Revert "[Renderer] Separate out RendererConfig from ModelConfig (#30145)" (#30199)

This commit is contained in:
Cyrus Leung
2025-12-07 00:00:22 -08:00
committed by GitHub
parent 27f4c2fd46
commit e83b7e379c
105 changed files with 797 additions and 969 deletions
+3 -4
View File
@@ -91,7 +91,6 @@ class AsyncLLM(EngineClient):
# Ensure we can serialize custom transformer configs
maybe_register_config_serialize_by_value()
self.renderer_config = vllm_config.renderer_config
self.model_config = vllm_config.model_config
self.vllm_config = vllm_config
self.observability_config = vllm_config.observability_config
@@ -109,15 +108,15 @@ class AsyncLLM(EngineClient):
"enabling logging without default stat loggers."
)
if self.renderer_config.skip_tokenizer_init:
if self.model_config.skip_tokenizer_init:
tokenizer = None
else:
tokenizer = init_tokenizer_from_config(self.renderer_config)
tokenizer = init_tokenizer_from_config(self.model_config)
self.input_processor = InputProcessor(self.vllm_config, tokenizer)
self.io_processor = get_io_processor(
self.vllm_config,
self.renderer_config.io_processor_plugin,
self.model_config.io_processor_plugin,
)
# OutputProcessor (converts EngineCoreOutputs --> RequestOutput).
+3 -4
View File
@@ -43,7 +43,6 @@ class InputProcessor:
mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
) -> None:
self.vllm_config = vllm_config
self.renderer_config = vllm_config.renderer_config
self.model_config = vllm_config.model_config
self.cache_config = vllm_config.cache_config
self.lora_config = vllm_config.lora_config
@@ -55,7 +54,7 @@ class InputProcessor:
self.mm_processor_cache = processor_cache_from_config(vllm_config, mm_registry)
self.input_preprocessor = InputPreprocessor(
self.renderer_config,
self.model_config,
tokenizer,
mm_registry,
mm_processor_cache=self.mm_processor_cache,
@@ -253,7 +252,7 @@ class InputProcessor:
if not params.structured_outputs or not self.structured_outputs_config:
return
if self.renderer_config.skip_tokenizer_init and params.structured_outputs:
if self.model_config.skip_tokenizer_init and params.structured_outputs:
raise ValueError(
"Structured outputs requires a tokenizer so it can't be used with 'skip_tokenizer_init'" # noqa: E501
)
@@ -583,7 +582,7 @@ class InputProcessor:
if prompt_type == "encoder" and model_config.is_multimodal_model:
mm_registry = self.input_preprocessor.mm_registry
mm_processor = mm_registry.create_processor(
self.renderer_config,
model_config,
tokenizer=tokenizer,
)
assert isinstance(mm_processor, EncDecMultiModalProcessor)
+3 -4
View File
@@ -60,7 +60,6 @@ class LLMEngine:
) -> None:
self.vllm_config = vllm_config
self.observability_config = vllm_config.observability_config
self.renderer_config = vllm_config.renderer_config
self.model_config = vllm_config.model_config
self.cache_config = vllm_config.cache_config
@@ -84,15 +83,15 @@ class LLMEngine:
self.dp_group = None
self.should_execute_dummy_batch = False
if self.renderer_config.skip_tokenizer_init:
if self.model_config.skip_tokenizer_init:
tokenizer = None
else:
tokenizer = init_tokenizer_from_config(self.renderer_config)
tokenizer = init_tokenizer_from_config(self.model_config)
self.input_processor = InputProcessor(self.vllm_config, tokenizer)
self.io_processor = get_io_processor(
self.vllm_config,
self.renderer_config.io_processor_plugin,
self.model_config.io_processor_plugin,
)
# OutputProcessor (convert EngineCoreOutputs --> RequestOutput).