Use CPU-only PyTorch build to reduce image size

part4_api.py

@@ -137,7 +137,9 @@ async def cache_stats() -> CacheStatsResponse:
 async def flush_cache() -> dict:
     state.cache.flush(); return {"status": "ok", "message": "Cache flushed."}
 
-@app.get("/health")
+# Allow HEAD as well as GET: uptime monitors (e.g. UptimeRobot) probe with
+# HEAD by default, and a GET-only route answers HEAD with 405 Method Not Allowed.
+@app.api_route("/health", methods=["GET", "HEAD"])
 async def health() -> dict:
     return {"status": "ok", "corpus_loaded": state.chroma_collection is not None, "fcm_loaded": state.cache._fcm is not None if state.cache else False, "cache_entries": len(state.cache) if state.cache else 0, "rag_ready": state.rag_graph is not None}
 

requirements.txt

@@ -1,3 +1,10 @@
+# Pull torch (and its deps) from PyTorch's CPU-only wheel index.
+# Without this, pip installs the default CUDA build of torch, which drags in
+# ~2.5 GB of unusable NVIDIA GPU libraries (cublas, cudnn, nccl, cufft, ...).
+# On a CPU-only Hugging Face Space that bloat makes the image too large to
+# schedule ("Scheduling failure: unable to schedule") and slows cold starts.
+--extra-index-url https://download.pytorch.org/whl/cpu
+
 fastapi>=0.111.0
 uvicorn[standard]>=0.29.0
 pydantic>=2.7.0
@@ -7,6 +14,9 @@ langchain-core>=0.2.0
 langchain-groq>=0.1.0
 langchain-huggingface>=0.0.3
 langgraph>=0.1.0
+# CPU-only torch build (the +cpu wheel declares no CUDA dependencies).
+# Same version the CUDA build already resolved to, so nothing else changes.
+torch==2.12.0+cpu
 sentence-transformers>=3.0.0
 chromadb>=1.0.0
 numpy>=1.26.0