update default args to include -ngl 99, improve startup sleep duration

.runpod/hub.json

@@ -14,10 +14,10 @@
             {
                 "key": "LLAMA_SERVER_CMD_ARGS",
                 "input": {
-                    "name": "Model Name",
+                    "name": "Command line arguments for llama-server",
                     "type": "string",
                     "description": "Launch command line arguments (argv) for the llama-server binary. Do not define the port.",
-                    "default": "-hf unsloth/gemma-3-270m-it-GGUF:Q6_K --ctx-size 4096",
+                    "default": "-hf unsloth/gemma-3-270m-it-GGUF:Q6_K --ctx-size 4096 -ngl 99",
                     "advanced": false
                 }
             },

src/engine.py

@@ -21,7 +21,6 @@ Typical usage:
 """
 
 import json
-import os
 
 from dotenv import load_dotenv
 from openai import OpenAI

src/start.sh

@@ -14,6 +14,12 @@ cleanup() {
     exit 0
 }
 
+# check if $LLAMA_SERVER_CMD_ARGS is set
+if [ -z "$LLAMA_SERVER_CMD_ARGS" ]; then
+    echo "start.sh: Warning: LLAMA_SERVER_CMD_ARGS is not set. Defaulting to -hf unsloth/gemma-3-270m-it-GGUF:Q6_K --ctx-size 4096"
+    LLAMA_SERVER_CMD_ARGS="-hf unsloth/gemma-3-270m-it-GGUF:Q6_K --ctx-size 4096 -ngl 99"
+fi
+
 # check if the substring /workspace is in LLAMA_SERVER_CMD_ARGS
 if [[ "$LLAMA_SERVER_CMD_ARGS" != *"/workspace"* ]]; then
     echo "start.sh: Tip: For reduced downloads and faster startup times, consider using a model stored in a network volume mounted to /workspace."
@@ -37,7 +43,7 @@ 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".
+# it contains a.e. "-hf modelname --ctx-size 4096 -ngl 99".
 
 echo "start.sh: Running llama-server $LLAMA_SERVER_CMD_ARGS --port 3098"
 
@@ -62,7 +68,8 @@ echo "start.sh: Waiting for llama-server to start..."
 
 # wait for the server to start
 while ! check_server_is_running; do
-    sleep 5
+    # we don't want to lose too much time, so we check very frequently
+    sleep 0.5
 done
 
 echo "start.sh: llama-server is up and running, delegating to the handler script."