimerPO on Time-RA RATs-Uni — 2-GPU train + eval launcher.
+#
+# Task : 15-class anomaly classification + reasoning (no localization).
+# Tuning: LoRA fine-tuning (GRPO) by default.
+#
+# Modes (STAGE):
+#   train       : train only (with periodic in-training validation)   [default]
+#   eval        : evaluate MODEL_PATH only (val_only)
+#   train_eval  : train -> auto-merge latest checkpoint -> evaluate it
+#
+# Examples:
+#   bash example/timerpo_trainer/run_rats_2gpu.sh                      # train
+#   STAGE=eval MODEL_PATH=Qwen/Qwen2.5-VL-3B-Instruct \
+#       bash example/timerpo_trainer/run_rats_2gpu.sh                  # zero-shot eval
+#   STAGE=train_eval bash example/timerpo_trainer/run_rats_2gpu.sh     # train then eval
+# =============================================================================
+set -euo pipefail
+
+# ---- scratch on /dev/shm (NOT /tmp) ----------------------------------------
+# The CIFS share (/mnt/share01) cannot host Ray's unix sockets, and the local root
+# disk (/, where /tmp and ~/.cache live) is ~100% full. So Ray's runtime/spill and
+# the torch/triton/vLLM/HF caches go to /dev/shm: local tmpfs, supports sockets,
+# ~120 GB free. Persistent outputs (checkpoints, logs) still go under the project.
+SCRATCH_ROOT="${SCRATCH_ROOT:-/dev/shm/anomseer}"
+export RAY_TMPDIR="${RAY_TMPDIR:-${SCRATCH_ROOT}/ray}"
+export TMPDIR="${TMPDIR:-${SCRATCH_ROOT}/tmp}"
+export XDG_CACHE_HOME="${XDG_CACHE_HOME:-${SCRATCH_ROOT}/cache}"
+export TORCHINDUCTOR_CACHE_DIR="${TORCHINDUCTOR_CACHE_DIR:-${SCRATCH_ROOT}/cache/torchinductor}"
+export TRITON_CACHE_DIR="${TRITON_CACHE_DIR:-${SCRATCH_ROOT}/cache/triton}"
+export VLLM_CACHE_ROOT="${VLLM_CACHE_ROOT:-${SCRATCH_ROOT}/cache/vllm}"
+export HF_HOME="${HF_HOME:-${SCRATCH_ROOT}/cache/hf}"
+mkdir -p "$RAY_TMPDIR" "$TMPDIR" "$XDG_CACHE_HOME" "$TORCHINDUCTOR_CACHE_DIR" \
+         "$TRITON_CACHE_DIR" "$VLLM_CACHE_ROOT" "$HF_HOME"
+
+# ---- knobs (override via env) ----------------------------------------------
+PYTHON_BIN="${PYTHON_BIN:-/home/suiqk/anaconda3/envs/scalerag-ts-v4/bin/python}"
+MODEL_PATH=${MODEL_PATH:-/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct}   # local 3B; set to the 7B path / HF id to switch
+STAGE=${STAGE:-train}                                   # train | eval | train_eval
+N_GPUS=${N_GPUS:-2}                                     # number of GPUs (this script targets 2)
+# TP = rollout tensor-parallel size; must divide N_GPUS. Default 2 splits vLLM's model
+# across both cards (safer on memory). Set TP=1 for data-parallel rollout (full replica
+# per card, no cross-card TP comm — faster on these no-NVLink cards; fine for 3B on 48 GB).
+TP=${TP:-2}
+GPU_MEM_UTIL=${GPU_MEM_UTIL:-0.4}                       # vllm KV-cache fraction (lower if OOM)
+TRAIN_BATCH=${TRAIN_BATCH:-16}                         # prompts fetched per training step (rollout batch)
+MICRO_BSZ=${MICRO_BSZ:-2}                               # actor ppo micro batch per GPU (lower if OOM)
+LOGP_MICRO_BSZ=${LOGP_MICRO_BSZ:-8}                     # log-prob micro batch per GPU
+PARAM_OFFLOAD=${PARAM_OFFLOAD:-False}                   # set True to offload actor params (saves VRAM)
+OPTIM_OFFLOAD=${OPTIM_OFFLOAD:-False}                   # set True to offload optimizer (saves VRAM)
+MAX_RESPONSE_LENGTH=${MAX_RESPONSE_LENGTH:-384}         # covers the expert explanations without runaway output
+KL_COEF=${KL_COEF:-0.01}                                # resist reward-driven mode collapse
+# remove-padding (rmpad) is a throughput optimization that monkey-patches Qwen2VL
+# FlashAttention2 internals. Those classes were removed in transformers>=4.52, so it
+# must stay False on this env (transformers 4.54.1). Set True only on transformers<=4.51.1.
+USE_RMPAD=${USE_RMPAD:-False}
+EPOCHS=${EPOCHS:-1}
+
+# ---- LoRA (optional) -------------------------------------------------------
+# LORA_RANK=16 (default) -> LoRA fine-tuning.
+# LORA_RANK=0            -> full-parameter fine-tuning.
+LORA_RANK=${LORA_RANK:-16}
+LORA_ALPHA=${LORA_ALPHA:-16}
+LORA_DROPOUT=${LORA_DROPOUT:-0.0}
+# NOTE: PEFT 'all-linear' tries to wrap whole Qwen2_5_VLVisionBlock modules and errors on
+# this VL model, so default to the explicit LLM (+vision MLP) projection names instead.
+LORA_TARGET_MODULES=${LORA_TARGET_MODULES:-q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj}
+
+# LoRA usually wants a larger LR than full fine-tuning; default accordingly (override with LR=).
+if [ "$LORA_RANK" -gt 0 ]; then
+  LR=${LR:-1e-5}
+else
+  LR=${LR:-1e-6}
+fi
+LOGGER=${LOGGER:-console}                               # 'console' or "console','wandb" (needs wandb login)
+
+PROJECT=${PROJECT:-anomseer}
+EXP=${EXP:-anomseer_rats_uni_2gpu}
+CKPT_ROOT=${CKPT_ROOT:-checkpoints/${PROJECT}/${EXP}}
+
+# Default trains on a random 1/4 subset (7566 samples) for faster runs; override with
+# TRAIN_FILE=./data/rats_uni_processed/train_full.parquet for the full 30266.
+TRAIN_FILE=${TRAIN_FILE:-./data/rats_uni_processed/train_quarter.parquet}
+# verl sends the WHOLE val set to vLLM in one batch, so the full 6034-sample test set
+# OOMs system RAM. Use a stratified 474-sample subset for in-training validation; for a
+# final full-test eval pass VAL_FILE=./data/rats_uni_processed/test_full.parquet.
+VAL_FILE=${VAL_FILE:-./data/rats_uni_processed/test_small.parquet}
+
+# ---- core launcher ---------------------------------------------------------
+# args: $1 = val_only (True/False)   $2 = model path
+run_verl () {
+  local VAL_ONLY="$1"; local MPATH="$2"
+  local ACTIVE_LORA_RANK="$LORA_RANK"
+  local RESUME_MODE="auto"
+  if [ "$VAL_ONLY" = "True" ]; then
+    # Evaluation consumes a complete HF model, including any already-merged adapter.
+    ACTIVE_LORA_RANK=0
+    RESUME_MODE="disable"
+  fi
+  "$PYTHON_BIN" -m verl.trainer.main_ppo \
+      algorithm.adv_estimator=grpo \
+      data.train_files="$TRAIN_FILE" \
+      data.val_files="$VAL_FILE" \
+      data.train_batch_size="$TRAIN_BATCH" \
+      data.max_prompt_length=1024 \
+      data.max_response_length="$MAX_RESPONSE_LENGTH" \
+      data.filter_overlong_prompts=True \
+      data.truncation='error' \
+      data.image_key=images \
+      actor_rollout_ref.model.path="$MPATH" \
+      actor_rollout_ref.model.lora_rank="$ACTIVE_LORA_RANK" \
+      actor_rollout_ref.model.lora_alpha="$LORA_ALPHA" \
+      actor_rollout_ref.model.lora_dropout="$LORA_DROPOUT" \
+      actor_rollout_ref.model.lora_target_modules="'$LORA_TARGET_MODULES'" \
+      actor_rollout_ref.actor.optim.lr="$LR" \
+      actor_rollout_ref.model.use_remove_padding="$USE_RMPAD" \
+      actor_rollout_ref.actor.ppo_mini_batch_size="$TRAIN_BATCH" \
+      actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu="$MICRO_BSZ" \
+      actor_rollout_ref.actor.use_kl_loss=True \
+      actor_rollout_ref.actor.kl_loss_coef="$KL_COEF" \
+      actor_rollout_ref.actor.kl_loss_type=low_var_kl \
+      actor_rollout_ref.model.enable_gradient_checkpointing=True \
+      actor_rollout_ref.actor.fsdp_config.param_offload="$PARAM_OFFLOAD" \
+      actor_rollout_ref.actor.fsdp_config.optimizer_offload="$OPTIM_OFFLOAD" \
+      actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu="$LOGP_MICRO_BSZ" \
+      actor_rollout_ref.rollout.tensor_model_parallel_size="$TP" \
+      actor_rollout_ref.rollout.name=vllm \
+      actor_rollout_ref.rollout.gpu_memory_utilization="$GPU_MEM_UTIL" \
+      actor_rollout_ref.rollout.stop='["</class>"]' \
+      actor_rollout_ref.rollout.include_stop_str_in_output=True \
+      actor_rollout_ref.rollout.enable_chunked_prefill=False \
+      actor_rollout_ref.rollout.enforce_eager=False \
+      actor_rollout_ref.rollout.free_cache_engine=False \
+      actor_rollout_ref.rollout.n=5 \
+      actor_rollout_ref.rollout.val_kwargs.do_sample=False \
+      actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu="$LOGP_MICRO_BSZ" \
+      actor_rollout_ref.ref.fsdp_config.param_offload=True \
+      algorithm.kl_ctrl.kl_coef="$KL_COEF" \
+      trainer.critic_warmup=0 \
+      trainer.logger="['${LOGGER}']" \
+      trainer.project_name="$PROJECT" \
+      trainer.experiment_name="$EXP" \
+      trainer.default_local_dir="$CKPT_ROOT" \
+      trainer.n_gpus_per_node="$N_GPUS" \
+      trainer.nnodes=1 \
+      trainer.save_freq=500 \
+      trainer.test_freq=10 \
+      trainer.val_only="$VAL_ONLY" \
+      trainer.val_before_train=True \
+      trainer.resume_mode="$RESUME_MODE" \
+      trainer.total_epochs="$EPOCHS" \
+      ts.use_sem_orth=True \
+      ts.adv_mix=0.3 \
+      ts.similarity_method=ot \
+      ts.ot_eps=0.08 \
+      ts.ot_n_iter=50 "${@:3}"
+}
+
+merge_latest_ckpt () {
+  local latest
+  latest=$(ls -d "${CKPT_ROOT}"/global_step_* 2>/dev/null | sort -t_ -k3 -n | tail -1 || true)
+  if [ -z "$latest" ]; then
+    echo "[ERROR] no checkpoint found under ${CKPT_ROOT}" >&2; exit 1
+  fi
+  if has_hf_weights "${latest}/actor/huggingface"; then
+    echo "[merge] reusing existing HF weights: ${latest}/actor/huggingface" >&2
+    echo "${latest}/actor/huggingface"
+    return
+  fi
+  echo "[merge] latest checkpoint: ${latest}/actor" >&2
+  "$PYTHON_BIN" scripts/model_merger.py \
+      --local_dir "${latest}/actor" \
+      --lora-alpha "$LORA_ALPHA" >&2
+  echo "${latest}/actor/huggingface"   # only the path goes to stdout (captured by caller)
+}
+
+has_hf_weights () {
+  local path="$1"
+  [ -f "${path}/model.safetensors" ] ||
+    [ -f "${path}/model.safetensors.index.json" ] ||
+    [ -f "${path}/pytorch_model.bin" ] ||
+    [ -f "${path}/pytorch_model.bin.index.json" ]
+}
+
+prepare_eval_model () {
+  local model_path="$1"
+  if has_hf_weights "$model_path"; then
+    echo "$model_path"
+    return
+  fi
+
+  local actor_dir=""
+  if [ "$(basename "$model_path")" = "huggingface" ]; then
+    actor_dir="$(dirname "$model_path")"
+  elif [ -d "${model_path}/huggingface" ]; then
+    actor_dir="$model_path"
+  fi
+
+  if [ -z "$actor_dir" ] ||
+     ! find "$actor_dir" -maxdepth 1 -name 'model_world_size_*_rank_0.pt' -print -quit | grep -q .; then
+    echo "[ERROR] '${model_path}' has no HF model weights and is not a mergeable FSDP checkpoint." >&2
+    return 1
+  fi
+
+  echo "[merge] HF weights missing; merging checkpoint ${actor_dir}" >&2
+  "$PYTHON_BIN" scripts/model_merger.py \
+      --local_dir "$actor_dir" \
+      --lora-alpha "$LORA_ALPHA" >&2
+
+  model_path="${actor_dir}/huggingface"
+  if ! has_hf_weights "$model_path"; then
+    echo "[ERROR] merge completed without producing HF model weights under '${model_path}'." >&2
+    return 1
+  fi
+  echo "$model_path"
+}
+
+# ---- log file --------------------------------------------------------------
+# Tee all output (terminal + file). Override path with LOG_FILE=, or LOG_DIR=.
+LOG_DIR=${LOG_DIR:-/mnt/share01/sqk/AnomSeer/logs}
+mkdir -p "$LOG_DIR"
+LOG_FILE=${LOG_FILE:-${LOG_DIR}/rats_${STAGE}$([ "$LORA_RANK" -gt 0 ] && echo _lora)_$(date +%Y%m%d_%H%M%S).log}
+exec > >(tee -a "$LOG_FILE") 2>&1
+echo "[log] saving full output to: $LOG_FILE"
+
+# ---- dispatch --------------------------------------------------------------
+if [ "$LORA_RANK" -gt 0 ]; then
+  FT_MODE="LoRA (r=${LORA_RANK}, alpha=${LORA_ALPHA}, target=${LORA_TARGET_MODULES})"
+else
+  FT_MODE="full-parameter fine-tuning"
+fi
+if [ "$TP" -le 1 ]; then PARALLEL="data-parallel (DP=${N_GPUS}, full replica per card)"; else PARALLEL="tensor-parallel (TP=${TP})"; fi
+echo "[config] tuning=${FT_MODE} | lr=${LR} | gpus=${N_GPUS} | ${PARALLEL}"
+echo "[config] model=${MODEL_PATH}"
+echo "[config] python=${PYTHON_BIN} | stage=${STAGE}"
+if [ "$LORA_RANK" -eq 0 ] && [ "$PARAM_OFFLOAD" != "True" ] && [[ "$MODEL_PATH" == *7[bB]* ]]; then
+  echo "[hint] full-parameter 7B on 2x48GB will likely OOM. Either:"
+  echo "       (a) LoRA:    LORA_RANK=16 bash $0"
+  echo "       (b) offload: PARAM_OFFLOAD=True OPTIM_OFFLOAD=True bash $0"
+fi
+
+case "$STAGE" in
+  train)
+    echo "[stage] TRAIN  (model=${MODEL_PATH}, gpus=${N_GPUS}, tp=${TP})"
+    run_verl False "$MODEL_PATH" "$@"
+    ;;
+  eval)
+    eval_model=$(prepare_eval_model "$MODEL_PATH")
+    echo "[stage] EVAL   (model=${eval_model}, gpus=${N_GPUS}, tp=${TP}, lora_rank=0)"
+    run_verl True "$eval_model" "$@"
+    ;;
+  train_eval)
+    echo "[stage] TRAIN  (model=${MODEL_PATH}, gpus=${N_GPUS}, tp=${TP})"
+    run_verl False "$MODEL_PATH"
+    merged=$(merge_latest_ckpt)
+    echo "[stage] EVAL   (merged checkpoint=${merged})"
+    run_verl True "$merged" "$@"
+    ;;
+  *)
+    echo "[ERROR] unknown STAGE='$STAGE' (use train | eval | train_eval)" >&2; exit 1
+    ;;
+esac

logs/anomseer_rats_uni_full_20260612_184427_exp3/rats_eval_lora_20260615_053549.log

@@ -0,0 +1,417 @@
+[log] saving full output to: /mnt/share01/sqk/AnomSeer/logs/anomseer_rats_uni_full_20260612_184427_exp3/rats_eval_lora_20260615_053549.log
+[config] tuning=LoRA (r=16, alpha=16, target=q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj) | lr=1e-4 | gpus=2 | tensor-parallel (TP=2)
+[config] model=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface
+[config] python=/home/suiqk/anaconda3/envs/scalerag-ts-v4/bin/python | stage=eval
+[stage] EVAL   (model=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface, gpus=2, tp=2)
+2026-06-15 05:35:59,961	INFO worker.py:2012 -- Started a local Ray instance.
+/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/worker.py:2051: FutureWarning: Tip: In future versions of Ray, Ray will no longer override accelerator visible devices env var if num_gpus=0 or num_gpus=None (default). To enable this behavior and turn off this error message, set RAY_ACCEL_ENV_VAR_OVERRIDE_ON_ZERO=0
+  warnings.warn(
+(TaskRunner pid=363397) {'actor_rollout_ref': {'actor': {'clip_ratio': 0.2,
+(TaskRunner pid=363397)                                  'entropy_coeff': 0.001,
+(TaskRunner pid=363397)                                  'fsdp_config': {'fsdp_size': -1,
+(TaskRunner pid=363397)                                                  'optimizer_offload': False,
+(TaskRunner pid=363397)                                                  'param_offload': False,
+(TaskRunner pid=363397)                                                  'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=363397)                                  'grad_clip': 1.0,
+(TaskRunner pid=363397)                                  'kl_loss_coef': 0.001,
+(TaskRunner pid=363397)                                  'kl_loss_type': 'low_var_kl',
+(TaskRunner pid=363397)                                  'optim': {'lr': 0.0001,
+(TaskRunner pid=363397)                                            'lr_warmup_steps': -1,
+(TaskRunner pid=363397)                                            'lr_warmup_steps_ratio': 0.0,
+(TaskRunner pid=363397)                                            'min_lr_ratio': None,
+(TaskRunner pid=363397)                                            'total_training_steps': -1,
+(TaskRunner pid=363397)                                            'warmup_style': 'constant'},
+(TaskRunner pid=363397)                                  'ppo_epochs': 1,
+(TaskRunner pid=363397)                                  'ppo_max_token_len_per_gpu': 16384,
+(TaskRunner pid=363397)                                  'ppo_micro_batch_size': None,
+(TaskRunner pid=363397)                                  'ppo_micro_batch_size_per_gpu': 2,
+(TaskRunner pid=363397)                                  'ppo_mini_batch_size': 16,
+(TaskRunner pid=363397)                                  'shuffle': False,
+(TaskRunner pid=363397)                                  'strategy': 'fsdp',
+(TaskRunner pid=363397)                                  'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=363397)                                  'use_dynamic_bsz': False,
+(TaskRunner pid=363397)                                  'use_kl_loss': True,
+(TaskRunner pid=363397)                                  'use_torch_compile': True},
+(TaskRunner pid=363397)                        'hybrid_engine': True,
+(TaskRunner pid=363397)                        'model': {'enable_gradient_checkpointing': True,
+(TaskRunner pid=363397)                                  'external_lib': None,
+(TaskRunner pid=363397)                                  'lora_alpha': 16,
+(TaskRunner pid=363397)                                  'lora_dropout': 0.0,
+(TaskRunner pid=363397)                                  'lora_rank': 16,
+(TaskRunner pid=363397)                                  'lora_target_modules': 'q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj',
+(TaskRunner pid=363397)                                  'override_config': {},
+(TaskRunner pid=363397)                                  'path': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface',
+(TaskRunner pid=363397)                                  'use_remove_padding': False},
+(TaskRunner pid=363397)                        'ref': {'fsdp_config': {'param_offload': True,
+(TaskRunner pid=363397)                                                'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=363397)                                'log_prob_max_token_len_per_gpu': 16384,
+(TaskRunner pid=363397)                                'log_prob_micro_batch_size': None,
+(TaskRunner pid=363397)                                'log_prob_micro_batch_size_per_gpu': 8,
+(TaskRunner pid=363397)                                'log_prob_use_dynamic_bsz': False,
+(TaskRunner pid=363397)                                'ulysses_sequence_parallel_size': 1},
+(TaskRunner pid=363397)                        'rollout': {'disable_log_stats': True,
+(TaskRunner pid=363397)                                    'do_sample': True,
+(TaskRunner pid=363397)                                    'dtype': 'bfloat16',
+(TaskRunner pid=363397)                                    'enable_chunked_prefill': False,
+(TaskRunner pid=363397)                                    'enforce_eager': False,
+(TaskRunner pid=363397)                                    'free_cache_engine': False,
+(TaskRunner pid=363397)                                    'gpu_memory_utilization': 0.4,
+(TaskRunner pid=363397)                                    'ignore_eos': False,
+(TaskRunner pid=363397)                                    'load_format': 'dummy_dtensor',
+(TaskRunner pid=363397)                                    'log_prob_max_token_len_per_gpu': 16384,
+(TaskRunner pid=363397)                                    'log_prob_micro_batch_size': None,
+(TaskRunner pid=363397)                                    'log_prob_micro_batch_size_per_gpu': 8,
+(TaskRunner pid=363397)                                    'log_prob_use_dynamic_bsz': False,
+(TaskRunner pid=363397)                                    'max_model_len': None,
+(TaskRunner pid=363397)                                    'max_num_batched_tokens': 8192,
+(TaskRunner pid=363397)                                    'max_num_seqs': 1024,
+(TaskRunner pid=363397)                                    'n': 5,
+(TaskRunner pid=363397)                                    'name': 'vllm',
+(TaskRunner pid=363397)                                    'prompt_length': 1024,
+(TaskRunner pid=363397)                                    'response_length': 1024,
+(TaskRunner pid=363397)                                    'temperature': 1.0,
+(TaskRunner pid=363397)                                    'tensor_model_parallel_size': 2,
+(TaskRunner pid=363397)                                    'top_k': -1,
+(TaskRunner pid=363397)                                    'top_p': 1,
+(TaskRunner pid=363397)                                    'use_fire_sampling': False,
+(TaskRunner pid=363397)                                    'val_kwargs': {'do_sample': False,
+(TaskRunner pid=363397)                                                   'n': 1,
+(TaskRunner pid=363397)                                                   'temperature': 0.4,
+(TaskRunner pid=363397)                                                   'top_k': -1,
+(TaskRunner pid=363397)                                                   'top_p': 1.0}}},
+(TaskRunner pid=363397)  'algorithm': {'adv_estimator': 'grpo',
+(TaskRunner pid=363397)                'gamma': 1.0,
+(TaskRunner pid=363397)                'kl_ctrl': {'kl_coef': 0.001, 'type': 'fixed'},
+(TaskRunner pid=363397)                'kl_penalty': 'kl',
+(TaskRunner pid=363397)                'lam': 1.0},
+(TaskRunner pid=363397)  'critic': {'cliprange_value': 0.5,
+(TaskRunner pid=363397)             'forward_max_token_len_per_gpu': 32768,
+(TaskRunner pid=363397)             'forward_micro_batch_size': None,
+(TaskRunner pid=363397)             'forward_micro_batch_size_per_gpu': None,
+(TaskRunner pid=363397)             'grad_clip': 1.0,
+(TaskRunner pid=363397)             'model': {'enable_gradient_checkpointing': True,
+(TaskRunner pid=363397)                       'external_lib': None,
+(TaskRunner pid=363397)                       'fsdp_config': {'fsdp_size': -1,
+(TaskRunner pid=363397)                                       'optimizer_offload': False,(TaskRunner pid=363397) DeprecationWarning: `ray.state.available_resources_per_node` is a private attribute and access will be removed in a future Ray version.
+
+(TaskRunner pid=363397)                                       'param_offload': False,
+(TaskRunner pid=363397)                                       'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=363397)                       'override_config': {},
+(TaskRunner pid=363397)                       'path': '~/models/deepseek-llm-7b-chat',
+(TaskRunner pid=363397)                       'tokenizer_path': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface',
+(TaskRunner pid=363397)                       'use_remove_padding': False},
+(TaskRunner pid=363397)             'optim': {'lr': 1e-05,
+(TaskRunner pid=363397)                       'lr_warmup_steps_ratio': 0.0,
+(TaskRunner pid=363397)                       'min_lr_ratio': None,
+(TaskRunner pid=363397)                       'total_training_steps': -1,
+(TaskRunner pid=363397)                       'warmup_style': 'constant'},
+(TaskRunner pid=363397)             'ppo_epochs': 1,
+(TaskRunner pid=363397)             'ppo_max_token_len_per_gpu': 32768,
+(TaskRunner pid=363397)             'ppo_micro_batch_size': None,
+(TaskRunner pid=363397)             'ppo_micro_batch_size_per_gpu': None,
+(TaskRunner pid=363397)             'ppo_mini_batch_size': 16,
+(TaskRunner pid=363397)             'shuffle': False,
+(TaskRunner pid=363397)             'strategy': 'fsdp',
+(TaskRunner pid=363397)             'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=363397)             'use_dynamic_bsz': False},
+(TaskRunner pid=363397)  'custom_reward_function': {'name': 'compute_score', 'path': None},
+(TaskRunner pid=363397)  'data': {'filter_overlong_prompts': True,
+(TaskRunner pid=363397)           'image_key': 'images',
+(TaskRunner pid=363397)           'max_prompt_length': 1024,
+(TaskRunner pid=363397)           'max_response_length': 1024,
+(TaskRunner pid=363397)           'prompt_key': 'prompt',
+(TaskRunner pid=363397)           'return_raw_chat': False,
+(TaskRunner pid=363397)           'return_raw_input_ids': False,
+(TaskRunner pid=363397)           'shuffle': True,
+(TaskRunner pid=363397)           'tokenizer': None,
+(TaskRunner pid=363397)           'train_batch_size': 16,
+(TaskRunner pid=363397)           'train_files': '/mnt/share01/sqk/AnomSeer/data/rats_uni_processed/train_full.parquet',
+(TaskRunner pid=363397)           'truncation': 'error',
+(TaskRunner pid=363397)           'val_batch_size': 64,
+(TaskRunner pid=363397)           'val_files': '/mnt/share01/sqk/AnomSeer/data/rats_uni_processed/test_full.parquet'},
+(TaskRunner pid=363397)  'reward_model': {'enable': False,
+(TaskRunner pid=363397)                   'forward_max_token_len_per_gpu': 32768,
+(TaskRunner pid=363397)                   'max_length': None,
+(TaskRunner pid=363397)                   'micro_batch_size': None,
+(TaskRunner pid=363397)                   'micro_batch_size_per_gpu': None,
+(TaskRunner pid=363397)                   'model': {'external_lib': None,
+(TaskRunner pid=363397)                             'fsdp_config': {'fsdp_size': -1,
+(TaskRunner pid=363397)                                             'param_offload': False,
+(TaskRunner pid=363397)                                             'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=363397)                             'input_tokenizer': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface',
+(TaskRunner pid=363397)                             'path': '~/models/FsfairX-LLaMA3-RM-v0.1',
+(TaskRunner pid=363397)                             'use_remove_padding': False},
+(TaskRunner pid=363397)                   'reward_manager': 'naive',
+(TaskRunner pid=363397)                   'strategy': 'fsdp',
+(TaskRunner pid=363397)                   'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=363397)                   'use_dynamic_bsz': False},
+(TaskRunner pid=363397)  'trainer': {'balance_batch': True,
+(TaskRunner pid=363397)              'critic_warmup': 0,
+(TaskRunner pid=363397)              'default_hdfs_dir': None,
+(TaskRunner pid=363397)              'default_local_dir': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3',
+(TaskRunner pid=363397)              'del_local_ckpt_after_load': False,
+(TaskRunner pid=363397)              'experiment_name': 'anomseer_rats_uni_full_20260612_184427_exp3_full_eval',
+(TaskRunner pid=363397)              'logger': ['console'],
+(TaskRunner pid=363397)              'n_gpus_per_node': 2,
+(TaskRunner pid=363397)              'nnodes': 1,
+(TaskRunner pid=363397)              'project_name': 'anomseer',
+(TaskRunner pid=363397)              'remove_previous_ckpt_in_save': False,
+(TaskRunner pid=363397)              'resume_from_path': False,
+(TaskRunner pid=363397)              'resume_mode': 'auto',
+(TaskRunner pid=363397)              'save_freq': 500,
+(TaskRunner pid=363397)              'test_freq': -1,
+(TaskRunner pid=363397)              'total_epochs': 1,
+(TaskRunner pid=363397)              'total_training_steps': None,
+(TaskRunner pid=363397)              'val_before_train': True,
+(TaskRunner pid=363397)              'val_generations_to_log_to_wandb': 0,
+(TaskRunner pid=363397)              'val_only': True},
+(TaskRunner pid=363397)  'ts': {'adv_mix': 0.3,
+(TaskRunner pid=363397)         'clip_pooling': 'mean',
+(TaskRunner pid=363397)         'clip_temperature': 1.0,
+(TaskRunner pid=363397)         'ot_eps': 0.08,
+(TaskRunner pid=363397)         'ot_n_iter': 50,
+(TaskRunner pid=363397)         'similarity_method': 'ot',
+(TaskRunner pid=363397)         'use_sem_orth': True}}
+(TaskRunner pid=363397) reward_manager_name: anomts
+(TaskRunner pid=363397) WARNING: val_batch_size is deprecated. Validation datasets are sent to inference engines as a whole batch, which will schedule the memory themselves.
+(TaskRunner pid=363397) [validate_config] All configuration checks passed successfully!
+(TaskRunner pid=363397) dataset len: 30266
+(TaskRunner pid=363397) filter dataset len: 30266
+(TaskRunner pid=363397) dataset len: 6034
+(TaskRunner pid=363397) filter dataset len: 6034
+(TaskRunner pid=363397) Size of train dataloader: 1891
+(TaskRunner pid=363397) Total training steps: 1891
+(WorkerDict pid=364993) Model config after override: Qwen2_5_VLConfig {
+(WorkerDict pid=364993)   "architectures": [
+(WorkerDict pid=364993)     "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=364993)   ],
+(WorkerDict pid=364993)   "attention_dropout": 0.0,
+(WorkerDict pid=364993)   "eos_token_id": 151645,
+(WorkerDict pid=364993)   "hidden_act": "silu",
+(WorkerDict pid=364993)   "hidden_size": 2048,
+(WorkerDict pid=364993)   "image_token_id": 151655,
+(WorkerDict pid=364993)   "initializer_range": 0.02,
+(WorkerDict pid=364993)   "intermediate_size": 11008,
+(WorkerDict pid=364993)   "max_position_embeddings": 128000,
+(WorkerDict pid=364993)   "max_window_layers": 70,
+(WorkerDict pid=364993)   "model_type": "qwen2_5_vl",
+(WorkerDict pid=364993)   "num_attention_heads": 16,
+(WorkerDict pid=364993)   "num_hidden_layers": 36,
+(WorkerDict pid=364993)   "num_key_value_heads": 2,
+(WorkerDict pid=364993)   "pad_token_id": 151643,
+(WorkerDict pid=364993)   "rms_norm_eps": 1e-06,
+(WorkerDict pid=364993)   "rope_scaling": {
+(WorkerDict pid=364993)     "mrope_section": [
+(WorkerDict pid=364993)       16,
+(WorkerDict pid=364993)       24,
+(WorkerDict pid=364993)       24
+(WorkerDict pid=364993)     ],
+(WorkerDict pid=364993)     "rope_type": "default",
+(WorkerDict pid=364993)     "type": "default"
+(WorkerDict pid=364993)   },
+(WorkerDict pid=364993)   "rope_theta": 1000000.0,
+(WorkerDict pid=364993)   "sliding_window": 32768,
+(WorkerDict pid=364993)   "text_config": {
+(WorkerDict pid=364993)     "architectures": [
+(WorkerDict pid=364993)       "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=364993)     ],
+(WorkerDict pid=364993)     "attention_dropout": 0.0,
+(WorkerDict pid=364993)     "bos_token_id": 151643,
+(WorkerDict pid=364993)     "eos_token_id": 151645,
+(WorkerDict pid=364993)     "hidden_act": "silu",
+(WorkerDict pid=364993)     "hidden_size": 2048,
+(WorkerDict pid=364993)     "image_token_id": null,
+(WorkerDict pid=364993)     "initializer_range": 0.02,
+(WorkerDict pid=364993)     "intermediate_size": 11008,
+(WorkerDict pid=364993)     "layer_types": [
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention",
+(WorkerDict pid=364993)       "full_attention"
+(WorkerDict pid=364993)     ],
+(WorkerDict pid=364993)     "max_position_embeddings": 128000,
+(WorkerDict pid=364993)     "max_window_layers": 70,
+(WorkerDict pid=364993)     "model_type": "qwen2_5_vl_text",
+(WorkerDict pid=364993)     "num_attention_heads": 16,
+(WorkerDict pid=364993)     "num_hidden_layers": 36,
+(WorkerDict pid=364993)     "num_key_value_heads": 2,
+(WorkerDict pid=364993)     "rms_norm_eps": 1e-06,
+(WorkerDict pid=364993)     "rope_scaling": {
+(WorkerDict pid=364993)       "mrope_section": [
+(WorkerDict pid=364993)         16,
+(WorkerDict pid=364993)         24,
+(WorkerDict pid=364993)         24
+(WorkerDict pid=364993)       ],
+(WorkerDict pid=364993)       "rope_type": "default",
+(WorkerDict pid=364993)       "type": "default"
+(WorkerDict pid=364993)     },
+(WorkerDict pid=364993)     "rope_theta": 1000000.0,
+(WorkerDict pid=364993)     "sliding_window": null,
+(WorkerDict pid=364993)     "tie_word_embeddings": true,
+(WorkerDict pid=364993)     "torch_dtype": "float32",
+(WorkerDict pid=364993)     "use_cache": true,
+(WorkerDict pid=364993)     "use_sliding_window": false,
+(WorkerDict pid=364993)     "video_token_id": null,
+(WorkerDict pid=364993)     "vision_end_token_id": 151653,
+(WorkerDict pid=364993)     "vision_start_token_id": 151652,
+(WorkerDict pid=364993)     "vision_token_id": 151654,
+(WorkerDict pid=364993)     "vocab_size": 151936
+(WorkerDict pid=364993)   },
+(WorkerDict pid=364993)   "torch_dtype": "float32",
+(WorkerDict pid=364993)   "transformers_version": "4.54.1",
+(WorkerDict pid=364993)   "use_cache": true,
+(WorkerDict pid=364993)   "use_sliding_window": false,
+(WorkerDict pid=364993)   "video_token_id": 151656,
+(WorkerDict pid=364993)   "vision_config": {
+(WorkerDict pid=364993)     "depth": 32,
+(WorkerDict pid=364993)     "fullatt_block_indexes": [
+(WorkerDict pid=364993)       7,
+(WorkerDict pid=364993)       15,
+(WorkerDict pid=364993)       23,
+(WorkerDict pid=364993)       31
+(WorkerDict pid=364993)     ],
+(WorkerDict pid=364993)     "hidden_act": "silu",
+(WorkerDict pid=364993)     "hidden_size": 1280,
+(WorkerDict pid=364993)     "in_channels": 3,
+(WorkerDict pid=364993)     "in_chans": 3,
+(WorkerDict pid=364993)     "initializer_range": 0.02,
+(WorkerDict pid=364993)     "intermediate_size": 3420,
+(WorkerDict pid=364993)     "model_type": "qwen2_5_vl",
+(WorkerDict pid=364993)     "num_heads": 16,
+(WorkerDict pid=364993)     "out_hidden_size": 2048,
+(WorkerDict pid=364993)     "patch_size": 14,
+(WorkerDict pid=364993)     "spatial_merge_size": 2,
+(WorkerDict pid=364993)     "spatial_patch_size": 14,
+(WorkerDict pid=364993)     "temporal_patch_size": 2,
+(WorkerDict pid=364993)     "tokens_per_second": 2,
+(WorkerDict pid=364993)     "torch_dtype": "float32",
+(WorkerDict pid=364993)     "window_size": 112
+(WorkerDict pid=364993)   },
+(WorkerDict pid=364993)   "vision_end_token_id": 151653,
+(WorkerDict pid=364993)   "vision_start_token_id": 151652,
+(WorkerDict pid=364993)   "vision_token_id": 151654,
+(WorkerDict pid=364993)   "vocab_size": 151936
+(WorkerDict pid=364993) }
+(WorkerDict pid=364993) 
+Error executing job with overrides: ['algorithm.adv_estimator=grpo', 'data.train_files=/mnt/share01/sqk/AnomSeer/data/rats_uni_processed/train_full.parquet', 'data.val_files=/mnt/share01/sqk/AnomSeer/data/rats_uni_processed/test_full.parquet', 'data.train_batch_size=16', 'data.max_prompt_length=1024', 'data.max_response_length=1024', 'data.filter_overlong_prompts=True', 'data.truncation=error', 'data.image_key=images', 'actor_rollout_ref.model.path=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface', 'actor_rollout_ref.model.lora_rank=16', 'actor_rollout_ref.model.lora_alpha=16', 'actor_rollout_ref.model.lora_dropout=0.0', "actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'", 'actor_rollout_ref.actor.optim.lr=1e-4', 'actor_rollout_ref.model.use_remove_padding=False', 'actor_rollout_ref.actor.ppo_mini_batch_size=16', 'actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2', 'actor_rollout_ref.actor.use_kl_loss=True', 'actor_rollout_ref.actor.kl_loss_coef=0.001', 'actor_rollout_ref.actor.kl_loss_type=low_var_kl', 'actor_rollout_ref.model.enable_gradient_checkpointing=True', 'actor_rollout_ref.actor.fsdp_config.param_offload=False', 'actor_rollout_ref.actor.fsdp_config.optimizer_offload=False', 'actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8', 'actor_rollout_ref.rollout.tensor_model_parallel_size=2', 'actor_rollout_ref.rollout.name=vllm', 'actor_rollout_ref.rollout.gpu_memory_utilization=0.4', 'actor_rollout_ref.rollout.enable_chunked_prefill=False', 'actor_rollout_ref.rollout.enforce_eager=False', 'actor_rollout_ref.rollout.free_cache_engine=False', 'actor_rollout_ref.rollout.n=5', 'actor_rollout_ref.rollout.val_kwargs.do_sample=False', 'actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8', 'actor_rollout_ref.ref.fsdp_config.param_offload=True', 'algorithm.kl_ctrl.kl_coef=0.001', 'trainer.critic_warmup=0', "trainer.logger=['console']", 'trainer.project_name=anomseer', 'trainer.experiment_name=anomseer_rats_uni_full_20260612_184427_exp3_full_eval', 'trainer.default_local_dir=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3', 'trainer.n_gpus_per_node=2', 'trainer.nnodes=1', 'trainer.save_freq=500', 'trainer.test_freq=10', 'trainer.val_only=True', 'trainer.val_before_train=True', 'trainer.total_epochs=1', 'ts.use_sem_orth=True', 'ts.adv_mix=0.3', 'ts.similarity_method=ot', 'ts.ot_eps=0.08', 'ts.ot_n_iter=50', 'data.val_batch_size=64', 'trainer.test_freq=-1', 'trainer.val_before_train=True']
+(TaskRunner pid=363397) Unhandled error (suppress with 'RAY_IGNORE_UNHANDLED_ERRORS=1'): ray::WorkerDict.ref_init_model() (pid=365521, ip=172.20.115.25, actor_id=69baeda02fb711b165f2623201000000, repr=<verl.single_controller.ray.base.WorkerDict object at 0x7e5ac98e7d90>)
+(TaskRunner pid=363397)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=363397)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=363397)   File "/mnt/share01/sqk/AnomSeer/verl/single_controller/ray/base.py", line 420, in func
+(TaskRunner pid=363397)     return getattr(self.worker_dict[key], name)(*args, **kwargs)
+(TaskRunner pid=363397)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=363397)   File "/mnt/share01/sqk/AnomSeer/verl/single_controller/base/decorator.py", line 404, in inner
+(TaskRunner pid=363397)     return func(*args, **kwargs)
+(TaskRunner pid=363397)            ^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=363397)   File "/mnt/share01/sqk/AnomSeer/verl/workers/fsdp_workers.py", line 466, in init_model
+(TaskRunner pid=363397)     self.ref_module_fsdp = self._build_model_optimizer(model_path=self.config.model.path,
+(TaskRunner pid=363397)                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=363397)   File "/mnt/share01/sqk/AnomSeer/verl/workers/fsdp_workers.py", line 217, in _build_model_optimizer
+(TaskRunner pid=363397)     actor_module = actor_module_class.from_pretrained(pretrained_model_name_or_path=local_path,
+(TaskRunner pid=363397)                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=363397)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/transformers/models/auto/modeling_auto.py", line 2165, in from_pretrained
+(TaskRunner pid=363397)     return super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+(TaskRunner pid=363397)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=363397)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/transformers/models/auto/auto_factory.py", line 600, in from_pretrained
+(TaskRunner pid=363397)     return model_class.from_pretrained(
+(TaskRunner pid=363397)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=363397)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/transformers/modeling_utils.py", line 315, in _wrapper
+(TaskRunner pid=363397)     return func(*args, **kwargs)
+(TaskRunner pid=363397)            ^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=363397)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/transformers/modeling_utils.py", line 4854, in from_pretrained
+(TaskRunner pid=363397)     checkpoint_files, sharded_metadata = _get_resolved_checkpoint_files(
+(TaskRunner pid=363397)                                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=363397)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/transformers/modeling_utils.py", line 1093, in _get_resolved_checkpoint_files
+(TaskRunner pid=363397)     raise OSError(
+(TaskRunner pid=363397) OSError: Error no file named pytorch_model.bin, model.safetensors, tf_model.h5, model.ckpt.index or flax_model.msgpack found in directory /mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface.
+Traceback (most recent call last):
+  File "/mnt/share01/sqk/AnomSeer/verl/trainer/main_ppo.py", line 68, in main
+    ray.get(runner.run.remote(config))
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/auto_init_hook.py", line 22, in auto_init_wrapper
+    return fn(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/client_mode_hook.py", line 107, in wrapper
+    return func(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/worker.py", line 2980, in get
+    values, debugger_breakpoint = worker.get_objects(
+                                  ^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/worker.py", line 1023, in get_objects
+    raise value.as_instanceof_cause()
+ray.exceptions.RayTaskError(OSError): ray::TaskRunner.run() (pid=363397, ip=172.20.115.25, actor_id=eb96e7c0e8f2d425a19fb2e901000000, repr=<main_ppo.TaskRunner object at 0x77b72dd69590>)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/trainer/main_ppo.py", line 172, in run
+    trainer.init_workers()
+  File "/mnt/share01/sqk/AnomSeer/verl/trainer/ppo/ray_trainer.py", line 2024, in init_workers
+    self.ref_policy_wg.init_model()
+  File "/mnt/share01/sqk/AnomSeer/verl/single_controller/ray/base.py", line 43, in func
+    output = ray.get(output)
+             ^^^^^^^^^^^^^^^
+           ^^^^^^^^^^^^^^^^^^^
+           ^^^^^^^^^^^^^^^^^^^^^
+                                  ^^^^^^^^^^^^^^^^^^^
+ray.exceptions.RayTaskError(OSError): ray::WorkerDict.ref_init_model() (pid=364993, ip=172.20.115.25, actor_id=30e087db86b71d1893cef58101000000, repr=<verl.single_controller.ray.base.WorkerDict object at 0x73530b26be10>)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/single_controller/ray/base.py", line 420, in func
+    return getattr(self.worker_dict[key], name)(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/single_controller/base/decorator.py", line 404, in inner
+    return func(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/workers/fsdp_workers.py", line 466, in init_model
+    self.ref_module_fsdp = self._build_model_optimizer(model_path=self.config.model.path,
+                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/workers/fsdp_workers.py", line 217, in _build_model_optimizer
+    actor_module = actor_module_class.from_pretrained(pretrained_model_name_or_path=local_path,
+                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/transformers/models/auto/modeling_auto.py", line 2165, in from_pretrained
+    return super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/transformers/models/auto/auto_factory.py", line 600, in from_pretrained
+    return model_class.from_pretrained(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/transformers/modeling_utils.py", line 315, in _wrapper
+    return func(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/transformers/modeling_utils.py", line 4854, in from_pretrained
+    checkpoint_files, sharded_metadata = _get_resolved_checkpoint_files(
+                                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/transformers/modeling_utils.py", line 1093, in _get_resolved_checkpoint_files
+    raise OSError(
+OSError: Error no file named pytorch_model.bin, model.safetensors, tf_model.h5, model.ckpt.index or flax_model.msgpack found in directory /mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface.
