EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Drop 0.11.0 support (#4377)

Browse Source 
There is a lot hack code for v0.11.0, which makes the code hard to
upgrade to newer vLLM version. Since v0.11.0 will release soon. Let's
drop v0.11.0 support first. Then we'll upgrade to v0.11.2 soon.


- vLLM version: v0.11.0
- vLLM main:
2918c1b49c

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-11-24 17:08:20 +08:00
committed by GitHub
parent 41ddb06554
commit a1f142b7ad
80 changed files with 467 additions and 1755 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
vllm_ascend/models/qwen2_5_vl.py
View File

@@ -40,14 +40,11 @@ from vllm.model_executor.models.qwen2_5_vl import (
Qwen2_5_VLDummyInputsBuilder, Qwen2_5_VLForConditionalGeneration,
Qwen2_5_VLMultiModalProcessor, Qwen2_5_VLProcessingInfo)
from vllm.model_executor.models.utils import maybe_prefix
from vllm.model_executor.models.vision import conv3d_to_linear_weight
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm_ascend.ascend_forward_context import set_ascend_forward_context
from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_ND, is_enable_nz,
vllm_version_is)
if not vllm_version_is("0.11.0"):
from vllm.model_executor.models.vision import conv3d_to_linear_weight
from vllm_ascend.utils import ACL_FORMAT_FRACTAL_ND, is_enable_nz
MIN_PAD_SIZE = 64 # min_size to pad weight
MAX_PAD_SIZE = 128 # max_size to pad weight
@@ -360,9 +357,8 @@ class AscendQwen2_5_VisionTransformer(Qwen2_5_VisionTransformer):
params_dict = dict(self.named_parameters(remove_duplicate=False))
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if not vllm_version_is("0.11.0"):
if name.endswith("patch_embed.proj.weight"):
loaded_weight = conv3d_to_linear_weight(loaded_weight)
if name.endswith("patch_embed.proj.weight"):
loaded_weight = conv3d_to_linear_weight(loaded_weight)
for (param_name, weight_name, shard_id) in stacked_params_mapping:
if weight_name not in name:
continue
@@ -537,11 +533,8 @@ class AscendQwen2_5_VLForConditionalGeneration(
image_embeds = image_input["image_embeds"].type(self.visual.dtype)
else:
pixel_values = image_input["pixel_values"].type(self.visual.dtype)
if vllm_version_is("0.11.0"):
with set_ascend_forward_context(None, self.vllm_config):
image_embeds = self.visual(pixel_values, grid_thw=grid_thw)
else:
with set_ascend_forward_context(None, self.vllm_config):
image_embeds = self.visual(pixel_values, grid_thw=grid_thw)
# Split concatenated embeddings for each image item.
merge_size = self.visual.spatial_merge_size
@@ -558,13 +551,9 @@ class AscendQwen2_5_VLForConditionalGeneration(
else:
pixel_values_videos = video_input["pixel_values_videos"].type(
self.visual.dtype)
if vllm_version_is("0.11.0"):
with set_ascend_forward_context(None, self.vllm_config):
video_embeds = self.visual(pixel_values_videos,
grid_thw=grid_thw)
else:
with set_ascend_forward_context(None, self.vllm_config):
video_embeds = self.visual(pixel_values_videos,
grid_thw=grid_thw)
# Split concatenated embeddings for each video item.
merge_size = self.visual.spatial_merge_size