minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[NPU]Bugfix:Set default values for npu_wrapper_preprocess parameters (#25130)

Browse Source
This commit is contained in:
AndyLi429
2026-05-14 14:21:11 +08:00
committed by GitHub
parent be156d6804
commit 3640116397
2 changed files with 12 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/hardware_backend/npu/attention/ascend_gdn_backend.py
+1 -1
View File
@@ -414,7 +414,7 @@ class AscendGDNAttnBackend(AscendMambaAttnBackendBase):
nv=num_value_heads, nv=num_value_heads,
intermediate_state=intermediate_state, intermediate_state=intermediate_state,
cache_indices=cache_indices, cache_indices=cache_indices,
num_accept_tokens=num_accept_tokens, num_accepted_tokens=num_accept_tokens,
g=g, g=g,
) )
python/sglang/srt/hardware_backend/npu/modules/qwen_vl_processor.py
+11 -16
View File
@@ -1,12 +1,7 @@
from typing import Optional
import torch import torch
import torchvision.transforms.v2.functional as tvF import torchvision.transforms.v2.functional as tvF
from transformers.image_processing_utils import BatchFeature from transformers.image_processing_utils import BatchFeature
from transformers.image_processing_utils_fast import ( from transformers.image_transforms import group_images_by_shape, reorder_images
group_images_by_shape,
reorder_images,
)
from transformers.image_utils import ( from transformers.image_utils import (
ChannelDimension, ChannelDimension,
PILImageResampling, PILImageResampling,
@@ -63,8 +58,8 @@ def transform_patches_to_flatten(
return flatten_patches return flatten_patches
# Func refers to transformers.models.qwen2_vl.image_processing_qwen2_vl_fast.py # Func refers to transformers.models.qwen2_vl.image_processing_qwen2_vl.py
# Qwen2VLImageProcessorFast._preprocess # Qwen2VLImageProcessor._preprocess
def npu_wrapper_preprocess(func): def npu_wrapper_preprocess(func):
def _preprocess( def _preprocess(
@@ -72,7 +67,7 @@ def npu_wrapper_preprocess(func):
images: list["torch.Tensor"], images: list["torch.Tensor"],
do_resize: bool, do_resize: bool,
size: SizeDict, size: SizeDict,
interpolation: Optional["tvF.InterpolationMode"], resample: "PILImageResampling | tvF.InterpolationMode | int | None",
do_rescale: bool, do_rescale: bool,
rescale_factor: float, rescale_factor: float,
do_normalize: bool, do_normalize: bool,
@@ -97,13 +92,13 @@ def npu_wrapper_preprocess(func):
height, height,
width, width,
factor=patch_size * merge_size, factor=patch_size * merge_size,
min_pixels=size["shortest_edge"], min_pixels=size.shortest_edge,
max_pixels=size["longest_edge"], max_pixels=size.longest_edge,
) )
stacked_images = self.resize( stacked_images = self.resize(
image=stacked_images, image=stacked_images,
size=SizeDict(height=resized_height, width=resized_width), size=SizeDict(height=resized_height, width=resized_width),
interpolation=interpolation, resample=resample,
) )
resized_images_grouped[shape] = stacked_images resized_images_grouped[shape] = stacked_images
resized_images = reorder_images(resized_images_grouped, grouped_images_index) resized_images = reorder_images(resized_images_grouped, grouped_images_index)
@@ -173,7 +168,7 @@ def npu_wrapper_preprocess(func):
# Func refers to transformers.models.qwen3_vl.video_processing_qwen3_vl.py # Func refers to transformers.models.qwen3_vl.video_processing_qwen3_vl.py
# Qwen3VLVideoProcessorFast._preprocess # Qwen3VLVideoProcessor._preprocess
def npu_wrapper_video_preprocess(func): def npu_wrapper_video_preprocess(func):
def _preprocess( def _preprocess(
@@ -182,7 +177,7 @@ def npu_wrapper_video_preprocess(func):
do_convert_rgb: bool = True, do_convert_rgb: bool = True,
do_resize: bool = True, do_resize: bool = True,
size: SizeDict | None = None, size: SizeDict | None = None,
interpolation: PILImageResampling = PILImageResampling.BICUBIC, resample: "PILImageResampling | tvF.InterpolationMode | int | None" = PILImageResampling.BICUBIC,
do_rescale: bool = True, do_rescale: bool = True,
rescale_factor: float = 1 / 255.0, rescale_factor: float = 1 / 255.0,
do_normalize: bool = True, do_normalize: bool = True,
@@ -214,7 +209,7 @@ def npu_wrapper_video_preprocess(func):
stacked_videos = self.resize( stacked_videos = self.resize(
stacked_videos, stacked_videos,
size=SizeDict(height=resized_height, width=resized_width), size=SizeDict(height=resized_height, width=resized_width),
interpolation=interpolation, resample=resample,
) )
stacked_videos = stacked_videos.view( stacked_videos = stacked_videos.view(
B, T, C, resized_height, resized_width B, T, C, resized_height, resized_width
@@ -297,7 +292,7 @@ def npu_apply_qwen_image_preprocess_patch():
if _npu_preprocess_patched: if _npu_preprocess_patched:
return return
apply_module_patch( apply_module_patch(
"transformers.models.qwen2_vl.image_processing_qwen2_vl_fast.Qwen2VLImageProcessorFast", "transformers.models.qwen2_vl.image_processing_qwen2_vl.Qwen2VLImageProcessor",
"_preprocess", "_preprocess",
[npu_wrapper_preprocess], [npu_wrapper_preprocess],
) )