[model][refactor] remove cuda hard code in models and layers (#13658)
This commit is contained in:
@@ -59,6 +59,7 @@ from vllm.multimodal.parse import (DictEmbeddingItems, ImageItem, ImageSize,
|
||||
from vllm.multimodal.processing import (BaseMultiModalProcessor,
|
||||
BaseProcessingInfo, PromptReplacement)
|
||||
from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.sequence import IntermediateTensors
|
||||
|
||||
from .idefics2_vision_model import Idefics2VisionTransformer
|
||||
@@ -1184,7 +1185,8 @@ class MiniCPMV2_0(MiniCPMVBaseModel):
|
||||
quant_config=quant_config,
|
||||
prefix=prefix)
|
||||
|
||||
return resampler.to(device="cuda", dtype=torch.get_default_dtype())
|
||||
return resampler.to(device=current_platform.device_type,
|
||||
dtype=torch.get_default_dtype())
|
||||
|
||||
def get_vision_embedding(
|
||||
self,
|
||||
@@ -1266,7 +1268,8 @@ class MiniCPMV2_5(MiniCPMVBaseModel, SupportsLoRA):
|
||||
quant_config=quant_config,
|
||||
prefix=prefix)
|
||||
|
||||
return resampler.to(device="cuda", dtype=torch.get_default_dtype())
|
||||
return resampler.to(device=current_platform.device_type,
|
||||
dtype=torch.get_default_dtype())
|
||||
|
||||
def get_vision_embedding(
|
||||
self,
|
||||
@@ -1360,7 +1363,8 @@ class MiniCPMV2_6(MiniCPMVBaseModel, SupportsLoRA):
|
||||
quant_config=quant_config,
|
||||
prefix=prefix)
|
||||
|
||||
return resampler.to(device="cuda", dtype=torch.get_default_dtype())
|
||||
return resampler.to(device=current_platform.device_type,
|
||||
dtype=torch.get_default_dtype())
|
||||
|
||||
def get_vision_embedding(
|
||||
self,
|
||||
|
||||
Reference in New Issue
Block a user