+
+Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

logs/anomseer_rats_uni_full_20260612_184427_exp3/rats_eval_lora_20260615_122910.log

@@ -0,0 +1,1196 @@
+[log] saving full output to: /mnt/share01/sqk/AnomSeer/logs/anomseer_rats_uni_full_20260612_184427_exp3/rats_eval_lora_20260615_122910.log
+[config] tuning=LoRA (r=16, alpha=16, target=q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj) | lr=1e-4 | gpus=2 | tensor-parallel (TP=2)
+[config] model=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface
+[config] python=/home/suiqk/anaconda3/envs/scalerag-ts-v4/bin/python | stage=eval
+[stage] EVAL   (model=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface, gpus=2, tp=2, lora_rank=0)
+2026-06-15 12:29:20,742	INFO worker.py:2012 -- Started a local Ray instance.
+/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/worker.py:2051: FutureWarning: Tip: In future versions of Ray, Ray will no longer override accelerator visible devices env var if num_gpus=0 or num_gpus=None (default). To enable this behavior and turn off this error message, set RAY_ACCEL_ENV_VAR_OVERRIDE_ON_ZERO=0
+  warnings.warn(
+(TaskRunner pid=1868415) {'actor_rollout_ref': {'actor': {'clip_ratio': 0.2,
+(TaskRunner pid=1868415)                                  'entropy_coeff': 0.001,
+(TaskRunner pid=1868415)                                  'fsdp_config': {'fsdp_size': -1,
+(TaskRunner pid=1868415)                                                  'optimizer_offload': False,
+(TaskRunner pid=1868415)                                                  'param_offload': False,
+(TaskRunner pid=1868415)                                                  'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=1868415)                                  'grad_clip': 1.0,
+(TaskRunner pid=1868415)                                  'kl_loss_coef': 0.001,
+(TaskRunner pid=1868415)                                  'kl_loss_type': 'low_var_kl',
+(TaskRunner pid=1868415)                                  'optim': {'lr': 0.0001,
+(TaskRunner pid=1868415)                                            'lr_warmup_steps': -1,
+(TaskRunner pid=1868415)                                            'lr_warmup_steps_ratio': 0.0,
+(TaskRunner pid=1868415)                                            'min_lr_ratio': None,
+(TaskRunner pid=1868415)                                            'total_training_steps': -1,
+(TaskRunner pid=1868415)                                            'warmup_style': 'constant'},
+(TaskRunner pid=1868415)                                  'ppo_epochs': 1,
+(TaskRunner pid=1868415)                                  'ppo_max_token_len_per_gpu': 16384,
+(TaskRunner pid=1868415)                                  'ppo_micro_batch_size': None,
+(TaskRunner pid=1868415)                                  'ppo_micro_batch_size_per_gpu': 2,
+(TaskRunner pid=1868415)                                  'ppo_mini_batch_size': 16,
+(TaskRunner pid=1868415)                                  'shuffle': False,
+(TaskRunner pid=1868415)                                  'strategy': 'fsdp',
+(TaskRunner pid=1868415)                                  'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=1868415)                                  'use_dynamic_bsz': False,
+(TaskRunner pid=1868415)                                  'use_kl_loss': True,
+(TaskRunner pid=1868415)                                  'use_torch_compile': True},
+(TaskRunner pid=1868415)                        'hybrid_engine': True,
+(TaskRunner pid=1868415)                        'model': {'enable_gradient_checkpointing': True,
+(TaskRunner pid=1868415)                                  'external_lib': None,
+(TaskRunner pid=1868415)                                  'lora_alpha': 16,
+(TaskRunner pid=1868415)                                  'lora_dropout': 0.0,
+(TaskRunner pid=1868415)                                  'lora_rank': 0,
+(TaskRunner pid=1868415)                                  'lora_target_modules': 'q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj',
+(TaskRunner pid=1868415)                                  'override_config': {},
+(TaskRunner pid=1868415)                                  'path': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface',
+(TaskRunner pid=1868415)                                  'use_remove_padding': False},
+(TaskRunner pid=1868415)                        'ref': {'fsdp_config': {'param_offload': True,
+(TaskRunner pid=1868415)                                                'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=1868415)                                'log_prob_max_token_len_per_gpu': 16384,
+(TaskRunner pid=1868415)                                'log_prob_micro_batch_size': None,
+(TaskRunner pid=1868415)                                'log_prob_micro_batch_size_per_gpu': 8,
+(TaskRunner pid=1868415)                                'log_prob_use_dynamic_bsz': False,
+(TaskRunner pid=1868415)                                'ulysses_sequence_parallel_size': 1},
+(TaskRunner pid=1868415)                        'rollout': {'disable_log_stats': True,
+(TaskRunner pid=1868415)                                    'do_sample': True,
+(TaskRunner pid=1868415)                                    'dtype': 'bfloat16',
+(TaskRunner pid=1868415)                                    'enable_chunked_prefill': False,
+(TaskRunner pid=1868415)                                    'enforce_eager': False,
+(TaskRunner pid=1868415)                                    'free_cache_engine': False,
+(TaskRunner pid=1868415)                                    'gpu_memory_utilization': 0.4,
+(TaskRunner pid=1868415)                                    'ignore_eos': False,
+(TaskRunner pid=1868415)                                    'load_format': 'dummy_dtensor',
+(TaskRunner pid=1868415)                                    'log_prob_max_token_len_per_gpu': 16384,
+(TaskRunner pid=1868415)                                    'log_prob_micro_batch_size': None,
+(TaskRunner pid=1868415)                                    'log_prob_micro_batch_size_per_gpu': 8,
+(TaskRunner pid=1868415)                                    'log_prob_use_dynamic_bsz': False,
+(TaskRunner pid=1868415)                                    'max_model_len': None,
+(TaskRunner pid=1868415)                                    'max_num_batched_tokens': 8192,
+(TaskRunner pid=1868415)                                    'max_num_seqs': 1024,
+(TaskRunner pid=1868415)                                    'n': 5,
+(TaskRunner pid=1868415)                                    'name': 'vllm',
+(TaskRunner pid=1868415)                                    'prompt_length': 1024,
+(TaskRunner pid=1868415)                                    'response_length': 1024,
+(TaskRunner pid=1868415)                                    'temperature': 1.0,
+(TaskRunner pid=1868415)                                    'tensor_model_parallel_size': 2,
+(TaskRunner pid=1868415)                                    'top_k': -1,
+(TaskRunner pid=1868415)                                    'top_p': 1,
+(TaskRunner pid=1868415)                                    'use_fire_sampling': False,
+(TaskRunner pid=1868415)                                    'val_kwargs': {'do_sample': False,
+(TaskRunner pid=1868415)                                                   'n': 1,
+(TaskRunner pid=1868415)                                                   'temperature': 0.4,
+(TaskRunner pid=1868415)                                                   'top_k': -1,
+(TaskRunner pid=1868415)                                                   'top_p': 1.0}}},
+(TaskRunner pid=1868415)  'algorithm': {'adv_estimator': 'grpo',
+(TaskRunner pid=1868415)                'gamma': 1.0,
+(TaskRunner pid=1868415)                'kl_ctrl': {'kl_coef': 0.001, 'type': 'fixed'},
+(TaskRunner pid=1868415)                'kl_penalty': 'kl',
+(TaskRunner pid=1868415)                'lam': 1.0},
+(TaskRunner pid=1868415)  'critic': {'cliprange_value': 0.5,
+(TaskRunner pid=1868415)             'forward_max_token_len_per_gpu': 32768,
+(TaskRunner pid=1868415)             'forward_micro_batch_size': None,
+(TaskRunner pid=1868415)             'forward_micro_batch_size_per_gpu': None,
+(TaskRunner pid=1868415)             'grad_clip': 1.0,
+(TaskRunner pid=1868415)             'model': {'enable_gradient_checkpointing': True,
+(TaskRunner pid=1868415)                       'external_lib': None,
+(TaskRunner pid=1868415)                       'fsdp_config': {'fsdp_size': -1,(TaskRunner pid=1868415) DeprecationWarning: `ray.state.available_resources_per_node` is a private attribute and access will be removed in a future Ray version.
+
+(TaskRunner pid=1868415)                                       'optimizer_offload': False,
+(TaskRunner pid=1868415)                                       'param_offload': False,
+(TaskRunner pid=1868415)                                       'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=1868415)                       'override_config': {},
+(TaskRunner pid=1868415)                       'path': '~/models/deepseek-llm-7b-chat',
+(TaskRunner pid=1868415)                       'tokenizer_path': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface',
+(TaskRunner pid=1868415)                       'use_remove_padding': False},
+(TaskRunner pid=1868415)             'optim': {'lr': 1e-05,
+(TaskRunner pid=1868415)                       'lr_warmup_steps_ratio': 0.0,
+(TaskRunner pid=1868415)                       'min_lr_ratio': None,
+(TaskRunner pid=1868415)                       'total_training_steps': -1,
+(TaskRunner pid=1868415)                       'warmup_style': 'constant'},
+(TaskRunner pid=1868415)             'ppo_epochs': 1,
+(TaskRunner pid=1868415)             'ppo_max_token_len_per_gpu': 32768,
+(TaskRunner pid=1868415)             'ppo_micro_batch_size': None,
+(TaskRunner pid=1868415)             'ppo_micro_batch_size_per_gpu': None,
+(TaskRunner pid=1868415)             'ppo_mini_batch_size': 16,
+(TaskRunner pid=1868415)             'shuffle': False,
+(TaskRunner pid=1868415)             'strategy': 'fsdp',
+(TaskRunner pid=1868415)             'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=1868415)             'use_dynamic_bsz': False},
+(TaskRunner pid=1868415)  'custom_reward_function': {'name': 'compute_score', 'path': None},
+(TaskRunner pid=1868415)  'data': {'filter_overlong_prompts': True,
+(TaskRunner pid=1868415)           'image_key': 'images',
+(TaskRunner pid=1868415)           'max_prompt_length': 1024,
+(TaskRunner pid=1868415)           'max_response_length': 1024,
+(TaskRunner pid=1868415)           'prompt_key': 'prompt',
+(TaskRunner pid=1868415)           'return_raw_chat': False,
+(TaskRunner pid=1868415)           'return_raw_input_ids': False,
+(TaskRunner pid=1868415)           'shuffle': True,
+(TaskRunner pid=1868415)           'tokenizer': None,
+(TaskRunner pid=1868415)           'train_batch_size': 16,
+(TaskRunner pid=1868415)           'train_files': '/mnt/share01/sqk/AnomSeer/data/rats_uni_processed/train_full.parquet',
+(TaskRunner pid=1868415)           'truncation': 'error',
+(TaskRunner pid=1868415)           'val_batch_size': 64,
+(TaskRunner pid=1868415)           'val_files': '/mnt/share01/sqk/AnomSeer/data/rats_uni_processed/test_full.parquet'},
+(TaskRunner pid=1868415)  'reward_model': {'enable': False,
+(TaskRunner pid=1868415)                   'forward_max_token_len_per_gpu': 32768,
+(TaskRunner pid=1868415)                   'max_length': None,
+(TaskRunner pid=1868415)                   'micro_batch_size': None,
+(TaskRunner pid=1868415)                   'micro_batch_size_per_gpu': None,
+(TaskRunner pid=1868415)                   'model': {'external_lib': None,
+(TaskRunner pid=1868415)                             'fsdp_config': {'fsdp_size': -1,
+(TaskRunner pid=1868415)                                             'param_offload': False,
+(TaskRunner pid=1868415)                                             'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=1868415)                             'input_tokenizer': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface',
+(TaskRunner pid=1868415)                             'path': '~/models/FsfairX-LLaMA3-RM-v0.1',
+(TaskRunner pid=1868415)                             'use_remove_padding': False},
+(TaskRunner pid=1868415)                   'reward_manager': 'naive',
+(TaskRunner pid=1868415)                   'strategy': 'fsdp',
+(TaskRunner pid=1868415)                   'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=1868415)                   'use_dynamic_bsz': False},
+(TaskRunner pid=1868415)  'trainer': {'balance_batch': True,
+(TaskRunner pid=1868415)              'critic_warmup': 0,
+(TaskRunner pid=1868415)              'default_hdfs_dir': None,
+(TaskRunner pid=1868415)              'default_local_dir': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3',
+(TaskRunner pid=1868415)              'del_local_ckpt_after_load': False,
+(TaskRunner pid=1868415)              'experiment_name': 'anomseer_rats_uni_full_20260612_184427_exp3_full_eval',
+(TaskRunner pid=1868415)              'logger': ['console'],
+(TaskRunner pid=1868415)              'n_gpus_per_node': 2,
+(TaskRunner pid=1868415)              'nnodes': 1,
+(TaskRunner pid=1868415)              'project_name': 'anomseer',
+(TaskRunner pid=1868415)              'remove_previous_ckpt_in_save': False,
+(TaskRunner pid=1868415)              'resume_from_path': False,
+(TaskRunner pid=1868415)              'resume_mode': 'disable',
+(TaskRunner pid=1868415)              'save_freq': 500,
+(TaskRunner pid=1868415)              'test_freq': -1,
+(TaskRunner pid=1868415)              'total_epochs': 1,
+(TaskRunner pid=1868415)              'total_training_steps': None,
+(TaskRunner pid=1868415)              'val_before_train': True,
+(TaskRunner pid=1868415)              'val_generations_to_log_to_wandb': 0,
+(TaskRunner pid=1868415)              'val_only': True},
+(TaskRunner pid=1868415)  'ts': {'adv_mix': 0.3,
+(TaskRunner pid=1868415)         'clip_pooling': 'mean',
+(TaskRunner pid=1868415)         'clip_temperature': 1.0,
+(TaskRunner pid=1868415)         'ot_eps': 0.08,
+(TaskRunner pid=1868415)         'ot_n_iter': 50,
+(TaskRunner pid=1868415)         'similarity_method': 'ot',
+(TaskRunner pid=1868415)         'use_sem_orth': True}}
+(TaskRunner pid=1868415) reward_manager_name: anomts
+(TaskRunner pid=1868415) WARNING: val_batch_size is deprecated. Validation datasets are sent to inference engines as a whole batch, which will schedule the memory themselves.
+(TaskRunner pid=1868415) [validate_config] All configuration checks passed successfully!
+(TaskRunner pid=1868415) dataset len: 30266
+(TaskRunner pid=1868415) filter dataset len: 30266
+(TaskRunner pid=1868415) dataset len: 6034
+(TaskRunner pid=1868415) filter dataset len: 6034
+(TaskRunner pid=1868415) Size of train dataloader: 1891
+(TaskRunner pid=1868415) Total training steps: 1891
+(WorkerDict pid=1875911) Model config after override: Qwen2_5_VLConfig {
+(WorkerDict pid=1875911)   "architectures": [
+(WorkerDict pid=1875911)     "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=1875911)   ],
+(WorkerDict pid=1875911)   "attention_dropout": 0.0,
+(WorkerDict pid=1875911)   "eos_token_id": 151645,
+(WorkerDict pid=1875911)   "hidden_act": "silu",
+(WorkerDict pid=1875911)   "hidden_size": 2048,
+(WorkerDict pid=1875911)   "image_token_id": 151655,
+(WorkerDict pid=1875911)   "initializer_range": 0.02,
+(WorkerDict pid=1875911)   "intermediate_size": 11008,
+(WorkerDict pid=1875911)   "max_position_embeddings": 128000,
+(WorkerDict pid=1875911)   "max_window_layers": 70,
+(WorkerDict pid=1875911)   "model_type": "qwen2_5_vl",
+(WorkerDict pid=1875911)   "num_attention_heads": 16,(WorkerDict pid=1875911) 
+(WorkerDict pid=1875911) 
+(WorkerDict pid=1877925) 
+(WorkerDict pid=1875911) 
+(WorkerDict pid=1877925) 
+(WorkerDict pid=1875911) [rank0]:[W615 12:30:45.599970048 ProcessGroupNCCL.cpp:4561] [PG ID 0 PG GUID 0 Rank 0]  using GPU 0 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. Specify device_ids in barrier() to force use of a particular device, or call init_process_group() with a device_id.
+
+(WorkerDict pid=1875911)   "num_hidden_layers": 36,
+(WorkerDict pid=1875911)   "num_key_value_heads": 2,
+(WorkerDict pid=1875911)   "pad_token_id": 151643,
+(WorkerDict pid=1875911)   "rms_norm_eps": 1e-06,
+(WorkerDict pid=1875911)   "rope_scaling": {
+(WorkerDict pid=1875911)     "mrope_section": [
+(WorkerDict pid=1875911)       16,
+(WorkerDict pid=1875911)       24,
+(WorkerDict pid=1875911)       24
+(WorkerDict pid=1875911)     ],
+(WorkerDict pid=1875911)     "rope_type": "default",
+(WorkerDict pid=1875911)     "type": "default"
+(WorkerDict pid=1875911)   },
+(WorkerDict pid=1875911)   "rope_theta": 1000000.0,
+(WorkerDict pid=1875911)   "sliding_window": 32768,
+(WorkerDict pid=1875911)   "text_config": {
+(WorkerDict pid=1875911)     "architectures": [
+(WorkerDict pid=1875911)       "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=1875911)     ],
+(WorkerDict pid=1875911)     "attention_dropout": 0.0,
+(WorkerDict pid=1875911)     "bos_token_id": 151643,
+(WorkerDict pid=1875911)     "eos_token_id": 151645,
+(WorkerDict pid=1875911)     "hidden_act": "silu",
+(WorkerDict pid=1875911)     "hidden_size": 2048,
+(WorkerDict pid=1875911)     "image_token_id": null,
+(WorkerDict pid=1875911)     "initializer_range": 0.02,
+(WorkerDict pid=1875911)     "intermediate_size": 11008,
+(WorkerDict pid=1875911)     "layer_types": [
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention"
+(WorkerDict pid=1875911)     ],
+(WorkerDict pid=1875911)     "max_position_embeddings": 128000,
+(WorkerDict pid=1875911)     "max_window_layers": 70,
+(WorkerDict pid=1875911)     "model_type": "qwen2_5_vl_text",
+(WorkerDict pid=1875911)     "num_attention_heads": 16,
+(WorkerDict pid=1875911)     "num_hidden_layers": 36,
+(WorkerDict pid=1875911)     "num_key_value_heads": 2,
+(WorkerDict pid=1875911)     "rms_norm_eps": 1e-06,
+(WorkerDict pid=1875911)     "rope_scaling": {
+(WorkerDict pid=1875911)       "mrope_section": [
+(WorkerDict pid=1875911)         16,
+(WorkerDict pid=1875911)         24,
+(WorkerDict pid=1875911)         24
+(WorkerDict pid=1875911)       ],
+(WorkerDict pid=1875911)       "rope_type": "default",
+(WorkerDict pid=1875911)       "type": "default"
+(WorkerDict pid=1875911)     },
+(WorkerDict pid=1875911)     "rope_theta": 1000000.0,
+(WorkerDict pid=1875911)     "sliding_window": null,
+(WorkerDict pid=1875911)     "tie_word_embeddings": true,
+(WorkerDict pid=1875911)     "torch_dtype": "float32",
+(WorkerDict pid=1875911)     "use_cache": true,
+(WorkerDict pid=1875911)     "use_sliding_window": false,
+(WorkerDict pid=1875911)     "video_token_id": null,
+(WorkerDict pid=1875911)     "vision_end_token_id": 151653,
+(WorkerDict pid=1875911)     "vision_start_token_id": 151652,
+(WorkerDict pid=1875911)     "vision_token_id": 151654,
+(WorkerDict pid=1875911)     "vocab_size": 151936
+(WorkerDict pid=1875911)   },
+(WorkerDict pid=1875911)   "torch_dtype": "float32",
+(WorkerDict pid=1875911)   "transformers_version": "4.54.1",
+(WorkerDict pid=1875911)   "use_cache": true,
+(WorkerDict pid=1875911)   "use_sliding_window": false,
+(WorkerDict pid=1875911)   "video_token_id": 151656,
+(WorkerDict pid=1875911)   "vision_config": {
+(WorkerDict pid=1875911)     "depth": 32,
+(WorkerDict pid=1875911)     "fullatt_block_indexes": [
+(WorkerDict pid=1875911)       7,
+(WorkerDict pid=1875911)       15,
+(WorkerDict pid=1875911)       23,
+(WorkerDict pid=1875911)       31
+(WorkerDict pid=1875911)     ],
+(WorkerDict pid=1875911)     "hidden_act": "silu",
+(WorkerDict pid=1875911)     "hidden_size": 1280,
+(WorkerDict pid=1875911)     "in_channels": 3,
+(WorkerDict pid=1875911)     "in_chans": 3,
+(WorkerDict pid=1875911)     "initializer_range": 0.02,
+(WorkerDict pid=1875911)     "intermediate_size": 3420,
+(WorkerDict pid=1875911)     "model_type": "qwen2_5_vl",
+(WorkerDict pid=1875911)     "num_heads": 16,
+(WorkerDict pid=1875911)     "out_hidden_size": 2048,
+(WorkerDict pid=1875911)     "patch_size": 14,
+(WorkerDict pid=1875911)     "spatial_merge_size": 2,
+(WorkerDict pid=1875911)     "spatial_patch_size": 14,
+(WorkerDict pid=1875911)     "temporal_patch_size": 2,
+(WorkerDict pid=1875911)     "tokens_per_second": 2,
+(WorkerDict pid=1875911)     "torch_dtype": "float32",
+(WorkerDict pid=1875911)     "window_size": 112
+(WorkerDict pid=1875911)   },
+(WorkerDict pid=1875911)   "vision_end_token_id": 151653,
+(WorkerDict pid=1875911)   "vision_start_token_id": 151652,
+(WorkerDict pid=1875911)   "vision_token_id": 151654,
+(WorkerDict pid=1875911)   "vocab_size": 151936
+(WorkerDict pid=1875911) }
+(WorkerDict pid=1875911) 
+(WorkerDict pid=1875911) NCCL version 2.21.5+cuda12.4
+(WorkerDict pid=1875911) Qwen2_5_VLForConditionalGeneration contains 3.75B parameters
+(WorkerDict pid=1875911) wrap_policy: functools.partial(<function _or_policy at 0x75feda0d8680>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x75feda0d8540>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])
+(WorkerDict pid=1875911) Actor use_remove_padding=False
+(WorkerDict pid=1877925) wrap_policy: functools.partial(<function _or_policy at 0x71cf4e4c8680>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x71cf4e4c8540>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])
+(WorkerDict pid=1875911) Model config after override: Qwen2_5_VLConfig {
+(WorkerDict pid=1875911)   "architectures": [
+(WorkerDict pid=1875911)     "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=1875911)   ],
+(WorkerDict pid=1875911)   "attention_dropout": 0.0,
+(WorkerDict pid=1875911)   "eos_token_id": 151645,
+(WorkerDict pid=1875911)   "hidden_act": "silu",
+(WorkerDict pid=1875911)   "hidden_size": 2048,
+(WorkerDict pid=1875911)   "image_token_id": 151655,
+(WorkerDict pid=1875911)   "initializer_range": 0.02,
+(WorkerDict pid=1875911)   "intermediate_size": 11008,
+(WorkerDict pid=1875911)   "max_position_embeddings": 128000,
+(WorkerDict pid=1875911)   "max_window_layers": 70,
+(WorkerDict pid=1875911)   "model_type": "qwen2_5_vl",
+(WorkerDict pid=1875911)   "num_attention_heads": 16,
+(WorkerDict pid=1875911)   "num_hidden_layers": 36,
+(WorkerDict pid=1875911)   "num_key_value_heads": 2,
+(WorkerDict pid=1875911)   "pad_token_id": 151643,
+(WorkerDict pid=1875911)   "rms_norm_eps": 1e-06,
+(WorkerDict pid=1875911)   "rope_scaling": {
+(WorkerDict pid=1875911)     "mrope_section": [
+(WorkerDict pid=1875911)       16,
+(WorkerDict pid=1875911)       24,
+(WorkerDict pid=1875911)       24
+(WorkerDict pid=1875911)     ],
+(WorkerDict pid=1875911)     "rope_type": "default",
+(WorkerDict pid=1875911)     "type": "default"
+(WorkerDict pid=1875911)   },
+(WorkerDict pid=1875911)   "rope_theta": 1000000.0,
+(WorkerDict pid=1875911)   "sliding_window": 32768,
+(WorkerDict pid=1875911)   "text_config": {
+(WorkerDict pid=1875911)     "architectures": [
+(WorkerDict pid=1875911)       "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=1875911)     ],
+(WorkerDict pid=1875911)     "attention_dropout": 0.0,
+(WorkerDict pid=1875911)     "bos_token_id": 151643,
+(WorkerDict pid=1875911)     "eos_token_id": 151645,
+(WorkerDict pid=1875911)     "hidden_act": "silu",
+(WorkerDict pid=1875911)     "hidden_size": 2048,
+(WorkerDict pid=1875911)     "image_token_id": null,
+(WorkerDict pid=1875911)     "initializer_range": 0.02,
+(WorkerDict pid=1875911)     "intermediate_size": 11008,
+(WorkerDict pid=1875911)     "layer_types": [
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention",
+(WorkerDict pid=1875911)       "full_attention"
+(WorkerDict pid=1875911)     ],
+(WorkerDict pid=1875911)     "max_position_embeddings": 128000,
+(WorkerDict pid=1875911)     "max_window_layers": 70,
+(WorkerDict pid=1875911)     "model_type": "qwen2_5_vl_text",
+(WorkerDict pid=1875911)     "num_attention_heads": 16,
+(WorkerDict pid=1875911)     "num_hidden_layers": 36,
+(WorkerDict pid=1875911)     "num_key_value_heads": 2,
+(WorkerDict pid=1875911)     "rms_norm_eps": 1e-06,
+(WorkerDict pid=1875911)     "rope_scaling": {
+(WorkerDict pid=1875911)       "mrope_section": [
+(WorkerDict pid=1875911)         16,
+(WorkerDict pid=1875911)         24,
+(WorkerDict pid=1875911)         24
+(WorkerDict pid=1875911)       ],
+(WorkerDict pid=1875911)       "rope_type": "default",
+(WorkerDict pid=1875911)       "type": "default"
+(WorkerDict pid=1875911)     },
+(WorkerDict pid=1875911)     "rope_theta": 1000000.0,
+(WorkerDict pid=1875911)     "sliding_window": null,
+(WorkerDict pid=1875911)     "tie_word_embeddings": true,
+(WorkerDict pid=1875911)     "torch_dtype": "float32",
+(WorkerDict pid=1875911)     "use_cache": true,
+(WorkerDict pid=1875911)     "use_sliding_window": false,
+(WorkerDict pid=1875911)     "video_token_id": null,
+(WorkerDict pid=1875911)     "vision_end_token_id": 151653,
+(WorkerDict pid=1875911)     "vision_start_token_id": 151652,
+(WorkerDict pid=1875911)     "vision_token_id": 151654,
+(WorkerDict pid=1875911)     "vocab_size": 151936
+(WorkerDict pid=1875911)   },
+(WorkerDict pid=1875911)   "torch_dtype": "float32",
+(WorkerDict pid=1875911)   "transformers_version": "4.54.1",
+(WorkerDict pid=1875911)   "use_cache": true,
+(WorkerDict pid=1875911)   "use_sliding_window": false,
+(WorkerDict pid=1875911)   "video_token_id": 151656,
+(WorkerDict pid=1875911)   "vision_config": {
+(WorkerDict pid=1875911)     "depth": 32,
+(WorkerDict pid=1875911)     "fullatt_block_indexes": [
+(WorkerDict pid=1875911)       7,
+(WorkerDict pid=1875911)       15,
+(WorkerDict pid=1875911)       23,
+(WorkerDict pid=1875911)       31
+(WorkerDict pid=1875911)     ],
+(WorkerDict pid=1875911)     "hidden_act": "silu",
+(WorkerDict pid=1875911)     "hidden_size": 1280,
+(WorkerDict pid=1875911)     "in_channels": 3,
+(WorkerDict pid=1875911)     "in_chans": 3,
+(WorkerDict pid=1875911)     "initializer_range": 0.02,
+(WorkerDict pid=1875911)     "intermediate_size": 3420,
+(WorkerDict pid=1875911)     "model_type": "qwen2_5_vl",
+(WorkerDict pid=1875911)     "num_heads": 16,
+(WorkerDict pid=1875911)     "out_hidden_size": 2048,
+(WorkerDict pid=1875911)     "patch_size": 14,
+(WorkerDict pid=1875911)     "spatial_merge_size": 2,
+(WorkerDict pid=1875911)     "spatial_patch_size": 14,
+(WorkerDict pid=1875911)     "temporal_patch_size": 2,
+(WorkerDict pid=1875911)     "tokens_per_second": 2,
+(WorkerDict pid=1875911)     "torch_dtype": "float32",
+(WorkerDict pid=1875911)     "window_size": 112
+(WorkerDict pid=1875911)   },
+(WorkerDict pid=1875911)   "vision_end_token_id": 151653,
+(WorkerDict pid=1875911)   "vision_start_token_id": 151652,
+(WorkerDict pid=1875911)   "vision_token_id": 151654,(WorkerDict pid=1875911) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLForConditionalGeneration is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=1875911) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=1875911) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VisionTransformerPretrainedModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=1875911) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLTextModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=1877925) 
+(WorkerDict pid=1877925) [rank1]:[W615 12:30:45.595731994 ProcessGroupNCCL.cpp:4561] [PG ID 0 PG GUID 0 Rank 1]  using GPU 0 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. Specify device_ids in barrier() to force use of a particular device, or call init_process_group() with a device_id.
+(WorkerDict pid=1875911) 
+(WorkerDict pid=1875911) 
+(WorkerDict pid=1877925) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLTextModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)` [repeated 4x across cluster] (Ray deduplicates logs by default. Set RAY_DEDUP_LOGS=0 to disable log deduplication, or see https://docs.ray.io/en/master/ray-observability/user-guides/configure-logging.html#log-deduplication for more options.)
+(WorkerDict pid=1877925) 
+(WorkerDict pid=1875911) 
+(WorkerDict pid=1877925) 
+(WorkerDict pid=1875911) 
+(WorkerDict pid=1877925) 
+(WorkerDict pid=1875911) 
+(WorkerDict pid=1875911) 
+(WorkerDict pid=1875911) 
+(WorkerDict pid=1875911) 
+(WorkerDict pid=1875911) /home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/torch/distributed/fsdp/fully_sharded_data_parallel.py:690: FutureWarning: FSDP.state_dict_type() and FSDP.set_state_dict_type() are being deprecated. Please use APIs, get_state_dict() and set_state_dict(), which can support different parallelisms, FSDP1, FSDP2, DDP. API doc: https://pytorch.org/docs/stable/distributed.checkpoint.html#torch.distributed.checkpoint.state_dict.get_state_dict .Tutorial: https://pytorch.org/tutorials/recipes/distributed_checkpoint_recipe.html .
+(WorkerDict pid=1875911)   warnings.warn(
+
+(WorkerDict pid=1875911)   "vocab_size": 151936
+(WorkerDict pid=1875911) }
+(WorkerDict pid=1875911) 
+(WorkerDict pid=1875911) Qwen2_5_VLForConditionalGeneration contains 3.75B parameters
+(WorkerDict pid=1875911) wrap_policy: functools.partial(<function _or_policy at 0x75feda0d8680>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x75feda0d8540>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])
+(WorkerDict pid=1877925) Actor use_remove_padding=False
+(WorkerDict pid=1875911) Total steps: 1891, num_warmup_steps: 0
+(WorkerDict pid=1875911) Actor use_remove_padding=False
+(WorkerDict pid=1875911) Before building vllm rollout, memory allocated (GB): 7.027174949645996, memory reserved (GB): 12.189453125
+(WorkerDict pid=1875911) WARNING 06-15 12:33:19 [utils.py:2522] Methods determine_num_available_blocks,device_config,get_cache_block_size_bytes,initialize_cache not implemented in <vllm.v1.worker.gpu_worker.Worker object at 0x75fad74a9210>
+(WorkerDict pid=1877925) wrap_policy: functools.partial(<function _or_policy at 0x71cf4e4c8680>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x71cf4e4c8540>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])
+(WorkerDict pid=1877925) Total steps: 1891, num_warmup_steps: 0
+(WorkerDict pid=1877925) Actor use_remove_padding=False
+(WorkerDict pid=1875911) WARNING 06-15 12:33:24 [topk_topp_sampler.py:69] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
+(WorkerDict pid=1877925) WARNING 06-15 12:33:19 [utils.py:2522] Methods determine_num_available_blocks,device_config,get_cache_block_size_bytes,initialize_cache not implemented in <vllm.v1.worker.gpu_worker.Worker object at 0x71cbd0b48850>
+(WorkerDict pid=1875911) kwargs: {'n': 5, 'logprobs': 0, 'max_tokens': 1024, 'detokenize': False, 'temperature': 1.0, 'top_k': -1, 'top_p': 1, 'ignore_eos': False}
+(WorkerDict pid=1875911) After building vllm rollout, memory allocated (GB): 13.633802890777588, memory reserved (GB): 15.04296875
+(WorkerDict pid=1875911) After building sharding manager, memory allocated (GB): 13.633802890777588, memory reserved (GB): 15.04296875
+(WorkerDict pid=1877925) WARNING 06-15 12:33:24 [topk_topp_sampler.py:69] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
+(TaskRunner pid=1868415) Using LocalLogger is deprecated. The constructor API will change 
+(TaskRunner pid=1868415) validation generation end
+(WorkerDict pid=1877925) kwargs: {'n': 5, 'logprobs': 0, 'max_tokens': 1024, 'detokenize': False, 'temperature': 1.0, 'top_k': -1, 'top_p': 1, 'ignore_eos': False}
+(TaskRunner pid=1868415) [prompt] <|im_start|>system
+(TaskRunner pid=1868415) You are a helpful assistant.<|im_end|>
+(TaskRunner pid=1868415) <|im_start|>user
+(TaskRunner pid=1868415) <|vision_start|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|vision_end|>
+(TaskRunner pid=1868415) You are an expert in univariate time-series anomaly detection. The figure shows a single-channel time series of length 128 from the "Healthcare-ECG" domain.
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) Decide whether the series is normal or contains an anomaly. If it is anomalous, choose the single most appropriate anomaly type from the following 15 categories (format `id: name — description`):
+(TaskRunner pid=1868415) 0: Normal Sequence — There are no abnormal situations in this time series.
+(TaskRunner pid=1868415) 1: Point Anomaly — A single data point significantly deviates from the local or global pattern of the sequence.
+(TaskRunner pid=1868415) 2: Periodic Change Anomaly — The original periodic pattern is disrupted, e.g. the period is broken or the amplitude becomes anomalous.
+(TaskRunner pid=1868415) 3: Trend Change Anomaly — A sudden change in the long-term trend of the time series.
+(TaskRunner pid=1868415) 4: Change Point Anomaly — Statistical properties (e.g. mean, variance) change abruptly at certain points.
+(TaskRunner pid=1868415) 5: Distributional Change Anomaly — The statistical distribution of the time series changes significantly.
+(TaskRunner pid=1868415) 6: Amplitude Anomaly — The amplitude of data points exceeds the normal upper and lower bounds.
+(TaskRunner pid=1868415) 7: Pattern Change Anomaly — The pattern of the time series suddenly changes from one form to another.
+(TaskRunner pid=1868415) 8: Sparse Anomaly — Isolated anomalous patterns occasionally appear in a long time series.
+(TaskRunner pid=1868415) 9: Repeated Value Anomaly — Continuous or intermittent repeated values disrupt the normal fluctuation pattern.
+(TaskRunner pid=1868415) 10: Sudden Flatline Anomaly — The time series suddenly becomes a flat line with no normal fluctuations.
+(TaskRunner pid=1868415) 11: Drift Anomaly — The data gradually drifts away from the normal level.
+(TaskRunner pid=1868415) 12: Sudden Spike Anomaly — The data suddenly spikes or drops within a short time and then returns to normal.
+(TaskRunner pid=1868415) 13: Continuous Segment Anomaly — A continuous segment of data points deviates from the normal pattern.
+(TaskRunner pid=1868415) 14: Nonlinear Pattern Anomaly — Nonlinear changes appear in the sequence, breaking the original linear rule.
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) Reason step by step inside <think>...</think> based on the visual shape of the series, then output exactly one line with your final answer:
+(TaskRunner pid=1868415) <class>one exact category name from the list above</class>
+(TaskRunner pid=1868415) If the series is normal, use <class>Normal Sequence</class>.<|im_end|>
+(TaskRunner pid=1868415) <|im_start|>assistant
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) [response] <think></think>
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think>
+(TaskRunner pid=1868415) Detected new reward function format (3 return values).
+(TaskRunner pid=1868415) validation generation end
+(TaskRunner pid=1868415) [prompt] <|im_start|>system
+(TaskRunner pid=1868415) You are a helpful assistant.<|im_end|>
+(TaskRunner pid=1868415) <|im_start|>user
+(TaskRunner pid=1868415) <|vision_start|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|vision_end|>
+(TaskRunner pid=1868415) You are an expert in univariate time-series anomaly detection. The figure shows a single-channel time series of length 32 from the "Server-YAHOO" domain.
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) Decide whether the series is normal or contains an anomaly. If it is anomalous, choose the single most appropriate anomaly type from the following 15 categories (format `id: name — description`):
+(TaskRunner pid=1868415) 0: Normal Sequence — There are no abnormal situations in this time series.
+(TaskRunner pid=1868415) 1: Point Anomaly — A single data point significantly deviates from the local or global pattern of the sequence.
+(TaskRunner pid=1868415) 2: Periodic Change Anomaly — The original periodic pattern is disrupted, e.g. the period is broken or the amplitude becomes anomalous.
+(TaskRunner pid=1868415) 3: Trend Change Anomaly — A sudden change in the long-term trend of the time series.
+(TaskRunner pid=1868415) 4: Change Point Anomaly — Statistical properties (e.g. mean, variance) change abruptly at certain points.
+(TaskRunner pid=1868415) 5: Distributional Change Anomaly — The statistical distribution of the time series changes significantly.
+(TaskRunner pid=1868415) 6: Amplitude Anomaly — The amplitude of data points exceeds the normal upper and lower bounds.
+(TaskRunner pid=1868415) 7: Pattern Change Anomaly — The pattern of the time series suddenly changes from one form to another.
+(TaskRunner pid=1868415) 8: Sparse Anomaly — Isolated anomalous patterns occasionally appear in a long time series.
+(TaskRunner pid=1868415) 9: Repeated Value Anomaly — Continuous or intermittent repeated values disrupt the normal fluctuation pattern.
+(TaskRunner pid=1868415) 10: Sudden Flatline Anomaly — The time series suddenly becomes a flat line with no normal fluctuations.
+(TaskRunner pid=1868415) 11: Drift Anomaly — The data gradually drifts away from the normal level.
+(TaskRunner pid=1868415) 12: Sudden Spike Anomaly — The data suddenly spikes or drops within a short time and then returns to normal.
+(TaskRunner pid=1868415) 13: Continuous Segment Anomaly — A continuous segment of data points deviates from the normal pattern.
+(TaskRunner pid=1868415) 14: Nonlinear Pattern Anomaly — Nonlinear changes appear in the sequence, breaking the original linear rule.
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) Reason step by step inside <think>...</think> based on the visual shape of the series, then output exactly one line with your final answer:
+(TaskRunner pid=1868415) <class>one exact category name from the list above</class>
+(TaskRunner pid=1868415) If the series is normal, use <class>Normal Sequence</class>.<|im_end|>
+(TaskRunner pid=1868415) <|im_start|>assistant
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) [response] <think></think>
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think>
+(TaskRunner pid=1868415) validation generation end
+(TaskRunner pid=1868415) [prompt] <|im_start|>system
+(TaskRunner pid=1868415) You are a helpful assistant.<|im_end|>
+(TaskRunner pid=1868415) <|im_start|>user
+(TaskRunner pid=1868415) <|vision_start|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|vision_end|>
+(TaskRunner pid=1868415) You are an expert in univariate time-series anomaly detection. The figure shows a single-channel time series of length 128 from the "Healthcare-ECG" domain.
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) Decide whether the series is normal or contains an anomaly. If it is anomalous, choose the single most appropriate anomaly type from the following 15 categories (format `id: name — description`):
+(TaskRunner pid=1868415) 0: Normal Sequence — There are no abnormal situations in this time series.
+(TaskRunner pid=1868415) 1: Point Anomaly — A single data point significantly deviates from the local or global pattern of the sequence.
+(TaskRunner pid=1868415) 2: Periodic Change Anomaly — The original periodic pattern is disrupted, e.g. the period is broken or the amplitude becomes anomalous.
+(TaskRunner pid=1868415) 3: Trend Change Anomaly — A sudden change in the long-term trend of the time series.
+(TaskRunner pid=1868415) 4: Change Point Anomaly — Statistical properties (e.g. mean, variance) change abruptly at certain points.
+(TaskRunner pid=1868415) 5: Distributional Change Anomaly — The statistical distribution of the time series changes significantly.
+(TaskRunner pid=1868415) 6: Amplitude Anomaly — The amplitude of data points exceeds the normal upper and lower bounds.
+(TaskRunner pid=1868415) 7: Pattern Change Anomaly — The pattern of the time series suddenly changes from one form to another.
+(TaskRunner pid=1868415) 8: Sparse Anomaly — Isolated anomalous patterns occasionally appear in a long time series.
+(TaskRunner pid=1868415) 9: Repeated Value Anomaly — Continuous or intermittent repeated values disrupt the normal fluctuation pattern.
+(TaskRunner pid=1868415) 10: Sudden Flatline Anomaly — The time series suddenly becomes a flat line with no normal fluctuations.
+(TaskRunner pid=1868415) 11: Drift Anomaly — The data gradually drifts away from the normal level.
+(TaskRunner pid=1868415) 12: Sudden Spike Anomaly — The data suddenly spikes or drops within a short time and then returns to normal.
+(TaskRunner pid=1868415) 13: Continuous Segment Anomaly — A continuous segment of data points deviates from the normal pattern.
+(TaskRunner pid=1868415) 14: Nonlinear Pattern Anomaly — Nonlinear changes appear in the sequence, breaking the original linear rule.
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) Reason step by step inside <think>...</think> based on the visual shape of the series, then output exactly one line with your final answer:
+(TaskRunner pid=1868415) <class>one exact category name from the list above</class>
+(TaskRunner pid=1868415) If the series is normal, use <class>Normal Sequence</class>.<|im_end|>
+(TaskRunner pid=1868415) <|im_start|>assistant
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) [response] <think></think>
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1868415) 
+(TaskRunner pid=1868415) <think></think> 

logs/anomseer_rats_uni_full_20260612_184427_exp3/rats_eval_lora_smoke_fixed_20260615.log

@@ -0,0 +1,635 @@
+[log] saving full output to: /mnt/share01/sqk/AnomSeer/logs/anomseer_rats_uni_full_20260612_184427_exp3/rats_eval_lora_smoke_fixed_20260615.log
+[config] tuning=LoRA (r=16, alpha=16, target=q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj) | lr=1e-4 | gpus=2 | tensor-parallel (TP=2)
+[config] model=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface
+[config] python=/home/suiqk/anaconda3/envs/scalerag-ts-v4/bin/python | stage=eval
+[stage] EVAL   (model=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface, gpus=2, tp=2, lora_rank=0)
+2026-06-15 11:18:02,410	INFO worker.py:2012 -- Started a local Ray instance.
