[Feature] Warm up readonly multimodal processor during renderer startup (#40797)

Signed-off-by: Chenguang ZHENG <645327136@qq.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Chenguang Zheng
2026-04-25 03:57:41 +00:00
committed by GitHub
co-authored by OpenAI Codex
parent 428b988c98
commit 07351e0883
7 changed files with 368 additions and 27 deletions
+8 -4
View File
@@ -136,7 +136,7 @@ class OpenAIServingRender:
"Beam search is not supported by the render endpoint"
)
result = await self.render_chat(request)
result = await self.render_chat(request, skip_mm_cache=True)
if isinstance(result, ErrorResponse):
return result
@@ -184,6 +184,8 @@ class OpenAIServingRender:
async def render_chat(
self,
request: ChatCompletionRequest,
*,
skip_mm_cache: bool = False,
) -> tuple[list[ConversationMessage], list[EngineInput]] | ErrorResponse:
"""Core preprocessing logic for chat requests (no model/engine check).
@@ -252,7 +254,7 @@ class OpenAIServingRender:
default_template_kwargs=self.default_chat_template_kwargs,
tool_dicts=tool_dicts,
tool_parser=tool_parser,
skip_mm_cache=True,
skip_mm_cache=skip_mm_cache,
reasoning_parser=self.reasoning_parser,
)
else:
@@ -276,7 +278,7 @@ class OpenAIServingRender:
error_check_ret = await self._check_model(request)
if error_check_ret is not None:
return error_check_ret
result = await self.render_completion(request)
result = await self.render_completion(request, skip_mm_cache=True)
if isinstance(result, ErrorResponse):
return result
generate_requests: list[GenerateRequest] = []
@@ -322,6 +324,8 @@ class OpenAIServingRender:
async def render_completion(
self,
request: CompletionRequest,
*,
skip_mm_cache: bool = False,
) -> list[EngineInput] | ErrorResponse:
"""Core preprocessing logic for completion requests (no model/engine check).
@@ -344,7 +348,7 @@ class OpenAIServingRender:
request,
prompt_input=request.prompt,
prompt_embeds=request.prompt_embeds,
skip_mm_cache=True,
skip_mm_cache=skip_mm_cache,
)
return engine_inputs