Repository navigation
Add NPU support for OmniMRotaryEmbedding - #3609
Merged
hsliuustc0106 merged 1 commit intoJun 18, 2026
Merged
Conversation
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
Wallbreazzz
force-pushed
the
feature/npu-mrope-support
branch
from
May 14, 2026 12:28
c39aadf to
f904d6c
Compare
Signed-off-by: Wallbreazzz <110282866+Wallbreazzz@users.noreply.github.com>
Wallbreazzz
force-pushed
the
feature/npu-mrope-support
branch
from
May 14, 2026 12:33
f904d6c to
fa8b15f
Compare
| from vllm.logger import init_logger | ||
| from vllm.model_executor.layers.rotary_embedding.mrope import MRotaryEmbedding | ||
|
|
||
| from vllm_omni.platforms import current_omni_platform |
Collaborator
There was a problem hiding this comment.
I'm a little bit confused about the plugin system, where are the entries of different hardware? @gcanlin
hsliuustc0106
approved these changes
Jun 17, 2026
alex-jw-brooks
pushed a commit
to alex-jw-brooks/vllm-omni
that referenced
this pull request
Jun 23, 2026
Signed-off-by: Wallbreazzz <110282866+Wallbreazzz@users.noreply.github.com>
khairulkabir1661
pushed a commit
to khairulkabir1661/vllm-omni
that referenced
this pull request
Sep 25, 2026
Signed-off-by: Wallbreazzz <110282866+Wallbreazzz@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
This PR adds NPU platform support for OmniMRotaryEmbedding to enable NPU-specific optimizations for Qwen3-TTS models, resulting in significant performance improvements on Ascend NPU.
Problem
OmniMRotaryEmbedding does not support NPU platform. When running Qwen3-TTS on NPU, it imports the default MRotaryEmbedding from upstream vLLM which lacks NPU-specific optimizations, leading to suboptimal performance.
Root Cause & Fix
The mrope.py file directly imports MRotaryEmbedding from upstream vLLM without platform detection.This prevents using NPU-optimized AscendMRotaryEmbedding from vllm-ascend. Add platform detection to dynamically select the appropriate base class.This enables OmniMRotaryEmbedding to inherit from AscendMRotaryEmbedding on NPU, ensuring proper NPU support while maintaining compatibility with other platforms.
Testing
Performance benchmark on Ascend 910B NPU shows significant improvements:
Single concurrent (max-concurrency=1):
10 concurrent (max-concurrency=10):