[BugFix]converting pa get_workspace back to capturing#5833
Merged
wangxiyuan merged 1 commit intovllm-project:mainfrom Jan 22, 2026
Merged
[BugFix]converting pa get_workspace back to capturing#5833wangxiyuan merged 1 commit intovllm-project:mainfrom
wangxiyuan merged 1 commit intovllm-project:mainfrom
Conversation
Contributor
|
👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:
If CI fails, you can run linting and testing checks locally according Contributing and Testing. |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request refactors the workspace management for paged attention in ACL graphs. It removes a runtime workspace calculation that was implemented as a workaround for a bug in _npu_paged_attention. The new approach relies on using a pre-captured workspace from graph_params.
While this simplifies the code, I have a critical concern about removing the workaround. The original code comments indicated that this safeguard was necessary because smaller sequence lengths could unexpectedly require a larger workspace. My review comment details this concern, highlighting the risk of re-introducing runtime errors if the underlying issue has not been resolved.
I am having trouble creating individual review comments. Click here to see my feedback.
vllm_ascend/compilation/acl_graph.py (240-257)
This change removes the runtime workspace calculation for _npu_paged_attention. This calculation was originally added as a workaround for a bug where smaller seq_lens could require a larger workspace than what's allocated during graph capture. The new implementation relies on graph_params.workspaces.get(runtime_shape) to provide the workspace.
The removed TODO comment indicated this workaround should only be removed once _npu_paged_attention is replaced by npu_fused_infer_attention_score. Since _npu_paged_attention is still in use, removing this safeguard is risky unless the underlying bug in torch_npu has been fixed, or the graph capture logic has been updated to determine a sufficiently large workspace for all cases.
Without confirmation that the underlying issue is resolved, this change may re-introduce the original bug, potentially leading to runtime errors.
yiz-liu
approved these changes
Jan 13, 2026
Contributor
|
This pull request has conflicts, please resolve those before we can evaluate the pull request. |
Angazenn
force-pushed
the
pa_fix
branch
2 times, most recently
from
df842aa to
54c4e59
Compare
Angazenn
changed the title
wangxiyuan
pushed a commit
that referenced
this pull request
Jan 22, 2026
…turing (#6108) ### What this PR does / why we need it? This cherry-picks #5833 . This helps to fix a bug in for pa get_workspace. In earlier implementation, we use `_npu_paged_attention_get_workspace` in `_update_pa_attn_params`. However, this might cause some potential memory problems as it dynamically allocate new memory for workspace when calling this api. Therefor, we move this back to capturing, and use a fixed `SEQ_LEN_WITH_MAX_PA_WORKSPACE` to get max workspace. --------- Signed-off-by: Angazenn <supperccell@163.com>
845473182
pushed a commit
to 845473182/vllm-ascend
that referenced
this pull request
Jan 22, 2026
…to qwen3next_rebase * 'main' of https://github.com/vllm-project/vllm-ascend: (51 commits) [Bugfix] Remove `use_aclgraph` in mtp_proposer and use `use_cuda_graph` (vllm-project#6032) [BugFix] fix 3vl dense model load quant weight (vllm-project#6100) [CP&SP] Integrate FIA operator in mla_cp._forward_decode (vllm-project#5641) [CI][Doc] Upgrade wheel building's CANN to 8.5.0 and update the Docs (vllm-project#6145) [CI]Install clang in dokerfile for triton ascend (vllm-project#4409) [Main] Upgrade PTA to 2.9.0 (vllm-project#6112) [Graph][Fusion] Add QKVNormRope and QKVNormRopeWithBias (vllm-project#5721) [P/D][PCP]bugfix pcp force free twice caused logger error (vllm-project#6124) [BugFix]converting pa get_workspace back to capturing (vllm-project#5833) [CI] optimize lint term (vllm-project#5986) [Bugfix] Fix Triton operator usage for multimodal models based on `the mrope_interleaved` parameter (vllm-project#6042) [bugfix][npugraph_ex]fix the model output type issue caused by manually modify FX graph (vllm-project#6015) [BugFix] Support setting tp=1 for the Eagle draft model to take effect (vllm-project#6097) [Misc] Bump mooncake version to v0.3.8.post1 (vllm-project#6110) [Feature]Enable DispatchGmmCombineDecode when eagle is moe with w8a8 or not moe [RFC: issue 5476] (vllm-project#5758) [bugfix] adapt_remote_request_id (vllm-project#6051) [Feature] Add support of new W4A4_LAOS_DYNAMIC quantization method (vllm-project#5143) [Feature] Support DSA-CP for Hybrid scenario (vllm-project#5702) [CI] Upgrade CANN to 8.5.0 (vllm-project#6070) Default enable MLAPO (vllm-project#5952) ...
