[VLM] Refactor MultiModalConfig initialization and profiling (#7530)

This commit is contained in:
Roger Wang
2024-08-17 13:30:55 -07:00
committed by GitHub
parent 1ef13cf92f
commit bbf55c4805
29 changed files with 143 additions and 190 deletions
+2 -4
View File
@@ -4,8 +4,8 @@ from typing import Any, Dict, List, Optional, Tuple, Type
import torch
from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
ModelConfig, MultiModalConfig, ObservabilityConfig,
ParallelConfig, PromptAdapterConfig, SchedulerConfig)
ModelConfig, ObservabilityConfig, ParallelConfig,
PromptAdapterConfig, SchedulerConfig)
from vllm.logger import init_logger
from vllm.model_executor.pooling_metadata import PoolingMetadata
from vllm.multimodal import MultiModalInputs
@@ -44,7 +44,6 @@ class EmbeddingModelRunner(
kv_cache_dtype: Optional[str] = "auto",
is_driver_worker: bool = False,
prompt_adapter_config: Optional[PromptAdapterConfig] = None,
multimodal_config: Optional[MultiModalConfig] = None,
observability_config: Optional[ObservabilityConfig] = None,
):
super().__init__(model_config,
@@ -57,7 +56,6 @@ class EmbeddingModelRunner(
kv_cache_dtype=kv_cache_dtype,
is_driver_worker=is_driver_worker,
prompt_adapter_config=prompt_adapter_config,
multimodal_config=multimodal_config,
observability_config=observability_config)
@torch.inference_mode()