Skip to content

fix qwen3vl mrope op #4484

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
+2 −6
Open

fix qwen3vl mrope op #4484

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
64fabac
fix qwen3vl mrope op
shaopeng-666 Nov 27, 2025
88f1c73
fix qwen3vl mrope op
shaopeng-666 Nov 28, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 2 additions & 6 deletions vllm_ascend/ops/rotary_embedding.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,6 @@
from vllm.model_executor.layers.rotary_embedding import (
DeepseekScalingRotaryEmbedding, MRotaryEmbedding, RotaryEmbedding,
YaRNScalingRotaryEmbedding)
from vllm.platforms import CpuArchEnum

from vllm_ascend.platform import NPUPlatform
from vllm_ascend.utils import (AscendDeviceType, enable_custom_op,
Expand Down Expand Up @@ -408,10 +407,7 @@ def forward_oot(
query: torch.Tensor,
key: torch.Tensor,
):
# TODO: This judgment will be removed once the mrope precision issue is fixed
if self.mrope_section != [
16, 24, 24
] or NPUPlatform.get_cpu_architecture() == CpuArchEnum.X86:
if self.mrope_section != [16, 24, 24]:
return super().forward_oot(positions, query, key)

import torch_npu
Expand All @@ -426,7 +422,7 @@ def forward_oot(
self.cos_sin_cache = self.cos_sin_cache.to( # type: ignore
query.dtype) # type: ignore

query, key = torch_npu.npu_mrope(positions,
query, key = torch_npu.npu_mrope(positions.contiguous(),
query.contiguous(),
key.contiguous(),
self.cos_sin_cache.contiguous(),
Expand Down
Loading