pin dependencies

this seems to have been introduced in a recent OpenAI SDK update

.runpod/hub.json

@@ -1,5 +1,5 @@
 {
-    "title": "llama.cpp inference",
+    "title": "llama.cpp",
     "description": "Run llama.cpp inference using serverless RunPod workers!",
     "type": "serverless",
     "category": "language",

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
--hf /runpod-volume/model.gguf --ctx-size 4096 # etc...
+-m /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/engine.py

@@ -28,7 +28,7 @@ from utils import JobInput
 
 client = OpenAI(
     base_url="http://localhost:3098/v1/",
-    api_key="",
+    api_key="none",
 )
 
 

src/find_cached.py

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

@@ -1,5 +1,3 @@
-runpod
+runpod==1.9.1
-python-dotenv
+python-dotenv==1.2.2
-openai
+openai==2.41.1
-huggingface_hub
-hf-transfer

src/start.sh

@@ -17,7 +17,13 @@ cleanup() {
 CACHED_LLAMA_ARGS=""
 
 find_cached_path() {
-    CACHED_LLAMA_ARGS="-m $(python ./find_cached.py $LLAMA_CACHED_MODEL $LLAMA_CACHED_GGUF_PATH)"
+    local model_path
+    model_path=$(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