fixing issue with the requirements

docs/cached.md

@@ -13,7 +13,7 @@ However, this will cause every worker to download the model from the Hugging Fac
 A naive way to cache the model would be to store it on a network volume in RunPod and reference the model files this way:
 
 ```bash
--m /runpod-volume/model.gguf --ctx-size 4096 # etc...
+-hf /runpod-volume/model.gguf --ctx-size 4096 # etc...
 ```
 
 Unfortunately, network volume performance is often not sufficient for loading large models, leading to long load times. RunPod introduced a [caching mechanism](https://docs.runpod.io/serverless/endpoints/model-caching) to solve this problem.

src/find_cached.py

@@ -1,64 +1,69 @@
 """
-Finds the full LLM GGUF path from the Hugging Face cache.
+Finds the full GGUF path from the Hugging Face cache using huggingface_hub.
 """
 
-import os
-import sys
 import argparse
-
+import os
-CACHE_DIR = "/runpod-volume/huggingface-cache/hub"
+from huggingface_hub import snapshot_download, scan_cache_dir
 
 
-def find_model_path(model_name, gguf_in_repo="model.gguf"):
+def find_model_path(model_name: str, gguf_in_repo: str) -> str | None:
     """
-    Find the path to a cached model.
+    Resolve the GGUF file path from the Hugging Face cache.
 
     Args:
-        model_name: The model name from Hugging Face
+        model_name: Hugging Face repo id (e.g. TheBloke/Mistral-7B-GGUF)
+        gguf_in_repo: Relative path to the GGUF file inside the repo
 
     Returns:
-        The full path to the cached model, or None if not found
+        Full filesystem path to the GGUF file, or None if not found
     """
 
-    cache_name = model_name.replace("/", "--").lower()
+    # First try the official way: resolve snapshot path from cache
-    snapshots_dir = os.path.join(
+    try:
-        CACHE_DIR, f"models--{cache_name}", "snapshots"
+        snapshot_path = snapshot_download(
-    )
+            repo_id=model_name,
+            local_files_only=True
+        )
+        candidate = os.path.join(snapshot_path, gguf_in_repo)
+        if os.path.isfile(candidate):
+            return candidate
+    except Exception:
+        pass
 
-    if os.path.exists(snapshots_dir):
+    # Fallback: scan cache metadata explicitly (no downloads)
-        snapshots = os.listdir(snapshots_dir)
+    cache = scan_cache_dir()
-
+    for repo in cache.repos:
-        if snapshots:
+        if repo.repo_id == model_name:
-            return os.path.join(snapshots_dir, snapshots[0], gguf_in_repo)
+            for revision in repo.revisions:
+                candidate = os.path.join(revision.snapshot_path, gguf_in_repo)
+                if os.path.isfile(candidate):
+                    return candidate
 
     return None
 
 
 def main():
-    """
-    Main function to find and print the model path.
-    """
-
     parser = argparse.ArgumentParser(
         description="Find the full GGUF path from the Hugging Face cache."
     )
     parser.add_argument(
-        "model", type=str, help="The model name from Hugging Face"
+        "model",
+        type=str,
+        help="Hugging Face model repo id (e.g. TheBloke/Mistral-7B-GGUF)",
     )
     parser.add_argument(
         "path",
         type=str,
-        help="The path to the GGUF file within the model repository",
+        help="Relative path to the GGUF file inside the repo",
     )
     args = parser.parse_args()
 
     model_path = find_model_path(args.model, args.path)
+
     if model_path is None:
-        print(
+        raise SystemExit("GGUF file not found in Hugging Face cache")
-            f"Error: Cached model not found. Model='{args.model}', GGUF='{args.path}', Cache dir='{CACHE_DIR}'",
+
-            file=sys.stderr,
-        )
-        sys.exit(1)
     print(model_path, end="")
 
 

src/requirements.txt

@@ -1,3 +1,5 @@
 runpod
 python-dotenv
 openai
+huggingface_hub
+hf-transfer

src/start.sh

@@ -17,13 +17,7 @@ cleanup() {
 CACHED_LLAMA_ARGS=""
 
 find_cached_path() {
-    local model_path
+    CACHED_LLAMA_ARGS="-m $(python ./find_cached.py $LLAMA_CACHED_MODEL $LLAMA_CACHED_GGUF_PATH)"
-    model_path=$(python ./find_cached.py "$LLAMA_CACHED_MODEL" "$LLAMA_CACHED_GGUF_PATH")
-    if [ $? -ne 0 ] || [ -z "$model_path" ]; then
-        echo "start.sh: Error: Could not resolve cached model path. Check that LLAMA_CACHED_MODEL and LLAMA_CACHED_GGUF_PATH are correct and the model is fully cached on the network volume."
-        exit 1
-    fi
-    CACHED_LLAMA_ARGS="-m $model_path"
 }
 
 # check if $LLAMA_CACHED_MODEL is set and not empty