fix(frontend): Update vllm processor for vllm 0.16 (#6799)

Signed-off-by: Graham King <grahamk@nvidia.com>

components/src/dynamo/frontend/prepost.py

@@ -11,6 +11,7 @@ from typing import Any
 from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
 from vllm.entrypoints.openai.engine.protocol import DeltaMessage, DeltaToolCall
 from vllm.reasoning import ReasoningParser
+from vllm.renderers import ChatParams
 from vllm.sampling_params import SamplingParams
 from vllm.tokenizers import TokenizerLike
 from vllm.tool_parsers import ToolParser
@@ -73,9 +74,7 @@ def _prepare_request(
     *,
     tokenizer: TokenizerLike,
     tool_parser_class: type[ToolParser] | None,
-) -> tuple[
-    ChatCompletionRequest, ToolParser | None, dict[str, Any], Any, dict[str, Any]
-]:
+) -> tuple[ChatCompletionRequest, ToolParser | None, dict[str, Any], Any, ChatParams]:
     """Validate request and build arguments for template rendering.
 
     Returns:
@@ -83,7 +82,7 @@ def _prepare_request(
         tool_parser: Instantiated tool parser, or None.
         chat_template_kwargs: Template kwargs (for PreprocessResult).
         messages_for_render: Messages to pass as first arg to render_messages.
-        render_kwargs: Keyword arguments for render_messages / render_messages_async.
+        chat_params: ChatParams for render_messages / render_messages_async.
     """
     if isinstance(request, ChatCompletionRequest):
         request_for_sampling = request
@@ -123,15 +122,17 @@ def _prepare_request(
         else request_for_sampling.messages
     )
 
-    render_kwargs = dict(
+    chat_params = ChatParams(
         chat_template=request_for_sampling.chat_template,
         chat_template_content_format="auto",
-        add_generation_prompt=request_for_sampling.add_generation_prompt,
-        continue_final_message=request_for_sampling.continue_final_message,
-        tools=tool_dicts,
-        documents=request_for_sampling.documents,
-        tokenize=tokenize_in_template,
-        **chat_template_kwargs,
+        chat_template_kwargs=dict(
+            add_generation_prompt=request_for_sampling.add_generation_prompt,
+            continue_final_message=request_for_sampling.continue_final_message,
+            tools=tool_dicts,
+            documents=request_for_sampling.documents,
+            tokenize=tokenize_in_template,
+            **chat_template_kwargs,
+        ),
     )
 
     return (
@@ -139,7 +140,7 @@ def _prepare_request(
         tool_parser,
         chat_template_kwargs,
         messages_for_render,
-        render_kwargs,
+        chat_params,
     )
 
 
@@ -155,12 +156,12 @@ async def preprocess_chat_request(
         tool_parser,
         chat_template_kwargs,
         messages,
-        render_kwargs,
+        chat_params,
     ) = _prepare_request(
         request, tokenizer=tokenizer, tool_parser_class=tool_parser_class
     )
 
-    _, engine_prompt = await renderer.render_messages_async(messages, **render_kwargs)
+    _, engine_prompt = await renderer.render_messages_async(messages, chat_params)
 
     if "prompt_token_ids" in engine_prompt:
         tokens = list(engine_prompt["prompt_token_ids"])
@@ -194,12 +195,12 @@ def preprocess_chat_request_sync(
         tool_parser,
         chat_template_kwargs,
         messages,
-        render_kwargs,
+        chat_params,
     ) = _prepare_request(
         request, tokenizer=tokenizer, tool_parser_class=tool_parser_class
     )
 
-    _, engine_prompt = renderer.render_messages(messages, **render_kwargs)
+    _, engine_prompt = renderer.render_messages(messages, chat_params)
 
     if "prompt_token_ids" in engine_prompt:
         tokens = list(engine_prompt["prompt_token_ids"])

components/src/dynamo/frontend/vllm_processor.py

@@ -83,8 +83,6 @@ def map_finish_reason(raw_reason: str | None) -> FinishReason | None:
 _w_input_processor: InputProcessor | None = None
 _w_tokenizer: Any = None
 _w_tool_parser_class: type[ToolParser] | None = None
-_w_reasoning_parser_class: type[ReasoningParser] | None = None
-_w_stream_interval: int = 20
 
 
 class _PreprocessError(Exception):
@@ -113,12 +111,9 @@ def _init_worker(
     config_format: str,
     load_format: str,
     tool_parser_name: str | None,
-    reasoning_parser_name: str | None,
-    stream_interval: int,
 ) -> None:
     """Initialize a worker process with its own VllmConfig and InputProcessor."""
     global _w_input_processor, _w_tokenizer, _w_tool_parser_class
-    global _w_reasoning_parser_class, _w_stream_interval
 
     model_config = ModelConfig(
         model=model_path,
@@ -139,14 +134,6 @@ def _init_worker(
     else:
         _w_tool_parser_class = None
 
-    if reasoning_parser_name:
-        _w_reasoning_parser_class = ReasoningParserManager.get_reasoning_parser(
-            reasoning_parser_name
-        )
-    else:
-        _w_reasoning_parser_class = None
-    _w_stream_interval = max(1, stream_interval)
-
 
 def _worker_warmup() -> bool:
     """Dummy task to ensure worker process is fully initialized."""
@@ -158,11 +145,7 @@ def _preprocess_worker(
     request_id: str,
     model_name: str,
 ) -> PreprocessWorkerResult:
-    """Preprocess a request in a worker process and return a picklable result.
-
-    This replaces _request_handler's Phase A.  No queues — errors propagate
-    naturally via the Future.
-    """
+    """Preprocess a request in a worker process and return a picklable result."""
     pre = preprocess_chat_request_sync(
         request,
         tokenizer=_w_tokenizer,
@@ -838,8 +821,6 @@ class EngineFactory:
                     config_format,
                     load_format,
                     tool_parser_name,
-                    reasoning_parser_name,
-                    self.stream_interval,
                 ),
             )
             # Warm up all workers to ensure initialization completes