Summary

• Implemented global offline mode for all HuggingFace-dependent modules to eliminate unauthenticated request warnings and prevent external connection attempts during startup and runtime.

   Changes
   • Global Offline Configuration: Added HF_HUB_OFFLINE=1 and TRANSFORMERS_OFFLINE=1 to the environment variables at the top of edit_api.py, embeddings.py, watcher.py, and pose_llm/pose_llm_api.py.
   • Explicit Local Files Only: Updated all huggingface_hub.hf_hub_download calls in edit_api.py to include local_files_only=True, ensuring they never attempt to reach the Hub if models are already cached.
   • LLM Service Optimization: Updated AutoTokenizer and AutoModelForCausalLM calls in pose_llm/pose_llm_api.py to use local_files_only=True.
   • Consistency: Ensured that shared modules like embeddings.py which uses open_clip are also locked to offline mode to prevent background downloads.

pose_llm/pose_llm_api.py

@@ -21,6 +21,8 @@ Env:
 import asyncio
 import json
 import os
+os.environ["HF_HUB_OFFLINE"] = "1"
+os.environ["TRANSFORMERS_OFFLINE"] = "1"
 import subprocess
 import threading
 import time
@@ -69,7 +71,7 @@ async def lifespan(app: FastAPI):
     # Boost GPU to high performance mode (avoids power-saving clock throttle)
     subprocess.run(["/opt/rocm/bin/rocm-smi", "--setperflevel", "high"], capture_output=True)
     print(f"Loading model {MODEL_ID} (gpu≤{MAX_GPU_MEM} cpu≤{MAX_CPU_MEM})...", flush=True)
-    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, local_files_only=True)
     if not tokenizer.chat_template:
         print("No chat_template found — applying Mistral [INST] fallback.", flush=True)
         tokenizer.chat_template = _MISTRAL_TEMPLATE
@@ -78,6 +80,7 @@ async def lifespan(app: FastAPI):
         torch_dtype=torch.float16,
         device_map="auto",
         max_memory={0: MAX_GPU_MEM, "cpu": MAX_CPU_MEM},
+        local_files_only=True,
     )
     model.eval()
     state["tokenizer"] = tokenizer