starmountain1997
pushed a commit
to starmountain1997/vllm-ascend
that referenced
this pull request
Jan 31, 2026
) ### What this PR does / why we need it? This helps to fix a bug in for pa get_workspace. In earlier implementation, we use `_npu_paged_attention_get_workspace` in `_update_pa_attn_params`. However, this might cause some potential memory problems as it dynamically allocate new memory for workspace when calling this api. Therefor, we move this back to capturing, and use a fixed `SEQ_LEN_WITH_MAX_PA_WORKSPACE` to get max workspace. ### Does this PR introduce _any_ user-facing change? ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: vllm-project/vllm@2f4e654 Signed-off-by: Angazenn <supperccell@163.com>
starmountain1997
pushed a commit
to starmountain1997/vllm-ascend
that referenced
this pull request
Jan 31, 2026
…turing (vllm-project#6108) ### What this PR does / why we need it? This cherry-picks vllm-project#5833 . This helps to fix a bug in for pa get_workspace. In earlier implementation, we use `_npu_paged_attention_get_workspace` in `_update_pa_attn_params`. However, this might cause some potential memory problems as it dynamically allocate new memory for workspace when calling this api. Therefor, we move this back to capturing, and use a fixed `SEQ_LEN_WITH_MAX_PA_WORKSPACE` to get max workspace. --------- Signed-off-by: Angazenn <supperccell@163.com>
starmountain1997
pushed a commit
to starmountain1997/vllm-ascend
that referenced
this pull request
Jan 31, 2026
) ### What this PR does / why we need it? This helps to fix a bug in for pa get_workspace. In earlier implementation, we use `_npu_paged_attention_get_workspace` in `_update_pa_attn_params`. However, this might cause some potential memory problems as it dynamically allocate new memory for workspace when calling this api. Therefor, we move this back to capturing, and use a fixed `SEQ_LEN_WITH_MAX_PA_WORKSPACE` to get max workspace. ### Does this PR introduce _any_ user-facing change? ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: vllm-project/vllm@2f4e654 Signed-off-by: Angazenn <supperccell@163.com>
ZRJ026
pushed a commit
to ZRJ026/vllm-ascend
that referenced
this pull request
Feb 28, 2026
) ### What this PR does / why we need it? This helps to fix a bug in for pa get_workspace. In earlier implementation, we use `_npu_paged_attention_get_workspace` in `_update_pa_attn_params`. However, this might cause some potential memory problems as it dynamically allocate new memory for workspace when calling this api. Therefor, we move this back to capturing, and use a fixed `SEQ_LEN_WITH_MAX_PA_WORKSPACE` to get max workspace. ### Does this PR introduce _any_ user-facing change? ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: vllm-project/vllm@2f4e654 Signed-off-by: Angazenn <supperccell@163.com> Signed-off-by: zrj026 <zhangrunjiang026@gmail.com>
maoxx241
pushed a commit
to maoxx241/vllm-ascend
that referenced
this pull request
Mar 2, 2026
) ### What this PR does / why we need it? This helps to fix a bug in for pa get_workspace. In earlier implementation, we use `_npu_paged_attention_get_workspace` in `_update_pa_attn_params`. However, this might cause some potential memory problems as it dynamically allocate new memory for workspace when calling this api. Therefor, we move this back to capturing, and use a fixed `SEQ_LEN_WITH_MAX_PA_WORKSPACE` to get max workspace. ### Does this PR introduce _any_ user-facing change? ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: vllm-project/vllm@2f4e654 Signed-off-by: Angazenn <supperccell@163.com>
ZRJ026
pushed a commit
to ZRJ026/vllm-ascend
that referenced
this pull request
Mar 4, 2026
) ### What this PR does / why we need it? This helps to fix a bug in for pa get_workspace. In earlier implementation, we use `_npu_paged_attention_get_workspace` in `_update_pa_attn_params`. However, this might cause some potential memory problems as it dynamically allocate new memory for workspace when calling this api. Therefor, we move this back to capturing, and use a fixed `SEQ_LEN_WITH_MAX_PA_WORKSPACE` to get max workspace. ### Does this PR introduce _any_ user-facing change? ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: vllm-project/vllm@2f4e654 Signed-off-by: Angazenn <supperccell@163.com> Signed-off-by: zrj026 <zhangrunjiang026@gmail.com>
LCAIZJ
pushed a commit
to LCAIZJ/vllm-ascend
that referenced
this pull request
Mar 7, 2026
) ### What this PR does / why we need it? This helps to fix a bug in for pa get_workspace. In earlier implementation, we use `_npu_paged_attention_get_workspace` in `_update_pa_attn_params`. However, this might cause some potential memory problems as it dynamically allocate new memory for workspace when calling this api. Therefor, we move this back to capturing, and use a fixed `SEQ_LEN_WITH_MAX_PA_WORKSPACE` to get max workspace. ### Does this PR introduce _any_ user-facing change? ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: vllm-project/vllm@2f4e654 Signed-off-by: Angazenn <supperccell@163.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.
What this PR does / why we need it?
This helps to fix a bug in for pa get_workspace. In earlier implementation, we use
_npu_paged_attention_get_workspacein_update_pa_attn_params. However, this might cause some potential memory problems as it dynamically allocate new memory for workspace when calling this api. Therefor, we move this back to capturing, and use a fixedSEQ_LEN_WITH_MAX_PA_WORKSPACEto get max workspace.Does this PR introduce any user-facing change?
How was this patch tested?