[Renderer] Introduce Renderer for processing chat messages (using RendererConfig)
#30198
Conversation
DarkLight1337
changed the title
mergify
bot
added
frontend
multi-modality
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces a Renderer abstraction to encapsulate chat template processing and tokenization logic. This is a significant and positive architectural refactoring, moving model-specific rendering logic out of the core engine and into dedicated renderer classes. The changes are extensive, touching many parts of the codebase to replace direct tokenizer usage with the new renderer interface. The implementation appears mostly correct and consistent. However, I found a bug in the MistralToolParser where an instance variable is not initialized, which will lead to an AttributeError.
| "the tokenizer!" | ||
| ) | ||
|
|
||
| self.prev_tool_call_arr: list[dict[str, Any]] |
There was a problem hiding this comment.
The instance variable prev_tool_call_arr is declared as a type hint but not initialized in the __init__ method. This will cause an AttributeError when it's accessed before being set. Please initialize it in __init__, for example: self.prev_tool_call_arr = [].
| return None | ||
|
|
||
| renderer = self._get_renderer(ctx.tokenizer) | ||
| renderer = self._get_renderer(self.renderer.tokenizer) |
There was a problem hiding this comment.
Note that vllm.renderers.Renderer (self.renderer) is currently for chat only, and is not to be confused with the Renderer inside vllm.entrypoints.renderer.CompletionRenderer (the result of self._get_renderer). The two implementations will be merged in a later PR.
|
|
||
| self.request_logger = request_logger | ||
| self.return_tokens_as_token_ids = return_tokens_as_token_ids | ||
| self._tokenizer_executor = ThreadPoolExecutor(max_workers=1) |
There was a problem hiding this comment.
This has been moved to the Mistral renderer
| created_time: int = field(default_factory=lambda: int(time.time())) | ||
| lora_request: LoRARequest | None = None | ||
|
|
||
| # Shared across most requests |
There was a problem hiding this comment.
Prefer using self.renderer to simplify the code.
|
This pull request has merge conflicts that must be resolved before it can be |
Signed-off-by: DarkLight1337 <[email protected]>
DarkLight1337
force-pushed
the
init-renderer
branch
from
0db05b1 to
47c1f05
Compare
DarkLight1337
changed the title
RendererConfig)
DarkLight1337
changed the title
RendererConfig)ModelConfig)
DarkLight1337
changed the title
ModelConfig)RendererConfig)
github-project-automation
bot
moved this from To Triage
to Done
in gpt-oss Issues & Enhancements
|
Superseded by #30200 |
1 participant
Purpose
vllm.renderers.RendererLike, to process chat messages into engine inputs.RendererRegistrywhich lazily registers renderers to avoid circular import problem..vllm.renderers.tokenizer_modetorenderer_mode, and use a specific tokenizer implementation for each renderer, deprecatingTokenizerRegistryin the process.Towards #22880 and #23873
Test Plan
Test Result
Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model.