Upload llm.py with huggingface_hub

llm.py

@@ -1,12 +1,13 @@
 """
 Multi-provider LLM Client for DS Theorist — rotating keys, OpenAI-compatible.
 
-Provider chain (tested 2026-03-31):
-  1. Groq         — llama-3.3-70b-versatile  (4 working keys, free)
-  2. Cerebras     — qwen-3-235b-a22b-instruct-2507 (2 keys, free, great for math)
-  3. NVIDIA       — meta/llama-3.3-70b-instruct (3 keys, free credits)
-  4. Mistral      — mistral-small-latest (1 key, free tier)
-  5. Inception    — mercury-2 (6 keys, free tier)
+Provider chain (updated 2026-04-03):
+  1. Groq         — llama-3.3-70b-versatile  (PRIMARY — GROQ_API_KEY or GROQ_KEY_1..8)
+  2. Cerebras     — qwen-3-235b-a22b-instruct-2507 (FALLBACK — CEREBRAS_API_KEY or CEREBRAS_KEY_1..4)
+  3. Sarvam       — sarvam-m (SECOND FALLBACK — SARVAM_API_KEY)
+  4. NVIDIA       — meta/llama-3.3-70b-instruct (legacy keys)
+  5. Mistral      — mistral-small-latest (legacy keys)
+  6. Inception    — mercury-2 (legacy keys)
 
 Optimized for mathematical reasoning and formal proofs.
 """
@@ -20,11 +21,27 @@ import httpx
 def _load_keys(prefix: str, count: int) -> list[str]:
     return [k for k in [os.getenv(f"{prefix}_{i}") for i in range(1, count + 1)] if k]
 
+def _load_single_or_multi(single_env: str, multi_prefix: str, count: int) -> list[str]:
+    """Load a single env var key first, then fall back to numbered keys."""
+    keys = []
+    single = os.getenv(single_env)
+    if single:
+        keys.append(single)
+    keys += _load_keys(multi_prefix, count)
+    # deduplicate while preserving order
+    seen = set()
+    result = []
+    for k in keys:
+        if k not in seen:
+            seen.add(k)
+            result.append(k)
+    return result
+
 
 PROVIDERS = []
 
-# 1. Groq (primary — fast, free)
-_groq_keys = _load_keys("GROQ_KEY", 8)
+# 1. Groq (PRIMARY — fast, free)
+_groq_keys = _load_single_or_multi("GROQ_API_KEY", "GROQ_KEY", 8)
 if _groq_keys:
     PROVIDERS.append({
         "name": "Groq",
@@ -37,8 +54,8 @@ if _groq_keys:
         "max_tokens_cap": 4096,
     })
 
-# 2. Cerebras (best for math — qwen-3-235b)
-_cerebras_keys = _load_keys("CEREBRAS_KEY", 4)
+# 2. Cerebras (FALLBACK — best for math, qwen-3-235b)
+_cerebras_keys = _load_single_or_multi("CEREBRAS_API_KEY", "CEREBRAS_KEY", 4)
 if _cerebras_keys:
     PROVIDERS.append({
         "name": "Cerebras",
@@ -51,7 +68,21 @@ if _cerebras_keys:
         "max_tokens_cap": 4096,
     })
 
-# 3. NVIDIA (free credits)
+# 3. Sarvam (SECOND FALLBACK)
+_sarvam_keys = [k for k in [os.getenv("SARVAM_API_KEY")] if k]
+if _sarvam_keys:
+    PROVIDERS.append({
+        "name": "Sarvam",
+        "base": "https://api.sarvam.ai/v1/chat/completions",
+        "model": "sarvam-m",
+        "model_fast": "sarvam-m",
+        "keys": _sarvam_keys,
+        "auth": "Bearer",
+        "timeout": 120.0,
+        "max_tokens_cap": 4096,
+    })
+
+# 4. NVIDIA (legacy free credits)
 _nvidia_keys = _load_keys("NVAPI_KEY", 4)
 if _nvidia_keys:
     PROVIDERS.append({
@@ -65,7 +96,7 @@ if _nvidia_keys:
         "max_tokens_cap": 4096,
     })
 
-# 4. Mistral (solid quality)
+# 5. Mistral (legacy — solid quality)
 _mistral_keys = _load_keys("MISTRAL_KEY", 4)
 if _mistral_keys:
     PROVIDERS.append({
@@ -79,7 +110,7 @@ if _mistral_keys:
         "max_tokens_cap": 4096,
     })
 
-# 5. Inception (mercury-2)
+# 6. Inception (legacy — mercury-2)
 _inception_keys = _load_keys("INCEPTION_KEY", 8)
 if _inception_keys:
     PROVIDERS.append({
@@ -125,7 +156,7 @@ def complete(
     Optimized for mathematical reasoning and formal proofs.
     """
     if not PROVIDERS:
-        raise RuntimeError("LLM: no providers configured — set GROQ_KEY_1 or CEREBRAS_KEY_1 in env")
+        raise RuntimeError("LLM: no providers configured — set GROQ_API_KEY, CEREBRAS_API_KEY, or SARVAM_API_KEY in env")
 
     last_error = "no providers"
     for provider in PROVIDERS:
@@ -134,6 +165,8 @@ def complete(
         capped_tokens = min(max_tokens, cap)
         if fast:
             capped_tokens = min(capped_tokens, 300)
+        min_tok = provider.get("min_tokens", 1)
+        capped_tokens = max(capped_tokens, min_tok)
 
         payload = {
             "model": model,