xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-24 04:55:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

Update to Transformers v4.56.2 (#24638)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Harry Mellor 2025-10-01 06:07:07 +01:00 committed by yewentao256
parent 7c795fdf41
commit fda819837e
5 changed files with 20 additions and 34 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
requirements/nightly_torch_test.txt
View File

@ -29,8 +29,8 @@ opencv-python-headless >= 4.11.0 # required for video test
datamodel_code_generator # required for minicpm3 test
lm-eval[api] @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d # required for model evaluation test
mteb>=1.38.11, <2 # required for mteb test
transformers==4.52.4
tokenizers==0.21.1
transformers==4.56.2
tokenizers==0.22.0
schemathesis>=3.39.15 # Required for openai schema test.
# quantization
bitsandbytes>=0.46.1

4
requirements/test.in
View File

@ -37,8 +37,8 @@ datamodel_code_generator # required for minicpm3 test
# TODO: Use lm-eval[api]==0.4.10 once released
lm-eval[api] @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d # required for model evaluation test
mteb[bm25s]>=1.38.11, <2 # required for mteb test
transformers==4.55.2
tokenizers==0.21.1
transformers==4.56.2
tokenizers==0.22.0
schemathesis>=3.39.15 # Required for openai schema test.
# quantization
bitsandbytes==0.46.1

4
requirements/test.txt
View File

@ -1072,7 +1072,7 @@ timm==1.0.17
# segmentation-models-pytorch
# terratorch
# torchgeo
tokenizers==0.21.1
tokenizers==0.22.0
# via
# -r requirements/test.in
# transformers
@ -1153,7 +1153,7 @@ tqdm==4.66.6
# transformers
tqdm-multiprocess==0.0.11
# via lm-eval
transformers==4.55.2
transformers==4.56.2
# via
# -r requirements/test.in
# genai-perf

4
tests/models/multimodal/generation/test_common.py
View File

@ -214,7 +214,9 @@ VLM_TEST_SETTINGS = {
vllm_runner_kwargs={
"model_impl": "transformers",
},
marks=[large_gpu_mark(min_gb=32)],
# FIXME: Investigate mrope issue
marks=[large_gpu_mark(min_gb=32),
pytest.mark.skip(reason="Mrope issue")],
),
#### Extended model tests
"aria": VLMTestInfo(

38
vllm/model_executor/models/transformers.py
View File

@ -51,7 +51,6 @@ from vllm.multimodal.processing import (BaseMultiModalProcessor,
BaseProcessingInfo)
from vllm.multimodal.profiling import BaseDummyInputsBuilder
from vllm.sequence import IntermediateTensors
from vllm.utils import is_list_of
from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
SupportsMultiModal, SupportsPP, SupportsQuant)
@ -217,9 +216,6 @@ def init_on_device_without_buffers(device: torch.device):
class MultiModalProcessingInfo(BaseProcessingInfo):
def get_hf_config(self):
return self.ctx.model_config.hf_config
def get_supported_mm_limits(self):
return {"image": None}
@ -784,6 +780,7 @@ def flatten_and_concat(x: list[torch.Tensor]) -> torch.Tensor:
},
enable_if=can_enable_torch_compile)
class TransformersForMultimodalLM(TransformersForCausalLM, SupportsMultiModal):
merge_by_field_config = True
# Backwards compatibility for prev released models. State dicts back then
# had different formats and cannot be loaded with `AutoModel` mapping as is
hf_to_vllm_mapper = WeightsMapper(
@ -828,40 +825,27 @@ class TransformersForMultimodalLM(TransformersForCausalLM, SupportsMultiModal):
return self.model
def get_multimodal_embeddings(self, **kwargs):
pixel_values = kwargs.pop("pixel_values", None)
pixel_values = pixel_values if pixel_values is not None else kwargs.pop(
"image_patches", None)
image_embeds = kwargs.pop("image_embeds", None)
pixel_values: Optional[torch.Tensor] = kwargs.pop("pixel_values", None)
image_embeds: Optional[torch.Tensor] = kwargs.pop("image_embeds", None)
# Model might use `image_patches` instead of `pixel_values`
if pixel_values is None:
pixel_values = kwargs.pop("image_patches", None)
if image_embeds is not None:
return image_embeds
if pixel_values is None and image_embeds is None:
if pixel_values is None:
return None
num_image_patches = kwargs.pop("num_image_patches")
if pixel_values is not None:
if isinstance(pixel_values, torch.Tensor):
pixel_values = flatten_bn(pixel_values).to(self.dtype)
elif is_list_of(pixel_values, torch.Tensor):
pixel_values = flatten_and_concat(pixel_values).to(self.dtype)
else:
raise ValueError(
f"Unsupported pixel_values type {type(pixel_values)}. "
"Expected `torch.Tensor` or list of `torch.Tensor`.")
if isinstance(num_image_patches, list):
num_image_patches = torch.cat(num_image_patches)
vision_embeddings = self.model.get_image_features(
pixel_values,
**{
k: v.flatten(0, 1)
for k, v in kwargs.items()
},
)
pixel_values, **kwargs)
if isinstance(vision_embeddings, torch.Tensor):
if isinstance(num_image_patches, list):
num_image_patches = torch.cat(num_image_patches)
if vision_embeddings.ndim == 2:
vision_embeddings = vision_embeddings.unsqueeze(0)