+/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/worker.py:2051: FutureWarning: Tip: In future versions of Ray, Ray will no longer override accelerator visible devices env var if num_gpus=0 or num_gpus=None (default). To enable this behavior and turn off this error message, set RAY_ACCEL_ENV_VAR_OVERRIDE_ON_ZERO=0
+  warnings.warn(
+(TaskRunner pid=1165231) {'actor_rollout_ref': {'actor': {'clip_ratio': 0.2,
+(TaskRunner pid=1165231)                                  'entropy_coeff': 0.001,
+(TaskRunner pid=1165231)                                  'fsdp_config': {'fsdp_size': -1,
+(TaskRunner pid=1165231)                                                  'optimizer_offload': False,
+(TaskRunner pid=1165231)                                                  'param_offload': False,
+(TaskRunner pid=1165231)                                                  'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=1165231)                                  'grad_clip': 1.0,
+(TaskRunner pid=1165231)                                  'kl_loss_coef': 0.001,
+(TaskRunner pid=1165231)                                  'kl_loss_type': 'low_var_kl',
+(TaskRunner pid=1165231)                                  'optim': {'lr': 0.0001,
+(TaskRunner pid=1165231)                                            'lr_warmup_steps': -1,
+(TaskRunner pid=1165231)                                            'lr_warmup_steps_ratio': 0.0,
+(TaskRunner pid=1165231)                                            'min_lr_ratio': None,
+(TaskRunner pid=1165231)                                            'total_training_steps': -1,
+(TaskRunner pid=1165231)                                            'warmup_style': 'constant'},
+(TaskRunner pid=1165231)                                  'ppo_epochs': 1,
+(TaskRunner pid=1165231)                                  'ppo_max_token_len_per_gpu': 16384,
+(TaskRunner pid=1165231)                                  'ppo_micro_batch_size': None,
+(TaskRunner pid=1165231)                                  'ppo_micro_batch_size_per_gpu': 2,
+(TaskRunner pid=1165231)                                  'ppo_mini_batch_size': 16,
+(TaskRunner pid=1165231)                                  'shuffle': False,
+(TaskRunner pid=1165231)                                  'strategy': 'fsdp',
+(TaskRunner pid=1165231)                                  'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=1165231)                                  'use_dynamic_bsz': False,
+(TaskRunner pid=1165231)                                  'use_kl_loss': True,
+(TaskRunner pid=1165231)                                  'use_torch_compile': True},
+(TaskRunner pid=1165231)                        'hybrid_engine': True,
+(TaskRunner pid=1165231)                        'model': {'enable_gradient_checkpointing': True,
+(TaskRunner pid=1165231)                                  'external_lib': None,
+(TaskRunner pid=1165231)                                  'lora_alpha': 16,
+(TaskRunner pid=1165231)                                  'lora_dropout': 0.0,
+(TaskRunner pid=1165231)                                  'lora_rank': 0,
+(TaskRunner pid=1165231)                                  'lora_target_modules': 'q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj',
+(TaskRunner pid=1165231)                                  'override_config': {},
+(TaskRunner pid=1165231)                                  'path': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface',
+(TaskRunner pid=1165231)                                  'use_remove_padding': False},
+(TaskRunner pid=1165231)                        'ref': {'fsdp_config': {'param_offload': True,
+(TaskRunner pid=1165231)                                                'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=1165231)                                'log_prob_max_token_len_per_gpu': 16384,
+(TaskRunner pid=1165231)                                'log_prob_micro_batch_size': None,
+(TaskRunner pid=1165231)                                'log_prob_micro_batch_size_per_gpu': 8,
+(TaskRunner pid=1165231)                                'log_prob_use_dynamic_bsz': False,
+(TaskRunner pid=1165231)                                'ulysses_sequence_parallel_size': 1},
+(TaskRunner pid=1165231)                        'rollout': {'disable_log_stats': True,
+(TaskRunner pid=1165231)                                    'do_sample': True,
+(TaskRunner pid=1165231)                                    'dtype': 'bfloat16',
+(TaskRunner pid=1165231)                                    'enable_chunked_prefill': False,
+(TaskRunner pid=1165231)                                    'enforce_eager': False,
+(TaskRunner pid=1165231)                                    'free_cache_engine': False,
+(TaskRunner pid=1165231)                                    'gpu_memory_utilization': 0.4,
+(TaskRunner pid=1165231)                                    'ignore_eos': False,
+(TaskRunner pid=1165231)                                    'load_format': 'dummy_dtensor',
+(TaskRunner pid=1165231)                                    'log_prob_max_token_len_per_gpu': 16384,
+(TaskRunner pid=1165231)                                    'log_prob_micro_batch_size': None,
+(TaskRunner pid=1165231)                                    'log_prob_micro_batch_size_per_gpu': 8,
+(TaskRunner pid=1165231)                                    'log_prob_use_dynamic_bsz': False,
+(TaskRunner pid=1165231)                                    'max_model_len': None,
+(TaskRunner pid=1165231)                                    'max_num_batched_tokens': 8192,
+(TaskRunner pid=1165231)                                    'max_num_seqs': 1024,
+(TaskRunner pid=1165231)                                    'n': 5,
+(TaskRunner pid=1165231)                                    'name': 'vllm',
+(TaskRunner pid=1165231)                                    'prompt_length': 1024,
+(TaskRunner pid=1165231)                                    'response_length': 1024,
+(TaskRunner pid=1165231)                                    'temperature': 1.0,
+(TaskRunner pid=1165231)                                    'tensor_model_parallel_size': 2,
+(TaskRunner pid=1165231)                                    'top_k': -1,
+(TaskRunner pid=1165231)                                    'top_p': 1,
+(TaskRunner pid=1165231)                                    'use_fire_sampling': False,
+(TaskRunner pid=1165231)                                    'val_kwargs': {'do_sample': False,
+(TaskRunner pid=1165231)                                                   'n': 1,
+(TaskRunner pid=1165231)                                                   'temperature': 0.4,
+(TaskRunner pid=1165231)                                                   'top_k': -1,
+(TaskRunner pid=1165231)                                                   'top_p': 1.0}}},
+(TaskRunner pid=1165231)  'algorithm': {'adv_estimator': 'grpo',
+(TaskRunner pid=1165231)                'gamma': 1.0,
+(TaskRunner pid=1165231)                'kl_ctrl': {'kl_coef': 0.001, 'type': 'fixed'},
+(TaskRunner pid=1165231)                'kl_penalty': 'kl',
+(TaskRunner pid=1165231)                'lam': 1.0},
+(TaskRunner pid=1165231)  'critic': {'cliprange_value': 0.5,
+(TaskRunner pid=1165231)             'forward_max_token_len_per_gpu': 32768,
+(TaskRunner pid=1165231)             'forward_micro_batch_size': None,
+(TaskRunner pid=1165231)             'forward_micro_batch_size_per_gpu': None,
+(TaskRunner pid=1165231)             'grad_clip': 1.0,
+(TaskRunner pid=1165231)             'model': {'enable_gradient_checkpointing': True,
+(TaskRunner pid=1165231)                       'external_lib': None,
+(TaskRunner pid=1165231)                       'fsdp_config': {'fsdp_size': -1,(TaskRunner pid=1165231) DeprecationWarning: `ray.state.available_resources_per_node` is a private attribute and access will be removed in a future Ray version.
+
+(TaskRunner pid=1165231)                                       'optimizer_offload': False,
+(TaskRunner pid=1165231)                                       'param_offload': False,
+(TaskRunner pid=1165231)                                       'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=1165231)                       'override_config': {},
+(TaskRunner pid=1165231)                       'path': '~/models/deepseek-llm-7b-chat',
+(TaskRunner pid=1165231)                       'tokenizer_path': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface',
+(TaskRunner pid=1165231)                       'use_remove_padding': False},
+(TaskRunner pid=1165231)             'optim': {'lr': 1e-05,
+(TaskRunner pid=1165231)                       'lr_warmup_steps_ratio': 0.0,
+(TaskRunner pid=1165231)                       'min_lr_ratio': None,
+(TaskRunner pid=1165231)                       'total_training_steps': -1,
+(TaskRunner pid=1165231)                       'warmup_style': 'constant'},
+(TaskRunner pid=1165231)             'ppo_epochs': 1,
+(TaskRunner pid=1165231)             'ppo_max_token_len_per_gpu': 32768,
+(TaskRunner pid=1165231)             'ppo_micro_batch_size': None,
+(TaskRunner pid=1165231)             'ppo_micro_batch_size_per_gpu': None,
+(TaskRunner pid=1165231)             'ppo_mini_batch_size': 16,
+(TaskRunner pid=1165231)             'shuffle': False,
+(TaskRunner pid=1165231)             'strategy': 'fsdp',
+(TaskRunner pid=1165231)             'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=1165231)             'use_dynamic_bsz': False},
+(TaskRunner pid=1165231)  'custom_reward_function': {'name': 'compute_score', 'path': None},
+(TaskRunner pid=1165231)  'data': {'filter_overlong_prompts': True,
+(TaskRunner pid=1165231)           'image_key': 'images',
+(TaskRunner pid=1165231)           'max_prompt_length': 1024,
+(TaskRunner pid=1165231)           'max_response_length': 1024,
+(TaskRunner pid=1165231)           'prompt_key': 'prompt',
+(TaskRunner pid=1165231)           'return_raw_chat': False,
+(TaskRunner pid=1165231)           'return_raw_input_ids': False,
+(TaskRunner pid=1165231)           'shuffle': True,
+(TaskRunner pid=1165231)           'tokenizer': None,
+(TaskRunner pid=1165231)           'train_batch_size': 16,
+(TaskRunner pid=1165231)           'train_files': '/mnt/share01/sqk/AnomSeer/data/rats_uni_processed/train_quarter.parquet',
+(TaskRunner pid=1165231)           'truncation': 'error',
+(TaskRunner pid=1165231)           'val_batch_size': 1,
+(TaskRunner pid=1165231)           'val_files': '/tmp/anomseer_eval_smoke.parquet'},
+(TaskRunner pid=1165231)  'reward_model': {'enable': False,
+(TaskRunner pid=1165231)                   'forward_max_token_len_per_gpu': 32768,
+(TaskRunner pid=1165231)                   'max_length': None,
+(TaskRunner pid=1165231)                   'micro_batch_size': None,
+(TaskRunner pid=1165231)                   'micro_batch_size_per_gpu': None,
+(TaskRunner pid=1165231)                   'model': {'external_lib': None,
+(TaskRunner pid=1165231)                             'fsdp_config': {'fsdp_size': -1,
+(TaskRunner pid=1165231)                                             'param_offload': False,
+(TaskRunner pid=1165231)                                             'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=1165231)                             'input_tokenizer': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface',
+(TaskRunner pid=1165231)                             'path': '~/models/FsfairX-LLaMA3-RM-v0.1',
+(TaskRunner pid=1165231)                             'use_remove_padding': False},
+(TaskRunner pid=1165231)                   'reward_manager': 'naive',
+(TaskRunner pid=1165231)                   'strategy': 'fsdp',
+(TaskRunner pid=1165231)                   'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=1165231)                   'use_dynamic_bsz': False},
+(TaskRunner pid=1165231)  'trainer': {'balance_batch': True,
+(TaskRunner pid=1165231)              'critic_warmup': 0,
+(TaskRunner pid=1165231)              'default_hdfs_dir': None,
+(TaskRunner pid=1165231)              'default_local_dir': 'checkpoints/anomseer/anomseer_rats_uni_2gpu',
+(TaskRunner pid=1165231)              'del_local_ckpt_after_load': False,
+(TaskRunner pid=1165231)              'experiment_name': 'anomseer_rats_uni_2gpu',
+(TaskRunner pid=1165231)              'logger': ['console'],
+(TaskRunner pid=1165231)              'n_gpus_per_node': 2,
+(TaskRunner pid=1165231)              'nnodes': 1,
+(TaskRunner pid=1165231)              'project_name': 'anomseer',
+(TaskRunner pid=1165231)              'remove_previous_ckpt_in_save': False,
+(TaskRunner pid=1165231)              'resume_from_path': False,
+(TaskRunner pid=1165231)              'resume_mode': 'disable',
+(TaskRunner pid=1165231)              'save_freq': 500,
+(TaskRunner pid=1165231)              'test_freq': -1,
+(TaskRunner pid=1165231)              'total_epochs': 1,
+(TaskRunner pid=1165231)              'total_training_steps': None,
+(TaskRunner pid=1165231)              'val_before_train': True,
+(TaskRunner pid=1165231)              'val_generations_to_log_to_wandb': 0,
+(TaskRunner pid=1165231)              'val_only': True},
+(TaskRunner pid=1165231)  'ts': {'adv_mix': 0.3,
+(TaskRunner pid=1165231)         'clip_pooling': 'mean',
+(TaskRunner pid=1165231)         'clip_temperature': 1.0,
+(TaskRunner pid=1165231)         'ot_eps': 0.08,
+(TaskRunner pid=1165231)         'ot_n_iter': 50,
+(TaskRunner pid=1165231)         'similarity_method': 'ot',
+(TaskRunner pid=1165231)         'use_sem_orth': True}}
+(TaskRunner pid=1165231) reward_manager_name: anomts
+(TaskRunner pid=1165231) WARNING: val_batch_size is deprecated. Validation datasets are sent to inference engines as a whole batch, which will schedule the memory themselves.
+(TaskRunner pid=1165231) [validate_config] All configuration checks passed successfully!
+(TaskRunner pid=1165231) dataset len: 7566
+(TaskRunner pid=1165231) filter dataset len: 7566
+(TaskRunner pid=1165231) dataset len: 1
+(TaskRunner pid=1165231) filter dataset len: 1
+(TaskRunner pid=1165231) Size of train dataloader: 472
+(TaskRunner pid=1165231) Total training steps: 472
+(WorkerDict pid=1168371) Model config after override: Qwen2_5_VLConfig {
+(WorkerDict pid=1168371)   "architectures": [
+(WorkerDict pid=1168371)     "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=1168371)   ],
+(WorkerDict pid=1168371)   "attention_dropout": 0.0,
+(WorkerDict pid=1168371)   "eos_token_id": 151645,
+(WorkerDict pid=1168371)   "hidden_act": "silu",
+(WorkerDict pid=1168371)   "hidden_size": 2048,
+(WorkerDict pid=1168371)   "image_token_id": 151655,
+(WorkerDict pid=1168371)   "initializer_range": 0.02,
+(WorkerDict pid=1168371)   "intermediate_size": 11008,
+(WorkerDict pid=1168371)   "max_position_embeddings": 128000,
+(WorkerDict pid=1168371)   "max_window_layers": 70,
+(WorkerDict pid=1168371)   "model_type": "qwen2_5_vl",
+(WorkerDict pid=1168371)   "num_attention_heads": 16,
+(WorkerDict pid=1168371)   "num_hidden_layers": 36,
+(WorkerDict pid=1168371)   "num_key_value_heads": 2,(WorkerDict pid=1168371) 
+(WorkerDict pid=1168371) 
+(WorkerDict pid=1170483) 
+(WorkerDict pid=1168371) 
+(WorkerDict pid=1170483) 
+(WorkerDict pid=1168371) [rank0]:[W615 11:19:02.122526228 ProcessGroupNCCL.cpp:4561] [PG ID 0 PG GUID 0 Rank 0]  using GPU 0 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. Specify device_ids in barrier() to force use of a particular device, or call init_process_group() with a device_id.
+
+(WorkerDict pid=1168371)   "pad_token_id": 151643,
+(WorkerDict pid=1168371)   "rms_norm_eps": 1e-06,
+(WorkerDict pid=1168371)   "rope_scaling": {
+(WorkerDict pid=1168371)     "mrope_section": [
+(WorkerDict pid=1168371)       16,
+(WorkerDict pid=1168371)       24,
+(WorkerDict pid=1168371)       24
+(WorkerDict pid=1168371)     ],
+(WorkerDict pid=1168371)     "rope_type": "default",
+(WorkerDict pid=1168371)     "type": "default"
+(WorkerDict pid=1168371)   },
+(WorkerDict pid=1168371)   "rope_theta": 1000000.0,
+(WorkerDict pid=1168371)   "sliding_window": 32768,
+(WorkerDict pid=1168371)   "text_config": {
+(WorkerDict pid=1168371)     "architectures": [
+(WorkerDict pid=1168371)       "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=1168371)     ],
+(WorkerDict pid=1168371)     "attention_dropout": 0.0,
+(WorkerDict pid=1168371)     "bos_token_id": 151643,
+(WorkerDict pid=1168371)     "eos_token_id": 151645,
+(WorkerDict pid=1168371)     "hidden_act": "silu",
+(WorkerDict pid=1168371)     "hidden_size": 2048,
+(WorkerDict pid=1168371)     "image_token_id": null,
+(WorkerDict pid=1168371)     "initializer_range": 0.02,
+(WorkerDict pid=1168371)     "intermediate_size": 11008,
+(WorkerDict pid=1168371)     "layer_types": [
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention"
+(WorkerDict pid=1168371)     ],
+(WorkerDict pid=1168371)     "max_position_embeddings": 128000,
+(WorkerDict pid=1168371)     "max_window_layers": 70,
+(WorkerDict pid=1168371)     "model_type": "qwen2_5_vl_text",
+(WorkerDict pid=1168371)     "num_attention_heads": 16,
+(WorkerDict pid=1168371)     "num_hidden_layers": 36,
+(WorkerDict pid=1168371)     "num_key_value_heads": 2,
+(WorkerDict pid=1168371)     "rms_norm_eps": 1e-06,
+(WorkerDict pid=1168371)     "rope_scaling": {
+(WorkerDict pid=1168371)       "mrope_section": [
+(WorkerDict pid=1168371)         16,
+(WorkerDict pid=1168371)         24,
+(WorkerDict pid=1168371)         24
+(WorkerDict pid=1168371)       ],
+(WorkerDict pid=1168371)       "rope_type": "default",
+(WorkerDict pid=1168371)       "type": "default"
+(WorkerDict pid=1168371)     },
+(WorkerDict pid=1168371)     "rope_theta": 1000000.0,
+(WorkerDict pid=1168371)     "sliding_window": null,
+(WorkerDict pid=1168371)     "tie_word_embeddings": true,
+(WorkerDict pid=1168371)     "torch_dtype": "float32",
+(WorkerDict pid=1168371)     "use_cache": true,
+(WorkerDict pid=1168371)     "use_sliding_window": false,
+(WorkerDict pid=1168371)     "video_token_id": null,
+(WorkerDict pid=1168371)     "vision_end_token_id": 151653,
+(WorkerDict pid=1168371)     "vision_start_token_id": 151652,
+(WorkerDict pid=1168371)     "vision_token_id": 151654,
+(WorkerDict pid=1168371)     "vocab_size": 151936
+(WorkerDict pid=1168371)   },
+(WorkerDict pid=1168371)   "torch_dtype": "float32",
+(WorkerDict pid=1168371)   "transformers_version": "4.54.1",
+(WorkerDict pid=1168371)   "use_cache": true,
+(WorkerDict pid=1168371)   "use_sliding_window": false,
+(WorkerDict pid=1168371)   "video_token_id": 151656,
+(WorkerDict pid=1168371)   "vision_config": {
+(WorkerDict pid=1168371)     "depth": 32,
+(WorkerDict pid=1168371)     "fullatt_block_indexes": [
+(WorkerDict pid=1168371)       7,
+(WorkerDict pid=1168371)       15,
+(WorkerDict pid=1168371)       23,
+(WorkerDict pid=1168371)       31
+(WorkerDict pid=1168371)     ],
+(WorkerDict pid=1168371)     "hidden_act": "silu",
+(WorkerDict pid=1168371)     "hidden_size": 1280,
+(WorkerDict pid=1168371)     "in_channels": 3,
+(WorkerDict pid=1168371)     "in_chans": 3,
+(WorkerDict pid=1168371)     "initializer_range": 0.02,
+(WorkerDict pid=1168371)     "intermediate_size": 3420,
+(WorkerDict pid=1168371)     "model_type": "qwen2_5_vl",
+(WorkerDict pid=1168371)     "num_heads": 16,
+(WorkerDict pid=1168371)     "out_hidden_size": 2048,
+(WorkerDict pid=1168371)     "patch_size": 14,
+(WorkerDict pid=1168371)     "spatial_merge_size": 2,
+(WorkerDict pid=1168371)     "spatial_patch_size": 14,
+(WorkerDict pid=1168371)     "temporal_patch_size": 2,
+(WorkerDict pid=1168371)     "tokens_per_second": 2,
+(WorkerDict pid=1168371)     "torch_dtype": "float32",
+(WorkerDict pid=1168371)     "window_size": 112
+(WorkerDict pid=1168371)   },
+(WorkerDict pid=1168371)   "vision_end_token_id": 151653,
+(WorkerDict pid=1168371)   "vision_start_token_id": 151652,
+(WorkerDict pid=1168371)   "vision_token_id": 151654,
+(WorkerDict pid=1168371)   "vocab_size": 151936
+(WorkerDict pid=1168371) }
+(WorkerDict pid=1168371) 
+(WorkerDict pid=1168371) NCCL version 2.21.5+cuda12.4
+(WorkerDict pid=1168371) Qwen2_5_VLForConditionalGeneration contains 3.75B parameters
+(WorkerDict pid=1168371) wrap_policy: functools.partial(<function _or_policy at 0x7af6274bc5e0>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x7af6274bc4a0>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>})])
+(WorkerDict pid=1170483) Actor use_remove_padding=False
+(WorkerDict pid=1170483) wrap_policy: functools.partial(<function _or_policy at 0x78426bee45e0>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x78426bee44a0>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])(WorkerDict pid=1168371) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLForConditionalGeneration is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=1168371) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=1168371) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VisionTransformerPretrainedModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=1170483) 
+(WorkerDict pid=1170483) [rank1]:[W615 11:19:02.127933277 ProcessGroupNCCL.cpp:4561] [PG ID 0 PG GUID 0 Rank 1]  using GPU 0 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. Specify device_ids in barrier() to force use of a particular device, or call init_process_group() with a device_id.
+
+(WorkerDict pid=1168371) Model config after override: Qwen2_5_VLConfig {
+(WorkerDict pid=1168371)   "architectures": [
+(WorkerDict pid=1168371)     "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=1168371)   ],
+(WorkerDict pid=1168371)   "attention_dropout": 0.0,
+(WorkerDict pid=1168371)   "eos_token_id": 151645,
+(WorkerDict pid=1168371)   "hidden_act": "silu",
+(WorkerDict pid=1168371)   "hidden_size": 2048,
+(WorkerDict pid=1168371)   "image_token_id": 151655,
+(WorkerDict pid=1168371)   "initializer_range": 0.02,
+(WorkerDict pid=1168371)   "intermediate_size": 11008,
+(WorkerDict pid=1168371)   "max_position_embeddings": 128000,
+(WorkerDict pid=1168371)   "max_window_layers": 70,
+(WorkerDict pid=1168371)   "model_type": "qwen2_5_vl",
+(WorkerDict pid=1168371)   "num_attention_heads": 16,
+(WorkerDict pid=1168371)   "num_hidden_layers": 36,
+(WorkerDict pid=1168371)   "num_key_value_heads": 2,
+(WorkerDict pid=1168371)   "pad_token_id": 151643,
+(WorkerDict pid=1168371)   "rms_norm_eps": 1e-06,
+(WorkerDict pid=1168371)   "rope_scaling": {
+(WorkerDict pid=1168371)     "mrope_section": [
+(WorkerDict pid=1168371)       16,
+(WorkerDict pid=1168371)       24,
+(WorkerDict pid=1168371)       24
+(WorkerDict pid=1168371)     ],
+(WorkerDict pid=1168371)     "rope_type": "default",
+(WorkerDict pid=1168371)     "type": "default"
+(WorkerDict pid=1168371)   },
+(WorkerDict pid=1168371)   "rope_theta": 1000000.0,
+(WorkerDict pid=1168371)   "sliding_window": 32768,
+(WorkerDict pid=1168371)   "text_config": {
+(WorkerDict pid=1168371)     "architectures": [
+(WorkerDict pid=1168371)       "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=1168371)     ],
+(WorkerDict pid=1168371)     "attention_dropout": 0.0,
+(WorkerDict pid=1168371)     "bos_token_id": 151643,
+(WorkerDict pid=1168371)     "eos_token_id": 151645,
+(WorkerDict pid=1168371)     "hidden_act": "silu",
+(WorkerDict pid=1168371)     "hidden_size": 2048,
+(WorkerDict pid=1168371)     "image_token_id": null,
+(WorkerDict pid=1168371)     "initializer_range": 0.02,
+(WorkerDict pid=1168371)     "intermediate_size": 11008,
+(WorkerDict pid=1168371)     "layer_types": [
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention",
+(WorkerDict pid=1168371)       "full_attention"
+(WorkerDict pid=1168371)     ],
+(WorkerDict pid=1168371)     "max_position_embeddings": 128000,
+(WorkerDict pid=1168371)     "max_window_layers": 70,
+(WorkerDict pid=1168371)     "model_type": "qwen2_5_vl_text",
+(WorkerDict pid=1168371)     "num_attention_heads": 16,
+(WorkerDict pid=1168371)     "num_hidden_layers": 36,
+(WorkerDict pid=1168371)     "num_key_value_heads": 2,
+(WorkerDict pid=1168371)     "rms_norm_eps": 1e-06,
+(WorkerDict pid=1168371)     "rope_scaling": {
+(WorkerDict pid=1168371)       "mrope_section": [
+(WorkerDict pid=1168371)         16,
+(WorkerDict pid=1168371)         24,
+(WorkerDict pid=1168371)         24
+(WorkerDict pid=1168371)       ],
+(WorkerDict pid=1168371)       "rope_type": "default",
+(WorkerDict pid=1168371)       "type": "default"
+(WorkerDict pid=1168371)     },
+(WorkerDict pid=1168371)     "rope_theta": 1000000.0,
+(WorkerDict pid=1168371)     "sliding_window": null,
+(WorkerDict pid=1168371)     "tie_word_embeddings": true,
+(WorkerDict pid=1168371)     "torch_dtype": "float32",
+(WorkerDict pid=1168371)     "use_cache": true,
+(WorkerDict pid=1168371)     "use_sliding_window": false,
+(WorkerDict pid=1168371)     "video_token_id": null,
+(WorkerDict pid=1168371)     "vision_end_token_id": 151653,
+(WorkerDict pid=1168371)     "vision_start_token_id": 151652,
+(WorkerDict pid=1168371)     "vision_token_id": 151654,
+(WorkerDict pid=1168371)     "vocab_size": 151936
+(WorkerDict pid=1168371)   },
+(WorkerDict pid=1168371)   "torch_dtype": "float32",
+(WorkerDict pid=1168371)   "transformers_version": "4.54.1",
+(WorkerDict pid=1168371)   "use_cache": true,
+(WorkerDict pid=1168371)   "use_sliding_window": false,
+(WorkerDict pid=1168371)   "video_token_id": 151656,
+(WorkerDict pid=1168371)   "vision_config": {
+(WorkerDict pid=1168371)     "depth": 32,
+(WorkerDict pid=1168371)     "fullatt_block_indexes": [
+(WorkerDict pid=1168371)       7,
+(WorkerDict pid=1168371)       15,
+(WorkerDict pid=1168371)       23,
+(WorkerDict pid=1168371)       31
+(WorkerDict pid=1168371)     ],
+(WorkerDict pid=1168371)     "hidden_act": "silu",
+(WorkerDict pid=1168371)     "hidden_size": 1280,
+(WorkerDict pid=1168371)     "in_channels": 3,
+(WorkerDict pid=1168371)     "in_chans": 3,
+(WorkerDict pid=1168371)     "initializer_range": 0.02,
+(WorkerDict pid=1168371)     "intermediate_size": 3420,
+(WorkerDict pid=1168371)     "model_type": "qwen2_5_vl",
+(WorkerDict pid=1168371)     "num_heads": 16,
+(WorkerDict pid=1168371)     "out_hidden_size": 2048,
+(WorkerDict pid=1168371)     "patch_size": 14,
+(WorkerDict pid=1168371)     "spatial_merge_size": 2,
+(WorkerDict pid=1168371)     "spatial_patch_size": 14,
+(WorkerDict pid=1168371)     "temporal_patch_size": 2,
+(WorkerDict pid=1168371)     "tokens_per_second": 2,
+(WorkerDict pid=1168371)     "torch_dtype": "float32",
+(WorkerDict pid=1168371)     "window_size": 112
+(WorkerDict pid=1168371)   },
+(WorkerDict pid=1168371)   "vision_end_token_id": 151653,
+(WorkerDict pid=1168371)   "vision_start_token_id": 151652,
+(WorkerDict pid=1168371)   "vision_token_id": 151654,(WorkerDict pid=1168371) 
+(WorkerDict pid=1168371) 
+(WorkerDict pid=1170483) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLTextModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)` [repeated 5x across cluster] (Ray deduplicates logs by default. Set RAY_DEDUP_LOGS=0 to disable log deduplication, or see https://docs.ray.io/en/master/ray-observability/user-guides/configure-logging.html#log-deduplication for more options.)
+(WorkerDict pid=1170483) 
+(WorkerDict pid=1168371) 
+(WorkerDict pid=1170483) 
+(WorkerDict pid=1168371) 
+(WorkerDict pid=1170483) 
+(WorkerDict pid=1168371) 
+(WorkerDict pid=1168371) 
+(WorkerDict pid=1168371) 
+(WorkerDict pid=1168371) 
+(WorkerDict pid=1168371) /home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/torch/distributed/fsdp/fully_sharded_data_parallel.py:690: FutureWarning: FSDP.state_dict_type() and FSDP.set_state_dict_type() are being deprecated. Please use APIs, get_state_dict() and set_state_dict(), which can support different parallelisms, FSDP1, FSDP2, DDP. API doc: https://pytorch.org/docs/stable/distributed.checkpoint.html#torch.distributed.checkpoint.state_dict.get_state_dict .Tutorial: https://pytorch.org/tutorials/recipes/distributed_checkpoint_recipe.html .
+(WorkerDict pid=1168371)   warnings.warn(
+(TaskRunner pid=1165231) Unhandled error (suppress with 'RAY_IGNORE_UNHANDLED_ERRORS=1'): ray::WorkerDict.actor_rollout_generate_sequences() (pid=1170483, ip=172.20.115.25, actor_id=33049b605f0612dfcfe2c38801000000, repr=<verl.single_controller.ray.base.WorkerDict object at 0x7842040bb090>)
+(TaskRunner pid=1165231)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=1165231)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=1165231)   File "/mnt/share01/sqk/AnomSeer/verl/single_controller/ray/base.py", line 420, in func
+(TaskRunner pid=1165231)     return getattr(self.worker_dict[key], name)(*args, **kwargs)
+(TaskRunner pid=1165231)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=1165231)   File "/mnt/share01/sqk/AnomSeer/verl/single_controller/base/decorator.py", line 404, in inner
+(TaskRunner pid=1165231)     return func(*args, **kwargs)
+(TaskRunner pid=1165231)            ^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=1165231)   File "/mnt/share01/sqk/AnomSeer/verl/workers/fsdp_workers.py", line 552, in generate_sequences
+(TaskRunner pid=1165231)     with self.rollout_sharding_manager:
+(TaskRunner pid=1165231)   File "/mnt/share01/sqk/AnomSeer/verl/workers/sharding_manager/fsdp_vllm.py", line 134, in __enter__
+(TaskRunner pid=1165231)     loaded_params = model.load_weights(
+(TaskRunner pid=1165231)                     ^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=1165231)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/qwen2_5_vl.py", line 1126, in load_weights
+(TaskRunner pid=1165231)     return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
+(TaskRunner pid=1165231)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=1165231)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 261, in load_weights
+(TaskRunner pid=1165231)     autoloaded_weights = set(self._load_module("", self.module, weights))
+(TaskRunner pid=1165231)                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=1165231)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 222, in _load_module
+(TaskRunner pid=1165231)     yield from self._load_module(prefix,
+(TaskRunner pid=1165231)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 195, in _load_module
+(TaskRunner pid=1165231)     loaded_params = module_load_weights(weights)
+(TaskRunner pid=1165231)                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=1165231)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/qwen2.py", line 486, in load_weights
+(TaskRunner pid=1165231)     return loader.load_weights(weights)
+(TaskRunner pid=1165231)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=1165231)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 261, in load_weights
+(TaskRunner pid=1165231)     autoloaded_weights = set(self._load_module("", self.module, weights))
+(TaskRunner pid=1165231)                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=1165231)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 222, in _load_module
+(TaskRunner pid=1165231)     yield from self._load_module(prefix,
+(TaskRunner pid=1165231)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 195, in _load_module
+(TaskRunner pid=1165231)     loaded_params = module_load_weights(weights)
+(TaskRunner pid=1165231)                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+(TaskRunner pid=1165231)   File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/qwen2.py", line 405, in load_weights
+(TaskRunner pid=1165231)     param = params_dict[name]
+(TaskRunner pid=1165231)             ~~~~~~~~~~~^^^^^^
+(TaskRunner pid=1165231) KeyError: 'visual.patch_embed.proj.weight'
+
+(WorkerDict pid=1168371)   "vocab_size": 151936
+(WorkerDict pid=1168371) }
+(WorkerDict pid=1168371) 
+(WorkerDict pid=1168371) Qwen2_5_VLForConditionalGeneration contains 3.75B parameters
+(WorkerDict pid=1168371) wrap_policy: functools.partial(<function _or_policy at 0x7af6274bc5e0>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x7af6274bc4a0>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>})])
+(WorkerDict pid=1168371) Actor use_remove_padding=False
+(WorkerDict pid=1168371) Total steps: 472, num_warmup_steps: 0
+(WorkerDict pid=1168371) Actor use_remove_padding=False
+(WorkerDict pid=1168371) Before building vllm rollout, memory allocated (GB): 7.027174949645996, memory reserved (GB): 12.189453125
+(WorkerDict pid=1168371) WARNING 06-15 11:21:42 [utils.py:2522] Methods determine_num_available_blocks,device_config,get_cache_block_size_bytes,initialize_cache not implemented in <vllm.v1.worker.gpu_worker.Worker object at 0x7af2b75a5d90>
+(WorkerDict pid=1170483) wrap_policy: functools.partial(<function _or_policy at 0x78426bee45e0>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x78426bee44a0>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])
+(WorkerDict pid=1170483) Total steps: 472, num_warmup_steps: 0
+(WorkerDict pid=1170483) Actor use_remove_padding=False
+(WorkerDict pid=1168371) WARNING 06-15 11:21:48 [topk_topp_sampler.py:69] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
+(WorkerDict pid=1170483) WARNING 06-15 11:21:42 [utils.py:2522] Methods determine_num_available_blocks,device_config,get_cache_block_size_bytes,initialize_cache not implemented in <vllm.v1.worker.gpu_worker.Worker object at 0x783e7118be90>
+(WorkerDict pid=1168371) kwargs: {'n': 5, 'logprobs': 0, 'max_tokens': 1024, 'detokenize': False, 'temperature': 1.0, 'top_k': -1, 'top_p': 1, 'ignore_eos': False}
+(WorkerDict pid=1168371) After building vllm rollout, memory allocated (GB): 13.633802890777588, memory reserved (GB): 15.04296875
+(WorkerDict pid=1168371) After building sharding manager, memory allocated (GB): 13.633802890777588, memory reserved (GB): 15.04296875
+(WorkerDict pid=1170483) WARNING 06-15 11:21:48 [topk_topp_sampler.py:69] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
+(TaskRunner pid=1165231) Using LocalLogger is deprecated. The constructor API will change 
+Error executing job with overrides: ['algorithm.adv_estimator=grpo', 'data.train_files=/mnt/share01/sqk/AnomSeer/data/rats_uni_processed/train_quarter.parquet', 'data.val_files=/tmp/anomseer_eval_smoke.parquet', 'data.train_batch_size=16', 'data.max_prompt_length=1024', 'data.max_response_length=1024', 'data.filter_overlong_prompts=True', 'data.truncation=error', 'data.image_key=images', 'actor_rollout_ref.model.path=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface', 'actor_rollout_ref.model.lora_rank=0', 'actor_rollout_ref.model.lora_alpha=16', 'actor_rollout_ref.model.lora_dropout=0.0', "actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'", 'actor_rollout_ref.actor.optim.lr=1e-4', 'actor_rollout_ref.model.use_remove_padding=False', 'actor_rollout_ref.actor.ppo_mini_batch_size=16', 'actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2', 'actor_rollout_ref.actor.use_kl_loss=True', 'actor_rollout_ref.actor.kl_loss_coef=0.001', 'actor_rollout_ref.actor.kl_loss_type=low_var_kl', 'actor_rollout_ref.model.enable_gradient_checkpointing=True', 'actor_rollout_ref.actor.fsdp_config.param_offload=False', 'actor_rollout_ref.actor.fsdp_config.optimizer_offload=False', 'actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8', 'actor_rollout_ref.rollout.tensor_model_parallel_size=2', 'actor_rollout_ref.rollout.name=vllm', 'actor_rollout_ref.rollout.gpu_memory_utilization=0.4', 'actor_rollout_ref.rollout.enable_chunked_prefill=False', 'actor_rollout_ref.rollout.enforce_eager=False', 'actor_rollout_ref.rollout.free_cache_engine=False', 'actor_rollout_ref.rollout.n=5', 'actor_rollout_ref.rollout.val_kwargs.do_sample=False', 'actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8', 'actor_rollout_ref.ref.fsdp_config.param_offload=True', 'algorithm.kl_ctrl.kl_coef=0.001', 'trainer.critic_warmup=0', "trainer.logger=['console']", 'trainer.project_name=anomseer', 'trainer.experiment_name=anomseer_rats_uni_2gpu', 'trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu', 'trainer.n_gpus_per_node=2', 'trainer.nnodes=1', 'trainer.save_freq=500', 'trainer.test_freq=10', 'trainer.val_only=True', 'trainer.val_before_train=True', 'trainer.resume_mode=disable', 'trainer.total_epochs=1', 'ts.use_sem_orth=True', 'ts.adv_mix=0.3', 'ts.similarity_method=ot', 'ts.ot_eps=0.08', 'ts.ot_n_iter=50', 'data.val_batch_size=1', 'trainer.test_freq=-1']
+Traceback (most recent call last):
+  File "/mnt/share01/sqk/AnomSeer/verl/trainer/main_ppo.py", line 68, in main
+    ray.get(runner.run.remote(config))
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/auto_init_hook.py", line 22, in auto_init_wrapper
+    return fn(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/client_mode_hook.py", line 107, in wrapper
+    return func(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/worker.py", line 2980, in get
+    values, debugger_breakpoint = worker.get_objects(
+                                  ^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/worker.py", line 1023, in get_objects
+    raise value.as_instanceof_cause()
+ray.exceptions.RayTaskError(KeyError): ray::TaskRunner.run() (pid=1165231, ip=172.20.115.25, actor_id=58caf0b1df3b7b7227854f3c01000000, repr=<main_ppo.TaskRunner object at 0x74f018438150>)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/trainer/main_ppo.py", line 173, in run
+    trainer.fit()
+  File "/mnt/share01/sqk/AnomSeer/verl/trainer/ppo/ray_trainer.py", line 2160, in fit
+    val_metrics = self._validate()
+                  ^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/trainer/ppo/ray_trainer.py", line 1849, in _validate
+    test_output_gen_batch_padded = self.actor_rollout_wg.generate_sequences(test_gen_batch_padded)
+                                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/single_controller/ray/base.py", line 43, in func
+    output = ray.get(output)
+             ^^^^^^^^^^^^^^^
+           ^^^^^^^^^^^^^^^^^^^
+           ^^^^^^^^^^^^^^^^^^^^^
+                                  ^^^^^^^^^^^^^^^^^^^
+ray.exceptions.RayTaskError(KeyError): ray::WorkerDict.actor_rollout_generate_sequences() (pid=1168371, ip=172.20.115.25, actor_id=ccb80723c0cfb29a26c02beb01000000, repr=<verl.single_controller.ray.base.WorkerDict object at 0x7af5bfb1a510>)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/single_controller/ray/base.py", line 420, in func
+    return getattr(self.worker_dict[key], name)(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/single_controller/base/decorator.py", line 404, in inner
+    return func(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/workers/fsdp_workers.py", line 552, in generate_sequences
+    with self.rollout_sharding_manager:
+  File "/mnt/share01/sqk/AnomSeer/verl/workers/sharding_manager/fsdp_vllm.py", line 134, in __enter__
+    loaded_params = model.load_weights(
+                    ^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/qwen2_5_vl.py", line 1126, in load_weights
+    return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 261, in load_weights
+    autoloaded_weights = set(self._load_module("", self.module, weights))
+                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 222, in _load_module
+    yield from self._load_module(prefix,
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 195, in _load_module
+    loaded_params = module_load_weights(weights)
+                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/qwen2.py", line 486, in load_weights
+    return loader.load_weights(weights)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 261, in load_weights
+    autoloaded_weights = set(self._load_module("", self.module, weights))
+                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 222, in _load_module
+    yield from self._load_module(prefix,
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 195, in _load_module
+    loaded_params = module_load_weights(weights)
+                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/vllm/model_executor/models/qwen2.py", line 405, in load_weights
+    param = params_dict[name]
+            ~~~~~~~~~~~^^^^^^
+KeyError: 'visual.patch_embed.proj.weight'
+
+Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.
+(WorkerDict pid=1170483) /home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/torch/distributed/fsdp/fully_sharded_data_parallel.py:690: FutureWarning: FSDP.state_dict_type() and FSDP.set_state_dict_type() are being deprecated. Please use APIs, get_state_dict() and set_state_dict(), which can support different parallelisms, FSDP1, FSDP2, DDP. API doc: https://pytorch.org/docs/stable/distributed.checkpoint.html#torch.distributed.checkpoint.state_dict.get_state_dict .Tutorial: https://pytorch.org/tutorials/recipes/distributed_checkpoint_recipe.html .
+(WorkerDict pid=1170483)   warnings.warn(
+(WorkerDict pid=1170483) kwargs: {'n': 5, 'logprobs': 0, 'max_tokens': 1024, 'detokenize': False, 'temperature': 1.0, 'top_k': -1, 'top_p': 1, 'ignore_eos': False}

logs/anomseer_rats_uni_full_20260612_184427_exp3/rats_eval_lora_smoke_fixed_v2_20260615.log

@@ -0,0 +1,803 @@
+[log] saving full output to: /mnt/share01/sqk/AnomSeer/logs/anomseer_rats_uni_full_20260612_184427_exp3/rats_eval_lora_smoke_fixed_v2_20260615.log
+[config] tuning=LoRA (r=16, alpha=16, target=q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj) | lr=1e-4 | gpus=2 | tensor-parallel (TP=2)
+[config] model=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface
+[config] python=/home/suiqk/anaconda3/envs/scalerag-ts-v4/bin/python | stage=eval
+[stage] EVAL   (model=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface, gpus=2, tp=2, lora_rank=0)
+2026-06-15 11:25:40,665	INFO worker.py:2012 -- Started a local Ray instance.
