Configs in vLLM

Some details on Config loading in vLLM

1. Configs hf_config: The config.json from HF
   hf_text_config: llm_config from config.json

# api_server.py -> run_server -> init_app -> serve_http
def init_app()
  ...
  engine = AsyncLLMEngine.from_engine_args()
class AsyncEngine:
  def from_engine_args(...)
    ...
    vllm_config = engine_args.create_engine_config(usage_context)
class EngineArgs:
  def create_engine_config(self):
      ...
      model_config = self.create_model_config()
      return VllmConfig(
        model_config=model_config,
        ...
        )
  def create_model_config(self):
    ...
    return ModelConfig(...)
class ModelConfig:
  def __post_init__(self):
    ...
    hf_config = get_config(...) # read from the config registered at vllm/transformers_utils/configs/
    # Read `auto_map` configs
    self.hf_text_config = get_hf_text_config(self.hf_config)
  def get_hf_text_config(config: PretrainedConfig):
      ...
      ## transformers lib
      text_config = config.get_text_config()
      # looking for decoder_possible_text_config_names = ("decoder", "generator", "text_config")
      # if found, return 
      # otherwise return config so that text_config == config is True

2 Preprocessor_config

preprocessor_config is read by AutoImageProcessor and it’s also implemented in the same file in vLLM vllm.transformers_utils .processor.get_image_processors

processor = AutoImageProcessor.from_pretrained(
    processor_name,
    *args,
    revision=revision,
    trust_remote_code=trust_remote_code,
    **kwargs)

3 Decorators

MultiModel Registration decorator

@MULTIMODAL_REGISTRY.register_processor(
    InternVLMultiModalProcessor,
    info=InternVLProcessingInfo,
    dummy_inputs=InternVLDummyInputsBuilder)

1. Firstly self.processing_info.get_allowed_mm_limits() is called from multimodal.profiler.MultiModalProfiler by profiler.get_mm_limits()
2. class InternVLProcessingInfo
   • class BaseInternVLProcessingInfo
     • class BaseProcessingInfo from vllm.multimodal.processing
3. class InternVLMultiModalProcessor
   • class BaseInternVLMultiModalProcessor[InternVLProcessingInfo]
     • class BaseMultiModalProcessor[_I] from vllm.multimodal.processing
4. class InternVLDummyInputsBuilder
   • class BaseInternVLDummyInputsBuilder[InternVLProcessingInfo]
     • class BaseDummyInputsBuilder[_I]) from vllm.multimodal.profiling
5. _I = TypeVar("_I", bound=BaseInternVLProcessingInfo)
6. class InternVLProcessor
   • class BaseInternVLProcessor(ABC)
   • This is used at InternVLProcessingInfo.get_hf_processor