Force local cache path

Dockerfile

@@ -6,10 +6,15 @@ RUN pip install --no-cache-dir huggingface_hub
 COPY requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
 
-# 2) Snapshot the model locally
+# 2) Snapshot the model into /app/model_cache
 RUN python - <<EOF
 from huggingface_hub import snapshot_download
-snapshot_download("numind/NuExtract-1.5-tiny", cache_dir="/app/model_cache")
+snapshot_download(
+    repo_id="numind/NuExtract-1.5-tiny",
+    cache_dir="/app/model_cache",
+    local_dir="numind/NuExtract-1.5-tiny",
+    local_dir_use_symlinks=False
+)
 EOF
 
 # 3) Copy your code

app.py

@@ -9,6 +9,7 @@ import os
 from dotenv import load_dotenv
 import tempfile
 from supabase import create_client
+from huggingface_hub import snapshot_download
 
 
 load_dotenv()
@@ -16,6 +17,11 @@ app = FastAPI()
 
 
 model_name = "numind/NuExtract-1.5-tiny"
+MODEL_CACHE_DIR = "/app/model_cache"
+model_cache_path = snapshot_download(
+    repo_id="numind/NuExtract-1.5-tiny",
+    cache_dir=MODEL_CACHE_DIR
+)
 device     = "gpu" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
 dtype = torch.float16 if device in ("mps", "gpu") else torch.float32
 
@@ -35,13 +41,15 @@ def load_model():
     # model_name, torch_dtype=dtype, trust_remote_code=True
     # )
     model = AutoModelForCausalLM.from_pretrained(
-    "/app/model_cache/numind/NuExtract-1.5-tiny",
+    model_cache_path,
+    local_files_only=True,
     torch_dtype=dtype,
     trust_remote_code=True
     ).to(device).eval()
     # tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
     tokenizer = AutoTokenizer.from_pretrained(
-    "/app/model_cache/numind/NuExtract-1.5-tiny",
+    model_cache_path,
+    local_files_only=True,
     trust_remote_code=True
 )
     print("Model and tokenizer loaded.", flush=True)