fix: fail fast custom template not found & more verbose vllm engine error (#3278)

Signed-off-by: Chia-Wei Chen <cchen562@wisc.edu>

components/backends/vllm/src/dynamo/vllm/args.py

@@ -156,6 +156,20 @@ def parse_args() -> Config:
     config.tool_call_parser = args.dyn_tool_call_parser
     config.reasoning_parser = args.dyn_reasoning_parser
     config.custom_jinja_template = args.custom_jinja_template
+
+    # Validate custom Jinja template file exists if provided
+    if config.custom_jinja_template is not None:
+        # Expand environment variables and user home (~) before validation
+        expanded_template_path = os.path.expanduser(
+            os.path.expandvars(config.custom_jinja_template)
+        )
+        config.custom_jinja_template = expanded_template_path
+        if not os.path.isfile(expanded_template_path):
+            raise FileNotFoundError(
+                f"Custom Jinja template file not found: {expanded_template_path}. "
+                f"Please ensure the file exists and the path is correct."
+            )
+
     # Check for conflicting flags
     has_kv_transfer_config = (
         hasattr(engine_args, "kv_transfer_config")

components/backends/vllm/src/dynamo/vllm/engine_monitor.py

@@ -4,6 +4,7 @@
 import asyncio
 import logging
 import os
+import traceback
 
 from vllm.v1.engine.async_llm import AsyncLLM
 from vllm.v1.engine.exceptions import EngineDeadError
@@ -49,6 +50,7 @@ class VllmEngineMonitor:
                 await self.engine_client.check_health()
                 await asyncio.sleep(HEALTH_CHECK_INTERVAL)
             except EngineDeadError as e:
+                logger.error(f"Traceback: {traceback.format_exc()}")
                 logger.error(f"vLLM AsyncLLM health check failed: {e}")
                 logger.warning("Initiating Dynamo Runtime shutdown.")
                 self.runtime.shutdown()