+/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/worker.py:2051: FutureWarning: Tip: In future versions of Ray, Ray will no longer override accelerator visible devices env var if num_gpus=0 or num_gpus=None (default). To enable this behavior and turn off this error message, set RAY_ACCEL_ENV_VAR_OVERRIDE_ON_ZERO=0
+  warnings.warn(
+(TaskRunner pid=1203084) {'actor_rollout_ref': {'actor': {'clip_ratio': 0.2,
+(TaskRunner pid=1203084)                                  'entropy_coeff': 0.001,
+(TaskRunner pid=1203084)                                  'fsdp_config': {'fsdp_size': -1,
+(TaskRunner pid=1203084)                                                  'optimizer_offload': False,
+(TaskRunner pid=1203084)                                                  'param_offload': False,
+(TaskRunner pid=1203084)                                                  'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=1203084)                                  'grad_clip': 1.0,
+(TaskRunner pid=1203084)                                  'kl_loss_coef': 0.001,
+(TaskRunner pid=1203084)                                  'kl_loss_type': 'low_var_kl',
+(TaskRunner pid=1203084)                                  'optim': {'lr': 0.0001,
+(TaskRunner pid=1203084)                                            'lr_warmup_steps': -1,
+(TaskRunner pid=1203084)                                            'lr_warmup_steps_ratio': 0.0,
+(TaskRunner pid=1203084)                                            'min_lr_ratio': None,
+(TaskRunner pid=1203084)                                            'total_training_steps': -1,
+(TaskRunner pid=1203084)                                            'warmup_style': 'constant'},
+(TaskRunner pid=1203084)                                  'ppo_epochs': 1,
+(TaskRunner pid=1203084)                                  'ppo_max_token_len_per_gpu': 16384,
+(TaskRunner pid=1203084)                                  'ppo_micro_batch_size': None,
+(TaskRunner pid=1203084)                                  'ppo_micro_batch_size_per_gpu': 2,
+(TaskRunner pid=1203084)                                  'ppo_mini_batch_size': 16,
+(TaskRunner pid=1203084)                                  'shuffle': False,
+(TaskRunner pid=1203084)                                  'strategy': 'fsdp',
+(TaskRunner pid=1203084)                                  'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=1203084)                                  'use_dynamic_bsz': False,
+(TaskRunner pid=1203084)                                  'use_kl_loss': True,
+(TaskRunner pid=1203084)                                  'use_torch_compile': True},
+(TaskRunner pid=1203084)                        'hybrid_engine': True,
+(TaskRunner pid=1203084)                        'model': {'enable_gradient_checkpointing': True,
+(TaskRunner pid=1203084)                                  'external_lib': None,
+(TaskRunner pid=1203084)                                  'lora_alpha': 16,
+(TaskRunner pid=1203084)                                  'lora_dropout': 0.0,
+(TaskRunner pid=1203084)                                  'lora_rank': 0,
+(TaskRunner pid=1203084)                                  'lora_target_modules': 'q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj',
+(TaskRunner pid=1203084)                                  'override_config': {},
+(TaskRunner pid=1203084)                                  'path': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface',
+(TaskRunner pid=1203084)                                  'use_remove_padding': False},
+(TaskRunner pid=1203084)                        'ref': {'fsdp_config': {'param_offload': True,
+(TaskRunner pid=1203084)                                                'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=1203084)                                'log_prob_max_token_len_per_gpu': 16384,
+(TaskRunner pid=1203084)                                'log_prob_micro_batch_size': None,
+(TaskRunner pid=1203084)                                'log_prob_micro_batch_size_per_gpu': 8,
+(TaskRunner pid=1203084)                                'log_prob_use_dynamic_bsz': False,
+(TaskRunner pid=1203084)                                'ulysses_sequence_parallel_size': 1},
+(TaskRunner pid=1203084)                        'rollout': {'disable_log_stats': True,
+(TaskRunner pid=1203084)                                    'do_sample': True,
+(TaskRunner pid=1203084)                                    'dtype': 'bfloat16',
+(TaskRunner pid=1203084)                                    'enable_chunked_prefill': False,
+(TaskRunner pid=1203084)                                    'enforce_eager': False,
+(TaskRunner pid=1203084)                                    'free_cache_engine': False,
+(TaskRunner pid=1203084)                                    'gpu_memory_utilization': 0.4,
+(TaskRunner pid=1203084)                                    'ignore_eos': False,
+(TaskRunner pid=1203084)                                    'load_format': 'dummy_dtensor',
+(TaskRunner pid=1203084)                                    'log_prob_max_token_len_per_gpu': 16384,
+(TaskRunner pid=1203084)                                    'log_prob_micro_batch_size': None,
+(TaskRunner pid=1203084)                                    'log_prob_micro_batch_size_per_gpu': 8,
+(TaskRunner pid=1203084)                                    'log_prob_use_dynamic_bsz': False,
+(TaskRunner pid=1203084)                                    'max_model_len': None,
+(TaskRunner pid=1203084)                                    'max_num_batched_tokens': 8192,
+(TaskRunner pid=1203084)                                    'max_num_seqs': 1024,
+(TaskRunner pid=1203084)                                    'n': 5,
+(TaskRunner pid=1203084)                                    'name': 'vllm',
+(TaskRunner pid=1203084)                                    'prompt_length': 1024,
+(TaskRunner pid=1203084)                                    'response_length': 1024,
+(TaskRunner pid=1203084)                                    'temperature': 1.0,
+(TaskRunner pid=1203084)                                    'tensor_model_parallel_size': 2,
+(TaskRunner pid=1203084)                                    'top_k': -1,
+(TaskRunner pid=1203084)                                    'top_p': 1,
+(TaskRunner pid=1203084)                                    'use_fire_sampling': False,
+(TaskRunner pid=1203084)                                    'val_kwargs': {'do_sample': False,
+(TaskRunner pid=1203084)                                                   'n': 1,
+(TaskRunner pid=1203084)                                                   'temperature': 0.4,
+(TaskRunner pid=1203084)                                                   'top_k': -1,
+(TaskRunner pid=1203084)                                                   'top_p': 1.0}}},
+(TaskRunner pid=1203084)  'algorithm': {'adv_estimator': 'grpo',
+(TaskRunner pid=1203084)                'gamma': 1.0,
+(TaskRunner pid=1203084)                'kl_ctrl': {'kl_coef': 0.001, 'type': 'fixed'},
+(TaskRunner pid=1203084)                'kl_penalty': 'kl',
+(TaskRunner pid=1203084)                'lam': 1.0},
+(TaskRunner pid=1203084)  'critic': {'cliprange_value': 0.5,
+(TaskRunner pid=1203084)             'forward_max_token_len_per_gpu': 32768,
+(TaskRunner pid=1203084)             'forward_micro_batch_size': None,
+(TaskRunner pid=1203084)             'forward_micro_batch_size_per_gpu': None,
+(TaskRunner pid=1203084)             'grad_clip': 1.0,
+(TaskRunner pid=1203084)             'model': {'enable_gradient_checkpointing': True,
+(TaskRunner pid=1203084)                       'external_lib': None,
+(TaskRunner pid=1203084)                       'fsdp_config': {'fsdp_size': -1,(TaskRunner pid=1203084) DeprecationWarning: `ray.state.available_resources_per_node` is a private attribute and access will be removed in a future Ray version.
+
+(TaskRunner pid=1203084)                                       'optimizer_offload': False,
+(TaskRunner pid=1203084)                                       'param_offload': False,
+(TaskRunner pid=1203084)                                       'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=1203084)                       'override_config': {},
+(TaskRunner pid=1203084)                       'path': '~/models/deepseek-llm-7b-chat',
+(TaskRunner pid=1203084)                       'tokenizer_path': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface',
+(TaskRunner pid=1203084)                       'use_remove_padding': False},
+(TaskRunner pid=1203084)             'optim': {'lr': 1e-05,
+(TaskRunner pid=1203084)                       'lr_warmup_steps_ratio': 0.0,
+(TaskRunner pid=1203084)                       'min_lr_ratio': None,
+(TaskRunner pid=1203084)                       'total_training_steps': -1,
+(TaskRunner pid=1203084)                       'warmup_style': 'constant'},
+(TaskRunner pid=1203084)             'ppo_epochs': 1,
+(TaskRunner pid=1203084)             'ppo_max_token_len_per_gpu': 32768,
+(TaskRunner pid=1203084)             'ppo_micro_batch_size': None,
+(TaskRunner pid=1203084)             'ppo_micro_batch_size_per_gpu': None,
+(TaskRunner pid=1203084)             'ppo_mini_batch_size': 16,
+(TaskRunner pid=1203084)             'shuffle': False,
+(TaskRunner pid=1203084)             'strategy': 'fsdp',
+(TaskRunner pid=1203084)             'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=1203084)             'use_dynamic_bsz': False},
+(TaskRunner pid=1203084)  'custom_reward_function': {'name': 'compute_score', 'path': None},
+(TaskRunner pid=1203084)  'data': {'filter_overlong_prompts': True,
+(TaskRunner pid=1203084)           'image_key': 'images',
+(TaskRunner pid=1203084)           'max_prompt_length': 1024,
+(TaskRunner pid=1203084)           'max_response_length': 1024,
+(TaskRunner pid=1203084)           'prompt_key': 'prompt',
+(TaskRunner pid=1203084)           'return_raw_chat': False,
+(TaskRunner pid=1203084)           'return_raw_input_ids': False,
+(TaskRunner pid=1203084)           'shuffle': True,
+(TaskRunner pid=1203084)           'tokenizer': None,
+(TaskRunner pid=1203084)           'train_batch_size': 16,
+(TaskRunner pid=1203084)           'train_files': '/mnt/share01/sqk/AnomSeer/data/rats_uni_processed/train_quarter.parquet',
+(TaskRunner pid=1203084)           'truncation': 'error',
+(TaskRunner pid=1203084)           'val_batch_size': 1,
+(TaskRunner pid=1203084)           'val_files': '/tmp/anomseer_eval_smoke.parquet'},
+(TaskRunner pid=1203084)  'reward_model': {'enable': False,
+(TaskRunner pid=1203084)                   'forward_max_token_len_per_gpu': 32768,
+(TaskRunner pid=1203084)                   'max_length': None,
+(TaskRunner pid=1203084)                   'micro_batch_size': None,
+(TaskRunner pid=1203084)                   'micro_batch_size_per_gpu': None,
+(TaskRunner pid=1203084)                   'model': {'external_lib': None,
+(TaskRunner pid=1203084)                             'fsdp_config': {'fsdp_size': -1,
+(TaskRunner pid=1203084)                                             'param_offload': False,
+(TaskRunner pid=1203084)                                             'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=1203084)                             'input_tokenizer': '/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface',
+(TaskRunner pid=1203084)                             'path': '~/models/FsfairX-LLaMA3-RM-v0.1',
+(TaskRunner pid=1203084)                             'use_remove_padding': False},
+(TaskRunner pid=1203084)                   'reward_manager': 'naive',
+(TaskRunner pid=1203084)                   'strategy': 'fsdp',
+(TaskRunner pid=1203084)                   'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=1203084)                   'use_dynamic_bsz': False},
+(TaskRunner pid=1203084)  'trainer': {'balance_batch': True,
+(TaskRunner pid=1203084)              'critic_warmup': 0,
+(TaskRunner pid=1203084)              'default_hdfs_dir': None,
+(TaskRunner pid=1203084)              'default_local_dir': 'checkpoints/anomseer/anomseer_rats_uni_2gpu',
+(TaskRunner pid=1203084)              'del_local_ckpt_after_load': False,
+(TaskRunner pid=1203084)              'experiment_name': 'anomseer_rats_uni_2gpu',
+(TaskRunner pid=1203084)              'logger': ['console'],
+(TaskRunner pid=1203084)              'n_gpus_per_node': 2,
+(TaskRunner pid=1203084)              'nnodes': 1,
+(TaskRunner pid=1203084)              'project_name': 'anomseer',
+(TaskRunner pid=1203084)              'remove_previous_ckpt_in_save': False,
+(TaskRunner pid=1203084)              'resume_from_path': False,
+(TaskRunner pid=1203084)              'resume_mode': 'disable',
+(TaskRunner pid=1203084)              'save_freq': 500,
+(TaskRunner pid=1203084)              'test_freq': -1,
+(TaskRunner pid=1203084)              'total_epochs': 1,
+(TaskRunner pid=1203084)              'total_training_steps': None,
+(TaskRunner pid=1203084)              'val_before_train': True,
+(TaskRunner pid=1203084)              'val_generations_to_log_to_wandb': 0,
+(TaskRunner pid=1203084)              'val_only': True},
+(TaskRunner pid=1203084)  'ts': {'adv_mix': 0.3,
+(TaskRunner pid=1203084)         'clip_pooling': 'mean',
+(TaskRunner pid=1203084)         'clip_temperature': 1.0,
+(TaskRunner pid=1203084)         'ot_eps': 0.08,
+(TaskRunner pid=1203084)         'ot_n_iter': 50,
+(TaskRunner pid=1203084)         'similarity_method': 'ot',
+(TaskRunner pid=1203084)         'use_sem_orth': True}}
+(TaskRunner pid=1203084) reward_manager_name: anomts
+(TaskRunner pid=1203084) WARNING: val_batch_size is deprecated. Validation datasets are sent to inference engines as a whole batch, which will schedule the memory themselves.
+(TaskRunner pid=1203084) [validate_config] All configuration checks passed successfully!
+(TaskRunner pid=1203084) dataset len: 7566
+(TaskRunner pid=1203084) filter dataset len: 7566
+(TaskRunner pid=1203084) dataset len: 1
+(TaskRunner pid=1203084) filter dataset len: 1
+(TaskRunner pid=1203084) Size of train dataloader: 472
+(TaskRunner pid=1203084) Total training steps: 472
+(WorkerDict pid=1206072) Model config after override: Qwen2_5_VLConfig {
+(WorkerDict pid=1206072)   "architectures": [
+(WorkerDict pid=1206072)     "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=1206072)   ],
+(WorkerDict pid=1206072)   "attention_dropout": 0.0,
+(WorkerDict pid=1206072)   "eos_token_id": 151645,
+(WorkerDict pid=1206072)   "hidden_act": "silu",
+(WorkerDict pid=1206072)   "hidden_size": 2048,
+(WorkerDict pid=1206072)   "image_token_id": 151655,
+(WorkerDict pid=1206072)   "initializer_range": 0.02,
+(WorkerDict pid=1206072)   "intermediate_size": 11008,
+(WorkerDict pid=1206072)   "max_position_embeddings": 128000,
+(WorkerDict pid=1206072)   "max_window_layers": 70,
+(WorkerDict pid=1206072)   "model_type": "qwen2_5_vl",
+(WorkerDict pid=1206072)   "num_attention_heads": 16,
+(WorkerDict pid=1206072)   "num_hidden_layers": 36,
+(WorkerDict pid=1206072)   "num_key_value_heads": 2,(WorkerDict pid=1208116) 
+(WorkerDict pid=1206072) 
+(WorkerDict pid=1206072) 
+(WorkerDict pid=1206072) 
+(WorkerDict pid=1208116) 
+(WorkerDict pid=1206072) [rank0]:[W615 11:26:41.949330793 ProcessGroupNCCL.cpp:4561] [PG ID 0 PG GUID 0 Rank 0]  using GPU 0 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. Specify device_ids in barrier() to force use of a particular device, or call init_process_group() with a device_id.
+
+(WorkerDict pid=1206072)   "pad_token_id": 151643,
+(WorkerDict pid=1206072)   "rms_norm_eps": 1e-06,
+(WorkerDict pid=1206072)   "rope_scaling": {
+(WorkerDict pid=1206072)     "mrope_section": [
+(WorkerDict pid=1206072)       16,
+(WorkerDict pid=1206072)       24,
+(WorkerDict pid=1206072)       24
+(WorkerDict pid=1206072)     ],
+(WorkerDict pid=1206072)     "rope_type": "default",
+(WorkerDict pid=1206072)     "type": "default"
+(WorkerDict pid=1206072)   },
+(WorkerDict pid=1206072)   "rope_theta": 1000000.0,
+(WorkerDict pid=1206072)   "sliding_window": 32768,
+(WorkerDict pid=1206072)   "text_config": {
+(WorkerDict pid=1206072)     "architectures": [
+(WorkerDict pid=1206072)       "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=1206072)     ],
+(WorkerDict pid=1206072)     "attention_dropout": 0.0,
+(WorkerDict pid=1206072)     "bos_token_id": 151643,
+(WorkerDict pid=1206072)     "eos_token_id": 151645,
+(WorkerDict pid=1206072)     "hidden_act": "silu",
+(WorkerDict pid=1206072)     "hidden_size": 2048,
+(WorkerDict pid=1206072)     "image_token_id": null,
+(WorkerDict pid=1206072)     "initializer_range": 0.02,
+(WorkerDict pid=1206072)     "intermediate_size": 11008,
+(WorkerDict pid=1206072)     "layer_types": [
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention"
+(WorkerDict pid=1206072)     ],
+(WorkerDict pid=1206072)     "max_position_embeddings": 128000,
+(WorkerDict pid=1206072)     "max_window_layers": 70,
+(WorkerDict pid=1206072)     "model_type": "qwen2_5_vl_text",
+(WorkerDict pid=1206072)     "num_attention_heads": 16,
+(WorkerDict pid=1206072)     "num_hidden_layers": 36,
+(WorkerDict pid=1206072)     "num_key_value_heads": 2,
+(WorkerDict pid=1206072)     "rms_norm_eps": 1e-06,
+(WorkerDict pid=1206072)     "rope_scaling": {
+(WorkerDict pid=1206072)       "mrope_section": [
+(WorkerDict pid=1206072)         16,
+(WorkerDict pid=1206072)         24,
+(WorkerDict pid=1206072)         24
+(WorkerDict pid=1206072)       ],
+(WorkerDict pid=1206072)       "rope_type": "default",
+(WorkerDict pid=1206072)       "type": "default"
+(WorkerDict pid=1206072)     },
+(WorkerDict pid=1206072)     "rope_theta": 1000000.0,
+(WorkerDict pid=1206072)     "sliding_window": null,
+(WorkerDict pid=1206072)     "tie_word_embeddings": true,
+(WorkerDict pid=1206072)     "torch_dtype": "float32",
+(WorkerDict pid=1206072)     "use_cache": true,
+(WorkerDict pid=1206072)     "use_sliding_window": false,
+(WorkerDict pid=1206072)     "video_token_id": null,
+(WorkerDict pid=1206072)     "vision_end_token_id": 151653,
+(WorkerDict pid=1206072)     "vision_start_token_id": 151652,
+(WorkerDict pid=1206072)     "vision_token_id": 151654,
+(WorkerDict pid=1206072)     "vocab_size": 151936
+(WorkerDict pid=1206072)   },
+(WorkerDict pid=1206072)   "torch_dtype": "float32",
+(WorkerDict pid=1206072)   "transformers_version": "4.54.1",
+(WorkerDict pid=1206072)   "use_cache": true,
+(WorkerDict pid=1206072)   "use_sliding_window": false,
+(WorkerDict pid=1206072)   "video_token_id": 151656,
+(WorkerDict pid=1206072)   "vision_config": {
+(WorkerDict pid=1206072)     "depth": 32,
+(WorkerDict pid=1206072)     "fullatt_block_indexes": [
+(WorkerDict pid=1206072)       7,
+(WorkerDict pid=1206072)       15,
+(WorkerDict pid=1206072)       23,
+(WorkerDict pid=1206072)       31
+(WorkerDict pid=1206072)     ],
+(WorkerDict pid=1206072)     "hidden_act": "silu",
+(WorkerDict pid=1206072)     "hidden_size": 1280,
+(WorkerDict pid=1206072)     "in_channels": 3,
+(WorkerDict pid=1206072)     "in_chans": 3,
+(WorkerDict pid=1206072)     "initializer_range": 0.02,
+(WorkerDict pid=1206072)     "intermediate_size": 3420,
+(WorkerDict pid=1206072)     "model_type": "qwen2_5_vl",
+(WorkerDict pid=1206072)     "num_heads": 16,
+(WorkerDict pid=1206072)     "out_hidden_size": 2048,
+(WorkerDict pid=1206072)     "patch_size": 14,
+(WorkerDict pid=1206072)     "spatial_merge_size": 2,
+(WorkerDict pid=1206072)     "spatial_patch_size": 14,
+(WorkerDict pid=1206072)     "temporal_patch_size": 2,
+(WorkerDict pid=1206072)     "tokens_per_second": 2,
+(WorkerDict pid=1206072)     "torch_dtype": "float32",
+(WorkerDict pid=1206072)     "window_size": 112
+(WorkerDict pid=1206072)   },
+(WorkerDict pid=1206072)   "vision_end_token_id": 151653,
+(WorkerDict pid=1206072)   "vision_start_token_id": 151652,
+(WorkerDict pid=1206072)   "vision_token_id": 151654,
+(WorkerDict pid=1206072)   "vocab_size": 151936
+(WorkerDict pid=1206072) }
+(WorkerDict pid=1206072) 
+(WorkerDict pid=1206072) NCCL version 2.21.5+cuda12.4
+(WorkerDict pid=1206072) Qwen2_5_VLForConditionalGeneration contains 3.75B parameters
+(WorkerDict pid=1206072) wrap_policy: functools.partial(<function _or_policy at 0x7cebba2d45e0>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x7cebba2d44a0>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])
+(WorkerDict pid=1206072) Actor use_remove_padding=False
+(WorkerDict pid=1208116) wrap_policy: functools.partial(<function _or_policy at 0x79a367cd45e0>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x79a367cd44a0>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])
+(WorkerDict pid=1206072) Model config after override: Qwen2_5_VLConfig {
+(WorkerDict pid=1206072)   "architectures": [
+(WorkerDict pid=1206072)     "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=1206072)   ],
+(WorkerDict pid=1206072)   "attention_dropout": 0.0,
+(WorkerDict pid=1206072)   "eos_token_id": 151645,
+(WorkerDict pid=1206072)   "hidden_act": "silu",
+(WorkerDict pid=1206072)   "hidden_size": 2048,
+(WorkerDict pid=1206072)   "image_token_id": 151655,
+(WorkerDict pid=1206072)   "initializer_range": 0.02,
+(WorkerDict pid=1206072)   "intermediate_size": 11008,
+(WorkerDict pid=1206072)   "max_position_embeddings": 128000,
+(WorkerDict pid=1206072)   "max_window_layers": 70,
+(WorkerDict pid=1206072)   "model_type": "qwen2_5_vl",
+(WorkerDict pid=1206072)   "num_attention_heads": 16,
+(WorkerDict pid=1206072)   "num_hidden_layers": 36,
+(WorkerDict pid=1206072)   "num_key_value_heads": 2,
+(WorkerDict pid=1206072)   "pad_token_id": 151643,
+(WorkerDict pid=1206072)   "rms_norm_eps": 1e-06,
+(WorkerDict pid=1206072)   "rope_scaling": {
+(WorkerDict pid=1206072)     "mrope_section": [
+(WorkerDict pid=1206072)       16,
+(WorkerDict pid=1206072)       24,
+(WorkerDict pid=1206072)       24
+(WorkerDict pid=1206072)     ],
+(WorkerDict pid=1206072)     "rope_type": "default",
+(WorkerDict pid=1206072)     "type": "default"
+(WorkerDict pid=1206072)   },
+(WorkerDict pid=1206072)   "rope_theta": 1000000.0,
+(WorkerDict pid=1206072)   "sliding_window": 32768,
+(WorkerDict pid=1206072)   "text_config": {
+(WorkerDict pid=1206072)     "architectures": [
+(WorkerDict pid=1206072)       "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=1206072)     ],
+(WorkerDict pid=1206072)     "attention_dropout": 0.0,
+(WorkerDict pid=1206072)     "bos_token_id": 151643,
+(WorkerDict pid=1206072)     "eos_token_id": 151645,
+(WorkerDict pid=1206072)     "hidden_act": "silu",
+(WorkerDict pid=1206072)     "hidden_size": 2048,
+(WorkerDict pid=1206072)     "image_token_id": null,
+(WorkerDict pid=1206072)     "initializer_range": 0.02,
+(WorkerDict pid=1206072)     "intermediate_size": 11008,
+(WorkerDict pid=1206072)     "layer_types": [
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention",
+(WorkerDict pid=1206072)       "full_attention"
+(WorkerDict pid=1206072)     ],
+(WorkerDict pid=1206072)     "max_position_embeddings": 128000,
+(WorkerDict pid=1206072)     "max_window_layers": 70,
+(WorkerDict pid=1206072)     "model_type": "qwen2_5_vl_text",
+(WorkerDict pid=1206072)     "num_attention_heads": 16,
+(WorkerDict pid=1206072)     "num_hidden_layers": 36,
+(WorkerDict pid=1206072)     "num_key_value_heads": 2,
+(WorkerDict pid=1206072)     "rms_norm_eps": 1e-06,
+(WorkerDict pid=1206072)     "rope_scaling": {
+(WorkerDict pid=1206072)       "mrope_section": [
+(WorkerDict pid=1206072)         16,
+(WorkerDict pid=1206072)         24,
+(WorkerDict pid=1206072)         24
+(WorkerDict pid=1206072)       ],
+(WorkerDict pid=1206072)       "rope_type": "default",
+(WorkerDict pid=1206072)       "type": "default"
+(WorkerDict pid=1206072)     },
+(WorkerDict pid=1206072)     "rope_theta": 1000000.0,
+(WorkerDict pid=1206072)     "sliding_window": null,
+(WorkerDict pid=1206072)     "tie_word_embeddings": true,
+(WorkerDict pid=1206072)     "torch_dtype": "float32",
+(WorkerDict pid=1206072)     "use_cache": true,
+(WorkerDict pid=1206072)     "use_sliding_window": false,
+(WorkerDict pid=1206072)     "video_token_id": null,
+(WorkerDict pid=1206072)     "vision_end_token_id": 151653,
+(WorkerDict pid=1206072)     "vision_start_token_id": 151652,
+(WorkerDict pid=1206072)     "vision_token_id": 151654,
+(WorkerDict pid=1206072)     "vocab_size": 151936
+(WorkerDict pid=1206072)   },
+(WorkerDict pid=1206072)   "torch_dtype": "float32",
+(WorkerDict pid=1206072)   "transformers_version": "4.54.1",
+(WorkerDict pid=1206072)   "use_cache": true,
+(WorkerDict pid=1206072)   "use_sliding_window": false,
+(WorkerDict pid=1206072)   "video_token_id": 151656,
+(WorkerDict pid=1206072)   "vision_config": {
+(WorkerDict pid=1206072)     "depth": 32,
+(WorkerDict pid=1206072)     "fullatt_block_indexes": [
+(WorkerDict pid=1206072)       7,
+(WorkerDict pid=1206072)       15,
+(WorkerDict pid=1206072)       23,
+(WorkerDict pid=1206072)       31
+(WorkerDict pid=1206072)     ],
+(WorkerDict pid=1206072)     "hidden_act": "silu",
+(WorkerDict pid=1206072)     "hidden_size": 1280,
+(WorkerDict pid=1206072)     "in_channels": 3,
+(WorkerDict pid=1206072)     "in_chans": 3,
+(WorkerDict pid=1206072)     "initializer_range": 0.02,
+(WorkerDict pid=1206072)     "intermediate_size": 3420,
+(WorkerDict pid=1206072)     "model_type": "qwen2_5_vl",
+(WorkerDict pid=1206072)     "num_heads": 16,
+(WorkerDict pid=1206072)     "out_hidden_size": 2048,
+(WorkerDict pid=1206072)     "patch_size": 14,
+(WorkerDict pid=1206072)     "spatial_merge_size": 2,
+(WorkerDict pid=1206072)     "spatial_patch_size": 14,
+(WorkerDict pid=1206072)     "temporal_patch_size": 2,
+(WorkerDict pid=1206072)     "tokens_per_second": 2,
+(WorkerDict pid=1206072)     "torch_dtype": "float32",
+(WorkerDict pid=1206072)     "window_size": 112
+(WorkerDict pid=1206072)   },
+(WorkerDict pid=1206072)   "vision_end_token_id": 151653,
+(WorkerDict pid=1206072)   "vision_start_token_id": 151652,
+(WorkerDict pid=1206072)   "vision_token_id": 151654,(WorkerDict pid=1206072) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLForConditionalGeneration is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=1206072) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=1206072) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VisionTransformerPretrainedModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=1206072) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLTextModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=1206072) 
+(WorkerDict pid=1208116) 
+(WorkerDict pid=1208116) [rank1]:[W615 11:26:41.958300854 ProcessGroupNCCL.cpp:4561] [PG ID 0 PG GUID 0 Rank 1]  using GPU 0 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. Specify device_ids in barrier() to force use of a particular device, or call init_process_group() with a device_id.
+(WorkerDict pid=1206072) 
+(WorkerDict pid=1208116) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLTextModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)` [repeated 4x across cluster] (Ray deduplicates logs by default. Set RAY_DEDUP_LOGS=0 to disable log deduplication, or see https://docs.ray.io/en/master/ray-observability/user-guides/configure-logging.html#log-deduplication for more options.)
+(WorkerDict pid=1208116) 
+(WorkerDict pid=1206072) 
+(WorkerDict pid=1208116) 
+(WorkerDict pid=1206072) 
+(WorkerDict pid=1208116) 
+(WorkerDict pid=1206072) 
+(WorkerDict pid=1206072) 
+(WorkerDict pid=1206072) 
+(WorkerDict pid=1206072) 
+(WorkerDict pid=1208116) /home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/torch/distributed/fsdp/fully_sharded_data_parallel.py:690: FutureWarning: FSDP.state_dict_type() and FSDP.set_state_dict_type() are being deprecated. Please use APIs, get_state_dict() and set_state_dict(), which can support different parallelisms, FSDP1, FSDP2, DDP. API doc: https://pytorch.org/docs/stable/distributed.checkpoint.html#torch.distributed.checkpoint.state_dict.get_state_dict .Tutorial: https://pytorch.org/tutorials/recipes/distributed_checkpoint_recipe.html .
+(WorkerDict pid=1208116)   warnings.warn(
+
+(WorkerDict pid=1206072)   "vocab_size": 151936
+(WorkerDict pid=1206072) }
+(WorkerDict pid=1206072) 
+(WorkerDict pid=1206072) Qwen2_5_VLForConditionalGeneration contains 3.75B parameters
+(WorkerDict pid=1206072) wrap_policy: functools.partial(<function _or_policy at 0x7cebba2d45e0>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x7cebba2d44a0>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])
+(WorkerDict pid=1208116) Actor use_remove_padding=False
+(WorkerDict pid=1206072) Total steps: 472, num_warmup_steps: 0
+(WorkerDict pid=1206072) Actor use_remove_padding=False
+(WorkerDict pid=1206072) Before building vllm rollout, memory allocated (GB): 7.027174949645996, memory reserved (GB): 12.189453125
+(WorkerDict pid=1208116) WARNING 06-15 11:29:28 [utils.py:2522] Methods determine_num_available_blocks,device_config,get_cache_block_size_bytes,initialize_cache not implemented in <vllm.v1.worker.gpu_worker.Worker object at 0x799fe047bf50>
+(WorkerDict pid=1208116) wrap_policy: functools.partial(<function _or_policy at 0x79a367cd45e0>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x79a367cd44a0>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])
+(WorkerDict pid=1208116) Total steps: 472, num_warmup_steps: 0
+(WorkerDict pid=1208116) Actor use_remove_padding=False
+(WorkerDict pid=1208116) WARNING 06-15 11:29:36 [topk_topp_sampler.py:69] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
+(WorkerDict pid=1206072) WARNING 06-15 11:29:28 [utils.py:2522] Methods determine_num_available_blocks,device_config,get_cache_block_size_bytes,initialize_cache not implemented in <vllm.v1.worker.gpu_worker.Worker object at 0x7ce8382d5cd0>
+(WorkerDict pid=1208116) kwargs: {'n': 5, 'logprobs': 0, 'max_tokens': 1024, 'detokenize': False, 'temperature': 1.0, 'top_k': -1, 'top_p': 1, 'ignore_eos': False}
+(WorkerDict pid=1206072) WARNING 06-15 11:29:37 [topk_topp_sampler.py:69] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
+(WorkerDict pid=1206072) After building vllm rollout, memory allocated (GB): 13.633802890777588, memory reserved (GB): 15.04296875
+(WorkerDict pid=1206072) After building sharding manager, memory allocated (GB): 13.633802890777588, memory reserved (GB): 15.04296875
+(TaskRunner pid=1203084) Using LocalLogger is deprecated. The constructor API will change 
+(TaskRunner pid=1203084) validation generation end
+(TaskRunner pid=1203084) [prompt] <|im_start|>system
+(TaskRunner pid=1203084) You are a helpful assistant.<|im_end|>
+(TaskRunner pid=1203084) <|im_start|>user
+(TaskRunner pid=1203084) <|vision_start|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|vision_end|>
+(TaskRunner pid=1203084) You are an expert in univariate time-series anomaly detection. The figure shows a single-channel time series of length 64 from the "Server-IOPS" domain.
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) Decide whether the series is normal or contains an anomaly. If it is anomalous, choose the single most appropriate anomaly type from the following 15 categories (format `id: name — description`):
+(TaskRunner pid=1203084) 0: Normal Sequence — There are no abnormal situations in this time series.
+(TaskRunner pid=1203084) 1: Point Anomaly — A single data point significantly deviates from the local or global pattern of the sequence.
+(TaskRunner pid=1203084) 2: Periodic Change Anomaly — The original periodic pattern is disrupted, e.g. the period is broken or the amplitude becomes anomalous.
+(TaskRunner pid=1203084) 3: Trend Change Anomaly — A sudden change in the long-term trend of the time series.
+(TaskRunner pid=1203084) 4: Change Point Anomaly — Statistical properties (e.g. mean, variance) change abruptly at certain points.
+(TaskRunner pid=1203084) 5: Distributional Change Anomaly — The statistical distribution of the time series changes significantly.
+(TaskRunner pid=1203084) 6: Amplitude Anomaly — The amplitude of data points exceeds the normal upper and lower bounds.
+(TaskRunner pid=1203084) 7: Pattern Change Anomaly — The pattern of the time series suddenly changes from one form to another.
+(TaskRunner pid=1203084) 8: Sparse Anomaly — Isolated anomalous patterns occasionally appear in a long time series.
+(TaskRunner pid=1203084) 9: Repeated Value Anomaly — Continuous or intermittent repeated values disrupt the normal fluctuation pattern.
+(TaskRunner pid=1203084) 10: Sudden Flatline Anomaly — The time series suddenly becomes a flat line with no normal fluctuations.
+(TaskRunner pid=1203084) 11: Drift Anomaly — The data gradually drifts away from the normal level.
+(TaskRunner pid=1203084) 12: Sudden Spike Anomaly — The data suddenly spikes or drops within a short time and then returns to normal.
+(TaskRunner pid=1203084) 13: Continuous Segment Anomaly — A continuous segment of data points deviates from the normal pattern.
+(TaskRunner pid=1203084) 14: Nonlinear Pattern Anomaly — Nonlinear changes appear in the sequence, breaking the original linear rule.
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) Reason step by step inside <think>...</think> based on the visual shape of the series, then output exactly one line with your final answer:
+(TaskRunner pid=1203084) <class>one exact category name from the list above</class>
+(TaskRunner pid=1203084) If the series is normal, use <class>Normal Sequence</class>.<|im_end|>
+(TaskRunner pid=1203084) <|im_start|>assistant
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) [response] <think></think>
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <class>Sudden Spike Anomaly</class> 
+(TaskRunner pid=1203084) 
+(TaskRunner pid=1203084) <think></think>
+(TaskRunner pid=1203084) Detected new reward function format (3 return values).
+(WorkerDict pid=1206072) kwargs: {'n': 5, 'logprobs': 0, 'max_tokens': 1024, 'detokenize': False, 'temperature': 1.0, 'top_k': -1, 'top_p': 1, 'ignore_eos': False}
+(WorkerDict pid=1206072) /home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/torch/distributed/fsdp/fully_sharded_data_parallel.py:690: FutureWarning: FSDP.state_dict_type() and FSDP.set_state_dict_type() are being deprecated. Please use APIs, get_state_dict() and set_state_dict(), which can support different parallelisms, FSDP1, FSDP2, DDP. API doc: https://pytorch.org/docs/stable/distributed.checkpoint.html#torch.distributed.checkpoint.state_dict.get_state_dict .Tutorial: https://pytorch.org/tutorials/recipes/distributed_checkpoint_recipe.html .
+(WorkerDict pid=1206072)   warnings.warn(
+(TaskRunner pid=1203084) ("Initial validation metrics: {'val/reward_score/timeseries_rats': "
+(TaskRunner pid=1203084)  "0.20000000298023224, 'val/metrics_overall/mean_affi precision': 0.0, "
+(TaskRunner pid=1203084)  "'val/metrics_overall/mean_affi recall': 0.0, 'val/metrics_overall/mean_affi "
+(TaskRunner pid=1203084)  "f1': 0.0, 'val/metrics_overall/mean_fmt_score': 1.0, "
+(TaskRunner pid=1203084)  "'val/metrics_overall/mean_class acc': 0.0, 'val/num_total_samples': 1, "
+(TaskRunner pid=1203084)  "'val/num_valid_samples': 1, 'val/num_invalid_predictions': 0, "
+(TaskRunner pid=1203084)  "'val/type_accuracy': 0.0, 'val/type_precision_macro': 0.0, "
+(TaskRunner pid=1203084)  "'val/type_recall_macro': 0.0, 'val/type_f1_macro': 0.0, "
+(TaskRunner pid=1203084)  "'val/binary_accuracy': 1.0, 'val/binary_precision_macro': 1.0, "
+(TaskRunner pid=1203084)  "'val/binary_recall_macro': 1.0, 'val/binary_f1_macro': 1.0}")
+(TaskRunner pid=1203084) step:0 - val/reward_score/timeseries_rats:0.200 - val/metrics_overall/mean_affi precision:0.000 - val/metrics_overall/mean_affi recall:0.000 - val/metrics_overall/mean_affi f1:0.000 - val/metrics_overall/mean_fmt_score:1.000 - val/metrics_overall/mean_class acc:0.000 - val/num_total_samples:1.000 - val/num_valid_samples:1.000 - val/num_invalid_predictions:0.000 - val/type_accuracy:0.000 - val/type_precision_macro:0.000 - val/type_recall_macro:0.000 - val/type_f1_macro:0.000 - val/binary_accuracy:1.000 - val/binary_precision_macro:1.000 - val/binary_recall_macro:1.000 - val/binary_f1_macro:1.000

logs/rats_eval_full.log

@@ -0,0 +1,569 @@
+[log] saving full output to: /mnt/share01/sqk/AnomSeer/logs/rats_eval_full.log
+[config] tuning=LoRA (r=16, alpha=16, target=q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj) | lr=1e-4 | gpus=2 | tensor-parallel (TP=2)
+[config] model=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct
+[config] python=/home/suiqk/anaconda3/envs/scalerag-ts-v4/bin/python | stage=eval
+[stage] EVAL   (model=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct, gpus=2, tp=2)
+2026-06-12 11:58:39,437	INFO worker.py:2012 -- Started a local Ray instance.
+/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/worker.py:2051: FutureWarning: Tip: In future versions of Ray, Ray will no longer override accelerator visible devices env var if num_gpus=0 or num_gpus=None (default). To enable this behavior and turn off this error message, set RAY_ACCEL_ENV_VAR_OVERRIDE_ON_ZERO=0
+  warnings.warn(
+(TaskRunner pid=2835960) {'actor_rollout_ref': {'actor': {'clip_ratio': 0.2,
+(TaskRunner pid=2835960)                                  'entropy_coeff': 0.001,
+(TaskRunner pid=2835960)                                  'fsdp_config': {'fsdp_size': -1,
+(TaskRunner pid=2835960)                                                  'optimizer_offload': False,
+(TaskRunner pid=2835960)                                                  'param_offload': False,
+(TaskRunner pid=2835960)                                                  'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=2835960)                                  'grad_clip': 1.0,
+(TaskRunner pid=2835960)                                  'kl_loss_coef': 0.001,
+(TaskRunner pid=2835960)                                  'kl_loss_type': 'low_var_kl',
+(TaskRunner pid=2835960)                                  'optim': {'lr': 0.0001,
+(TaskRunner pid=2835960)                                            'lr_warmup_steps': -1,
+(TaskRunner pid=2835960)                                            'lr_warmup_steps_ratio': 0.0,
+(TaskRunner pid=2835960)                                            'min_lr_ratio': None,
+(TaskRunner pid=2835960)                                            'total_training_steps': -1,
+(TaskRunner pid=2835960)                                            'warmup_style': 'constant'},
+(TaskRunner pid=2835960)                                  'ppo_epochs': 1,
+(TaskRunner pid=2835960)                                  'ppo_max_token_len_per_gpu': 16384,
+(TaskRunner pid=2835960)                                  'ppo_micro_batch_size': None,
+(TaskRunner pid=2835960)                                  'ppo_micro_batch_size_per_gpu': 2,
+(TaskRunner pid=2835960)                                  'ppo_mini_batch_size': 16,
+(TaskRunner pid=2835960)                                  'shuffle': False,
+(TaskRunner pid=2835960)                                  'strategy': 'fsdp',
+(TaskRunner pid=2835960)                                  'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=2835960)                                  'use_dynamic_bsz': False,
+(TaskRunner pid=2835960)                                  'use_kl_loss': True,
+(TaskRunner pid=2835960)                                  'use_torch_compile': True},
+(TaskRunner pid=2835960)                        'hybrid_engine': True,
+(TaskRunner pid=2835960)                        'model': {'enable_gradient_checkpointing': True,
+(TaskRunner pid=2835960)                                  'external_lib': None,
+(TaskRunner pid=2835960)                                  'lora_alpha': 16,
+(TaskRunner pid=2835960)                                  'lora_dropout': 0.0,
+(TaskRunner pid=2835960)                                  'lora_rank': 16,
+(TaskRunner pid=2835960)                                  'lora_target_modules': 'q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj',
+(TaskRunner pid=2835960)                                  'override_config': {},
+(TaskRunner pid=2835960)                                  'path': '/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct',
+(TaskRunner pid=2835960)                                  'use_remove_padding': False},
+(TaskRunner pid=2835960)                        'ref': {'fsdp_config': {'param_offload': True,
+(TaskRunner pid=2835960)                                                'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=2835960)                                'log_prob_max_token_len_per_gpu': 16384,
+(TaskRunner pid=2835960)                                'log_prob_micro_batch_size': None,
+(TaskRunner pid=2835960)                                'log_prob_micro_batch_size_per_gpu': 8,
+(TaskRunner pid=2835960)                                'log_prob_use_dynamic_bsz': False,
+(TaskRunner pid=2835960)                                'ulysses_sequence_parallel_size': 1},
+(TaskRunner pid=2835960)                        'rollout': {'disable_log_stats': True,
+(TaskRunner pid=2835960)                                    'do_sample': True,
+(TaskRunner pid=2835960)                                    'dtype': 'bfloat16',
+(TaskRunner pid=2835960)                                    'enable_chunked_prefill': False,
+(TaskRunner pid=2835960)                                    'enforce_eager': False,
+(TaskRunner pid=2835960)                                    'free_cache_engine': False,
+(TaskRunner pid=2835960)                                    'gpu_memory_utilization': 0.4,
+(TaskRunner pid=2835960)                                    'ignore_eos': False,
+(TaskRunner pid=2835960)                                    'load_format': 'dummy_dtensor',
+(TaskRunner pid=2835960)                                    'log_prob_max_token_len_per_gpu': 16384,
+(TaskRunner pid=2835960)                                    'log_prob_micro_batch_size': None,
+(TaskRunner pid=2835960)                                    'log_prob_micro_batch_size_per_gpu': 8,
+(TaskRunner pid=2835960)                                    'log_prob_use_dynamic_bsz': False,
+(TaskRunner pid=2835960)                                    'max_model_len': None,
+(TaskRunner pid=2835960)                                    'max_num_batched_tokens': 8192,
+(TaskRunner pid=2835960)                                    'max_num_seqs': 1024,
+(TaskRunner pid=2835960)                                    'n': 5,
+(TaskRunner pid=2835960)                                    'name': 'vllm',
+(TaskRunner pid=2835960)                                    'prompt_length': 1024,
+(TaskRunner pid=2835960)                                    'response_length': 1024,
+(TaskRunner pid=2835960)                                    'temperature': 1.0,
+(TaskRunner pid=2835960)                                    'tensor_model_parallel_size': 2,
+(TaskRunner pid=2835960)                                    'top_k': -1,
+(TaskRunner pid=2835960)                                    'top_p': 1,
+(TaskRunner pid=2835960)                                    'use_fire_sampling': False,
+(TaskRunner pid=2835960)                                    'val_kwargs': {'do_sample': False,
+(TaskRunner pid=2835960)                                                   'n': 1,
+(TaskRunner pid=2835960)                                                   'temperature': 0.4,
+(TaskRunner pid=2835960)                                                   'top_k': -1,
+(TaskRunner pid=2835960)                                                   'top_p': 1.0}}},
+(TaskRunner pid=2835960)  'algorithm': {'adv_estimator': 'grpo',
+(TaskRunner pid=2835960)                'gamma': 1.0,
+(TaskRunner pid=2835960)                'kl_ctrl': {'kl_coef': 0.001, 'type': 'fixed'},
+(TaskRunner pid=2835960)                'kl_penalty': 'kl',
+(TaskRunner pid=2835960)                'lam': 1.0},
+(TaskRunner pid=2835960)  'critic': {'cliprange_value': 0.5,
+(TaskRunner pid=2835960)             'forward_max_token_len_per_gpu': 32768,
+(TaskRunner pid=2835960)             'forward_micro_batch_size': None,
+(TaskRunner pid=2835960)             'forward_micro_batch_size_per_gpu': None,
+(TaskRunner pid=2835960)             'grad_clip': 1.0,
+(TaskRunner pid=2835960)             'model': {'enable_gradient_checkpointing': True,
+(TaskRunner pid=2835960)                       'external_lib': None,
+(TaskRunner pid=2835960)                       'fsdp_config': {'fsdp_size': -1,
+(TaskRunner pid=2835960)                                       'optimizer_offload': False,(TaskRunner pid=2835960) You have video processor config saved in `preprocessor.json` file which is deprecated. Video processor configs should be saved in their own `video_preprocessor.json` file. You can rename the file or load and save the processor back which renames it automatically. Loading from `preprocessor.json` will be removed in v5.0.
+(TaskRunner pid=2835960) DeprecationWarning: `ray.state.available_resources_per_node` is a private attribute and access will be removed in a future Ray version.
+(WorkerDict pid=2839719) The image processor of type `Qwen2VLImageProcessor` is now loaded as a fast processor by default, even if the model checkpoint was saved with a slow processor. This is a breaking change and may produce slightly different outputs. To continue using the slow processor, instantiate this class with `use_fast=False`. Note that this behavior will be extended to all models in a future release.
+(WorkerDict pid=2839719) You have video processor config saved in `preprocessor.json` file which is deprecated. Video processor configs should be saved in their own `video_preprocessor.json` file. You can rename the file or load and save the processor back which renames it automatically. Loading from `preprocessor.json` will be removed in v5.0.
