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",

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

@@ -3,6 +3,7 @@ Finds the full LLM GGUF path from the Hugging Face cache.
 """
 
 import os
+import sys
 import argparse
 
 CACHE_DIR = "/runpod-volume/huggingface-cache/hub"
@@ -19,7 +20,7 @@ def find_model_path(model_name, gguf_in_repo="model.gguf"):
         The full path to the cached model, or None if not found
     """
 
-    cache_name = model_name.replace("/", "--")
+    cache_name = model_name.replace("/", "--").lower()
     snapshots_dir = os.path.join(
         CACHE_DIR, f"models--{cache_name}", "snapshots"
     )
@@ -52,6 +53,12 @@ def main():
     args = parser.parse_args()
 
     model_path = find_model_path(args.model, args.path)
+    if model_path is None:
+        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

@@ -1,3 +1,3 @@
-runpod
+runpod==1.9.1
-python-dotenv
+python-dotenv==1.2.2
-openai
+openai==2.41.1

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