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",
@@ -53,4 +53,4 @@
             }
         ]
     }
-}
+}

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

@@ -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
@@ -56,7 +62,7 @@ echo "start.sh: Stopping existing llama-server instances (if any)..."
 # we have a string with all the command line arguments in the env var LLAMA_SERVER_CMD_ARGS;
 # it contains a.e. "-hf modelname --ctx-size 4096 -ngl 999".
 
-echo "start.sh: Running llama-server $LLAMA_SERVER_CMD_ARGS --port 3098"
+echo "start.sh: Running /app/llama-server $CACHED_LLAMA_ARGS $LLAMA_SERVER_CMD_ARGS --port 3098"
 
 touch llama.server.log
 
@@ -65,6 +71,8 @@ LD_LIBRARY_PATH=/app /app/llama-server $CACHED_LLAMA_ARGS $LLAMA_SERVER_CMD_ARGS
 
 LLAMA_SERVER_PID=$! # store the process ID (PID) of the background command
 
+tries_so_far=0
+
 check_server_is_running() {
     echo "start.sh: Checking if llama-server is done initializing..."
 
@@ -74,6 +82,13 @@ check_server_is_running() {
         return 1 # failure
     fi
 
+    tries_so_far=$((tries_so_far + 1))
+
+    if [ $tries_so_far -ge 120 ]; then
+        echo "start.sh: Error: llama-server did not start within 60 seconds."
+        exit 1
+    fi
+
     # check if the process is still running
     if ! kill -0 $LLAMA_SERVER_PID 2>/dev/null; then
         echo "start.sh: Error: llama-server process has exited unexpectedly."