+
+(TaskRunner pid=2835960)                                       'param_offload': False,
+(TaskRunner pid=2835960)                                       'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=2835960)                       'override_config': {},
+(TaskRunner pid=2835960)                       'path': '~/models/deepseek-llm-7b-chat',
+(TaskRunner pid=2835960)                       'tokenizer_path': '/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct',
+(TaskRunner pid=2835960)                       'use_remove_padding': False},
+(TaskRunner pid=2835960)             'optim': {'lr': 1e-05,
+(TaskRunner pid=2835960)                       'lr_warmup_steps_ratio': 0.0,
+(TaskRunner pid=2835960)                       'min_lr_ratio': None,
+(TaskRunner pid=2835960)                       'total_training_steps': -1,
+(TaskRunner pid=2835960)                       'warmup_style': 'constant'},
+(TaskRunner pid=2835960)             'ppo_epochs': 1,
+(TaskRunner pid=2835960)             'ppo_max_token_len_per_gpu': 32768,
+(TaskRunner pid=2835960)             'ppo_micro_batch_size': None,
+(TaskRunner pid=2835960)             'ppo_micro_batch_size_per_gpu': None,
+(TaskRunner pid=2835960)             'ppo_mini_batch_size': 16,
+(TaskRunner pid=2835960)             'shuffle': False,
+(TaskRunner pid=2835960)             'strategy': 'fsdp',
+(TaskRunner pid=2835960)             'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=2835960)             'use_dynamic_bsz': False},
+(TaskRunner pid=2835960)  'custom_reward_function': {'name': 'compute_score', 'path': None},
+(TaskRunner pid=2835960)  'data': {'filter_overlong_prompts': True,
+(TaskRunner pid=2835960)           'image_key': 'images',
+(TaskRunner pid=2835960)           'max_prompt_length': 1024,
+(TaskRunner pid=2835960)           'max_response_length': 1024,
+(TaskRunner pid=2835960)           'prompt_key': 'prompt',
+(TaskRunner pid=2835960)           'return_raw_chat': False,
+(TaskRunner pid=2835960)           'return_raw_input_ids': False,
+(TaskRunner pid=2835960)           'shuffle': True,
+(TaskRunner pid=2835960)           'tokenizer': None,
+(TaskRunner pid=2835960)           'train_batch_size': 16,
+(TaskRunner pid=2835960)           'train_files': './data/rats_uni_processed/train_quarter.parquet',
+(TaskRunner pid=2835960)           'truncation': 'error',
+(TaskRunner pid=2835960)           'val_batch_size': None,
+(TaskRunner pid=2835960)           'val_files': './data/rats_uni_processed/test_full.parquet'},
+(TaskRunner pid=2835960)  'reward_model': {'enable': False,
+(TaskRunner pid=2835960)                   'forward_max_token_len_per_gpu': 32768,
+(TaskRunner pid=2835960)                   'max_length': None,
+(TaskRunner pid=2835960)                   'micro_batch_size': None,
+(TaskRunner pid=2835960)                   'micro_batch_size_per_gpu': None,
+(TaskRunner pid=2835960)                   'model': {'external_lib': None,
+(TaskRunner pid=2835960)                             'fsdp_config': {'fsdp_size': -1,
+(TaskRunner pid=2835960)                                             'param_offload': False,
+(TaskRunner pid=2835960)                                             'wrap_policy': {'min_num_params': 0}},
+(TaskRunner pid=2835960)                             'input_tokenizer': '/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct',
+(TaskRunner pid=2835960)                             'path': '~/models/FsfairX-LLaMA3-RM-v0.1',
+(TaskRunner pid=2835960)                             'use_remove_padding': False},
+(TaskRunner pid=2835960)                   'reward_manager': 'naive',
+(TaskRunner pid=2835960)                   'strategy': 'fsdp',
+(TaskRunner pid=2835960)                   'ulysses_sequence_parallel_size': 1,
+(TaskRunner pid=2835960)                   'use_dynamic_bsz': False},
+(TaskRunner pid=2835960)  'trainer': {'balance_batch': True,
+(TaskRunner pid=2835960)              'critic_warmup': 0,
+(TaskRunner pid=2835960)              'default_hdfs_dir': None,
+(TaskRunner pid=2835960)              'default_local_dir': 'checkpoints/anomseer/anomseer_rats_uni_2gpu',
+(TaskRunner pid=2835960)              'del_local_ckpt_after_load': False,
+(TaskRunner pid=2835960)              'experiment_name': 'anomseer_rats_uni_2gpu',
+(TaskRunner pid=2835960)              'logger': ['console'],
+(TaskRunner pid=2835960)              'n_gpus_per_node': 2,
+(TaskRunner pid=2835960)              'nnodes': 1,
+(TaskRunner pid=2835960)              'project_name': 'anomseer',
+(TaskRunner pid=2835960)              'remove_previous_ckpt_in_save': False,
+(TaskRunner pid=2835960)              'resume_from_path': False,
+(TaskRunner pid=2835960)              'resume_mode': 'auto',
+(TaskRunner pid=2835960)              'save_freq': 500,
+(TaskRunner pid=2835960)              'test_freq': 10,
+(TaskRunner pid=2835960)              'total_epochs': 1,
+(TaskRunner pid=2835960)              'total_training_steps': None,
+(TaskRunner pid=2835960)              'val_before_train': True,
+(TaskRunner pid=2835960)              'val_generations_to_log_to_wandb': 0,
+(TaskRunner pid=2835960)              'val_only': True},
+(TaskRunner pid=2835960)  'ts': {'adv_mix': 0.3,
+(TaskRunner pid=2835960)         'clip_pooling': 'mean',
+(TaskRunner pid=2835960)         'clip_temperature': 1.0,
+(TaskRunner pid=2835960)         'ot_eps': 0.08,
+(TaskRunner pid=2835960)         'ot_n_iter': 50,
+(TaskRunner pid=2835960)         'similarity_method': 'ot',
+(TaskRunner pid=2835960)         'use_sem_orth': True}}
+(TaskRunner pid=2835960) reward_manager_name: anomts
+(TaskRunner pid=2835960) [validate_config] All configuration checks passed successfully!
+(TaskRunner pid=2835960) dataset len: 7566
+(TaskRunner pid=2835960) filter dataset len: 7566
+(TaskRunner pid=2835960) dataset len: 6034
+(TaskRunner pid=2835960) filter dataset len: 6034
+(TaskRunner pid=2835960) Size of train dataloader: 472
+(TaskRunner pid=2835960) Total training steps: 472
+(WorkerDict pid=2839719) Model config after override: Qwen2_5_VLConfig {
+(WorkerDict pid=2839719)   "architectures": [
+(WorkerDict pid=2839719)     "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=2839719)   ],
+(WorkerDict pid=2839719)   "attention_dropout": 0.0,
+(WorkerDict pid=2839719)   "eos_token_id": 151645,
+(WorkerDict pid=2839719)   "hidden_act": "silu",
+(WorkerDict pid=2839719)   "hidden_size": 2048,
+(WorkerDict pid=2839719)   "image_token_id": 151655,
+(WorkerDict pid=2839719)   "initializer_range": 0.02,
+(WorkerDict pid=2839719)   "intermediate_size": 11008,
+(WorkerDict pid=2839719)   "max_position_embeddings": 128000,
+(WorkerDict pid=2839719)   "max_window_layers": 70,
+(WorkerDict pid=2839719)   "model_type": "qwen2_5_vl",
+(WorkerDict pid=2839719)   "num_attention_heads": 16,
+(WorkerDict pid=2839719)   "num_hidden_layers": 36,
+(WorkerDict pid=2839719)   "num_key_value_heads": 2,
+(WorkerDict pid=2839719)   "pad_token_id": 151643,
+(WorkerDict pid=2839719)   "rms_norm_eps": 1e-06,
+(WorkerDict pid=2839719)   "rope_scaling": {
+(WorkerDict pid=2839719)     "mrope_section": [
+(WorkerDict pid=2839719)       16,
+(WorkerDict pid=2839719)       24,
+(WorkerDict pid=2839719)       24
+(WorkerDict pid=2839719)     ],
+(WorkerDict pid=2839719)     "rope_type": "default",(WorkerDict pid=2839719) 
+(WorkerDict pid=2839719) 
+(WorkerDict pid=2841587) The image processor of type `Qwen2VLImageProcessor` is now loaded as a fast processor by default, even if the model checkpoint was saved with a slow processor. This is a breaking change and may produce slightly different outputs. To continue using the slow processor, instantiate this class with `use_fast=False`. Note that this behavior will be extended to all models in a future release.
+(WorkerDict pid=2841587) You have video processor config saved in `preprocessor.json` file which is deprecated. Video processor configs should be saved in their own `video_preprocessor.json` file. You can rename the file or load and save the processor back which renames it automatically. Loading from `preprocessor.json` will be removed in v5.0.
+(WorkerDict pid=2841587) 
+(WorkerDict pid=2839719) 
+(WorkerDict pid=2841587) 
+(WorkerDict pid=2839719) [rank0]:[W612 11:59:36.925129932 ProcessGroupNCCL.cpp:4561] [PG ID 0 PG GUID 0 Rank 0]  using GPU 0 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. Specify device_ids in barrier() to force use of a particular device, or call init_process_group() with a device_id.
+(WorkerDict pid=2841587) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLForConditionalGeneration is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=2841587) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=2841587) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VisionTransformerPretrainedModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)`
+(WorkerDict pid=2841587) 
+(WorkerDict pid=2841587) [rank1]:[W612 11:59:36.938180582 ProcessGroupNCCL.cpp:4561] [PG ID 0 PG GUID 0 Rank 1]  using GPU 0 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. Specify device_ids in barrier() to force use of a particular device, or call init_process_group() with a device_id.
+
+(WorkerDict pid=2839719)     "type": "default"
+(WorkerDict pid=2839719)   },
+(WorkerDict pid=2839719)   "rope_theta": 1000000.0,
+(WorkerDict pid=2839719)   "sliding_window": 32768,
+(WorkerDict pid=2839719)   "text_config": {
+(WorkerDict pid=2839719)     "architectures": [
+(WorkerDict pid=2839719)       "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=2839719)     ],
+(WorkerDict pid=2839719)     "attention_dropout": 0.0,
+(WorkerDict pid=2839719)     "bos_token_id": 151643,
+(WorkerDict pid=2839719)     "eos_token_id": 151645,
+(WorkerDict pid=2839719)     "hidden_act": "silu",
+(WorkerDict pid=2839719)     "hidden_size": 2048,
+(WorkerDict pid=2839719)     "image_token_id": null,
+(WorkerDict pid=2839719)     "initializer_range": 0.02,
+(WorkerDict pid=2839719)     "intermediate_size": 11008,
+(WorkerDict pid=2839719)     "layer_types": [
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention"
+(WorkerDict pid=2839719)     ],
+(WorkerDict pid=2839719)     "max_position_embeddings": 128000,
+(WorkerDict pid=2839719)     "max_window_layers": 70,
+(WorkerDict pid=2839719)     "model_type": "qwen2_5_vl_text",
+(WorkerDict pid=2839719)     "num_attention_heads": 16,
+(WorkerDict pid=2839719)     "num_hidden_layers": 36,
+(WorkerDict pid=2839719)     "num_key_value_heads": 2,
+(WorkerDict pid=2839719)     "rms_norm_eps": 1e-06,
+(WorkerDict pid=2839719)     "rope_scaling": {
+(WorkerDict pid=2839719)       "mrope_section": [
+(WorkerDict pid=2839719)         16,
+(WorkerDict pid=2839719)         24,
+(WorkerDict pid=2839719)         24
+(WorkerDict pid=2839719)       ],
+(WorkerDict pid=2839719)       "rope_type": "default",
+(WorkerDict pid=2839719)       "type": "default"
+(WorkerDict pid=2839719)     },
+(WorkerDict pid=2839719)     "rope_theta": 1000000.0,
+(WorkerDict pid=2839719)     "sliding_window": null,
+(WorkerDict pid=2839719)     "tie_word_embeddings": true,
+(WorkerDict pid=2839719)     "torch_dtype": "bfloat16",
+(WorkerDict pid=2839719)     "use_cache": true,
+(WorkerDict pid=2839719)     "use_sliding_window": false,
+(WorkerDict pid=2839719)     "video_token_id": null,
+(WorkerDict pid=2839719)     "vision_end_token_id": 151653,
+(WorkerDict pid=2839719)     "vision_start_token_id": 151652,
+(WorkerDict pid=2839719)     "vision_token_id": 151654,
+(WorkerDict pid=2839719)     "vocab_size": 151936
+(WorkerDict pid=2839719)   },
+(WorkerDict pid=2839719)   "torch_dtype": "bfloat16",
+(WorkerDict pid=2839719)   "transformers_version": "4.54.1",
+(WorkerDict pid=2839719)   "use_cache": true,
+(WorkerDict pid=2839719)   "use_sliding_window": false,
+(WorkerDict pid=2839719)   "video_token_id": 151656,
+(WorkerDict pid=2839719)   "vision_config": {
+(WorkerDict pid=2839719)     "depth": 32,
+(WorkerDict pid=2839719)     "fullatt_block_indexes": [
+(WorkerDict pid=2839719)       7,
+(WorkerDict pid=2839719)       15,
+(WorkerDict pid=2839719)       23,
+(WorkerDict pid=2839719)       31
+(WorkerDict pid=2839719)     ],
+(WorkerDict pid=2839719)     "hidden_act": "silu",
+(WorkerDict pid=2839719)     "hidden_size": 1280,
+(WorkerDict pid=2839719)     "in_channels": 3,
+(WorkerDict pid=2839719)     "in_chans": 3,
+(WorkerDict pid=2839719)     "initializer_range": 0.02,
+(WorkerDict pid=2839719)     "intermediate_size": 3420,
+(WorkerDict pid=2839719)     "model_type": "qwen2_5_vl",
+(WorkerDict pid=2839719)     "num_heads": 16,
+(WorkerDict pid=2839719)     "out_hidden_size": 2048,
+(WorkerDict pid=2839719)     "patch_size": 14,
+(WorkerDict pid=2839719)     "spatial_merge_size": 2,
+(WorkerDict pid=2839719)     "spatial_patch_size": 14,
+(WorkerDict pid=2839719)     "temporal_patch_size": 2,
+(WorkerDict pid=2839719)     "tokens_per_second": 2,
+(WorkerDict pid=2839719)     "window_size": 112
+(WorkerDict pid=2839719)   },
+(WorkerDict pid=2839719)   "vision_end_token_id": 151653,
+(WorkerDict pid=2839719)   "vision_start_token_id": 151652,
+(WorkerDict pid=2839719)   "vision_token_id": 151654,
+(WorkerDict pid=2839719)   "vocab_size": 151936
+(WorkerDict pid=2839719) }
+(WorkerDict pid=2839719) 
+(WorkerDict pid=2839719) NCCL version 2.21.5+cuda12.4
+(WorkerDict pid=2839719) Qwen2_5_VLForConditionalGeneration contains 3.75B parameters
+(WorkerDict pid=2839719) wrap_policy: functools.partial(<function _or_policy at 0x7e64dc2c84a0>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x7e64dc2c8360>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])
+(WorkerDict pid=2839719) Actor use_remove_padding=False
+(WorkerDict pid=2841587) wrap_policy: functools.partial(<function _or_policy at 0x7bba8c2c44a0>, policies=[functools.partial(<function transformer_auto_wrap_policy at 0x7bba8c2c4360>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])
+(WorkerDict pid=2839719) Model config after override: Qwen2_5_VLConfig {
+(WorkerDict pid=2839719)   "architectures": [
+(WorkerDict pid=2839719)     "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=2839719)   ],
+(WorkerDict pid=2839719)   "attention_dropout": 0.0,
+(WorkerDict pid=2839719)   "eos_token_id": 151645,
+(WorkerDict pid=2839719)   "hidden_act": "silu",(WorkerDict pid=2839719) 
+(WorkerDict pid=2839719) 
+(WorkerDict pid=2841587) Flash Attention 2 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in Qwen2_5_VLTextModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="flash_attention_2", torch_dtype=torch.float16)` [repeated 5x across cluster] (Ray deduplicates logs by default. Set RAY_DEDUP_LOGS=0 to disable log deduplication, or see https://docs.ray.io/en/master/ray-observability/user-guides/configure-logging.html#log-deduplication for more options.)
+(WorkerDict pid=2841587) 
+(WorkerDict pid=2839719) 
+(WorkerDict pid=2841587) 
+
+(WorkerDict pid=2839719)   "hidden_size": 2048,
+(WorkerDict pid=2839719)   "image_token_id": 151655,
+(WorkerDict pid=2839719)   "initializer_range": 0.02,
+(WorkerDict pid=2839719)   "intermediate_size": 11008,
+(WorkerDict pid=2839719)   "max_position_embeddings": 128000,
+(WorkerDict pid=2839719)   "max_window_layers": 70,
+(WorkerDict pid=2839719)   "model_type": "qwen2_5_vl",
+(WorkerDict pid=2839719)   "num_attention_heads": 16,
+(WorkerDict pid=2839719)   "num_hidden_layers": 36,
+(WorkerDict pid=2839719)   "num_key_value_heads": 2,
+(WorkerDict pid=2839719)   "pad_token_id": 151643,
+(WorkerDict pid=2839719)   "rms_norm_eps": 1e-06,
+(WorkerDict pid=2839719)   "rope_scaling": {
+(WorkerDict pid=2839719)     "mrope_section": [
+(WorkerDict pid=2839719)       16,
+(WorkerDict pid=2839719)       24,
+(WorkerDict pid=2839719)       24
+(WorkerDict pid=2839719)     ],
+(WorkerDict pid=2839719)     "rope_type": "default",
+(WorkerDict pid=2839719)     "type": "default"
+(WorkerDict pid=2839719)   },
+(WorkerDict pid=2839719)   "rope_theta": 1000000.0,
+(WorkerDict pid=2839719)   "sliding_window": 32768,
+(WorkerDict pid=2839719)   "text_config": {
+(WorkerDict pid=2839719)     "architectures": [
+(WorkerDict pid=2839719)       "Qwen2_5_VLForConditionalGeneration"
+(WorkerDict pid=2839719)     ],
+(WorkerDict pid=2839719)     "attention_dropout": 0.0,
+(WorkerDict pid=2839719)     "bos_token_id": 151643,
+(WorkerDict pid=2839719)     "eos_token_id": 151645,
+(WorkerDict pid=2839719)     "hidden_act": "silu",
+(WorkerDict pid=2839719)     "hidden_size": 2048,
+(WorkerDict pid=2839719)     "image_token_id": null,
+(WorkerDict pid=2839719)     "initializer_range": 0.02,
+(WorkerDict pid=2839719)     "intermediate_size": 11008,
+(WorkerDict pid=2839719)     "layer_types": [
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention",
+(WorkerDict pid=2839719)       "full_attention"
+(WorkerDict pid=2839719)     ],
+(WorkerDict pid=2839719)     "max_position_embeddings": 128000,
+(WorkerDict pid=2839719)     "max_window_layers": 70,
+(WorkerDict pid=2839719)     "model_type": "qwen2_5_vl_text",
+(WorkerDict pid=2839719)     "num_attention_heads": 16,
+(WorkerDict pid=2839719)     "num_hidden_layers": 36,
+(WorkerDict pid=2839719)     "num_key_value_heads": 2,
+(WorkerDict pid=2839719)     "rms_norm_eps": 1e-06,
+(WorkerDict pid=2839719)     "rope_scaling": {
+(WorkerDict pid=2839719)       "mrope_section": [
+(WorkerDict pid=2839719)         16,
+(WorkerDict pid=2839719)         24,
+(WorkerDict pid=2839719)         24
+(WorkerDict pid=2839719)       ],
+(WorkerDict pid=2839719)       "rope_type": "default",
+(WorkerDict pid=2839719)       "type": "default"
+(WorkerDict pid=2839719)     },
+(WorkerDict pid=2839719)     "rope_theta": 1000000.0,
+(WorkerDict pid=2839719)     "sliding_window": null,
+(WorkerDict pid=2839719)     "tie_word_embeddings": true,
+(WorkerDict pid=2839719)     "torch_dtype": "bfloat16",
+(WorkerDict pid=2839719)     "use_cache": true,
+(WorkerDict pid=2839719)     "use_sliding_window": false,
+(WorkerDict pid=2839719)     "video_token_id": null,
+(WorkerDict pid=2839719)     "vision_end_token_id": 151653,
+(WorkerDict pid=2839719)     "vision_start_token_id": 151652,
+(WorkerDict pid=2839719)     "vision_token_id": 151654,
+(WorkerDict pid=2839719)     "vocab_size": 151936
+(WorkerDict pid=2839719)   },
+(WorkerDict pid=2839719)   "torch_dtype": "bfloat16",
+(WorkerDict pid=2839719)   "transformers_version": "4.54.1",
+(WorkerDict pid=2839719)   "use_cache": true,
+(WorkerDict pid=2839719)   "use_sliding_window": false,
+(WorkerDict pid=2839719)   "video_token_id": 151656,
+(WorkerDict pid=2839719)   "vision_config": {
+(WorkerDict pid=2839719)     "depth": 32,
+(WorkerDict pid=2839719)     "fullatt_block_indexes": [
+(WorkerDict pid=2839719)       7,
+(WorkerDict pid=2839719)       15,
+(WorkerDict pid=2839719)       23,
+(WorkerDict pid=2839719)       31
+(WorkerDict pid=2839719)     ],
+(WorkerDict pid=2839719)     "hidden_act": "silu",
+(WorkerDict pid=2839719)     "hidden_size": 1280,
+(WorkerDict pid=2839719)     "in_channels": 3,
+(WorkerDict pid=2839719)     "in_chans": 3,
+(WorkerDict pid=2839719)     "initializer_range": 0.02,
+(WorkerDict pid=2839719)     "intermediate_size": 3420,
+(WorkerDict pid=2839719)     "model_type": "qwen2_5_vl",
+(WorkerDict pid=2839719)     "num_heads": 16,
+(WorkerDict pid=2839719)     "out_hidden_size": 2048,
+(WorkerDict pid=2839719)     "patch_size": 14,
+(WorkerDict pid=2839719)     "spatial_merge_size": 2,
+(WorkerDict pid=2839719)     "spatial_patch_size": 14,
+(WorkerDict pid=2839719)     "temporal_patch_size": 2,
+(WorkerDict pid=2839719)     "tokens_per_second": 2,
+(WorkerDict pid=2839719)     "window_size": 112
+(WorkerDict pid=2839719)   },
+(WorkerDict pid=2839719)   "vision_end_token_id": 151653,
+(WorkerDict pid=2839719)   "vision_start_token_id": 151652,
+(WorkerDict pid=2839719)   "vision_token_id": 151654,
+(WorkerDict pid=2839719)   "vocab_size": 151936
+(WorkerDict pid=2839719) }
+(WorkerDict pid=2839719) 
+(WorkerDict pid=2839719) trainable params: 37,152,768 || all params: 3,791,775,744 || trainable%: 0.9798
+(WorkerDict pid=2839719) PeftModelForCausalLM contains 3.79B parameters
+(WorkerDict pid=2839719) wrap_policy: functools.partial(<function _or_policy at 0x7e64dc2c84a0>, policies=[functools.partial(<function lambda_auto_wrap_policy at 0x7e64dc29be20>, lambda_fn=<function get_fsdp_wrap_policy.<locals>.lambda_policy_fn at 0x7e63661714e0>), functools.partial(<function transformer_auto_wrap_policy at 0x7e64dc2c8360>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])(WorkerDict pid=2839719) 
+(WorkerDict pid=2841587) 
+(WorkerDict pid=2839719) 
+(WorkerDict pid=2839719) 
+(WorkerDict pid=2839719) 
+(WorkerDict pid=2839719) 
+(WorkerDict pid=2839719) /home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/torch/distributed/fsdp/fully_sharded_data_parallel.py:690: FutureWarning: FSDP.state_dict_type() and FSDP.set_state_dict_type() are being deprecated. Please use APIs, get_state_dict() and set_state_dict(), which can support different parallelisms, FSDP1, FSDP2, DDP. API doc: https://pytorch.org/docs/stable/distributed.checkpoint.html#torch.distributed.checkpoint.state_dict.get_state_dict .Tutorial: https://pytorch.org/tutorials/recipes/distributed_checkpoint_recipe.html .
+(WorkerDict pid=2839719)   warnings.warn(
+(raylet) [2026-06-12 12:09:38,824 E 2832550 2832550] (raylet) node_manager.cc:3303: 1 Workers (tasks / actors) killed due to memory pressure (OOM), 0 Workers crashed due to other reasons at node (ID: 49eb9cdf575181869c147cbd458d0f76c939299db90454b4d5aa27eb, IP: 172.20.115.25) over the last time period. To see more information about the Workers killed on this node, use `ray logs raylet.out -ip 172.20.115.25`
+(raylet) 
+(raylet) Refer to the documentation on how to address the out of memory issue: https://docs.ray.io/en/latest/ray-core/scheduling/ray-oom-prevention.html. Consider provisioning more memory on this node or reducing task parallelism by requesting more CPUs per task. To adjust the kill threshold, set the environment variable `RAY_memory_usage_threshold` when starting Ray. To disable worker killing, set the environment variable `RAY_memory_monitor_refresh_ms` to zero.
+(WorkerDict pid=2841587) /home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/torch/distributed/fsdp/fully_sharded_data_parallel.py:690: FutureWarning: FSDP.state_dict_type() and FSDP.set_state_dict_type() are being deprecated. Please use APIs, get_state_dict() and set_state_dict(), which can support different parallelisms, FSDP1, FSDP2, DDP. API doc: https://pytorch.org/docs/stable/distributed.checkpoint.html#torch.distributed.checkpoint.state_dict.get_state_dict .Tutorial: https://pytorch.org/tutorials/recipes/distributed_checkpoint_recipe.html .
+(WorkerDict pid=2841587)   warnings.warn(
+
+(WorkerDict pid=2841587) Actor use_remove_padding=False
+(WorkerDict pid=2839719) Total steps: 472, num_warmup_steps: 0
+(WorkerDict pid=2839719) Actor use_remove_padding=False
+(WorkerDict pid=2839719) Before building vllm rollout, memory allocated (GB): 7.096388816833496, memory reserved (GB): 12.259765625
+(WorkerDict pid=2839719) WARNING 06-12 12:01:58 [utils.py:2522] Methods determine_num_available_blocks,device_config,get_cache_block_size_bytes,initialize_cache not implemented in <vllm.v1.worker.gpu_worker.Worker object at 0x7e62a0c00f50>
+(WorkerDict pid=2841587) wrap_policy: functools.partial(<function _or_policy at 0x7bba8c2c44a0>, policies=[functools.partial(<function lambda_auto_wrap_policy at 0x7bba8c297e20>, lambda_fn=<function get_fsdp_wrap_policy.<locals>.lambda_policy_fn at 0x7bb9b43e3d80>), functools.partial(<function transformer_auto_wrap_policy at 0x7bba8c2c4360>, transformer_layer_cls={<class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionBlock'>, <class 'transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLDecoderLayer'>})])
+(WorkerDict pid=2841587) Total steps: 472, num_warmup_steps: 0
+(WorkerDict pid=2841587) Actor use_remove_padding=False
+(WorkerDict pid=2839719) WARNING 06-12 12:02:03 [topk_topp_sampler.py:69] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
+(WorkerDict pid=2841587) WARNING 06-12 12:01:58 [utils.py:2522] Methods determine_num_available_blocks,device_config,get_cache_block_size_bytes,initialize_cache not implemented in <vllm.v1.worker.gpu_worker.Worker object at 0x7bb84b9fb090>
+(WorkerDict pid=2839719) kwargs: {'n': 5, 'logprobs': 0, 'max_tokens': 1024, 'detokenize': False, 'temperature': 1.0, 'top_k': -1, 'top_p': 1, 'ignore_eos': False}
+(WorkerDict pid=2839719) After building vllm rollout, memory allocated (GB): 13.632704257965088, memory reserved (GB): 15.044921875
+(WorkerDict pid=2839719) After building sharding manager, memory allocated (GB): 13.632704257965088, memory reserved (GB): 15.044921875
+(WorkerDict pid=2841587) WARNING 06-12 12:02:03 [topk_topp_sampler.py:69] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
+(TaskRunner pid=2835960) Using LocalLogger is deprecated. The constructor API will change 
+(TaskRunner pid=2835960) Found checkpoint: %s /mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_2gpu/global_step_472
+(TaskRunner pid=2835960) Load from checkpoint folder: /mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_2gpu/global_step_472
+(TaskRunner pid=2835960) Setting global step to 472
+(TaskRunner pid=2835960) Resuming from /mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_2gpu/global_step_472
+(WorkerDict pid=2839719) [rank-0]: Loading from /mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_2gpu/global_step_472/actor/model_world_size_2_rank_0.pt and /mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_2gpu/global_step_472/actor/optim_world_size_2_rank_0.pt and /mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_2gpu/global_step_472/actor/extra_state_world_size_2_rank_0.pt
+(TaskRunner pid=2835960) test_gen_batch meta info: {'eos_token_id': 151645, 'pad_token_id': 151643, 'recompute_log_prob': False, 'do_sample': False, 'validate': True}
+(WorkerDict pid=2841587) kwargs: {'n': 5, 'logprobs': 0, 'max_tokens': 1024, 'detokenize': False, 'temperature': 1.0, 'top_k': -1, 'top_p': 1, 'ignore_eos': False}
+(WorkerDict pid=2841587) [rank-1]: Loading from /mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_2gpu/global_step_472/actor/model_world_size_2_rank_1.pt and /mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_2gpu/global_step_472/actor/optim_world_size_2_rank_1.pt and /mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_2gpu/global_step_472/actor/extra_state_world_size_2_rank_1.pt
+Error executing job with overrides: ['algorithm.adv_estimator=grpo', 'data.train_files=./data/rats_uni_processed/train_quarter.parquet', 'data.val_files=./data/rats_uni_processed/test_full.parquet', 'data.train_batch_size=16', 'data.max_prompt_length=1024', 'data.max_response_length=1024', 'data.filter_overlong_prompts=True', 'data.truncation=error', 'data.image_key=images', 'actor_rollout_ref.model.path=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct', 'actor_rollout_ref.model.lora_rank=16', 'actor_rollout_ref.model.lora_alpha=16', 'actor_rollout_ref.model.lora_dropout=0.0', "actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'", 'actor_rollout_ref.actor.optim.lr=1e-4', 'actor_rollout_ref.model.use_remove_padding=False', 'actor_rollout_ref.actor.ppo_mini_batch_size=16', 'actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2', 'actor_rollout_ref.actor.use_kl_loss=True', 'actor_rollout_ref.actor.kl_loss_coef=0.001', 'actor_rollout_ref.actor.kl_loss_type=low_var_kl', 'actor_rollout_ref.model.enable_gradient_checkpointing=True', 'actor_rollout_ref.actor.fsdp_config.param_offload=False', 'actor_rollout_ref.actor.fsdp_config.optimizer_offload=False', 'actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8', 'actor_rollout_ref.rollout.tensor_model_parallel_size=2', 'actor_rollout_ref.rollout.name=vllm', 'actor_rollout_ref.rollout.gpu_memory_utilization=0.4', 'actor_rollout_ref.rollout.enable_chunked_prefill=False', 'actor_rollout_ref.rollout.enforce_eager=False', 'actor_rollout_ref.rollout.free_cache_engine=False', 'actor_rollout_ref.rollout.n=5', 'actor_rollout_ref.rollout.val_kwargs.do_sample=False', 'actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8', 'actor_rollout_ref.ref.fsdp_config.param_offload=True', 'algorithm.kl_ctrl.kl_coef=0.001', 'trainer.critic_warmup=0', "trainer.logger=['console']", 'trainer.project_name=anomseer', 'trainer.experiment_name=anomseer_rats_uni_2gpu', 'trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu', 'trainer.n_gpus_per_node=2', 'trainer.nnodes=1', 'trainer.save_freq=500', 'trainer.test_freq=10', 'trainer.val_only=True', 'trainer.val_before_train=True', 'trainer.total_epochs=1', 'ts.use_sem_orth=True', 'ts.adv_mix=0.3', 'ts.similarity_method=ot', 'ts.ot_eps=0.08', 'ts.ot_n_iter=50']
+Traceback (most recent call last):
+  File "/mnt/share01/sqk/AnomSeer/verl/trainer/main_ppo.py", line 68, in main
+    ray.get(runner.run.remote(config))
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/auto_init_hook.py", line 22, in auto_init_wrapper
+    return fn(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/client_mode_hook.py", line 107, in wrapper
+    return func(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/worker.py", line 2980, in get
+    values, debugger_breakpoint = worker.get_objects(
+                                  ^^^^^^^^^^^^^^^^^^^
+  File "/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/_private/worker.py", line 1023, in get_objects
+    raise value.as_instanceof_cause()
+ray.exceptions.RayTaskError(OutOfMemoryError): ray::TaskRunner.run() (pid=2835960, ip=172.20.115.25, actor_id=4827f4aae735b6de5961046b01000000, repr=<main_ppo.TaskRunner object at 0x735c213ffa50>)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/trainer/main_ppo.py", line 173, in run
+    trainer.fit()
+  File "/mnt/share01/sqk/AnomSeer/verl/trainer/ppo/ray_trainer.py", line 2196, in fit
+    val_metrics = self._validate()
+                  ^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/trainer/ppo/ray_trainer.py", line 1840, in _validate
+    test_output_gen_batch_padded = self.actor_rollout_wg.generate_sequences(test_gen_batch_padded)
+                                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/mnt/share01/sqk/AnomSeer/verl/single_controller/ray/base.py", line 43, in func
+    output = ray.get(output)
+             ^^^^^^^^^^^^^^^
+           ^^^^^^^^^^^^^^^^^^^
+           ^^^^^^^^^^^^^^^^^^^^^
+                                  ^^^^^^^^^^^^^^^^^^^
+ray.exceptions.OutOfMemoryError: 1 worker(s) were killed due to the node running low on memory. Memory on the node (IP: 172.20.115.25, ID: 49eb9cdf575181869c147cbd458d0f76c939299db90454b4d5aa27eb) was 239.55GB / 251.21GB (0.953616), which exceeds the memory usage threshold of 0.950000; Object store memory usage: [- objects spillable: 0; - bytes spillable: 0; - objects unsealed: 0; - bytes unsealed: 0; - objects in use: 1; - bytes in use: 20367164421; - objects evictable: 0; - bytes evictable: 0; ; - objects created by worker: 1; - bytes created by worker: 20367164421; - objects restored: 0; - bytes restored: 0; - objects received: 0; - bytes received: 0; - objects errored: 0; - bytes errored: 0; ; Eviction Stats:; (global lru) capacity: 72791529062; (global lru) used: 0%; (global lru) num objects: 0; (global lru) num evictions: 0; (global lru) bytes evicted: 0]; Ray killed 1 worker(s) based on the killing policy: [(Actor(4f145c2210673ecf4c40aa0d01000000): job ID=01000000, lease ID=03000000d21587b00dcc5a97a787fed0e40fafc5b956e0f280f921aed06b8784, task name=R4qskTWorkerDict_0:1:WorkerDict.__init__, required resources={GPU_group_218e351d7bf7afdb768364c8f98301000000: 1, bundle_group_1_218e351d7bf7afdb768364c8f98301000000: 0.001, CPU_group_218e351d7bf7afdb768364c8f98301000000: 1, GPU_group_1_218e351d7bf7afdb768364c8f98301000000: 1, CPU_group_1_218e351d7bf7afdb768364c8f98301000000: 1, bundle_group_218e351d7bf7afdb768364c8f98301000000: 0.001}, pid=2841587, actual memory used=14.59GB, worker ID=1af15360d55724f19e6b7f6b7eaf205cab912f858e2e6ef3b4745534)]; To see more information about memory usage on this node, use `ray logs raylet.out -ip 172.20.115.25`; Top 10 memory users: PID	MEM(GB)	COMMAND, 2835960	78.85	ray::TaskRunner.run, 2839719	50.54	ray::WorkerDict.actor_rollout_generate_sequences, 2874386	47.90	ray::TaskRunner.run, 2841587	14.59	ray::WorkerDict, 2349684	0.55	/home/suiqk/.vscode-server/cli/servers/Stable-1b50d58d73426c9171299ec4037d01365d995b78/server/node /..., 2830946	0.46	/home/suiqk/anaconda3/envs/scalerag-ts-v4/bin/python -m verl.trainer.main_ppo algorithm.adv_estimato..., 2133445	0.41	/home/suiqk/.vscode-server/cli/servers/Stable-1b50d58d73426c9171299ec4037d01365d995b78/server/node -..., 2133423	0.35	/home/suiqk/.vscode-server/cli/servers/Stable-1b50d58d73426c9171299ec4037d01365d995b78/server/node /..., 2832066	0.31	/home/suiqk/anaconda3/envs/scalerag-ts-v4/lib/python3.11/site-packages/ray/core/src/ray/gcs/gcs_serv..., 1402406	0.18	/usr/bin/python3 /usr/bin/update-manager --no-update --no-focus-on-map, Refer to the documentation on how to address the out of memory issue: https://docs.ray.io/en/latest/ray-core/scheduling/ray-oom-prevention.html. Consider provisioning more memory on this node or reducing task parallelism by requesting more CPUs per task. Set max_restarts and max_task_retries to enable retry when the task crashes due to OOM. To adjust the kill threshold, set the environment variable `RAY_memory_usage_threshold` when starting Ray. To disable worker killing, set the environment variable `RAY_memory_monitor_refresh_ms` to zero.
+
+Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

logs/rats_eval_full2_metrics.json

@@ -0,0 +1,18 @@
+{
+  "checkpoint": "checkpoints/anomseer/anomseer_rats_uni_2gpu/global_step_472",
+  "val_file": "data/rats_uni_processed/test_full.parquet",
+  "decoding": "greedy (do_sample=False, val_kwargs.n=1)",
+  "num_total_samples": 6034,
+  "num_valid_samples": 6034,
+  "num_invalid_predictions": 0,
+  "type_accuracy": 0.14368578057673184,
+  "type_precision_macro": 0.07011846062585,
+  "type_recall_macro": 0.15190385921696692,
+  "type_f1_macro": 0.06658531363554307,
+  "binary_accuracy": 0.8367583692409678,
+  "binary_precision_macro": 0.7156772435960252,
+  "binary_recall_macro": 0.6171428305303734,
+  "binary_f1_macro": 0.6403655346950877,
+  "reward_score": 0.2986244636942485,
+  "mean_fmt_score": 1.0
+}

multimodal_data_processing/anom.py

@@ -0,0 +1,596 @@
+# -*- coding: utf-8 -*-
+
+import os
+import re
+import json
+import glob
+import argparse
+import pickle
+from typing import List, Tuple, Optional, Dict
+
+import numpy as np
+import pandas as pd
+import stumpy
+from PIL import Image
+from datasets import Dataset
+from scipy.signal import savgol_filter
+from scipy.stats import zscore
+
+
+def _dominant_period(x: np.ndarray) -> float:
+    """Computes the dominant period of a time series using FFT."""
+    x = np.asarray(x, np.float32)
+    n = x.size
+    if n < 8: return float('inf')
+    xdm = x - x.mean()
+    Y = np.fft.rfft(xdm)
+    P = np.abs(Y)**2
+    if P.size > 0: P[0] = 0.0
+    freqs = np.fft.rfftfreq(n, d=1.0)
+    if P.size <= 1: return float('inf')
+    k = int(np.argmax(P))
+    f = float(freqs[k])
+    return float(1.0 / f) if f > 0 else float('inf')
+
+def vector_to_intervals(vec: np.ndarray) -> List[List[int]]:
+    """Converts a binary (0/1) vector to a list of [start, end) intervals."""
+    vec = np.asarray(vec).astype(int).flatten()
+    intervals = []
+    in_seg = False
+    start = 0
+    for i, v in enumerate(vec):
+        if v == 1 and not in_seg:
+            start = i
+            in_seg = True
+        elif v == 0 and in_seg:
+            intervals.append([start, i])
+            in_seg = False
+    if in_seg:
+        intervals.append([start, len(vec)])
+    return intervals
+
+
+
+def _try_load_npy(path: str) -> Optional[np.ndarray]:
+    try:
+        if os.path.isfile(path):
+            return np.load(path, allow_pickle=True)
+    except Exception:
+        return None
+
+def _find_series_and_gt(root_task_dir: str, split: str, idx: str) -> Tuple[Optional[np.ndarray], List[List[int]]]:
+    """Finds raw series and ground truth intervals from common directory patterns."""
+    series_path = os.path.join(root_task_dir, split, "series", f"{idx}.npy")
+    labels_path = os.path.join(root_task_dir, split, "labels", f"{idx}.npy")
+    
+    ts = _try_load_npy(series_path)
+    gt_vec = _try_load_npy(labels_path)
+    
+    intervals = []
+    if gt_vec is not None:
+        try:
+            intervals = vector_to_intervals(gt_vec.astype(int))
+        except Exception:
+            pass
+            
+    return ts, intervals
+
+def _parse_idx_from_png(png_path: str) -> str:
+    """Extracts the numerical index from a PNG filename."""
+    base = os.path.basename(png_path)
+    m = re.search(r"(\d+)", base)
+    return m.group(1) if m else os.path.splitext(base)[0]
+
+
+
+
+def detect_range_anomalies(ts: np.ndarray, k: float = 3.0) -> Tuple[List[List[int]], str, Dict]:
+    """Uses k-sigma rule for range anomalies and generates a CoT-style text."""
+    if ts.size < 2:
+        return [], "Series too short.", {}
+    
+    mu, std = np.mean(ts), np.std(ts)
+    if std < 1e-6:
+        return [], f"The series is constant (value={mu:.2f}), so no range anomalies can be detected.", {}
+        
+    upper, lower = mu + k * std, mu - k * std
+    is_anomaly = (ts > upper) | (ts < lower)
+    intervals = vector_to_intervals(is_anomaly)
+    
+    analysis_text = (
+        f"To find range anomalies, I first calculated the global statistics: "
+        f"the mean is {mu:.3f} and the standard deviation is {std:.3f}. "
+        f"This establishes a normal range of [{lower:.3f}, {upper:.3f}] using a {k}-sigma rule. "
+        f"Scanning the series for points outside this boundary resulted in finding {len(intervals)} anomalous interval(s)."
+    )
+    metrics = {"range_mean": mu, "range_std": std, "range_upper": upper, "range_lower": lower}
+    return intervals, analysis_text, metrics
+
+def detect_trend_anomalies(ts: np.ndarray, window: int = 21, k: float = 3.0) -> Tuple[List[List[int]], str, Dict]:
+    """Uses smoothed gradient for trend anomalies and generates a CoT-style text."""
+    if ts.size < window * 2:
+        return [], "Series too short for trend analysis.", {}
+
+    smooth_ts = savgol_filter(ts, window_length=window, polyorder=2)
+    gradient = np.gradient(smooth_ts)
+    grad_mu, grad_std = np.mean(gradient), np.std(gradient)
+    if grad_std < 1e-6:
+         return [], f"The trend appears to be constant (gradient={grad_mu:.4f}), so no significant changes were detected.", {}
+
+    threshold = k * grad_std
+    is_anomaly = np.abs(gradient - grad_mu) > threshold
+    intervals = vector_to_intervals(is_anomaly)
+    
+    analysis_text = (
+        f"For trend analysis, I focused on the rate of change. After smoothing the series with a window of {window}, "
+        f"I computed its gradient. The normal gradient fluctuates around a mean of {grad_mu:.4f} with a std of {grad_std:.4f}. "
+        f"A significant trend shift should cause a gradient deviation larger than the {k}-sigma threshold of {threshold:.4f}. "
+        f"This check identified {len(intervals)} interval(s) meeting the condition."
+    )
+    metrics = {"trend_grad_mean": grad_mu, "trend_grad_std": grad_std, "trend_grad_threshold": threshold}
+    return intervals, analysis_text, metrics
+
+def detect_point_anomalies(ts: np.ndarray, window: int = 50, k: float = 3.5) -> Tuple[List[List[int]], str, Dict]:
+    """Uses Matrix Profile (STUMPY) for point anomalies and generates a CoT-style text."""
