Repository navigation
[Bugfix][NPU] Fix talker_mtp ACL graph capture missing force_uniform_decode - #7008
Merged
hsliuustc0106 merged 4 commits intoSep 17, 2026
Merged
hsliuustc0106 merged 4 commits into
hsliuustc0106 merged 4 commits into
Conversation
NPU's _determine_batch_execution_and_padding checks has_initial_state (num_computed_tokens > 0) for uniform_decode, which is False during graph capture (no real requests) but True at runtime (decode requests have computed tokens). This BatchDescriptor.uniform mismatch causes ACLGraphWrapper cache miss → lazy capture attempt → RuntimeError 'CUDA graph capturing detected at an inappropriate time'. Signed-off-by: Wallbreazzz <wallbreazzz@users.noreply.github.com> Signed-off-by: Wallbreazzz <110282866+Wallbreazzz@users.noreply.github.com>
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
This PR appears to belong to: docs/design/module/execution_platforms.md. Module owners: @gcanlin @tjtanaa @xuechendi @Wallbreazzz, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer. Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment. |
Omni ReviewBot triage noteAutomated triage of commit
These are automated triage suggestions only — the final decision belongs to the maintainers. |
hsliuustc0106
approved these changes
Sep 3, 2026
gcanlin
added
ready
hsliuustc0106
added
ready
hsliuustc0106
enabled auto-merge (squash)
September 17, 2026 03:47
mlaneuville
pushed a commit
to mlaneuville/vllm-omni
that referenced
this pull request
Sep 22, 2026
…decode (vllm-project#7008) Signed-off-by: Wallbreazzz <wallbreazzz@users.noreply.github.com> Signed-off-by: Wallbreazzz <110282866+Wallbreazzz@users.noreply.github.com> Co-authored-by: Canlin Guo <canlinguosdu@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Signed-off-by: Matthieu Laneuville <matthieu.laneuville@surf.nl>
6 tasks done
khairulkabir1661
pushed a commit
to khairulkabir1661/vllm-omni
that referenced
this pull request
Sep 25, 2026
…decode (vllm-project#7008) Signed-off-by: Wallbreazzz <wallbreazzz@users.noreply.github.com> Signed-off-by: Wallbreazzz <110282866+Wallbreazzz@users.noreply.github.com> Co-authored-by: Canlin Guo <canlinguosdu@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.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
6 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.
What
NPU TTS models (e.g. Moss-TTS) crash on first request with
RuntimeError: CUDA graph capturing detected at an inappropriate time.Root cause
_capture_talker_mtp_graphscalls_determine_batch_execution_and_paddingwithoutforce_uniform_decode. On NPU, this method checkshas_initial_state = np.all(num_computed_tokens_cpu > 0), which isFalseduring capture (no real requests) butTrueat runtime (decode requests have computed tokens). This makesBatchDescriptor.uniformdiffer (Falseat capture vsTrueat runtime), causing anACLGraphWrappercache miss → lazy capture attempt → error.GPU is unaffected because base vLLM's
_is_uniform_decodedoes not checkhas_initial_state.Fix
Pass
force_uniform_decode=Trueduring capture so the descriptor matches runtime.Verified on Ascend 910B with Moss-TTS-Local-Transformer-v1.5 (v0.28.0 image).