[Models]: Make Multimodal config implicit in ViT implementation (#31972)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
@@ -10,7 +10,7 @@ from typing import Final, Generic, Literal, Protocol, TypeAlias, TypeVar
|
||||
import torch
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.config import MultiModalConfig, VllmConfig, get_current_vllm_config
|
||||
from vllm.distributed import (
|
||||
get_tensor_model_parallel_rank,
|
||||
get_tensor_model_parallel_world_size,
|
||||
@@ -79,7 +79,7 @@ def get_vision_encoder_info(hf_config: VisionLanguageConfig) -> VisionEncoderInf
|
||||
raise NotImplementedError(msg)
|
||||
|
||||
|
||||
def get_vit_attn_backend(
|
||||
def _get_vit_attn_backend(
|
||||
head_size: int,
|
||||
dtype: torch.dtype,
|
||||
*,
|
||||
@@ -95,6 +95,52 @@ def get_vit_attn_backend(
|
||||
)
|
||||
|
||||
|
||||
def get_vit_attn_backend(
|
||||
head_size: int,
|
||||
dtype: torch.dtype,
|
||||
) -> AttentionBackendEnum:
|
||||
"""
|
||||
Get the attention backend for Vision Transformer.
|
||||
"""
|
||||
try:
|
||||
vllm_config: VllmConfig = get_current_vllm_config()
|
||||
multimodal_config: MultiModalConfig | None = (
|
||||
vllm_config.model_config.multimodal_config
|
||||
)
|
||||
except AssertionError:
|
||||
multimodal_config = None
|
||||
|
||||
attn_backend_override = (
|
||||
multimodal_config.mm_encoder_attn_backend
|
||||
if multimodal_config is not None
|
||||
else None
|
||||
)
|
||||
attn_backend = _get_vit_attn_backend(
|
||||
head_size,
|
||||
dtype,
|
||||
attn_backend_override=attn_backend_override,
|
||||
)
|
||||
return attn_backend
|
||||
|
||||
|
||||
def is_vit_use_data_parallel():
|
||||
"""
|
||||
Get the tensor parallel type for Vision Transformer.
|
||||
"""
|
||||
try:
|
||||
vllm_config: VllmConfig = get_current_vllm_config()
|
||||
multimodal_config: MultiModalConfig | None = (
|
||||
vllm_config.model_config.multimodal_config
|
||||
)
|
||||
except AssertionError:
|
||||
multimodal_config = None
|
||||
|
||||
mm_encoder_tp_mode = (
|
||||
multimodal_config.mm_encoder_tp_mode if multimodal_config is not None else None
|
||||
)
|
||||
return mm_encoder_tp_mode == "data"
|
||||
|
||||
|
||||
def should_torch_compile_mm_vit(vllm_config: VllmConfig) -> bool:
|
||||
"""Callable to be passed to `@support_torch_compile`'s `enable_if` argument."""
|
||||
return vllm_config.compilation_config.compile_mm_encoder
|
||||
|
||||
Reference in New Issue
Block a user