+    if ts.size < window * 2:
+        return [], "Series too short for discord detection.", {}
+        
+
+    mp = stumpy.stump(ts, m=window, ignore_trivial=True)
+
+
+
+    raw_mp_distances = mp[:, 0]
+    mp_distances = np.zeros(raw_mp_distances.shape, dtype=np.float64)
+    for i, val in enumerate(raw_mp_distances):
+        try:
+            float_val = float(val)
+            mp_distances[i] = float_val
+        except (ValueError, TypeError):
+            mp_distances[i] = 0.0
+
+    finite_mask = np.isfinite(mp_distances)
+    if np.any(finite_mask):
+        max_finite_val = np.max(mp_distances[finite_mask])
+        mp_distances[~finite_mask] = max_finite_val
+    else:
+        return [], "Matrix Profile computation resulted in non-finite values only.", {}
+
+    mp_zscores = zscore(mp_distances)
+    discord_idx = np.argmax(mp_distances)
+    max_zscore = mp_zscores[discord_idx]
+    
+    intervals = []
+    if max_zscore > k:
+        intervals = [[int(discord_idx), int(discord_idx + window)]]
+
+    analysis_text = (
+        f"To detect contextual point anomalies, I used the Matrix Profile (window={window}, ignoring trivial matches) to find the most unusual subsequence (discord). "
+        f"The computation revealed that the highest discord score is {mp_distances[discord_idx]:.2f}, located at index {discord_idx}. "
+        f"This score corresponds to a z-score of {max_zscore:.2f}, which is above my threshold of {k}, indicating a significant anomaly. "
+        f"Thus, {len(intervals)} anomaly was identified."
+    )
+    metrics = {"point_discord_idx": int(discord_idx), "point_discord_zscore": float(max_zscore)}
+    return intervals, analysis_text, metrics
+
+def detect_freq_anomalies(ts: np.ndarray, window: int = 100, k: float = 3.0) -> Tuple[List[List[int]], str, Dict]:
+    """Uses dominant period changes for frequency anomalies and generates a CoT-style text."""
+    if ts.size < window * 3:
+        return [], "Series too short for frequency analysis.", {}
+        
+    periods = [_dominant_period(ts[i : i + window]) for i in range(ts.size - window)]
+    periods = np.array([p if np.isfinite(p) else -1 for p in periods])
+    valid_periods = periods[periods != -1]
+    
+    if valid_periods.size < 10:
+        return [], "Could not determine a stable dominant period.", {}
+
+    per_mu, per_std = np.mean(valid_periods), np.std(valid_periods)
+    if per_std < 1.0: # If period is very stable
+        return [], f"The dominant period is highly stable around {per_mu:.2f}, so no frequency anomalies detected.", {}
+
+    threshold_upper = per_mu + k * per_std
+    threshold_lower = per_mu - k * per_std
+    is_anomaly_periods = (periods > threshold_upper) | ((periods < threshold_lower) & (periods != -1))
+    
+    is_anomaly_ts = np.zeros_like(ts, dtype=bool)
+    for i, is_anom in enumerate(is_anomaly_periods):
+        if is_anom:
+            is_anomaly_ts[i : i + window] = True
+            
+    intervals = vector_to_intervals(is_anomaly_ts)
+    
+    analysis_text = (
+        f"To find frequency anomalies, I calculated the dominant period over a sliding window of size {window}. "
+        f"The typical period is around {per_mu:.2f} with a std of {per_std:.2f}. "
+        f"I'm looking for regions where the period significantly deviates from the normal range of [{threshold_lower:.2f}, {threshold_upper:.2f}]. "
+        f"This analysis pointed to {len(intervals)} interval(s) with clear frequency shifts."
+    )
+    metrics = {"freq_period_mean": per_mu, "freq_period_std": per_std}
+    return intervals, analysis_text, metrics
+
+
+def _json_default(o):
+    import numpy as _np
+    if isinstance(o, (_np.integer,)):
+        return int(o)
+    if isinstance(o, (_np.floating,)):
+        return float(o)
+    if isinstance(o, _np.ndarray):
+        return o.tolist()
+    return str(o)
+
+def _intervals_str(intervals):
+    if not intervals:
+        return "[]"
+    return ", ".join([f"[{int(s)}, {int(e)}]" for s, e in intervals])
+
+def _normalize_class(name: str) -> str:
+    m = (name or "").lower()
+    if m in ["global","global point","out-of-range","range","noisy-range"]:
+        return "global point"
+    if m in ["contextual","contextual point","point","noisy-point"]:
+        return "contextual point"
+    if m in ["trend","trend shift","noisy-trend","flat-trend"]:
+        return "trend"
+    if m in ["seasonal","frequency","seasonal/frequency deviation","freq","noisy-freq"]:
+        return "seasonal"
+    if m in ["shapelet","subsequence","shapelet/subsequence"]:
+        return "shapelet"
+    return "normal"
+
+
+def _get_expert_reasoning_flow(ts: np.ndarray, task_hint: str) -> Tuple[str, callable]:
+    """
+    Simulates an expert's diagnostic process to select the right tool.
+    Returns a reasoning text and the selected detection function.
+    """
+    # --- Step 1: Global Scan 
+    global_mu, global_std = np.mean(ts), np.std(ts)
+    max_zscore = 0
+    if global_std > 1e-6:
+        z_scores = np.abs((ts - global_mu) / global_std)
+        max_zscore = np.max(z_scores)
+
+    if max_zscore > 5.0: # A very high Z-score suggests a simple range anomaly is likely.
+        reasoning = (
+            "My initial check reveals extreme values. The global mean is "
+            f"{global_mu:.3f} and std is {global_std:.3f}, but some points have a z-score as high as {max_zscore:.2f}. "
+            "This strongly suggests a range-based anomaly. I will now apply a k-sigma rule to formalize this."
+        )
+        return reasoning, detect_range_anomalies
+
+    # --- Step 2: Structural Scan
+    
+    # Check for stable trend
+    gradient = np.gradient(ts)
+    grad_mu, grad_std = np.mean(gradient), np.std(gradient)
+    # A low gradient std relative to the overall data std might indicate a stable trend
+    is_trend_stable = (grad_std / global_std) < 0.1 if global_std > 1e-6 else True
+    
+    # Check for stable frequency
+    dominant_p = _dominant_period(ts)
+    is_freq_stable = dominant_p != float('inf') and dominant_p > 8 # has a detectable period
+
+    # Decision based on structure and task hint
+    task_lc = task_hint.lower()
+    if 'trend' in task_lc:
+        reasoning = (
+            f"The global values seem normal (max z-score={max_zscore:.2f}), so I'll check the trend. "
+            "The gradient of the series appears unstable. This suggests a potential trend anomaly. "
+            "I will analyze the smoothed gradient to confirm any significant shifts."
+        )
+        return reasoning, detect_trend_anomalies
+
+    if 'freq' in task_lc:
+        reasoning = (
+            f"Global values and trend seem stable. However, the signal appears periodic. "
+            "An unstable period can indicate a frequency anomaly. "
+            "I will use a sliding window analysis to check for significant changes in the dominant period."
+        )
+        return reasoning, detect_freq_anomalies
+
+    # --- Step 3: Pattern Scan (If all else seems normal, look for unique patterns) ---
+    # Corresponds to Matrix Profile logic. This is the default for 'point' or when other checks fail.
+    reasoning = (
+        f"The series does not exhibit obvious global outliers (max z-score={max_zscore:.2f}) or clear structural instability. "
+        "The anomalies are likely subtle and contextual. This requires a pattern-based approach. "
+        "I will use the Matrix Profile to find the most dissimilar subsequence (a discord), which is the standard method for such cases."
+    )
+    return reasoning, detect_point_anomalies
+
+    
+def _link_gt_feature(gt_intervals: List[List[int]], class_name: str) -> str:
+    iv_str = _intervals_str(gt_intervals)
+    if iv_str == "[]":
+        return f"I did not observe any {class_name} anomaly in the series. "
+
+    if class_name == "global point":
+        return f"I observed that the values within {iv_str} exhibit clear out-of-range behavior, with sharp spikes deviating from the global distribution. "
+    if class_name == "contextual point":
+        return f"I observed that the subsequence {iv_str} appears inconsistent with its local temporal context, breaking the continuity of surrounding patterns. "
+    if class_name == "trend":
+        return f"I observed that the segment within {iv_str} shows a clear trend shift, with the long-term slope undergoing a marked change. "
+    if class_name == "seasonal":
+        return f"I observed that the oscillations within {iv_str} display frequency deviation, with periodic structure misaligned from the baseline cycle. "
+    if class_name == "shapelet":
+        return f"I observed that the subsequence within {iv_str} deviates in waveform shape, differing notably from typical local motifs. "
+    return f"I observed that the segment {iv_str} is annotated as {class_name}. "
+
+
+def build_prompt_and_expcot(ts: Optional[np.ndarray],
+                             task: str,
+                             L: int,
+                             gt_intervals: List[List[int]],
+                             gt_type: str) -> Tuple[str, str, Dict]:
+
+    prompt = (
+        "<image>\n"
+        f"You are a time series analysis expert. A time series plot of length {L} is provided. "
+        "Identify anomalous intervals along the x-axis and infer the most plausible anomaly type from "
+        "[\"contextual point\", \"global point\", \"seasonal\", \"trend\", \"shapelet\"].\n\n"
+        "Begin detailed reasoning inside <think>...</think>.\n"
+        "Then output:\n"
+        "<answer>[[start, end], ...]</answer>\n"
+        "<class>one of {contextual point, global point, seasonal, trend, shapelet, normal}</class>\n"
+        "If no anomalies, return <answer>[]</answer> and <class>normal</class>.\n"
+    )
+
+  
+    class_by_task = _normalize_class(task)
+    class_by_gt   = _normalize_class(gt_type)
+    class_name    = class_by_gt if class_by_gt != "normal" else class_by_task
+
+    if ts is None or ts.size < 20:
+        if gt_intervals:
+            expcot = (
+                f"\\textbf{{Observation}} — Ground truth marks { _intervals_str(gt_intervals) } "
+                f"as \\textit{{{class_name}}}, but raw series is unavailable/too short for verification.\n"
+                f"\\textbf{{Conclusion}} — We retain the GT label and intervals for supervision."
+            )
+        else:
+            expcot = (
+                "\\textbf{Observation} — Ground truth indicates no anomaly; raw series unavailable/too short.\n"
+                "\\textbf{Conclusion} — Treated as normal."
+            )
+        return prompt, expcot, {"intervals_gt": gt_intervals, "class": class_name}
+
+
+    x = ts if ts.ndim == 1 else ts.mean(axis=1)
+    pre = ""
+    if "noisy" in task.lower():
+        wl = max(5, min(31, (len(x)//4)|1))
+        x  = savgol_filter(x, window_length=wl, polyorder=2)
+        pre = f"[Preprocess] Applied Savitzky–Golay denoising (win={wl}). "
+
+    initial_reasoning, selected_detector = _get_expert_reasoning_flow(x, task)
+    det_intervals, final_evidence, detection_metrics = selected_detector(x)
+
+
+    obs = (
+        f"Ground truth labels { _intervals_str(gt_intervals) } as \\textit{{{class_name}}}. "
+        f"{initial_reasoning}"
+    )
+
+
+    val = (
+    _link_gt_feature(gt_intervals, class_name)
+    + final_evidence.replace("To find", "The analysis").replace("For", "The analysis")
+)
+
+
+  
+    if gt_intervals:
+        concl = (
+            f"We report the GT interval(s) { _intervals_str(gt_intervals) } "
+            f"as the final localization for supervision. "
+            f"For reference, the detector proposed { _intervals_str(det_intervals) }."
+        )
+        out_intervals = gt_intervals  # 
+    else:
+        concl = (
+            "No ground-truth anomalies are present; the series is treated as normal. "
+            f"For reference, the detector proposed { _intervals_str(det_intervals) }."
+        )
+        out_intervals = []  # normal
+
+    expcot = (
+        f"{pre}"
+        f"\\textbf{{Observation}} — {obs}\n"
+        f"\\textbf{{Reasoning \\& Validation}} — {val}\n"
+        f"\\textbf{{Conclusion}} — {concl}"
+    ).strip()
+
+  
+    return prompt, expcot, {
+        "intervals_gt": gt_intervals,
+        "intervals_pred": det_intervals,
+        "class": class_name,
+        **detection_metrics
+    }
+
+# --- CORRECTED DATA LOADING HELPERS ---
+def _load_split_pkl(task_dir: str, split: str) -> Tuple[Optional[list], Optional[list]]:
+
+    pkl_path = os.path.join(task_dir, split, "data.pkl")
+    if not os.path.isfile(pkl_path):
+        return None, None
+    try:
+        with open(pkl_path, "rb") as f:
+            obj = pickle.load(f)
+        series = obj.get("series", None)
+        anom = obj.get("anom", None)
+        return series, anom
+    except Exception:
+        return None, None
+
+def _pick_index_from_png_idx(idx_str: str, n: int) -> Optional[int]:
+
+    try:
+        k = int(idx_str)
+    except Exception:
+        return None
+    for cand in (k-1, k):
+        if 0 <= cand < n:
+            return cand
+    return None
+
+def _intervals_from_anom_entry(anom_entry) -> List[List[int]]:
+
+    intervals = []
+    if not anom_entry: return intervals
+    for ch_list in anom_entry:
+        for pair in ch_list:
+            if isinstance(pair, (list, tuple)) and len(pair) == 2:
+                s, e = int(pair[0]), int(pair[1])
+                if e > s: intervals.append([s, e])
+    intervals.sort(key=lambda z: (z[0], z[1]))
+    return intervals
+
+def _parse_idx_from_png(png_path: str) -> str:
+
+    base = os.path.basename(png_path)
+    m = re.search(r"(\d+)", base)
+    return m.group(1) if m else os.path.splitext(base)[0]
+
+def process_task(root_dir: str, task: str, out_dir: str) -> Tuple[str, Optional[str]]:
+
+    task_dir = os.path.join(root_dir, task)
+    results = {"train": [], "eval": []}
+
+    for split in ["train", "eval"]:
+        figs_dir = os.path.join(task_dir, split, "figs")
+        if not os.path.isdir(figs_dir):
+            print(f"Warning: Directory not found for {task}/{split}/figs. Skipping.")
+            continue
+
+      
+        series_list, anom_list = _load_split_pkl(task_dir, split)
+        n_samples = len(series_list) if isinstance(series_list, list) else 0
+        use_pkl = (series_list is not None) and (anom_list is not None) and (n_samples > 0)
+        
+        pngs = sorted(glob.glob(os.path.join(figs_dir, "*.png")))
+        for i, png_path in enumerate(pngs):
+            idx_str = _parse_idx_from_png(png_path)
+            
+            ts, gt_intervals = None, []
+            
+     
+            if use_pkl:
+                pkl_i = _pick_index_from_png_idx(idx_str, n_samples)
+                if pkl_i is not None:
+                    ts = np.asarray(series_list[pkl_i])
+                    gt_intervals = _intervals_from_anom_entry(anom_list[pkl_i])
+
+
+            def map_task_to_anomaly(task_name: str) -> str:
+                if task_name in ["range", "noisy-range"]:
+                    return "global"
+                elif task_name in ["freq", "noise-freq"]:
+                    return "seasonal"
+                elif task_name in ["point", "noise-point"]:
+                    return "contextual"
+                elif task_name in ["trend", "noise-trend"]:
+                    return "trend"
+                else:
+                    return "unknown" 
+            L = len(ts) if ts is not None else 0
+
+            anomaly_type = map_task_to_anomaly(task)
+            if gt_intervals == []:
+                anomaly_type = 'normal'
+            gt_type = anomaly_type if anomaly_type else _normalize_class(task)
+            prompt_text, expcot, detection_metrics = build_prompt_and_expcot(ts, task, L, gt_intervals, gt_type)
+
+            image = Image.open(png_path).convert("RGBA")
+                
+
+
+            row = {
+                "data_source": "timeseries_anol",
+                "prompt": [{"role": "user", "content": prompt_text}],
+                "images": [image],
+                "ability": "time_series_anomaly_detection",
+                "reward_model": {"style": "rule", "ground_truth": gt_intervals},
+                "extra_info": {
+                    "category": task,
+                    "split": split,
+                    "instance_index": i,
+                    "image_path": png_path,
+                    "expcot": expcot,
+                    "detection_metrics": detection_metrics,
+                    "gt_intervals": gt_intervals,
+                    "series_length": len(ts) if ts is not None else 0,
+                    "index": pkl_i,
+                    "anomaly_type": anomaly_type   
+                }
+            }
+            results[split].append(row)
+
+    os.makedirs(out_dir, exist_ok=True)
+    train_path = os.path.join(out_dir, f"{task}_train.parquet")
+    test_path  = os.path.join(out_dir, f"{task}_test.parquet")
+
+    if results["train"]:
+        train_ds = Dataset.from_list(results["train"])
+        train_ds.to_parquet(train_path)
+    else:
+        train_path = ""
+
+    if results["eval"]:
+        eval_ds = Dataset.from_list(results["eval"])
+        eval_ds.to_parquet(test_path)
+    else:
+        test_path = None
+
+    return train_path, test_path
+
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Process synthetic time series anomaly data into parquet files.")
+    parser.add_argument("--root_dir", type=str,
+                        default="./data/anomllm/data/synthetic",
+                        help="Root directory containing the task folders.")
+    parser.add_argument("--out_dir", type=str,
+                        default="./data/anol_processed_mllm_data",
+                        help="Output directory to store the final parquet files.")
+    args = parser.parse_args()
+
+    tasks = ["trend", "freq", "point", "range", "noisy-trend", "noisy-freq", "noisy-point", "flat-trend"]
+
+    print(f"Input data root: {args.root_dir}")
+    print(f"Output directory: {args.out_dir}")
+    os.makedirs(args.out_dir, exist_ok=True)
+
+    all_train_dfs = []
+    all_test_dfs = []
+
+    for task in tasks:
+        print(f"\nProcessing task: {task}...")
+
+        tr_path, te_path = process_task(args.root_dir, task, args.out_dir)
+        if tr_path and os.path.exists(tr_path):
+            print(f"  ✅ Saved train data to -> {tr_path}")
+            all_train_dfs.append(pd.read_parquet(tr_path))
+        if te_path and os.path.exists(te_path):
+            print(f"  ✅ Saved test data to  -> {te_path}")
+            all_test_dfs.append(pd.read_parquet(te_path))
+        if not tr_path and not te_path:
+            print(f"  ⚠️ No data found for task '{task}'. Check directory structure.")
+
+
+    if all_train_dfs:
+        full_train_df = pd.concat(all_train_dfs, ignore_index=True)
+        full_train_path = os.path.join(args.out_dir, "train_full.parquet")
+        full_train_df.to_parquet(full_train_path, index=False)
+        print(f"\n📦 Successfully merged and saved all training data to -> {full_train_path}")
+
+    if all_test_dfs:
+        full_test_df = pd.concat(all_test_dfs, ignore_index=True)
+        full_test_path = os.path.join(args.out_dir, "test_full.parquet")
+        full_test_df.to_parquet(full_test_path, index=False)
+        print(f"📦 Successfully merged and saved all testing data to  -> {full_test_path}")
+
+if __name__ == "__main__":
+    main()

multimodal_data_processing/rats_uni.py

@@ -0,0 +1,268 @@
+# -*- coding: utf-8 -*-
+"""
+Convert the Time-RA RATs-Uni (univariate) reasoning dataset into the parquet
+format consumed by AnomSeer's veRL training/eval pipeline.
+
+Unlike AnomLLM (see ``anom.py``), the Time-RA univariate dataset does **not**
+provide anomaly *intervals* — it provides, per univariate series:
+
+  * ``Observation``  : the raw time series (list of floats, length 32/64/128)
+  * ``FigurePath``   : a rendered plot (``figures/{train,test}/{idx}.jpg``)
+  * ``ActionID``     : the anomaly class id (0-14, 0 == normal)
+  * ``Action``       : the human-readable class name
+  * ``Label``        : "Normal" / "Anomaly"
+  * ``Thought``      : an expert chain-of-thought explanation
+
+We therefore adapt AnomSeer to a **classification + reasoning** task (no
+localization). Each output row matches the schema expected by
+``verl.utils.dataset.rl_dataset.RLHFDataset`` and is scored by the
+``timeseries_rats`` reward (``verl/utils/reward_score/rats.py``):
+
+  data_source  : "timeseries_rats"
+  prompt       : [{"role": "user", "content": "<image>\\n ... <class> ..."}]
+  images       : [{"bytes": <jpg bytes>, "path": <abs path>}]
+  ability      : "time_series_anomaly_detection"
+  reward_model : {"style": "rule", "ground_truth": <canonical class name>}
+  extra_info   : {category, source, split, anomaly_type, action_id, label,
+                  series_length, image_path, expcot, numtext, index}
+
+The ``numtext`` field is the expert CoT and is **required** by TimerPO's
+semantic-alignment term (``compute_hidden_states_of_hint`` in fsdp_workers.py).
+
+Usage
+-----
+    python multimodal_data_processing/rats_uni.py \
+        --json_path /path/to/RATs40K/RATs-Uni-TSImage_Reason.json \
+        --out_dir   ./data/rats_uni_processed
+
+Produces ``train_full.parquet`` and ``test_full.parquet`` in ``--out_dir``.
+"""
+
+import os
+import io
+import json
+import argparse
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+import pandas as pd
+from PIL import Image
+
+
+# Canonical 15-class taxonomy (id -> (name, description)), matching
+# Time-RA/eval_utils.py::ANOMALY_DICT exactly.
+ANOMALY_DICT: Dict[int, Tuple[str, str]] = {
+    0:  ("Normal Sequence", "There are no abnormal situations in this time series."),
+    1:  ("Point Anomaly", "A single data point significantly deviates from the local or global pattern of the sequence."),
+    2:  ("Periodic Change Anomaly", "The original periodic pattern is disrupted, e.g. the period is broken or the amplitude becomes anomalous."),
+    3:  ("Trend Change Anomaly", "A sudden change in the long-term trend of the time series."),
+    4:  ("Change Point Anomaly", "Statistical properties (e.g. mean, variance) change abruptly at certain points."),
+    5:  ("Distributional Change Anomaly", "The statistical distribution of the time series changes significantly."),
+    6:  ("Amplitude Anomaly", "The amplitude of data points exceeds the normal upper and lower bounds."),
+    7:  ("Pattern Change Anomaly", "The pattern of the time series suddenly changes from one form to another."),
+    8:  ("Sparse Anomaly", "Isolated anomalous patterns occasionally appear in a long time series."),
+    9:  ("Repeated Value Anomaly", "Continuous or intermittent repeated values disrupt the normal fluctuation pattern."),
+    10: ("Sudden Flatline Anomaly", "The time series suddenly becomes a flat line with no normal fluctuations."),
+    11: ("Drift Anomaly", "The data gradually drifts away from the normal level."),
+    12: ("Sudden Spike Anomaly", "The data suddenly spikes or drops within a short time and then returns to normal."),
+    13: ("Continuous Segment Anomaly", "A continuous segment of data points deviates from the normal pattern."),
+    14: ("Nonlinear Pattern Anomaly", "Nonlinear changes appear in the sequence, breaking the original linear rule."),
+}
+
+NAME_BY_ID = {i: name for i, (name, _) in ANOMALY_DICT.items()}
+ID_BY_NAME = {name.lower(): i for i, (name, _) in ANOMALY_DICT.items()}
+
+# Splits in the JSON -> output split name used by AnomSeer ("eval" == val/test).
+SPLIT_MAP = {"TSAD_train": "train", "TSAD_test": "eval"}
+
+
+def _build_taxonomy_block() -> str:
+    lines = []
+    for i in range(len(ANOMALY_DICT)):
+        name, desc = ANOMALY_DICT[i]
+        lines.append(f"{i}: {name} — {desc}")
+    return "\n".join(lines)
+
+
+_TAXONOMY_BLOCK = _build_taxonomy_block()
+_CLASS_NAMES = ", ".join(NAME_BY_ID[i] for i in range(len(NAME_BY_ID)))
+
+
+def build_prompt(length: int, source: str) -> str:
+    """Classification + reasoning prompt (no interval localization)."""
+    return (
+        "<image>\n"
+        f"You are an expert in univariate time-series anomaly detection. The figure shows a "
+        f"single-channel time series of length {length} from the \"{source}\" domain.\n\n"
+        "Decide whether the series is normal or contains an anomaly. If it is anomalous, choose the "
+        "single most appropriate anomaly type from the following 15 categories "
+        "(format `id: name — description`):\n"
+        f"{_TAXONOMY_BLOCK}\n\n"
+        "Reason step by step inside <think>...</think> based on the visual shape of the series, "
+        "then output exactly one line with your final answer:\n"
+        "<class>one exact category name from the list above</class>\n"
+        "If the series is normal, use <class>Normal Sequence</class>."
+    )
+
+
+def _canonical_class(action: Optional[str], action_id: Optional[int]) -> Tuple[str, int]:
+    """Resolve the canonical (name, id) from the raw Action / ActionID fields.
+
+    ActionID is treated as authoritative; Action is used only as a fallback.
+    """
+    if isinstance(action_id, (int, np.integer)) and int(action_id) in NAME_BY_ID:
+        cid = int(action_id)
+        return NAME_BY_ID[cid], cid
+    if isinstance(action, str) and action.strip().lower() in ID_BY_NAME:
+        cid = ID_BY_NAME[action.strip().lower()]
+        return NAME_BY_ID[cid], cid
+    # Unknown -> treat as normal (id 0) so downstream typing stays consistent.
+    return NAME_BY_ID[0], 0
+
+
+def _read_image_bytes(path: str) -> Optional[bytes]:
+    """Return raw image bytes, re-encoding to a clean RGB JPEG if needed."""
+    if not os.path.isfile(path):
+        return None
+    try:
+        with open(path, "rb") as f:
+            raw = f.read()
+        # Validate; re-encode non-RGB to avoid downstream surprises.
+        with Image.open(io.BytesIO(raw)) as im:
+            if im.mode == "RGB":
+                return raw
+            buf = io.BytesIO()
+            im.convert("RGB").save(buf, format="JPEG", quality=95)
+            return buf.getvalue()
+    except Exception as exc:  # noqa: BLE001
+        print(f"[WARN] failed to read image {path}: {exc}")
+        return None
+
+
+# Some test entries point FigurePath at a ``.pdf`` that was never exported; the
+# rendered raster lives next to it as a same-index ``.jpg``. Fall back across
+# common raster extensions before giving up.
+_IMG_EXT_FALLBACKS = (".jpg", ".png", ".jpeg")
+
+
+def _resolve_figure_path(data_root: str, figure_path: str) -> str:
+    """FigurePath is stored relative to the dataset root (e.g. figures/train/0.jpg).
+
+    Returns the first existing file, trying the stored path first and then the
+    same stem with a raster extension (handles ``.pdf`` FigurePaths)."""
+    base = figure_path if os.path.isabs(figure_path) else os.path.join(data_root, figure_path)
+    if os.path.isfile(base):
+        return base
+    stem, _ = os.path.splitext(base)
+    for ext in _IMG_EXT_FALLBACKS:
+        cand = stem + ext
+        if os.path.isfile(cand):
+            return cand
+    return base  # non-existent; caller reports it as a skip
+
+
+def process_split(records: dict, split_name: str, data_root: str,
+                  max_samples: Optional[int] = None) -> List[dict]:
+    rows: List[dict] = []
+    n_skipped_none = 0
+    n_skipped_img = 0
+
+    # Iterate in numeric key order for reproducibility.
+    keys = sorted(records.keys(), key=lambda k: int(k) if str(k).isdigit() else k)
+    for out_idx, key in enumerate(keys):
+        if max_samples is not None and len(rows) >= max_samples:
+            break
+
+        entry = records[key]
+        if entry is None:  # the JSON contains a handful of null test entries
+            n_skipped_none += 1
+            continue
+
+        obs = entry.get("Observation") or []
+        length = len(obs)
+        source = entry.get("Source", "unknown")
+        figure_path = entry.get("FigurePath", "")
+        thought = entry.get("Thought", "") or ""
+        label = entry.get("Label", "Anomaly")
+
+        img_path = _resolve_figure_path(data_root, figure_path)
+        img_bytes = _read_image_bytes(img_path)
+        if img_bytes is None:
+            n_skipped_img += 1
+            continue
+
+        class_name, class_id = _canonical_class(entry.get("Action"), entry.get("ActionID"))
+        prompt_text = build_prompt(length, source)
+
+        rows.append({
+            "data_source": "timeseries_rats",
+            "prompt": [{"role": "user", "content": prompt_text}],
+            "images": [{"bytes": img_bytes, "path": img_path}],
+            "ability": "time_series_anomaly_detection",
+            "reward_model": {"style": "rule", "ground_truth": class_name},
+            "extra_info": {
+                "index": int(key) if str(key).isdigit() else out_idx,
+                "category": class_name,          # per-class metric grouping
+                "source": source,                # domain grouping
+                "split": split_name,
+                "anomaly_type": class_name,       # read by the reward as GT class
+                "action_id": int(class_id),
+                "label": label,
+                "series_length": int(length),
+                "image_path": img_path,
+                "expcot": thought,                # expert CoT
+                "numtext": thought,               # required by TimerPO hint term
+            },
+        })
+
+    print(f"  [{split_name}] kept={len(rows)} "
+          f"skipped(null)={n_skipped_none} skipped(missing-image)={n_skipped_img}")
+    return rows
+
+
+def write_parquet(rows: List[dict], path: str) -> None:
+    if not rows:
+        print(f"  [skip] no rows for {path}")
+        return
+    df = pd.DataFrame(rows)
+    df.to_parquet(path, index=False, engine="pyarrow")
+    print(f"  ✅ wrote {len(df)} rows -> {path}")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Convert Time-RA RATs-Uni JSON into AnomSeer parquet (classification + reasoning).")
+    parser.add_argument("--json_path", type=str,
+                        default="/mnt/share01/sqk/datasets/RATs40K/RATs-Uni-TSImage_Reason.json",
+                        help="Path to RATs-Uni-TSImage_Reason.json")
+    parser.add_argument("--data_root", type=str, default=None,
+                        help="Dataset root that FigurePath is relative to "
+                             "(defaults to the directory of --json_path).")
+    parser.add_argument("--out_dir", type=str, default="./data/rats_uni_processed",
+                        help="Output directory for the parquet files.")
+    parser.add_argument("--max_samples", type=int, default=None,
+                        help="Optional cap on samples per split (for quick tests).")
+    args = parser.parse_args()
+
+    data_root = args.data_root or os.path.dirname(os.path.abspath(args.json_path))
+    os.makedirs(args.out_dir, exist_ok=True)
+
+    print(f"Reading {args.json_path}")
+    print(f"Figure root: {data_root}")
+    with open(args.json_path, "r") as f:
+        data = json.load(f)
+
+    for json_split, out_split in SPLIT_MAP.items():
+        if json_split not in data:
+            print(f"[WARN] split '{json_split}' not found in JSON; skipping.")
+            continue
+        print(f"\nProcessing {json_split} -> {out_split}")
+        rows = process_split(data[json_split], out_split, data_root, args.max_samples)
+        fname = "train_full.parquet" if out_split == "train" else "test_full.parquet"
+        write_parquet(rows, os.path.join(args.out_dir, fname))
+
+    print("\nDone.")
+
+
+if __name__ == "__main__":
+    main()

outputs/2026-06-09/15-29-24/.hydra/config.yaml

@@ -0,0 +1,194 @@
+data:
+  tokenizer: null
+  train_files: ./data/rats_uni_processed/train_full.parquet
+  val_files: ./data/rats_uni_processed/test_full.parquet
+  prompt_key: prompt
+  max_prompt_length: 1024
+  max_response_length: 1024
+  train_batch_size: 128
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  shuffle: true
+  filter_overlong_prompts: true
+  truncation: error
+  image_key: images
+actor_rollout_ref:
+  hybrid_engine: true
+  model:
+    path: /mnt/share01/sqk/models/qwen2.5-vl-7b-instruct
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    use_remove_padding: true
+    lora_rank: 16
+    lora_alpha: 16
+    lora_dropout: 0.0
+    lora_target_modules: all-linear
+  actor:
+    strategy: fsdp
+    ppo_mini_batch_size: 128
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: 2
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    grad_clip: 1.0
+    clip_ratio: 0.2
+    entropy_coeff: 0.001
+    use_kl_loss: true
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: low_var_kl
+    ppo_epochs: 1
+    shuffle: false
+    ulysses_sequence_parallel_size: 1
+    optim:
+      lr: 0.0001
+      lr_warmup_steps: -1
+      lr_warmup_steps_ratio: 0.0
+      min_lr_ratio: null
+      warmup_style: constant
+      total_training_steps: -1
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      fsdp_size: -1
+  ref:
+    fsdp_config:
+      param_offload: true
+      wrap_policy:
+        min_num_params: 0
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    ulysses_sequence_parallel_size: ${actor_rollout_ref.actor.ulysses_sequence_parallel_size}
+  rollout:
+    name: vllm
+    temperature: 1.0
+    top_k: -1
+    top_p: 1
+    use_fire_sampling: false
+    prompt_length: ${data.max_prompt_length}
+    response_length: ${data.max_response_length}
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: false
+    free_cache_engine: false
+    load_format: dummy_dtensor
+    tensor_model_parallel_size: 2
+    max_num_batched_tokens: 8192
+    max_model_len: null
+    max_num_seqs: 1024
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    disable_log_stats: true
+    enable_chunked_prefill: false
+    do_sample: true
+    'n': 5
+    val_kwargs:
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.4
+      'n': 1
+      do_sample: true
+critic:
+  strategy: fsdp
+  optim:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    min_lr_ratio: null
+    warmup_style: constant
+    total_training_steps: -1
+  model:
+    path: ~/models/deepseek-llm-7b-chat
+    tokenizer_path: ${actor_rollout_ref.model.path}
+    override_config: {}
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    fsdp_config:
+      param_offload: false
+      optimizer_offload: false
+      wrap_policy:
+        min_num_params: 0
+      fsdp_size: -1
+  ppo_mini_batch_size: ${actor_rollout_ref.actor.ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: null
+  forward_micro_batch_size: ${critic.ppo_micro_batch_size}
+  forward_micro_batch_size_per_gpu: ${critic.ppo_micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${critic.ppo_max_token_len_per_gpu}
+  ulysses_sequence_parallel_size: 1
+  ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
+  shuffle: ${actor_rollout_ref.actor.shuffle}
+  grad_clip: 1.0
+  cliprange_value: 0.5
+reward_model:
+  enable: false
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    use_remove_padding: false
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      fsdp_size: -1
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  ulysses_sequence_parallel_size: 1
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: grpo
+  kl_penalty: kl
+  kl_ctrl:
+    type: fixed
+    kl_coef: 0.001
+trainer:
+  balance_batch: true
+  total_epochs: 10
+  total_training_steps: null
+  project_name: anomseer
+  experiment_name: anomseer_rats_uni_2gpu
+  logger:
+  - console
+  val_generations_to_log_to_wandb: 0
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: 500
+  resume_mode: auto
+  resume_from_path: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  remove_previous_ckpt_in_save: false
+  del_local_ckpt_after_load: false
+  default_local_dir: checkpoints/anomseer/anomseer_rats_uni_2gpu
+  val_only: false
+  val_before_train: true
+ts:
+  adv_mix: 0.3
+  use_sem_orth: true
+  similarity_method: ot
+  clip_temperature: 1.0
+  clip_pooling: mean
+  ot_eps: 0.08
+  ot_n_iter: 50

outputs/2026-06-09/15-29-24/.hydra/hydra.yaml

@@ -0,0 +1,206 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - algorithm.adv_estimator=grpo
+    - data.train_files=./data/rats_uni_processed/train_full.parquet
+    - data.val_files=./data/rats_uni_processed/test_full.parquet
+    - data.train_batch_size=128
+    - data.max_prompt_length=1024
+    - data.max_response_length=1024
+    - data.filter_overlong_prompts=True
+    - data.truncation=error
+    - data.image_key=images
+    - actor_rollout_ref.model.path=/mnt/share01/sqk/models/qwen2.5-vl-7b-instruct
+    - actor_rollout_ref.model.lora_rank=16
+    - actor_rollout_ref.model.lora_alpha=16
+    - actor_rollout_ref.model.lora_dropout=0.0
+    - actor_rollout_ref.model.lora_target_modules=all-linear
+    - actor_rollout_ref.actor.optim.lr=1e-4
+    - actor_rollout_ref.model.use_remove_padding=True
+    - actor_rollout_ref.actor.ppo_mini_batch_size=128
+    - actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+    - actor_rollout_ref.actor.use_kl_loss=True
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_type=low_var_kl
+    - actor_rollout_ref.model.enable_gradient_checkpointing=True
+    - actor_rollout_ref.actor.fsdp_config.param_offload=False
+    - actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+    - actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.rollout.tensor_model_parallel_size=2
+    - actor_rollout_ref.rollout.name=vllm
+    - actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+    - actor_rollout_ref.rollout.enable_chunked_prefill=False
+    - actor_rollout_ref.rollout.enforce_eager=False
+    - actor_rollout_ref.rollout.free_cache_engine=False
+    - actor_rollout_ref.rollout.n=5
+    - actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.ref.fsdp_config.param_offload=True
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - trainer.critic_warmup=0
+    - trainer.logger=['console']
+    - trainer.project_name=anomseer
+    - trainer.experiment_name=anomseer_rats_uni_2gpu
+    - trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+    - trainer.n_gpus_per_node=2
+    - trainer.nnodes=1
+    - trainer.save_freq=500
+    - trainer.test_freq=10
+    - trainer.val_only=False
+    - trainer.val_before_train=True
+    - trainer.total_epochs=10
+    - ts.use_sem_orth=True
+    - ts.adv_mix=0.3
+    - ts.similarity_method=ot
+    - ts.ot_eps=0.08
+    - ts.ot_n_iter=50
+  job:
+    name: main_ppo
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.fsdp_config.optimizer_offload=False,actor_rollout_ref.actor.fsdp_config.param_offload=False,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.actor.kl_loss_type=low_var_kl,actor_rollout_ref.actor.optim.lr=1e-4,actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2,actor_rollout_ref.actor.ppo_mini_batch_size=128,actor_rollout_ref.actor.use_kl_loss=True,actor_rollout_ref.model.enable_gradient_checkpointing=True,actor_rollout_ref.model.lora_alpha=16,actor_rollout_ref.model.lora_dropout=0.0,actor_rollout_ref.model.lora_rank=16,actor_rollout_ref.model.lora_target_modules=all-linear,actor_rollout_ref.model.path=/mnt/share01/sqk/models/qwen2.5-vl-7b-instruct,actor_rollout_ref.model.use_remove_padding=True,actor_rollout_ref.ref.fsdp_config.param_offload=True,actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.enable_chunked_prefill=False,actor_rollout_ref.rollout.enforce_eager=False,actor_rollout_ref.rollout.free_cache_engine=False,actor_rollout_ref.rollout.gpu_memory_utilization=0.4,actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.n=5,actor_rollout_ref.rollout.name=vllm,actor_rollout_ref.rollout.tensor_model_parallel_size=2,algorithm.adv_estimator=grpo,algorithm.kl_ctrl.kl_coef=0.001,data.filter_overlong_prompts=True,data.image_key=images,data.max_prompt_length=1024,data.max_response_length=1024,data.train_batch_size=128,data.train_files=./data/rats_uni_processed/train_full.parquet,data.truncation=error,data.val_files=./data/rats_uni_processed/test_full.parquet,trainer.critic_warmup=0,trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu,trainer.experiment_name=anomseer_rats_uni_2gpu,trainer.logger=['console'],trainer.n_gpus_per_node=2,trainer.nnodes=1,trainer.project_name=anomseer,trainer.save_freq=500,trainer.test_freq=10,trainer.total_epochs=10,trainer.val_before_train=True,trainer.val_only=False,ts.adv_mix=0.3,ts.ot_eps=0.08,ts.ot_n_iter=50,ts.similarity_method=ot,ts.use_sem_orth=True
+    id: ???
+    num: ???
