[VLM] Refactor MultiModalConfig initialization and profiling (#7530)
This commit is contained in:
@@ -7,8 +7,8 @@ import torch.distributed
|
||||
import vllm.envs as envs
|
||||
from vllm.attention import get_attn_backend
|
||||
from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
|
||||
ModelConfig, MultiModalConfig, ParallelConfig,
|
||||
PromptAdapterConfig, SchedulerConfig)
|
||||
ModelConfig, ParallelConfig, PromptAdapterConfig,
|
||||
SchedulerConfig)
|
||||
from vllm.distributed import (ensure_model_parallel_initialized,
|
||||
init_distributed_environment)
|
||||
from vllm.logger import init_logger
|
||||
@@ -132,7 +132,6 @@ class CPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
|
||||
rank: int,
|
||||
distributed_init_method: str,
|
||||
lora_config: Optional[LoRAConfig] = None,
|
||||
multimodal_config: Optional[MultiModalConfig] = None,
|
||||
kv_cache_dtype: Optional[str] = "auto",
|
||||
prompt_adapter_config: Optional[PromptAdapterConfig] = None,
|
||||
is_driver_worker: bool = False,
|
||||
@@ -148,7 +147,6 @@ class CPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
|
||||
self.distributed_init_method = distributed_init_method
|
||||
self.lora_config = lora_config
|
||||
self.prompt_adapter_config = prompt_adapter_config
|
||||
self.multimodal_config = multimodal_config
|
||||
self.is_driver_worker = is_driver_worker
|
||||
if self.is_driver_worker:
|
||||
assert self.rank == 0, "The driver worker must have rank 0."
|
||||
@@ -173,7 +171,6 @@ class CPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
|
||||
cache_config,
|
||||
load_config=self.load_config,
|
||||
lora_config=self.lora_config,
|
||||
multimodal_config=self.multimodal_config,
|
||||
kv_cache_dtype=kv_cache_dtype,
|
||||
prompt_adapter_config=self.prompt_adapter_config,
|
||||
is_driver_worker=is_driver_worker)
|
||||
|
||||
Reference in New Issue
Block a user