+    config_name: ppo_trainer
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.0
+    version_base: '1.3'
+    cwd: /mnt/share01/sqk/AnomSeer
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/share01/sqk/AnomSeer/verl/trainer/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/share01/sqk/AnomSeer/outputs/2026-06-09/15-29-24
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

outputs/2026-06-09/15-29-24/.hydra/overrides.yaml

@@ -0,0 +1,52 @@
+- algorithm.adv_estimator=grpo
+- data.train_files=./data/rats_uni_processed/train_full.parquet
+- data.val_files=./data/rats_uni_processed/test_full.parquet
+- data.train_batch_size=128
+- data.max_prompt_length=1024
+- data.max_response_length=1024
+- data.filter_overlong_prompts=True
+- data.truncation=error
+- data.image_key=images
+- actor_rollout_ref.model.path=/mnt/share01/sqk/models/qwen2.5-vl-7b-instruct
+- actor_rollout_ref.model.lora_rank=16
+- actor_rollout_ref.model.lora_alpha=16
+- actor_rollout_ref.model.lora_dropout=0.0
+- actor_rollout_ref.model.lora_target_modules=all-linear
+- actor_rollout_ref.actor.optim.lr=1e-4
+- actor_rollout_ref.model.use_remove_padding=True
+- actor_rollout_ref.actor.ppo_mini_batch_size=128
+- actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+- actor_rollout_ref.actor.use_kl_loss=True
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.kl_loss_type=low_var_kl
+- actor_rollout_ref.model.enable_gradient_checkpointing=True
+- actor_rollout_ref.actor.fsdp_config.param_offload=False
+- actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+- actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.rollout.tensor_model_parallel_size=2
+- actor_rollout_ref.rollout.name=vllm
+- actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+- actor_rollout_ref.rollout.enable_chunked_prefill=False
+- actor_rollout_ref.rollout.enforce_eager=False
+- actor_rollout_ref.rollout.free_cache_engine=False
+- actor_rollout_ref.rollout.n=5
+- actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.ref.fsdp_config.param_offload=True
+- algorithm.kl_ctrl.kl_coef=0.001
+- trainer.critic_warmup=0
+- trainer.logger=['console']
+- trainer.project_name=anomseer
+- trainer.experiment_name=anomseer_rats_uni_2gpu
+- trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+- trainer.n_gpus_per_node=2
+- trainer.nnodes=1
+- trainer.save_freq=500
+- trainer.test_freq=10
+- trainer.val_only=False
+- trainer.val_before_train=True
+- trainer.total_epochs=10
+- ts.use_sem_orth=True
+- ts.adv_mix=0.3
+- ts.similarity_method=ot
+- ts.ot_eps=0.08
+- ts.ot_n_iter=50

outputs/2026-06-09/16-19-04/.hydra/config.yaml

@@ -0,0 +1,194 @@
+data:
+  tokenizer: null
+  train_files: ./data/rats_uni_processed/train_full.parquet
+  val_files: ./data/rats_uni_processed/test_full.parquet
+  prompt_key: prompt
+  max_prompt_length: 1024
+  max_response_length: 1024
+  train_batch_size: 128
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  shuffle: true
+  filter_overlong_prompts: true
+  truncation: error
+  image_key: images
+actor_rollout_ref:
+  hybrid_engine: true
+  model:
+    path: /mnt/share01/sqk/models/qwen2.5-vl-7b-instruct
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    lora_rank: 16
+    lora_alpha: 16
+    lora_dropout: 0.0
+    lora_target_modules: all-linear
+  actor:
+    strategy: fsdp
+    ppo_mini_batch_size: 128
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: 2
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    grad_clip: 1.0
+    clip_ratio: 0.2
+    entropy_coeff: 0.001
+    use_kl_loss: true
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: low_var_kl
+    ppo_epochs: 1
+    shuffle: false
+    ulysses_sequence_parallel_size: 1
+    optim:
+      lr: 0.0001
+      lr_warmup_steps: -1
+      lr_warmup_steps_ratio: 0.0
+      min_lr_ratio: null
+      warmup_style: constant
+      total_training_steps: -1
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      fsdp_size: -1
+  ref:
+    fsdp_config:
+      param_offload: true
+      wrap_policy:
+        min_num_params: 0
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    ulysses_sequence_parallel_size: ${actor_rollout_ref.actor.ulysses_sequence_parallel_size}
+  rollout:
+    name: vllm
+    temperature: 1.0
+    top_k: -1
+    top_p: 1
+    use_fire_sampling: false
+    prompt_length: ${data.max_prompt_length}
+    response_length: ${data.max_response_length}
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: false
+    free_cache_engine: false
+    load_format: dummy_dtensor
+    tensor_model_parallel_size: 2
+    max_num_batched_tokens: 8192
+    max_model_len: null
+    max_num_seqs: 1024
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    disable_log_stats: true
+    enable_chunked_prefill: false
+    do_sample: true
+    'n': 5
+    val_kwargs:
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.4
+      'n': 1
+      do_sample: true
+critic:
+  strategy: fsdp
+  optim:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    min_lr_ratio: null
+    warmup_style: constant
+    total_training_steps: -1
+  model:
+    path: ~/models/deepseek-llm-7b-chat
+    tokenizer_path: ${actor_rollout_ref.model.path}
+    override_config: {}
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    fsdp_config:
+      param_offload: false
+      optimizer_offload: false
+      wrap_policy:
+        min_num_params: 0
+      fsdp_size: -1
+  ppo_mini_batch_size: ${actor_rollout_ref.actor.ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: null
+  forward_micro_batch_size: ${critic.ppo_micro_batch_size}
+  forward_micro_batch_size_per_gpu: ${critic.ppo_micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${critic.ppo_max_token_len_per_gpu}
+  ulysses_sequence_parallel_size: 1
+  ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
+  shuffle: ${actor_rollout_ref.actor.shuffle}
+  grad_clip: 1.0
+  cliprange_value: 0.5
+reward_model:
+  enable: false
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    use_remove_padding: false
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      fsdp_size: -1
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  ulysses_sequence_parallel_size: 1
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: grpo
+  kl_penalty: kl
+  kl_ctrl:
+    type: fixed
+    kl_coef: 0.001
+trainer:
+  balance_batch: true
+  total_epochs: 10
+  total_training_steps: null
+  project_name: anomseer
+  experiment_name: anomseer_rats_uni_2gpu
+  logger:
+  - console
+  val_generations_to_log_to_wandb: 0
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: 500
+  resume_mode: auto
+  resume_from_path: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  remove_previous_ckpt_in_save: false
+  del_local_ckpt_after_load: false
+  default_local_dir: checkpoints/anomseer/anomseer_rats_uni_2gpu
+  val_only: false
+  val_before_train: true
+ts:
+  adv_mix: 0.3
+  use_sem_orth: true
+  similarity_method: ot
+  clip_temperature: 1.0
+  clip_pooling: mean
+  ot_eps: 0.08
+  ot_n_iter: 50

outputs/2026-06-09/16-19-04/.hydra/hydra.yaml

@@ -0,0 +1,206 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - algorithm.adv_estimator=grpo
+    - data.train_files=./data/rats_uni_processed/train_full.parquet
+    - data.val_files=./data/rats_uni_processed/test_full.parquet
+    - data.train_batch_size=128
+    - data.max_prompt_length=1024
+    - data.max_response_length=1024
+    - data.filter_overlong_prompts=True
+    - data.truncation=error
+    - data.image_key=images
+    - actor_rollout_ref.model.path=/mnt/share01/sqk/models/qwen2.5-vl-7b-instruct
+    - actor_rollout_ref.model.lora_rank=16
+    - actor_rollout_ref.model.lora_alpha=16
+    - actor_rollout_ref.model.lora_dropout=0.0
+    - actor_rollout_ref.model.lora_target_modules=all-linear
+    - actor_rollout_ref.actor.optim.lr=1e-4
+    - actor_rollout_ref.model.use_remove_padding=False
+    - actor_rollout_ref.actor.ppo_mini_batch_size=128
+    - actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+    - actor_rollout_ref.actor.use_kl_loss=True
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_type=low_var_kl
+    - actor_rollout_ref.model.enable_gradient_checkpointing=True
+    - actor_rollout_ref.actor.fsdp_config.param_offload=False
+    - actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+    - actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.rollout.tensor_model_parallel_size=2
+    - actor_rollout_ref.rollout.name=vllm
+    - actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+    - actor_rollout_ref.rollout.enable_chunked_prefill=False
+    - actor_rollout_ref.rollout.enforce_eager=False
+    - actor_rollout_ref.rollout.free_cache_engine=False
+    - actor_rollout_ref.rollout.n=5
+    - actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.ref.fsdp_config.param_offload=True
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - trainer.critic_warmup=0
+    - trainer.logger=['console']
+    - trainer.project_name=anomseer
+    - trainer.experiment_name=anomseer_rats_uni_2gpu
+    - trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+    - trainer.n_gpus_per_node=2
+    - trainer.nnodes=1
+    - trainer.save_freq=500
+    - trainer.test_freq=10
+    - trainer.val_only=False
+    - trainer.val_before_train=True
+    - trainer.total_epochs=10
+    - ts.use_sem_orth=True
+    - ts.adv_mix=0.3
+    - ts.similarity_method=ot
+    - ts.ot_eps=0.08
+    - ts.ot_n_iter=50
+  job:
+    name: main_ppo
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.fsdp_config.optimizer_offload=False,actor_rollout_ref.actor.fsdp_config.param_offload=False,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.actor.kl_loss_type=low_var_kl,actor_rollout_ref.actor.optim.lr=1e-4,actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2,actor_rollout_ref.actor.ppo_mini_batch_size=128,actor_rollout_ref.actor.use_kl_loss=True,actor_rollout_ref.model.enable_gradient_checkpointing=True,actor_rollout_ref.model.lora_alpha=16,actor_rollout_ref.model.lora_dropout=0.0,actor_rollout_ref.model.lora_rank=16,actor_rollout_ref.model.lora_target_modules=all-linear,actor_rollout_ref.model.path=/mnt/share01/sqk/models/qwen2.5-vl-7b-instruct,actor_rollout_ref.model.use_remove_padding=False,actor_rollout_ref.ref.fsdp_config.param_offload=True,actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.enable_chunked_prefill=False,actor_rollout_ref.rollout.enforce_eager=False,actor_rollout_ref.rollout.free_cache_engine=False,actor_rollout_ref.rollout.gpu_memory_utilization=0.4,actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.n=5,actor_rollout_ref.rollout.name=vllm,actor_rollout_ref.rollout.tensor_model_parallel_size=2,algorithm.adv_estimator=grpo,algorithm.kl_ctrl.kl_coef=0.001,data.filter_overlong_prompts=True,data.image_key=images,data.max_prompt_length=1024,data.max_response_length=1024,data.train_batch_size=128,data.train_files=./data/rats_uni_processed/train_full.parquet,data.truncation=error,data.val_files=./data/rats_uni_processed/test_full.parquet,trainer.critic_warmup=0,trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu,trainer.experiment_name=anomseer_rats_uni_2gpu,trainer.logger=['console'],trainer.n_gpus_per_node=2,trainer.nnodes=1,trainer.project_name=anomseer,trainer.save_freq=500,trainer.test_freq=10,trainer.total_epochs=10,trainer.val_before_train=True,trainer.val_only=False,ts.adv_mix=0.3,ts.ot_eps=0.08,ts.ot_n_iter=50,ts.similarity_method=ot,ts.use_sem_orth=True
+    id: ???
+    num: ???
+    config_name: ppo_trainer
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.0
+    version_base: '1.3'
+    cwd: /mnt/share01/sqk/AnomSeer
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/share01/sqk/AnomSeer/verl/trainer/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/share01/sqk/AnomSeer/outputs/2026-06-09/16-19-04
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

outputs/2026-06-09/16-19-04/.hydra/overrides.yaml

@@ -0,0 +1,52 @@
+- algorithm.adv_estimator=grpo
+- data.train_files=./data/rats_uni_processed/train_full.parquet
+- data.val_files=./data/rats_uni_processed/test_full.parquet
+- data.train_batch_size=128
+- data.max_prompt_length=1024
+- data.max_response_length=1024
+- data.filter_overlong_prompts=True
+- data.truncation=error
+- data.image_key=images
+- actor_rollout_ref.model.path=/mnt/share01/sqk/models/qwen2.5-vl-7b-instruct
+- actor_rollout_ref.model.lora_rank=16
+- actor_rollout_ref.model.lora_alpha=16
+- actor_rollout_ref.model.lora_dropout=0.0
+- actor_rollout_ref.model.lora_target_modules=all-linear
+- actor_rollout_ref.actor.optim.lr=1e-4
+- actor_rollout_ref.model.use_remove_padding=False
+- actor_rollout_ref.actor.ppo_mini_batch_size=128
+- actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+- actor_rollout_ref.actor.use_kl_loss=True
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.kl_loss_type=low_var_kl
+- actor_rollout_ref.model.enable_gradient_checkpointing=True
+- actor_rollout_ref.actor.fsdp_config.param_offload=False
+- actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+- actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.rollout.tensor_model_parallel_size=2
+- actor_rollout_ref.rollout.name=vllm
+- actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+- actor_rollout_ref.rollout.enable_chunked_prefill=False
+- actor_rollout_ref.rollout.enforce_eager=False
+- actor_rollout_ref.rollout.free_cache_engine=False
+- actor_rollout_ref.rollout.n=5
+- actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.ref.fsdp_config.param_offload=True
+- algorithm.kl_ctrl.kl_coef=0.001
+- trainer.critic_warmup=0
+- trainer.logger=['console']
+- trainer.project_name=anomseer
+- trainer.experiment_name=anomseer_rats_uni_2gpu
+- trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+- trainer.n_gpus_per_node=2
+- trainer.nnodes=1
+- trainer.save_freq=500
+- trainer.test_freq=10
+- trainer.val_only=False
+- trainer.val_before_train=True
+- trainer.total_epochs=10
+- ts.use_sem_orth=True
+- ts.adv_mix=0.3
+- ts.similarity_method=ot
+- ts.ot_eps=0.08
+- ts.ot_n_iter=50

outputs/2026-06-09/16-56-49/.hydra/config.yaml

@@ -0,0 +1,194 @@
+data:
+  tokenizer: null
+  train_files: ./data/rats_uni_processed/train_full.parquet
+  val_files: ./data/rats_uni_processed/test_full.parquet
+  prompt_key: prompt
+  max_prompt_length: 1024
+  max_response_length: 1024
+  train_batch_size: 128
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  shuffle: true
+  filter_overlong_prompts: true
+  truncation: error
+  image_key: images
+actor_rollout_ref:
+  hybrid_engine: true
+  model:
+    path: /mnt/share01/sqk/models/qwen2.5-vl-7b-instruct
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    lora_rank: 16
+    lora_alpha: 16
+    lora_dropout: 0.0
+    lora_target_modules: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
+  actor:
+    strategy: fsdp
+    ppo_mini_batch_size: 128
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: 2
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    grad_clip: 1.0
+    clip_ratio: 0.2
+    entropy_coeff: 0.001
+    use_kl_loss: true
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: low_var_kl
+    ppo_epochs: 1
+    shuffle: false
+    ulysses_sequence_parallel_size: 1
+    optim:
+      lr: 0.0001
+      lr_warmup_steps: -1
+      lr_warmup_steps_ratio: 0.0
+      min_lr_ratio: null
+      warmup_style: constant
+      total_training_steps: -1
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      fsdp_size: -1
+  ref:
+    fsdp_config:
+      param_offload: true
+      wrap_policy:
+        min_num_params: 0
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    ulysses_sequence_parallel_size: ${actor_rollout_ref.actor.ulysses_sequence_parallel_size}
+  rollout:
+    name: vllm
+    temperature: 1.0
+    top_k: -1
+    top_p: 1
+    use_fire_sampling: false
+    prompt_length: ${data.max_prompt_length}
+    response_length: ${data.max_response_length}
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: false
+    free_cache_engine: false
+    load_format: dummy_dtensor
+    tensor_model_parallel_size: 2
+    max_num_batched_tokens: 8192
+    max_model_len: null
+    max_num_seqs: 1024
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    disable_log_stats: true
+    enable_chunked_prefill: false
+    do_sample: true
+    'n': 5
+    val_kwargs:
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.4
+      'n': 1
+      do_sample: true
+critic:
+  strategy: fsdp
+  optim:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    min_lr_ratio: null
+    warmup_style: constant
+    total_training_steps: -1
+  model:
+    path: ~/models/deepseek-llm-7b-chat
+    tokenizer_path: ${actor_rollout_ref.model.path}
+    override_config: {}
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    fsdp_config:
+      param_offload: false
+      optimizer_offload: false
+      wrap_policy:
+        min_num_params: 0
+      fsdp_size: -1
+  ppo_mini_batch_size: ${actor_rollout_ref.actor.ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: null
+  forward_micro_batch_size: ${critic.ppo_micro_batch_size}
+  forward_micro_batch_size_per_gpu: ${critic.ppo_micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${critic.ppo_max_token_len_per_gpu}
+  ulysses_sequence_parallel_size: 1
+  ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
+  shuffle: ${actor_rollout_ref.actor.shuffle}
+  grad_clip: 1.0
+  cliprange_value: 0.5
+reward_model:
+  enable: false
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    use_remove_padding: false
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      fsdp_size: -1
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  ulysses_sequence_parallel_size: 1
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: grpo
+  kl_penalty: kl
+  kl_ctrl:
+    type: fixed
+    kl_coef: 0.001
+trainer:
+  balance_batch: true
+  total_epochs: 10
+  total_training_steps: null
+  project_name: anomseer
+  experiment_name: anomseer_rats_uni_2gpu
+  logger:
+  - console
+  val_generations_to_log_to_wandb: 0
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: 500
+  resume_mode: auto
+  resume_from_path: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  remove_previous_ckpt_in_save: false
+  del_local_ckpt_after_load: false
+  default_local_dir: checkpoints/anomseer/anomseer_rats_uni_2gpu
+  val_only: false
+  val_before_train: true
+ts:
+  adv_mix: 0.3
+  use_sem_orth: true
+  similarity_method: ot
+  clip_temperature: 1.0
+  clip_pooling: mean
+  ot_eps: 0.08
+  ot_n_iter: 50

outputs/2026-06-09/16-56-49/.hydra/hydra.yaml

@@ -0,0 +1,206 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - algorithm.adv_estimator=grpo
+    - data.train_files=./data/rats_uni_processed/train_full.parquet
+    - data.val_files=./data/rats_uni_processed/test_full.parquet
+    - data.train_batch_size=128
+    - data.max_prompt_length=1024
+    - data.max_response_length=1024
+    - data.filter_overlong_prompts=True
+    - data.truncation=error
+    - data.image_key=images
+    - actor_rollout_ref.model.path=/mnt/share01/sqk/models/qwen2.5-vl-7b-instruct
+    - actor_rollout_ref.model.lora_rank=16
+    - actor_rollout_ref.model.lora_alpha=16
+    - actor_rollout_ref.model.lora_dropout=0.0
+    - actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+    - actor_rollout_ref.actor.optim.lr=1e-4
+    - actor_rollout_ref.model.use_remove_padding=False
+    - actor_rollout_ref.actor.ppo_mini_batch_size=128
+    - actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+    - actor_rollout_ref.actor.use_kl_loss=True
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_type=low_var_kl
+    - actor_rollout_ref.model.enable_gradient_checkpointing=True
+    - actor_rollout_ref.actor.fsdp_config.param_offload=False
+    - actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+    - actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.rollout.tensor_model_parallel_size=2
+    - actor_rollout_ref.rollout.name=vllm
+    - actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+    - actor_rollout_ref.rollout.enable_chunked_prefill=False
+    - actor_rollout_ref.rollout.enforce_eager=False
+    - actor_rollout_ref.rollout.free_cache_engine=False
+    - actor_rollout_ref.rollout.n=5
+    - actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.ref.fsdp_config.param_offload=True
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - trainer.critic_warmup=0
+    - trainer.logger=['console']
+    - trainer.project_name=anomseer
+    - trainer.experiment_name=anomseer_rats_uni_2gpu
+    - trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+    - trainer.n_gpus_per_node=2
+    - trainer.nnodes=1
+    - trainer.save_freq=500
+    - trainer.test_freq=10
+    - trainer.val_only=False
+    - trainer.val_before_train=True
+    - trainer.total_epochs=10
+    - ts.use_sem_orth=True
+    - ts.adv_mix=0.3
+    - ts.similarity_method=ot
+    - ts.ot_eps=0.08
+    - ts.ot_n_iter=50
+  job:
+    name: main_ppo
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.fsdp_config.optimizer_offload=False,actor_rollout_ref.actor.fsdp_config.param_offload=False,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.actor.kl_loss_type=low_var_kl,actor_rollout_ref.actor.optim.lr=1e-4,actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2,actor_rollout_ref.actor.ppo_mini_batch_size=128,actor_rollout_ref.actor.use_kl_loss=True,actor_rollout_ref.model.enable_gradient_checkpointing=True,actor_rollout_ref.model.lora_alpha=16,actor_rollout_ref.model.lora_dropout=0.0,actor_rollout_ref.model.lora_rank=16,actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj',actor_rollout_ref.model.path=/mnt/share01/sqk/models/qwen2.5-vl-7b-instruct,actor_rollout_ref.model.use_remove_padding=False,actor_rollout_ref.ref.fsdp_config.param_offload=True,actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.enable_chunked_prefill=False,actor_rollout_ref.rollout.enforce_eager=False,actor_rollout_ref.rollout.free_cache_engine=False,actor_rollout_ref.rollout.gpu_memory_utilization=0.4,actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.n=5,actor_rollout_ref.rollout.name=vllm,actor_rollout_ref.rollout.tensor_model_parallel_size=2,algorithm.adv_estimator=grpo,algorithm.kl_ctrl.kl_coef=0.001,data.filter_overlong_prompts=True,data.image_key=images,data.max_prompt_length=1024,data.max_response_length=1024,data.train_batch_size=128,data.train_files=./data/rats_uni_processed/train_full.parquet,data.truncation=error,data.val_files=./data/rats_uni_processed/test_full.parquet,trainer.critic_warmup=0,trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu,trainer.experiment_name=anomseer_rats_uni_2gpu,trainer.logger=['console'],trainer.n_gpus_per_node=2,trainer.nnodes=1,trainer.project_name=anomseer,trainer.save_freq=500,trainer.test_freq=10,trainer.total_epochs=10,trainer.val_before_train=True,trainer.val_only=False,ts.adv_mix=0.3,ts.ot_eps=0.08,ts.ot_n_iter=50,ts.similarity_method=ot,ts.use_sem_orth=True
+    id: ???
+    num: ???
+    config_name: ppo_trainer
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.0
+    version_base: '1.3'
+    cwd: /mnt/share01/sqk/AnomSeer
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/share01/sqk/AnomSeer/verl/trainer/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/share01/sqk/AnomSeer/outputs/2026-06-09/16-56-49
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

outputs/2026-06-09/16-56-49/.hydra/overrides.yaml

@@ -0,0 +1,52 @@
+- algorithm.adv_estimator=grpo
+- data.train_files=./data/rats_uni_processed/train_full.parquet
+- data.val_files=./data/rats_uni_processed/test_full.parquet
+- data.train_batch_size=128
+- data.max_prompt_length=1024
+- data.max_response_length=1024
+- data.filter_overlong_prompts=True
+- data.truncation=error
+- data.image_key=images
+- actor_rollout_ref.model.path=/mnt/share01/sqk/models/qwen2.5-vl-7b-instruct
+- actor_rollout_ref.model.lora_rank=16
+- actor_rollout_ref.model.lora_alpha=16
+- actor_rollout_ref.model.lora_dropout=0.0
+- actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+- actor_rollout_ref.actor.optim.lr=1e-4
+- actor_rollout_ref.model.use_remove_padding=False
+- actor_rollout_ref.actor.ppo_mini_batch_size=128
+- actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+- actor_rollout_ref.actor.use_kl_loss=True
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.kl_loss_type=low_var_kl
+- actor_rollout_ref.model.enable_gradient_checkpointing=True
+- actor_rollout_ref.actor.fsdp_config.param_offload=False
+- actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+- actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.rollout.tensor_model_parallel_size=2
+- actor_rollout_ref.rollout.name=vllm
+- actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+- actor_rollout_ref.rollout.enable_chunked_prefill=False
+- actor_rollout_ref.rollout.enforce_eager=False
+- actor_rollout_ref.rollout.free_cache_engine=False
+- actor_rollout_ref.rollout.n=5
+- actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.ref.fsdp_config.param_offload=True
+- algorithm.kl_ctrl.kl_coef=0.001
+- trainer.critic_warmup=0
+- trainer.logger=['console']
+- trainer.project_name=anomseer
+- trainer.experiment_name=anomseer_rats_uni_2gpu
+- trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+- trainer.n_gpus_per_node=2
+- trainer.nnodes=1
+- trainer.save_freq=500
+- trainer.test_freq=10
+- trainer.val_only=False
+- trainer.val_before_train=True
+- trainer.total_epochs=10
+- ts.use_sem_orth=True
+- ts.adv_mix=0.3
+- ts.similarity_method=ot
+- ts.ot_eps=0.08
+- ts.ot_n_iter=50

outputs/2026-06-09/17-08-35/.hydra/config.yaml

@@ -0,0 +1,194 @@
+data:
+  tokenizer: null
+  train_files: ./data/rats_uni_processed/train_full.parquet
+  val_files: ./data/rats_uni_processed/test_full.parquet
+  prompt_key: prompt
+  max_prompt_length: 1024
+  max_response_length: 1024
+  train_batch_size: 128
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  shuffle: true
+  filter_overlong_prompts: true
+  truncation: error
+  image_key: images
+actor_rollout_ref:
+  hybrid_engine: true
+  model:
+    path: /mnt/share01/sqk/models/qwen2.5-vl-7b-instruct
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    lora_rank: 16
+    lora_alpha: 16
+    lora_dropout: 0.0
+    lora_target_modules: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
+  actor:
+    strategy: fsdp
+    ppo_mini_batch_size: 128
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: 2
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    grad_clip: 1.0
+    clip_ratio: 0.2
+    entropy_coeff: 0.001
+    use_kl_loss: true
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: low_var_kl
+    ppo_epochs: 1
+    shuffle: false
+    ulysses_sequence_parallel_size: 1
+    optim:
+      lr: 0.0001
+      lr_warmup_steps: -1
+      lr_warmup_steps_ratio: 0.0
+      min_lr_ratio: null
+      warmup_style: constant
+      total_training_steps: -1
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      fsdp_size: -1
+  ref:
+    fsdp_config:
+      param_offload: true
+      wrap_policy:
+        min_num_params: 0
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    ulysses_sequence_parallel_size: ${actor_rollout_ref.actor.ulysses_sequence_parallel_size}
+  rollout:
+    name: vllm
+    temperature: 1.0
+    top_k: -1
+    top_p: 1
+    use_fire_sampling: false
+    prompt_length: ${data.max_prompt_length}
+    response_length: ${data.max_response_length}
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: false
+    free_cache_engine: false
+    load_format: dummy_dtensor
+    tensor_model_parallel_size: 2
+    max_num_batched_tokens: 8192
+    max_model_len: null
+    max_num_seqs: 1024
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    disable_log_stats: true
+    enable_chunked_prefill: false
+    do_sample: true
+    'n': 5
+    val_kwargs:
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.4
+      'n': 1
+      do_sample: true
+critic:
+  strategy: fsdp
+  optim:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    min_lr_ratio: null
+    warmup_style: constant
+    total_training_steps: -1
+  model:
+    path: ~/models/deepseek-llm-7b-chat
+    tokenizer_path: ${actor_rollout_ref.model.path}
+    override_config: {}
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    fsdp_config:
+      param_offload: false
+      optimizer_offload: false
+      wrap_policy:
+        min_num_params: 0
+      fsdp_size: -1
+  ppo_mini_batch_size: ${actor_rollout_ref.actor.ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: null
+  forward_micro_batch_size: ${critic.ppo_micro_batch_size}
+  forward_micro_batch_size_per_gpu: ${critic.ppo_micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${critic.ppo_max_token_len_per_gpu}
+  ulysses_sequence_parallel_size: 1
+  ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
+  shuffle: ${actor_rollout_ref.actor.shuffle}
+  grad_clip: 1.0
+  cliprange_value: 0.5
+reward_model:
+  enable: false
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    use_remove_padding: false
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      fsdp_size: -1
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  ulysses_sequence_parallel_size: 1
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: grpo
+  kl_penalty: kl
+  kl_ctrl:
+    type: fixed
+    kl_coef: 0.001
+trainer:
+  balance_batch: true
+  total_epochs: 10
+  total_training_steps: null
+  project_name: anomseer
+  experiment_name: anomseer_rats_uni_2gpu
+  logger:
+  - console
+  val_generations_to_log_to_wandb: 0
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: 500
+  resume_mode: auto
+  resume_from_path: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  remove_previous_ckpt_in_save: false
+  del_local_ckpt_after_load: false
+  default_local_dir: checkpoints/anomseer/anomseer_rats_uni_2gpu
+  val_only: false
+  val_before_train: true
+ts:
+  adv_mix: 0.3
+  use_sem_orth: true
+  similarity_method: ot
+  clip_temperature: 1.0
+  clip_pooling: mean
+  ot_eps: 0.08
+  ot_n_iter: 50

outputs/2026-06-09/17-08-35/.hydra/hydra.yaml

@@ -0,0 +1,206 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - algorithm.adv_estimator=grpo
+    - data.train_files=./data/rats_uni_processed/train_full.parquet
+    - data.val_files=./data/rats_uni_processed/test_full.parquet
+    - data.train_batch_size=128
+    - data.max_prompt_length=1024
+    - data.max_response_length=1024
+    - data.filter_overlong_prompts=True
+    - data.truncation=error
+    - data.image_key=images
+    - actor_rollout_ref.model.path=/mnt/share01/sqk/models/qwen2.5-vl-7b-instruct
+    - actor_rollout_ref.model.lora_rank=16
+    - actor_rollout_ref.model.lora_alpha=16
+    - actor_rollout_ref.model.lora_dropout=0.0
+    - actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+    - actor_rollout_ref.actor.optim.lr=1e-4
+    - actor_rollout_ref.model.use_remove_padding=False
+    - actor_rollout_ref.actor.ppo_mini_batch_size=128
+    - actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+    - actor_rollout_ref.actor.use_kl_loss=True
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_type=low_var_kl
+    - actor_rollout_ref.model.enable_gradient_checkpointing=True
+    - actor_rollout_ref.actor.fsdp_config.param_offload=False
+    - actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+    - actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.rollout.tensor_model_parallel_size=2
+    - actor_rollout_ref.rollout.name=vllm
+    - actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+    - actor_rollout_ref.rollout.enable_chunked_prefill=False
+    - actor_rollout_ref.rollout.enforce_eager=False
+    - actor_rollout_ref.rollout.free_cache_engine=False
+    - actor_rollout_ref.rollout.n=5
+    - actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.ref.fsdp_config.param_offload=True
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - trainer.critic_warmup=0
+    - trainer.logger=['console']
+    - trainer.project_name=anomseer
+    - trainer.experiment_name=anomseer_rats_uni_2gpu
+    - trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+    - trainer.n_gpus_per_node=2
+    - trainer.nnodes=1
+    - trainer.save_freq=500
+    - trainer.test_freq=10
+    - trainer.val_only=False
+    - trainer.val_before_train=True
+    - trainer.total_epochs=10
+    - ts.use_sem_orth=True
+    - ts.adv_mix=0.3
+    - ts.similarity_method=ot
+    - ts.ot_eps=0.08
+    - ts.ot_n_iter=50
+  job:
+    name: main_ppo
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.fsdp_config.optimizer_offload=False,actor_rollout_ref.actor.fsdp_config.param_offload=False,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.actor.kl_loss_type=low_var_kl,actor_rollout_ref.actor.optim.lr=1e-4,actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2,actor_rollout_ref.actor.ppo_mini_batch_size=128,actor_rollout_ref.actor.use_kl_loss=True,actor_rollout_ref.model.enable_gradient_checkpointing=True,actor_rollout_ref.model.lora_alpha=16,actor_rollout_ref.model.lora_dropout=0.0,actor_rollout_ref.model.lora_rank=16,actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj',actor_rollout_ref.model.path=/mnt/share01/sqk/models/qwen2.5-vl-7b-instruct,actor_rollout_ref.model.use_remove_padding=False,actor_rollout_ref.ref.fsdp_config.param_offload=True,actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.enable_chunked_prefill=False,actor_rollout_ref.rollout.enforce_eager=False,actor_rollout_ref.rollout.free_cache_engine=False,actor_rollout_ref.rollout.gpu_memory_utilization=0.4,actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.n=5,actor_rollout_ref.rollout.name=vllm,actor_rollout_ref.rollout.tensor_model_parallel_size=2,algorithm.adv_estimator=grpo,algorithm.kl_ctrl.kl_coef=0.001,data.filter_overlong_prompts=True,data.image_key=images,data.max_prompt_length=1024,data.max_response_length=1024,data.train_batch_size=128,data.train_files=./data/rats_uni_processed/train_full.parquet,data.truncation=error,data.val_files=./data/rats_uni_processed/test_full.parquet,trainer.critic_warmup=0,trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu,trainer.experiment_name=anomseer_rats_uni_2gpu,trainer.logger=['console'],trainer.n_gpus_per_node=2,trainer.nnodes=1,trainer.project_name=anomseer,trainer.save_freq=500,trainer.test_freq=10,trainer.total_epochs=10,trainer.val_before_train=True,trainer.val_only=False,ts.adv_mix=0.3,ts.ot_eps=0.08,ts.ot_n_iter=50,ts.similarity_method=ot,ts.use_sem_orth=True
+    id: ???
+    num: ???
+    config_name: ppo_trainer
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.0
+    version_base: '1.3'
+    cwd: /mnt/share01/sqk/AnomSeer
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/share01/sqk/AnomSeer/verl/trainer/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/share01/sqk/AnomSeer/outputs/2026-06-09/17-08-35
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

outputs/2026-06-09/17-08-35/.hydra/overrides.yaml

@@ -0,0 +1,52 @@
+- algorithm.adv_estimator=grpo
+- data.train_files=./data/rats_uni_processed/train_full.parquet
+- data.val_files=./data/rats_uni_processed/test_full.parquet
+- data.train_batch_size=128
+- data.max_prompt_length=1024
+- data.max_response_length=1024
+- data.filter_overlong_prompts=True
+- data.truncation=error
+- data.image_key=images
+- actor_rollout_ref.model.path=/mnt/share01/sqk/models/qwen2.5-vl-7b-instruct
+- actor_rollout_ref.model.lora_rank=16
+- actor_rollout_ref.model.lora_alpha=16
+- actor_rollout_ref.model.lora_dropout=0.0
+- actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+- actor_rollout_ref.actor.optim.lr=1e-4
+- actor_rollout_ref.model.use_remove_padding=False
+- actor_rollout_ref.actor.ppo_mini_batch_size=128
+- actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+- actor_rollout_ref.actor.use_kl_loss=True
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.kl_loss_type=low_var_kl
+- actor_rollout_ref.model.enable_gradient_checkpointing=True
+- actor_rollout_ref.actor.fsdp_config.param_offload=False
+- actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+- actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.rollout.tensor_model_parallel_size=2
+- actor_rollout_ref.rollout.name=vllm
+- actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+- actor_rollout_ref.rollout.enable_chunked_prefill=False
+- actor_rollout_ref.rollout.enforce_eager=False
+- actor_rollout_ref.rollout.free_cache_engine=False
+- actor_rollout_ref.rollout.n=5
+- actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.ref.fsdp_config.param_offload=True
+- algorithm.kl_ctrl.kl_coef=0.001
+- trainer.critic_warmup=0
+- trainer.logger=['console']
+- trainer.project_name=anomseer
+- trainer.experiment_name=anomseer_rats_uni_2gpu
+- trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+- trainer.n_gpus_per_node=2
+- trainer.nnodes=1
+- trainer.save_freq=500
+- trainer.test_freq=10
+- trainer.val_only=False
+- trainer.val_before_train=True
+- trainer.total_epochs=10
+- ts.use_sem_orth=True
+- ts.adv_mix=0.3
+- ts.similarity_method=ot
+- ts.ot_eps=0.08
+- ts.ot_n_iter=50

outputs/2026-06-09/17-25-05/.hydra/config.yaml

@@ -0,0 +1,194 @@
+data:
+  tokenizer: null
+  train_files: ./data/rats_uni_processed/train_full.parquet
+  val_files: ./data/rats_uni_processed/test_full.parquet
+  prompt_key: prompt
+  max_prompt_length: 1024
+  max_response_length: 1024
+  train_batch_size: 128
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  shuffle: true
+  filter_overlong_prompts: true
+  truncation: error
+  image_key: images
+actor_rollout_ref:
+  hybrid_engine: true
+  model:
+    path: /mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    lora_rank: 16
+    lora_alpha: 16
+    lora_dropout: 0.0
+    lora_target_modules: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
+  actor:
+    strategy: fsdp
+    ppo_mini_batch_size: 128
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: 2
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    grad_clip: 1.0
+    clip_ratio: 0.2
+    entropy_coeff: 0.001
+    use_kl_loss: true
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: low_var_kl
+    ppo_epochs: 1
+    shuffle: false
+    ulysses_sequence_parallel_size: 1
+    optim:
+      lr: 0.0001
+      lr_warmup_steps: -1
+      lr_warmup_steps_ratio: 0.0
+      min_lr_ratio: null
+      warmup_style: constant
+      total_training_steps: -1
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      fsdp_size: -1
+  ref:
+    fsdp_config:
+      param_offload: true
+      wrap_policy:
+        min_num_params: 0
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    ulysses_sequence_parallel_size: ${actor_rollout_ref.actor.ulysses_sequence_parallel_size}
+  rollout:
+    name: vllm
+    temperature: 1.0
+    top_k: -1
+    top_p: 1
+    use_fire_sampling: false
+    prompt_length: ${data.max_prompt_length}
+    response_length: ${data.max_response_length}
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: false
+    free_cache_engine: false
+    load_format: dummy_dtensor
+    tensor_model_parallel_size: 2
+    max_num_batched_tokens: 8192
+    max_model_len: null
+    max_num_seqs: 1024
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    disable_log_stats: true
+    enable_chunked_prefill: false
+    do_sample: true
+    'n': 5
+    val_kwargs:
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.4
+      'n': 1
+      do_sample: true
+critic:
+  strategy: fsdp
+  optim:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    min_lr_ratio: null
+    warmup_style: constant
+    total_training_steps: -1
+  model:
+    path: ~/models/deepseek-llm-7b-chat
+    tokenizer_path: ${actor_rollout_ref.model.path}
+    override_config: {}
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    fsdp_config:
+      param_offload: false
+      optimizer_offload: false
+      wrap_policy:
+        min_num_params: 0
+      fsdp_size: -1
+  ppo_mini_batch_size: ${actor_rollout_ref.actor.ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: null
+  forward_micro_batch_size: ${critic.ppo_micro_batch_size}
+  forward_micro_batch_size_per_gpu: ${critic.ppo_micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${critic.ppo_max_token_len_per_gpu}
+  ulysses_sequence_parallel_size: 1
+  ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
+  shuffle: ${actor_rollout_ref.actor.shuffle}
+  grad_clip: 1.0
+  cliprange_value: 0.5
+reward_model:
+  enable: false
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    use_remove_padding: false
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      fsdp_size: -1
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  ulysses_sequence_parallel_size: 1
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: grpo
+  kl_penalty: kl
+  kl_ctrl:
+    type: fixed
+    kl_coef: 0.001
+trainer:
+  balance_batch: true
+  total_epochs: 10
+  total_training_steps: null
+  project_name: anomseer
+  experiment_name: anomseer_rats_uni_2gpu
+  logger:
+  - console
+  val_generations_to_log_to_wandb: 0
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: 500
+  resume_mode: auto
+  resume_from_path: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  remove_previous_ckpt_in_save: false
+  del_local_ckpt_after_load: false
+  default_local_dir: checkpoints/anomseer/anomseer_rats_uni_2gpu
+  val_only: false
+  val_before_train: true
+ts:
+  adv_mix: 0.3
+  use_sem_orth: true
+  similarity_method: ot
+  clip_temperature: 1.0
+  clip_pooling: mean
+  ot_eps: 0.08
+  ot_n_iter: 50

outputs/2026-06-09/17-25-05/.hydra/hydra.yaml

@@ -0,0 +1,206 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - algorithm.adv_estimator=grpo
+    - data.train_files=./data/rats_uni_processed/train_full.parquet
+    - data.val_files=./data/rats_uni_processed/test_full.parquet
+    - data.train_batch_size=128
+    - data.max_prompt_length=1024
+    - data.max_response_length=1024
+    - data.filter_overlong_prompts=True
+    - data.truncation=error
+    - data.image_key=images
+    - actor_rollout_ref.model.path=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct
+    - actor_rollout_ref.model.lora_rank=16
+    - actor_rollout_ref.model.lora_alpha=16
+    - actor_rollout_ref.model.lora_dropout=0.0
+    - actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+    - actor_rollout_ref.actor.optim.lr=1e-4
+    - actor_rollout_ref.model.use_remove_padding=False
+    - actor_rollout_ref.actor.ppo_mini_batch_size=128
+    - actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+    - actor_rollout_ref.actor.use_kl_loss=True
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_type=low_var_kl
+    - actor_rollout_ref.model.enable_gradient_checkpointing=True
+    - actor_rollout_ref.actor.fsdp_config.param_offload=False
+    - actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+    - actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.rollout.tensor_model_parallel_size=2
+    - actor_rollout_ref.rollout.name=vllm
+    - actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+    - actor_rollout_ref.rollout.enable_chunked_prefill=False
+    - actor_rollout_ref.rollout.enforce_eager=False
+    - actor_rollout_ref.rollout.free_cache_engine=False
+    - actor_rollout_ref.rollout.n=5
+    - actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.ref.fsdp_config.param_offload=True
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - trainer.critic_warmup=0
+    - trainer.logger=['console']
+    - trainer.project_name=anomseer
+    - trainer.experiment_name=anomseer_rats_uni_2gpu
+    - trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+    - trainer.n_gpus_per_node=2
+    - trainer.nnodes=1
+    - trainer.save_freq=500
+    - trainer.test_freq=10
+    - trainer.val_only=False
+    - trainer.val_before_train=True
+    - trainer.total_epochs=10
+    - ts.use_sem_orth=True
+    - ts.adv_mix=0.3
+    - ts.similarity_method=ot
+    - ts.ot_eps=0.08
+    - ts.ot_n_iter=50
+  job:
+    name: main_ppo
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.fsdp_config.optimizer_offload=False,actor_rollout_ref.actor.fsdp_config.param_offload=False,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.actor.kl_loss_type=low_var_kl,actor_rollout_ref.actor.optim.lr=1e-4,actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2,actor_rollout_ref.actor.ppo_mini_batch_size=128,actor_rollout_ref.actor.use_kl_loss=True,actor_rollout_ref.model.enable_gradient_checkpointing=True,actor_rollout_ref.model.lora_alpha=16,actor_rollout_ref.model.lora_dropout=0.0,actor_rollout_ref.model.lora_rank=16,actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj',actor_rollout_ref.model.path=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct,actor_rollout_ref.model.use_remove_padding=False,actor_rollout_ref.ref.fsdp_config.param_offload=True,actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.enable_chunked_prefill=False,actor_rollout_ref.rollout.enforce_eager=False,actor_rollout_ref.rollout.free_cache_engine=False,actor_rollout_ref.rollout.gpu_memory_utilization=0.4,actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.n=5,actor_rollout_ref.rollout.name=vllm,actor_rollout_ref.rollout.tensor_model_parallel_size=2,algorithm.adv_estimator=grpo,algorithm.kl_ctrl.kl_coef=0.001,data.filter_overlong_prompts=True,data.image_key=images,data.max_prompt_length=1024,data.max_response_length=1024,data.train_batch_size=128,data.train_files=./data/rats_uni_processed/train_full.parquet,data.truncation=error,data.val_files=./data/rats_uni_processed/test_full.parquet,trainer.critic_warmup=0,trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu,trainer.experiment_name=anomseer_rats_uni_2gpu,trainer.logger=['console'],trainer.n_gpus_per_node=2,trainer.nnodes=1,trainer.project_name=anomseer,trainer.save_freq=500,trainer.test_freq=10,trainer.total_epochs=10,trainer.val_before_train=True,trainer.val_only=False,ts.adv_mix=0.3,ts.ot_eps=0.08,ts.ot_n_iter=50,ts.similarity_method=ot,ts.use_sem_orth=True
+    id: ???
+    num: ???
+    config_name: ppo_trainer
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.0
+    version_base: '1.3'
+    cwd: /mnt/share01/sqk/AnomSeer
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/share01/sqk/AnomSeer/verl/trainer/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/share01/sqk/AnomSeer/outputs/2026-06-09/17-25-05
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

outputs/2026-06-09/17-25-05/.hydra/overrides.yaml

@@ -0,0 +1,52 @@
+- algorithm.adv_estimator=grpo
+- data.train_files=./data/rats_uni_processed/train_full.parquet
+- data.val_files=./data/rats_uni_processed/test_full.parquet
+- data.train_batch_size=128
+- data.max_prompt_length=1024
+- data.max_response_length=1024
+- data.filter_overlong_prompts=True
+- data.truncation=error
+- data.image_key=images
+- actor_rollout_ref.model.path=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct
+- actor_rollout_ref.model.lora_rank=16
+- actor_rollout_ref.model.lora_alpha=16
+- actor_rollout_ref.model.lora_dropout=0.0
+- actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+- actor_rollout_ref.actor.optim.lr=1e-4
+- actor_rollout_ref.model.use_remove_padding=False
+- actor_rollout_ref.actor.ppo_mini_batch_size=128
+- actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+- actor_rollout_ref.actor.use_kl_loss=True
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.kl_loss_type=low_var_kl
+- actor_rollout_ref.model.enable_gradient_checkpointing=True
+- actor_rollout_ref.actor.fsdp_config.param_offload=False
+- actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+- actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.rollout.tensor_model_parallel_size=2
+- actor_rollout_ref.rollout.name=vllm
+- actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+- actor_rollout_ref.rollout.enable_chunked_prefill=False
+- actor_rollout_ref.rollout.enforce_eager=False
+- actor_rollout_ref.rollout.free_cache_engine=False
+- actor_rollout_ref.rollout.n=5
+- actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.ref.fsdp_config.param_offload=True
+- algorithm.kl_ctrl.kl_coef=0.001
+- trainer.critic_warmup=0
+- trainer.logger=['console']
+- trainer.project_name=anomseer
+- trainer.experiment_name=anomseer_rats_uni_2gpu
+- trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+- trainer.n_gpus_per_node=2
+- trainer.nnodes=1
+- trainer.save_freq=500
+- trainer.test_freq=10
+- trainer.val_only=False
+- trainer.val_before_train=True
+- trainer.total_epochs=10
+- ts.use_sem_orth=True
+- ts.adv_mix=0.3
+- ts.similarity_method=ot
+- ts.ot_eps=0.08
+- ts.ot_n_iter=50

outputs/2026-06-09/17-44-32/.hydra/config.yaml

@@ -0,0 +1,194 @@
+data:
+  tokenizer: null
+  train_files: ./data/rats_uni_processed/train_full.parquet
+  val_files: ./data/rats_uni_processed/test_small.parquet
+  prompt_key: prompt
+  max_prompt_length: 1024
+  max_response_length: 1024
+  train_batch_size: 128
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  shuffle: true
+  filter_overlong_prompts: true
+  truncation: error
+  image_key: images
+actor_rollout_ref:
+  hybrid_engine: true
+  model:
+    path: /mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    lora_rank: 16
+    lora_alpha: 16
+    lora_dropout: 0.0
+    lora_target_modules: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
+  actor:
+    strategy: fsdp
+    ppo_mini_batch_size: 128
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: 2
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    grad_clip: 1.0
+    clip_ratio: 0.2
+    entropy_coeff: 0.001
+    use_kl_loss: true
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: low_var_kl
+    ppo_epochs: 1
+    shuffle: false
+    ulysses_sequence_parallel_size: 1
+    optim:
+      lr: 0.0001
+      lr_warmup_steps: -1
+      lr_warmup_steps_ratio: 0.0
+      min_lr_ratio: null
+      warmup_style: constant
+      total_training_steps: -1
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      fsdp_size: -1
+  ref:
+    fsdp_config:
+      param_offload: true
+      wrap_policy:
+        min_num_params: 0
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    ulysses_sequence_parallel_size: ${actor_rollout_ref.actor.ulysses_sequence_parallel_size}
+  rollout:
+    name: vllm
+    temperature: 1.0
+    top_k: -1
+    top_p: 1
+    use_fire_sampling: false
+    prompt_length: ${data.max_prompt_length}
+    response_length: ${data.max_response_length}
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: false
+    free_cache_engine: false
+    load_format: dummy_dtensor
+    tensor_model_parallel_size: 2
+    max_num_batched_tokens: 8192
+    max_model_len: null
+    max_num_seqs: 1024
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    disable_log_stats: true
+    enable_chunked_prefill: false
+    do_sample: true
+    'n': 5
+    val_kwargs:
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.4
+      'n': 1
+      do_sample: false
+critic:
+  strategy: fsdp
+  optim:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    min_lr_ratio: null
+    warmup_style: constant
+    total_training_steps: -1
+  model:
+    path: ~/models/deepseek-llm-7b-chat
+    tokenizer_path: ${actor_rollout_ref.model.path}
+    override_config: {}
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    fsdp_config:
+      param_offload: false
+      optimizer_offload: false
+      wrap_policy:
+        min_num_params: 0
+      fsdp_size: -1
+  ppo_mini_batch_size: ${actor_rollout_ref.actor.ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: null
+  forward_micro_batch_size: ${critic.ppo_micro_batch_size}
+  forward_micro_batch_size_per_gpu: ${critic.ppo_micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${critic.ppo_max_token_len_per_gpu}
+  ulysses_sequence_parallel_size: 1
+  ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
+  shuffle: ${actor_rollout_ref.actor.shuffle}
+  grad_clip: 1.0
+  cliprange_value: 0.5
+reward_model:
+  enable: false
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    use_remove_padding: false
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      fsdp_size: -1
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  ulysses_sequence_parallel_size: 1
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: grpo
+  kl_penalty: kl
+  kl_ctrl:
+    type: fixed
+    kl_coef: 0.001
+trainer:
+  balance_batch: true
+  total_epochs: 10
+  total_training_steps: null
+  project_name: anomseer
+  experiment_name: anomseer_rats_uni_2gpu
+  logger:
+  - console
+  val_generations_to_log_to_wandb: 0
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: 500
+  resume_mode: auto
+  resume_from_path: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  remove_previous_ckpt_in_save: false
+  del_local_ckpt_after_load: false
+  default_local_dir: checkpoints/anomseer/anomseer_rats_uni_2gpu
+  val_only: false
+  val_before_train: true
+ts:
+  adv_mix: 0.3
+  use_sem_orth: true
+  similarity_method: ot
+  clip_temperature: 1.0
+  clip_pooling: mean
+  ot_eps: 0.08
+  ot_n_iter: 50

outputs/2026-06-09/17-44-32/.hydra/hydra.yaml

@@ -0,0 +1,207 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - algorithm.adv_estimator=grpo
+    - data.train_files=./data/rats_uni_processed/train_full.parquet
+    - data.val_files=./data/rats_uni_processed/test_small.parquet
+    - data.train_batch_size=128
+    - data.max_prompt_length=1024
+    - data.max_response_length=1024
+    - data.filter_overlong_prompts=True
+    - data.truncation=error
+    - data.image_key=images
+    - actor_rollout_ref.model.path=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct
+    - actor_rollout_ref.model.lora_rank=16
+    - actor_rollout_ref.model.lora_alpha=16
+    - actor_rollout_ref.model.lora_dropout=0.0
+    - actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+    - actor_rollout_ref.actor.optim.lr=1e-4
+    - actor_rollout_ref.model.use_remove_padding=False
+    - actor_rollout_ref.actor.ppo_mini_batch_size=128
+    - actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+    - actor_rollout_ref.actor.use_kl_loss=True
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_type=low_var_kl
+    - actor_rollout_ref.model.enable_gradient_checkpointing=True
+    - actor_rollout_ref.actor.fsdp_config.param_offload=False
+    - actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+    - actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.rollout.tensor_model_parallel_size=2
+    - actor_rollout_ref.rollout.name=vllm
+    - actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+    - actor_rollout_ref.rollout.enable_chunked_prefill=False
+    - actor_rollout_ref.rollout.enforce_eager=False
+    - actor_rollout_ref.rollout.free_cache_engine=False
+    - actor_rollout_ref.rollout.n=5
+    - actor_rollout_ref.rollout.val_kwargs.do_sample=False
+    - actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.ref.fsdp_config.param_offload=True
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - trainer.critic_warmup=0
+    - trainer.logger=['console']
+    - trainer.project_name=anomseer
+    - trainer.experiment_name=anomseer_rats_uni_2gpu
+    - trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+    - trainer.n_gpus_per_node=2
+    - trainer.nnodes=1
+    - trainer.save_freq=500
+    - trainer.test_freq=10
+    - trainer.val_only=False
+    - trainer.val_before_train=True
+    - trainer.total_epochs=10
+    - ts.use_sem_orth=True
+    - ts.adv_mix=0.3
+    - ts.similarity_method=ot
+    - ts.ot_eps=0.08
+    - ts.ot_n_iter=50
+  job:
+    name: main_ppo
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.fsdp_config.optimizer_offload=False,actor_rollout_ref.actor.fsdp_config.param_offload=False,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.actor.kl_loss_type=low_var_kl,actor_rollout_ref.actor.optim.lr=1e-4,actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2,actor_rollout_ref.actor.ppo_mini_batch_size=128,actor_rollout_ref.actor.use_kl_loss=True,actor_rollout_ref.model.enable_gradient_checkpointing=True,actor_rollout_ref.model.lora_alpha=16,actor_rollout_ref.model.lora_dropout=0.0,actor_rollout_ref.model.lora_rank=16,actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj',actor_rollout_ref.model.path=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct,actor_rollout_ref.model.use_remove_padding=False,actor_rollout_ref.ref.fsdp_config.param_offload=True,actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.enable_chunked_prefill=False,actor_rollout_ref.rollout.enforce_eager=False,actor_rollout_ref.rollout.free_cache_engine=False,actor_rollout_ref.rollout.gpu_memory_utilization=0.4,actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.n=5,actor_rollout_ref.rollout.name=vllm,actor_rollout_ref.rollout.tensor_model_parallel_size=2,actor_rollout_ref.rollout.val_kwargs.do_sample=False,algorithm.adv_estimator=grpo,algorithm.kl_ctrl.kl_coef=0.001,data.filter_overlong_prompts=True,data.image_key=images,data.max_prompt_length=1024,data.max_response_length=1024,data.train_batch_size=128,data.train_files=./data/rats_uni_processed/train_full.parquet,data.truncation=error,data.val_files=./data/rats_uni_processed/test_small.parquet,trainer.critic_warmup=0,trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu,trainer.experiment_name=anomseer_rats_uni_2gpu,trainer.logger=['console'],trainer.n_gpus_per_node=2,trainer.nnodes=1,trainer.project_name=anomseer,trainer.save_freq=500,trainer.test_freq=10,trainer.total_epochs=10,trainer.val_before_train=True,trainer.val_only=False,ts.adv_mix=0.3,ts.ot_eps=0.08,ts.ot_n_iter=50,ts.similarity_method=ot,ts.use_sem_orth=True
+    id: ???
+    num: ???
+    config_name: ppo_trainer
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.0
+    version_base: '1.3'
+    cwd: /mnt/share01/sqk/AnomSeer
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/share01/sqk/AnomSeer/verl/trainer/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/share01/sqk/AnomSeer/outputs/2026-06-09/17-44-32
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

outputs/2026-06-09/17-44-32/.hydra/overrides.yaml

@@ -0,0 +1,53 @@
+- algorithm.adv_estimator=grpo
+- data.train_files=./data/rats_uni_processed/train_full.parquet
+- data.val_files=./data/rats_uni_processed/test_small.parquet
+- data.train_batch_size=128
+- data.max_prompt_length=1024
+- data.max_response_length=1024
+- data.filter_overlong_prompts=True
+- data.truncation=error
+- data.image_key=images
+- actor_rollout_ref.model.path=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct
+- actor_rollout_ref.model.lora_rank=16
+- actor_rollout_ref.model.lora_alpha=16
+- actor_rollout_ref.model.lora_dropout=0.0
+- actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+- actor_rollout_ref.actor.optim.lr=1e-4
+- actor_rollout_ref.model.use_remove_padding=False
+- actor_rollout_ref.actor.ppo_mini_batch_size=128
+- actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+- actor_rollout_ref.actor.use_kl_loss=True
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.kl_loss_type=low_var_kl
+- actor_rollout_ref.model.enable_gradient_checkpointing=True
+- actor_rollout_ref.actor.fsdp_config.param_offload=False
+- actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+- actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.rollout.tensor_model_parallel_size=2
+- actor_rollout_ref.rollout.name=vllm
+- actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+- actor_rollout_ref.rollout.enable_chunked_prefill=False
+- actor_rollout_ref.rollout.enforce_eager=False
+- actor_rollout_ref.rollout.free_cache_engine=False
+- actor_rollout_ref.rollout.n=5
+- actor_rollout_ref.rollout.val_kwargs.do_sample=False
+- actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.ref.fsdp_config.param_offload=True
+- algorithm.kl_ctrl.kl_coef=0.001
+- trainer.critic_warmup=0
+- trainer.logger=['console']
+- trainer.project_name=anomseer
+- trainer.experiment_name=anomseer_rats_uni_2gpu
+- trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+- trainer.n_gpus_per_node=2
+- trainer.nnodes=1
+- trainer.save_freq=500
+- trainer.test_freq=10
+- trainer.val_only=False
+- trainer.val_before_train=True
+- trainer.total_epochs=10
+- ts.use_sem_orth=True
+- ts.adv_mix=0.3
+- ts.similarity_method=ot
+- ts.ot_eps=0.08
+- ts.ot_n_iter=50

outputs/2026-06-09/18-11-50/.hydra/config.yaml

@@ -0,0 +1,194 @@
+data:
+  tokenizer: null
+  train_files: ./data/rats_uni_processed/train_full.parquet
+  val_files: /tmp/anomseer_rats_one.parquet
+  prompt_key: prompt
+  max_prompt_length: 1024
+  max_response_length: 16
+  train_batch_size: 128
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  shuffle: true
+  filter_overlong_prompts: true
+  truncation: error
+  image_key: images
+actor_rollout_ref:
+  hybrid_engine: true
+  model:
+    path: /mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    lora_rank: 16
+    lora_alpha: 16
+    lora_dropout: 0.0
+    lora_target_modules: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
+  actor:
+    strategy: fsdp
+    ppo_mini_batch_size: 128
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: 2
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    grad_clip: 1.0
+    clip_ratio: 0.2
+    entropy_coeff: 0.001
+    use_kl_loss: true
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: low_var_kl
+    ppo_epochs: 1
+    shuffle: false
+    ulysses_sequence_parallel_size: 1
+    optim:
+      lr: 0.0001
+      lr_warmup_steps: -1
+      lr_warmup_steps_ratio: 0.0
+      min_lr_ratio: null
+      warmup_style: constant
+      total_training_steps: -1
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      fsdp_size: -1
+  ref:
+    fsdp_config:
+      param_offload: true
+      wrap_policy:
+        min_num_params: 0
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    ulysses_sequence_parallel_size: ${actor_rollout_ref.actor.ulysses_sequence_parallel_size}
+  rollout:
+    name: vllm
+    temperature: 1.0
+    top_k: -1
+    top_p: 1
+    use_fire_sampling: false
+    prompt_length: ${data.max_prompt_length}
+    response_length: 16
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: false
+    free_cache_engine: false
+    load_format: dummy_dtensor
+    tensor_model_parallel_size: 2
+    max_num_batched_tokens: 8192
+    max_model_len: null
+    max_num_seqs: 1024
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    disable_log_stats: true
+    enable_chunked_prefill: false
+    do_sample: true
+    'n': 1
+    val_kwargs:
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.4
+      'n': 1
+      do_sample: false
+critic:
+  strategy: fsdp
+  optim:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    min_lr_ratio: null
+    warmup_style: constant
+    total_training_steps: -1
+  model:
+    path: ~/models/deepseek-llm-7b-chat
+    tokenizer_path: ${actor_rollout_ref.model.path}
+    override_config: {}
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    fsdp_config:
+      param_offload: false
+      optimizer_offload: false
+      wrap_policy:
+        min_num_params: 0
+      fsdp_size: -1
+  ppo_mini_batch_size: ${actor_rollout_ref.actor.ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: null
+  forward_micro_batch_size: ${critic.ppo_micro_batch_size}
+  forward_micro_batch_size_per_gpu: ${critic.ppo_micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${critic.ppo_max_token_len_per_gpu}
+  ulysses_sequence_parallel_size: 1
+  ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
+  shuffle: ${actor_rollout_ref.actor.shuffle}
+  grad_clip: 1.0
+  cliprange_value: 0.5
+reward_model:
+  enable: false
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    use_remove_padding: false
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      fsdp_size: -1
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  ulysses_sequence_parallel_size: 1
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: grpo
+  kl_penalty: kl
+  kl_ctrl:
+    type: fixed
+    kl_coef: 0.001
+trainer:
+  balance_batch: true
+  total_epochs: 10
+  total_training_steps: null
+  project_name: anomseer
+  experiment_name: anomseer_lora_smoke
+  logger:
+  - console
+  val_generations_to_log_to_wandb: 0
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: 500
+  resume_mode: auto
+  resume_from_path: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  remove_previous_ckpt_in_save: false
+  del_local_ckpt_after_load: false
+  default_local_dir: checkpoints/anomseer/anomseer_lora_smoke
+  val_only: true
+  val_before_train: true
+ts:
+  adv_mix: 0.3
+  use_sem_orth: true
+  similarity_method: ot
+  clip_temperature: 1.0
+  clip_pooling: mean
+  ot_eps: 0.08
+  ot_n_iter: 50

outputs/2026-06-09/18-11-50/.hydra/hydra.yaml

@@ -0,0 +1,210 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - algorithm.adv_estimator=grpo
+    - data.train_files=./data/rats_uni_processed/train_full.parquet
+    - data.val_files=/tmp/anomseer_rats_one.parquet
+    - data.train_batch_size=128
+    - data.max_prompt_length=1024
+    - data.max_response_length=1024
+    - data.filter_overlong_prompts=True
+    - data.truncation=error
+    - data.image_key=images
+    - actor_rollout_ref.model.path=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct
+    - actor_rollout_ref.model.lora_rank=16
+    - actor_rollout_ref.model.lora_alpha=16
+    - actor_rollout_ref.model.lora_dropout=0.0
+    - actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+    - actor_rollout_ref.actor.optim.lr=1e-4
+    - actor_rollout_ref.model.use_remove_padding=False
+    - actor_rollout_ref.actor.ppo_mini_batch_size=128
+    - actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+    - actor_rollout_ref.actor.use_kl_loss=True
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_type=low_var_kl
+    - actor_rollout_ref.model.enable_gradient_checkpointing=True
+    - actor_rollout_ref.actor.fsdp_config.param_offload=False
+    - actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+    - actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.rollout.tensor_model_parallel_size=2
+    - actor_rollout_ref.rollout.name=vllm
+    - actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+    - actor_rollout_ref.rollout.enable_chunked_prefill=False
+    - actor_rollout_ref.rollout.enforce_eager=False
+    - actor_rollout_ref.rollout.free_cache_engine=False
+    - actor_rollout_ref.rollout.n=5
+    - actor_rollout_ref.rollout.val_kwargs.do_sample=False
+    - actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.ref.fsdp_config.param_offload=True
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - trainer.critic_warmup=0
+    - trainer.logger=['console']
+    - trainer.project_name=anomseer
+    - trainer.experiment_name=anomseer_lora_smoke
+    - trainer.default_local_dir=checkpoints/anomseer/anomseer_lora_smoke
+    - trainer.n_gpus_per_node=2
+    - trainer.nnodes=1
+    - trainer.save_freq=500
+    - trainer.test_freq=10
+    - trainer.val_only=True
+    - trainer.val_before_train=True
+    - trainer.total_epochs=10
+    - ts.use_sem_orth=True
+    - ts.adv_mix=0.3
+    - ts.similarity_method=ot
+    - ts.ot_eps=0.08
+    - ts.ot_n_iter=50
+    - data.max_response_length=16
+    - actor_rollout_ref.rollout.response_length=16
+    - actor_rollout_ref.rollout.n=1
+  job:
+    name: main_ppo
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.fsdp_config.optimizer_offload=False,actor_rollout_ref.actor.fsdp_config.param_offload=False,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.actor.kl_loss_type=low_var_kl,actor_rollout_ref.actor.optim.lr=1e-4,actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2,actor_rollout_ref.actor.ppo_mini_batch_size=128,actor_rollout_ref.actor.use_kl_loss=True,actor_rollout_ref.model.enable_gradient_checkpointing=True,actor_rollout_ref.model.lora_alpha=16,actor_rollout_ref.model.lora_dropout=0.0,actor_rollout_ref.model.lora_rank=16,actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj',actor_rollout_ref.model.path=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct,actor_rollout_ref.model.use_remove_padding=False,actor_rollout_ref.ref.fsdp_config.param_offload=True,actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.enable_chunked_prefill=False,actor_rollout_ref.rollout.enforce_eager=False,actor_rollout_ref.rollout.free_cache_engine=False,actor_rollout_ref.rollout.gpu_memory_utilization=0.4,actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.n=1,actor_rollout_ref.rollout.n=5,actor_rollout_ref.rollout.name=vllm,actor_rollout_ref.rollout.response_length=16,actor_rollout_ref.rollout.tensor_model_parallel_size=2,actor_rollout_ref.rollout.val_kwargs.do_sample=False,algorithm.adv_estimator=grpo,algorithm.kl_ctrl.kl_coef=0.001,data.filter_overlong_prompts=True,data.image_key=images,data.max_prompt_length=1024,data.max_response_length=1024,data.max_response_length=16,data.train_batch_size=128,data.train_files=./data/rats_uni_processed/train_full.parquet,data.truncation=error,data.val_files=/tmp/anomseer_rats_one.parquet,trainer.critic_warmup=0,trainer.default_local_dir=checkpoints/anomseer/anomseer_lora_smoke,trainer.experiment_name=anomseer_lora_smoke,trainer.logger=['console'],trainer.n_gpus_per_node=2,trainer.nnodes=1,trainer.project_name=anomseer,trainer.save_freq=500,trainer.test_freq=10,trainer.total_epochs=10,trainer.val_before_train=True,trainer.val_only=True,ts.adv_mix=0.3,ts.ot_eps=0.08,ts.ot_n_iter=50,ts.similarity_method=ot,ts.use_sem_orth=True
+    id: ???
+    num: ???
+    config_name: ppo_trainer
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.0
+    version_base: '1.3'
+    cwd: /mnt/share01/sqk/AnomSeer
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/share01/sqk/AnomSeer/verl/trainer/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/share01/sqk/AnomSeer/outputs/2026-06-09/18-11-50
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

outputs/2026-06-09/18-11-50/.hydra/overrides.yaml

@@ -0,0 +1,56 @@
+- algorithm.adv_estimator=grpo
+- data.train_files=./data/rats_uni_processed/train_full.parquet
+- data.val_files=/tmp/anomseer_rats_one.parquet
+- data.train_batch_size=128
+- data.max_prompt_length=1024
+- data.max_response_length=1024
+- data.filter_overlong_prompts=True
+- data.truncation=error
+- data.image_key=images
+- actor_rollout_ref.model.path=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct
+- actor_rollout_ref.model.lora_rank=16
+- actor_rollout_ref.model.lora_alpha=16
+- actor_rollout_ref.model.lora_dropout=0.0
+- actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+- actor_rollout_ref.actor.optim.lr=1e-4
+- actor_rollout_ref.model.use_remove_padding=False
+- actor_rollout_ref.actor.ppo_mini_batch_size=128
+- actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+- actor_rollout_ref.actor.use_kl_loss=True
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.kl_loss_type=low_var_kl
+- actor_rollout_ref.model.enable_gradient_checkpointing=True
+- actor_rollout_ref.actor.fsdp_config.param_offload=False
+- actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+- actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.rollout.tensor_model_parallel_size=2
+- actor_rollout_ref.rollout.name=vllm
+- actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+- actor_rollout_ref.rollout.enable_chunked_prefill=False
+- actor_rollout_ref.rollout.enforce_eager=False
+- actor_rollout_ref.rollout.free_cache_engine=False
+- actor_rollout_ref.rollout.n=5
+- actor_rollout_ref.rollout.val_kwargs.do_sample=False
+- actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.ref.fsdp_config.param_offload=True
+- algorithm.kl_ctrl.kl_coef=0.001
+- trainer.critic_warmup=0
+- trainer.logger=['console']
+- trainer.project_name=anomseer
+- trainer.experiment_name=anomseer_lora_smoke
+- trainer.default_local_dir=checkpoints/anomseer/anomseer_lora_smoke
+- trainer.n_gpus_per_node=2
+- trainer.nnodes=1
+- trainer.save_freq=500
+- trainer.test_freq=10
+- trainer.val_only=True
+- trainer.val_before_train=True
+- trainer.total_epochs=10
+- ts.use_sem_orth=True
+- ts.adv_mix=0.3
+- ts.similarity_method=ot
+- ts.ot_eps=0.08
+- ts.ot_n_iter=50
+- data.max_response_length=16
+- actor_rollout_ref.rollout.response_length=16
+- actor_rollout_ref.rollout.n=1

outputs/2026-06-09/18-20-25/.hydra/config.yaml

@@ -0,0 +1,194 @@
+data:
+  tokenizer: null
+  train_files: ./data/rats_uni_processed/train_full.parquet
+  val_files: ./data/rats_uni_processed/test_small.parquet
+  prompt_key: prompt
+  max_prompt_length: 1024
+  max_response_length: 1024
+  train_batch_size: 128
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  shuffle: true
+  filter_overlong_prompts: true
+  truncation: error
+  image_key: images
+actor_rollout_ref:
+  hybrid_engine: true
+  model:
+    path: /mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    lora_rank: 16
+    lora_alpha: 16
+    lora_dropout: 0.0
+    lora_target_modules: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
+  actor:
+    strategy: fsdp
+    ppo_mini_batch_size: 128
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: 2
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    grad_clip: 1.0
+    clip_ratio: 0.2
+    entropy_coeff: 0.001
+    use_kl_loss: true
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: low_var_kl
+    ppo_epochs: 1
+    shuffle: false
+    ulysses_sequence_parallel_size: 1
+    optim:
+      lr: 0.0001
+      lr_warmup_steps: -1
+      lr_warmup_steps_ratio: 0.0
+      min_lr_ratio: null
+      warmup_style: constant
+      total_training_steps: -1
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      fsdp_size: -1
+  ref:
+    fsdp_config:
+      param_offload: true
+      wrap_policy:
+        min_num_params: 0
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    ulysses_sequence_parallel_size: ${actor_rollout_ref.actor.ulysses_sequence_parallel_size}
+  rollout:
+    name: vllm
+    temperature: 1.0
+    top_k: -1
+    top_p: 1
+    use_fire_sampling: false
+    prompt_length: ${data.max_prompt_length}
+    response_length: ${data.max_response_length}
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: false
+    free_cache_engine: false
+    load_format: dummy_dtensor
+    tensor_model_parallel_size: 2
+    max_num_batched_tokens: 8192
+    max_model_len: null
+    max_num_seqs: 1024
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    disable_log_stats: true
+    enable_chunked_prefill: false
+    do_sample: true
+    'n': 5
+    val_kwargs:
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.4
+      'n': 1
+      do_sample: false
+critic:
+  strategy: fsdp
+  optim:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    min_lr_ratio: null
+    warmup_style: constant
+    total_training_steps: -1
+  model:
+    path: ~/models/deepseek-llm-7b-chat
+    tokenizer_path: ${actor_rollout_ref.model.path}
+    override_config: {}
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    fsdp_config:
+      param_offload: false
+      optimizer_offload: false
+      wrap_policy:
+        min_num_params: 0
+      fsdp_size: -1
+  ppo_mini_batch_size: ${actor_rollout_ref.actor.ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: null
+  forward_micro_batch_size: ${critic.ppo_micro_batch_size}
+  forward_micro_batch_size_per_gpu: ${critic.ppo_micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${critic.ppo_max_token_len_per_gpu}
+  ulysses_sequence_parallel_size: 1
+  ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
+  shuffle: ${actor_rollout_ref.actor.shuffle}
+  grad_clip: 1.0
+  cliprange_value: 0.5
+reward_model:
+  enable: false
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    use_remove_padding: false
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      fsdp_size: -1
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  ulysses_sequence_parallel_size: 1
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: grpo
+  kl_penalty: kl
+  kl_ctrl:
+    type: fixed
+    kl_coef: 0.001
+trainer:
+  balance_batch: true
+  total_epochs: 10
+  total_training_steps: null
+  project_name: anomseer
+  experiment_name: anomseer_rats_uni_2gpu
+  logger:
+  - console
+  val_generations_to_log_to_wandb: 0
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: 500
+  resume_mode: auto
+  resume_from_path: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  remove_previous_ckpt_in_save: false
+  del_local_ckpt_after_load: false
+  default_local_dir: checkpoints/anomseer/anomseer_rats_uni_2gpu
+  val_only: false
+  val_before_train: true
+ts:
+  adv_mix: 0.3
+  use_sem_orth: true
+  similarity_method: ot
+  clip_temperature: 1.0
+  clip_pooling: mean
+  ot_eps: 0.08
+  ot_n_iter: 50

outputs/2026-06-09/18-20-25/.hydra/overrides.yaml

@@ -0,0 +1,53 @@
+- algorithm.adv_estimator=grpo
+- data.train_files=./data/rats_uni_processed/train_full.parquet
+- data.val_files=./data/rats_uni_processed/test_small.parquet
+- data.train_batch_size=128
+- data.max_prompt_length=1024
+- data.max_response_length=1024
+- data.filter_overlong_prompts=True
+- data.truncation=error
+- data.image_key=images
+- actor_rollout_ref.model.path=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct
+- actor_rollout_ref.model.lora_rank=16
+- actor_rollout_ref.model.lora_alpha=16
+- actor_rollout_ref.model.lora_dropout=0.0
+- actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+- actor_rollout_ref.actor.optim.lr=1e-4
+- actor_rollout_ref.model.use_remove_padding=False
+- actor_rollout_ref.actor.ppo_mini_batch_size=128
+- actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+- actor_rollout_ref.actor.use_kl_loss=True
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.kl_loss_type=low_var_kl
+- actor_rollout_ref.model.enable_gradient_checkpointing=True
+- actor_rollout_ref.actor.fsdp_config.param_offload=False
+- actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+- actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.rollout.tensor_model_parallel_size=2
+- actor_rollout_ref.rollout.name=vllm
+- actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+- actor_rollout_ref.rollout.enable_chunked_prefill=False
+- actor_rollout_ref.rollout.enforce_eager=False
+- actor_rollout_ref.rollout.free_cache_engine=False
+- actor_rollout_ref.rollout.n=5
+- actor_rollout_ref.rollout.val_kwargs.do_sample=False
+- actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.ref.fsdp_config.param_offload=True
+- algorithm.kl_ctrl.kl_coef=0.001
+- trainer.critic_warmup=0
+- trainer.logger=['console']
+- trainer.project_name=anomseer
+- trainer.experiment_name=anomseer_rats_uni_2gpu
+- trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+- trainer.n_gpus_per_node=2
+- trainer.nnodes=1
+- trainer.save_freq=500
+- trainer.test_freq=10
+- trainer.val_only=False
+- trainer.val_before_train=True
+- trainer.total_epochs=10
+- ts.use_sem_orth=True
+- ts.adv_mix=0.3
+- ts.similarity_method=ot
+- ts.ot_eps=0.08
+- ts.ot_n_iter=50

outputs/2026-06-09/18-33-56/.hydra/hydra.yaml

@@ -0,0 +1,216 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - algorithm.adv_estimator=grpo
+    - data.train_files=/tmp/anomseer_rats_train_one.parquet
+    - data.val_files=/tmp/anomseer_rats_val_one.parquet
+    - data.train_batch_size=128
+    - data.max_prompt_length=1024
+    - data.max_response_length=1024
+    - data.filter_overlong_prompts=True
+    - data.truncation=error
+    - data.image_key=images
+    - actor_rollout_ref.model.path=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct
+    - actor_rollout_ref.model.lora_rank=16
+    - actor_rollout_ref.model.lora_alpha=16
+    - actor_rollout_ref.model.lora_dropout=0.0
+    - actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+    - actor_rollout_ref.actor.optim.lr=1e-4
+    - actor_rollout_ref.model.use_remove_padding=False
+    - actor_rollout_ref.actor.ppo_mini_batch_size=128
+    - actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1
+    - actor_rollout_ref.actor.use_kl_loss=True
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_type=low_var_kl
+    - actor_rollout_ref.model.enable_gradient_checkpointing=True
+    - actor_rollout_ref.actor.fsdp_config.param_offload=False
+    - actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+    - actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1
+    - actor_rollout_ref.rollout.tensor_model_parallel_size=2
+    - actor_rollout_ref.rollout.name=vllm
+    - actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+    - actor_rollout_ref.rollout.enable_chunked_prefill=False
+    - actor_rollout_ref.rollout.enforce_eager=False
+    - actor_rollout_ref.rollout.free_cache_engine=False
+    - actor_rollout_ref.rollout.n=5
+    - actor_rollout_ref.rollout.val_kwargs.do_sample=False
+    - actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=1
+    - actor_rollout_ref.ref.fsdp_config.param_offload=True
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - trainer.critic_warmup=0
+    - trainer.logger=['console']
+    - trainer.project_name=anomseer
+    - trainer.experiment_name=anomseer_lora_train_smoke
+    - trainer.default_local_dir=checkpoints/anomseer/anomseer_lora_train_smoke
+    - trainer.n_gpus_per_node=2
+    - trainer.nnodes=1
+    - trainer.save_freq=500
+    - trainer.test_freq=10
+    - trainer.val_only=False
+    - trainer.val_before_train=True
+    - trainer.total_epochs=1
+    - ts.use_sem_orth=True
+    - ts.adv_mix=0.3
+    - ts.similarity_method=ot
+    - ts.ot_eps=0.08
+    - ts.ot_n_iter=50
+    - data.train_batch_size=1
+    - data.max_response_length=16
+    - actor_rollout_ref.rollout.response_length=16
+    - actor_rollout_ref.rollout.n=1
+    - actor_rollout_ref.actor.ppo_mini_batch_size=1
+    - trainer.val_before_train=False
+    - trainer.total_training_steps=1
+    - trainer.save_freq=-1
+    - trainer.test_freq=-1
+  job:
+    name: main_ppo
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.fsdp_config.optimizer_offload=False,actor_rollout_ref.actor.fsdp_config.param_offload=False,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.actor.kl_loss_type=low_var_kl,actor_rollout_ref.actor.optim.lr=1e-4,actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1,actor_rollout_ref.actor.ppo_mini_batch_size=1,actor_rollout_ref.actor.ppo_mini_batch_size=128,actor_rollout_ref.actor.use_kl_loss=True,actor_rollout_ref.model.enable_gradient_checkpointing=True,actor_rollout_ref.model.lora_alpha=16,actor_rollout_ref.model.lora_dropout=0.0,actor_rollout_ref.model.lora_rank=16,actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj',actor_rollout_ref.model.path=/mnt/share01/sqk/models/Qwen2.5-VL-3B-Instruct,actor_rollout_ref.model.use_remove_padding=False,actor_rollout_ref.ref.fsdp_config.param_offload=True,actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=1,actor_rollout_ref.rollout.enable_chunked_prefill=False,actor_rollout_ref.rollout.enforce_eager=False,actor_rollout_ref.rollout.free_cache_engine=False,actor_rollout_ref.rollout.gpu_memory_utilization=0.4,actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1,actor_rollout_ref.rollout.n=1,actor_rollout_ref.rollout.n=5,actor_rollout_ref.rollout.name=vllm,actor_rollout_ref.rollout.response_length=16,actor_rollout_ref.rollout.tensor_model_parallel_size=2,actor_rollout_ref.rollout.val_kwargs.do_sample=False,algorithm.adv_estimator=grpo,algorithm.kl_ctrl.kl_coef=0.001,data.filter_overlong_prompts=True,data.image_key=images,data.max_prompt_length=1024,data.max_response_length=1024,data.max_response_length=16,data.train_batch_size=1,data.train_batch_size=128,data.train_files=/tmp/anomseer_rats_train_one.parquet,data.truncation=error,data.val_files=/tmp/anomseer_rats_val_one.parquet,trainer.critic_warmup=0,trainer.default_local_dir=checkpoints/anomseer/anomseer_lora_train_smoke,trainer.experiment_name=anomseer_lora_train_smoke,trainer.logger=['console'],trainer.n_gpus_per_node=2,trainer.nnodes=1,trainer.project_name=anomseer,trainer.save_freq=-1,trainer.save_freq=500,trainer.test_freq=-1,trainer.test_freq=10,trainer.total_epochs=1,trainer.total_training_steps=1,trainer.val_before_train=False,trainer.val_before_train=True,trainer.val_only=False,ts.adv_mix=0.3,ts.ot_eps=0.08,ts.ot_n_iter=50,ts.similarity_method=ot,ts.use_sem_orth=True
+    id: ???
+    num: ???
+    config_name: ppo_trainer
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.0
+    version_base: '1.3'
+    cwd: /mnt/share01/sqk/AnomSeer
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/share01/sqk/AnomSeer/verl/trainer/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/share01/sqk/AnomSeer/outputs/2026-06-09/18-33-56
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

outputs/2026-06-15/11-17-58/.hydra/hydra.yaml

@@ -0,0 +1,210 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - algorithm.adv_estimator=grpo
+    - data.train_files=/mnt/share01/sqk/AnomSeer/data/rats_uni_processed/train_quarter.parquet
+    - data.val_files=/tmp/anomseer_eval_smoke.parquet
+    - data.train_batch_size=16
+    - data.max_prompt_length=1024
+    - data.max_response_length=1024
+    - data.filter_overlong_prompts=True
+    - data.truncation=error
+    - data.image_key=images
+    - actor_rollout_ref.model.path=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface
+    - actor_rollout_ref.model.lora_rank=0
+    - actor_rollout_ref.model.lora_alpha=16
+    - actor_rollout_ref.model.lora_dropout=0.0
+    - actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+    - actor_rollout_ref.actor.optim.lr=1e-4
+    - actor_rollout_ref.model.use_remove_padding=False
+    - actor_rollout_ref.actor.ppo_mini_batch_size=16
+    - actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+    - actor_rollout_ref.actor.use_kl_loss=True
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_type=low_var_kl
+    - actor_rollout_ref.model.enable_gradient_checkpointing=True
+    - actor_rollout_ref.actor.fsdp_config.param_offload=False
+    - actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+    - actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.rollout.tensor_model_parallel_size=2
+    - actor_rollout_ref.rollout.name=vllm
+    - actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+    - actor_rollout_ref.rollout.enable_chunked_prefill=False
+    - actor_rollout_ref.rollout.enforce_eager=False
+    - actor_rollout_ref.rollout.free_cache_engine=False
+    - actor_rollout_ref.rollout.n=5
+    - actor_rollout_ref.rollout.val_kwargs.do_sample=False
+    - actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+    - actor_rollout_ref.ref.fsdp_config.param_offload=True
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - trainer.critic_warmup=0
+    - trainer.logger=['console']
+    - trainer.project_name=anomseer
+    - trainer.experiment_name=anomseer_rats_uni_2gpu
+    - trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+    - trainer.n_gpus_per_node=2
+    - trainer.nnodes=1
+    - trainer.save_freq=500
+    - trainer.test_freq=10
+    - trainer.val_only=True
+    - trainer.val_before_train=True
+    - trainer.resume_mode=disable
+    - trainer.total_epochs=1
+    - ts.use_sem_orth=True
+    - ts.adv_mix=0.3
+    - ts.similarity_method=ot
+    - ts.ot_eps=0.08
+    - ts.ot_n_iter=50
+    - data.val_batch_size=1
+    - trainer.test_freq=-1
+  job:
+    name: main_ppo
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.fsdp_config.optimizer_offload=False,actor_rollout_ref.actor.fsdp_config.param_offload=False,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.actor.kl_loss_type=low_var_kl,actor_rollout_ref.actor.optim.lr=1e-4,actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2,actor_rollout_ref.actor.ppo_mini_batch_size=16,actor_rollout_ref.actor.use_kl_loss=True,actor_rollout_ref.model.enable_gradient_checkpointing=True,actor_rollout_ref.model.lora_alpha=16,actor_rollout_ref.model.lora_dropout=0.0,actor_rollout_ref.model.lora_rank=0,actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj',actor_rollout_ref.model.path=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface,actor_rollout_ref.model.use_remove_padding=False,actor_rollout_ref.ref.fsdp_config.param_offload=True,actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.enable_chunked_prefill=False,actor_rollout_ref.rollout.enforce_eager=False,actor_rollout_ref.rollout.free_cache_engine=False,actor_rollout_ref.rollout.gpu_memory_utilization=0.4,actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8,actor_rollout_ref.rollout.n=5,actor_rollout_ref.rollout.name=vllm,actor_rollout_ref.rollout.tensor_model_parallel_size=2,actor_rollout_ref.rollout.val_kwargs.do_sample=False,algorithm.adv_estimator=grpo,algorithm.kl_ctrl.kl_coef=0.001,data.filter_overlong_prompts=True,data.image_key=images,data.max_prompt_length=1024,data.max_response_length=1024,data.train_batch_size=16,data.train_files=/mnt/share01/sqk/AnomSeer/data/rats_uni_processed/train_quarter.parquet,data.truncation=error,data.val_batch_size=1,data.val_files=/tmp/anomseer_eval_smoke.parquet,trainer.critic_warmup=0,trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu,trainer.experiment_name=anomseer_rats_uni_2gpu,trainer.logger=['console'],trainer.n_gpus_per_node=2,trainer.nnodes=1,trainer.project_name=anomseer,trainer.resume_mode=disable,trainer.save_freq=500,trainer.test_freq=-1,trainer.test_freq=10,trainer.total_epochs=1,trainer.val_before_train=True,trainer.val_only=True,ts.adv_mix=0.3,ts.ot_eps=0.08,ts.ot_n_iter=50,ts.similarity_method=ot,ts.use_sem_orth=True
+    id: ???
+    num: ???
+    config_name: ppo_trainer
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.0
+    version_base: '1.3'
+    cwd: /mnt/share01/sqk/AnomSeer
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/share01/sqk/AnomSeer/verl/trainer/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/share01/sqk/AnomSeer/outputs/2026-06-15/11-17-58
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

outputs/2026-06-15/11-25-36/.hydra/config.yaml

@@ -0,0 +1,194 @@
+data:
+  tokenizer: null
+  train_files: /mnt/share01/sqk/AnomSeer/data/rats_uni_processed/train_quarter.parquet
+  val_files: /tmp/anomseer_eval_smoke.parquet
+  prompt_key: prompt
+  max_prompt_length: 1024
+  max_response_length: 1024
+  train_batch_size: 16
+  val_batch_size: 1
+  return_raw_input_ids: false
+  return_raw_chat: false
+  shuffle: true
+  filter_overlong_prompts: true
+  truncation: error
+  image_key: images
+actor_rollout_ref:
+  hybrid_engine: true
+  model:
+    path: /mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    lora_rank: 0
+    lora_alpha: 16
+    lora_dropout: 0.0
+    lora_target_modules: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
+  actor:
+    strategy: fsdp
+    ppo_mini_batch_size: 16
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: 2
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    grad_clip: 1.0
+    clip_ratio: 0.2
+    entropy_coeff: 0.001
+    use_kl_loss: true
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: low_var_kl
+    ppo_epochs: 1
+    shuffle: false
+    ulysses_sequence_parallel_size: 1
+    optim:
+      lr: 0.0001
+      lr_warmup_steps: -1
+      lr_warmup_steps_ratio: 0.0
+      min_lr_ratio: null
+      warmup_style: constant
+      total_training_steps: -1
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      fsdp_size: -1
+  ref:
+    fsdp_config:
+      param_offload: true
+      wrap_policy:
+        min_num_params: 0
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    ulysses_sequence_parallel_size: ${actor_rollout_ref.actor.ulysses_sequence_parallel_size}
+  rollout:
+    name: vllm
+    temperature: 1.0
+    top_k: -1
+    top_p: 1
+    use_fire_sampling: false
+    prompt_length: ${data.max_prompt_length}
+    response_length: ${data.max_response_length}
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: false
+    free_cache_engine: false
+    load_format: dummy_dtensor
+    tensor_model_parallel_size: 2
+    max_num_batched_tokens: 8192
+    max_model_len: null
+    max_num_seqs: 1024
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    disable_log_stats: true
+    enable_chunked_prefill: false
+    do_sample: true
+    'n': 5
+    val_kwargs:
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.4
+      'n': 1
+      do_sample: false
+critic:
+  strategy: fsdp
+  optim:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    min_lr_ratio: null
+    warmup_style: constant
+    total_training_steps: -1
+  model:
+    path: ~/models/deepseek-llm-7b-chat
+    tokenizer_path: ${actor_rollout_ref.model.path}
+    override_config: {}
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    fsdp_config:
+      param_offload: false
+      optimizer_offload: false
+      wrap_policy:
+        min_num_params: 0
+      fsdp_size: -1
+  ppo_mini_batch_size: ${actor_rollout_ref.actor.ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: null
+  forward_micro_batch_size: ${critic.ppo_micro_batch_size}
+  forward_micro_batch_size_per_gpu: ${critic.ppo_micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${critic.ppo_max_token_len_per_gpu}
+  ulysses_sequence_parallel_size: 1
+  ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
+  shuffle: ${actor_rollout_ref.actor.shuffle}
+  grad_clip: 1.0
+  cliprange_value: 0.5
+reward_model:
+  enable: false
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    use_remove_padding: false
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      fsdp_size: -1
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  ulysses_sequence_parallel_size: 1
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: grpo
+  kl_penalty: kl
+  kl_ctrl:
+    type: fixed
+    kl_coef: 0.001
+trainer:
+  balance_batch: true
+  total_epochs: 1
+  total_training_steps: null
+  project_name: anomseer
+  experiment_name: anomseer_rats_uni_2gpu
+  logger:
+  - console
+  val_generations_to_log_to_wandb: 0
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: 500
+  resume_mode: disable
+  resume_from_path: false
+  test_freq: -1
+  critic_warmup: 0
+  default_hdfs_dir: null
+  remove_previous_ckpt_in_save: false
+  del_local_ckpt_after_load: false
+  default_local_dir: checkpoints/anomseer/anomseer_rats_uni_2gpu
+  val_only: true
+  val_before_train: true
+ts:
+  adv_mix: 0.3
+  use_sem_orth: true
+  similarity_method: ot
+  clip_temperature: 1.0
+  clip_pooling: mean
+  ot_eps: 0.08
+  ot_n_iter: 50

outputs/2026-06-15/11-25-36/.hydra/overrides.yaml

@@ -0,0 +1,56 @@
+- algorithm.adv_estimator=grpo
+- data.train_files=/mnt/share01/sqk/AnomSeer/data/rats_uni_processed/train_quarter.parquet
+- data.val_files=/tmp/anomseer_eval_smoke.parquet
+- data.train_batch_size=16
+- data.max_prompt_length=1024
+- data.max_response_length=1024
+- data.filter_overlong_prompts=True
+- data.truncation=error
+- data.image_key=images
+- actor_rollout_ref.model.path=/mnt/share01/sqk/AnomSeer/checkpoints/anomseer/anomseer_rats_uni_full_20260612_184427_exp3/global_step_1891/actor/huggingface
+- actor_rollout_ref.model.lora_rank=0
+- actor_rollout_ref.model.lora_alpha=16
+- actor_rollout_ref.model.lora_dropout=0.0
+- actor_rollout_ref.model.lora_target_modules='q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj'
+- actor_rollout_ref.actor.optim.lr=1e-4
+- actor_rollout_ref.model.use_remove_padding=False
+- actor_rollout_ref.actor.ppo_mini_batch_size=16
+- actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2
+- actor_rollout_ref.actor.use_kl_loss=True
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.kl_loss_type=low_var_kl
+- actor_rollout_ref.model.enable_gradient_checkpointing=True
+- actor_rollout_ref.actor.fsdp_config.param_offload=False
+- actor_rollout_ref.actor.fsdp_config.optimizer_offload=False
+- actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.rollout.tensor_model_parallel_size=2
+- actor_rollout_ref.rollout.name=vllm
+- actor_rollout_ref.rollout.gpu_memory_utilization=0.4
+- actor_rollout_ref.rollout.enable_chunked_prefill=False
+- actor_rollout_ref.rollout.enforce_eager=False
+- actor_rollout_ref.rollout.free_cache_engine=False
+- actor_rollout_ref.rollout.n=5
+- actor_rollout_ref.rollout.val_kwargs.do_sample=False
+- actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8
+- actor_rollout_ref.ref.fsdp_config.param_offload=True
+- algorithm.kl_ctrl.kl_coef=0.001
+- trainer.critic_warmup=0
+- trainer.logger=['console']
+- trainer.project_name=anomseer
+- trainer.experiment_name=anomseer_rats_uni_2gpu
+- trainer.default_local_dir=checkpoints/anomseer/anomseer_rats_uni_2gpu
+- trainer.n_gpus_per_node=2
+- trainer.nnodes=1
+- trainer.save_freq=500
+- trainer.test_freq=10
+- trainer.val_only=True
+- trainer.val_before_train=True
+- trainer.resume_mode=disable
+- trainer.total_epochs=1
+- ts.use_sem_orth=True
+- ts.adv_mix=0.3
+- ts.similarity_method=ot
+- ts.ot_eps=0.08
+- ts.ot_n_iter=50
+- data.val_batch_size=1
+- trainer.test_freq=-1