[Bugfix] Reset all unused positions to prevent out-of-bounds in GatherV3 #1416
Conversation
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
…rror Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #1416 +/- ##
==========================================
- Coverage 27.39% 27.21% -0.19%
==========================================
Files 56 56
Lines 6191 6214 +23
==========================================
- Hits 1696 1691 -5
- Misses 4495 4523 +28
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
Yikun
changed the title
Yikun
approved these changes
Jun 25, 2025
Yikun
reviewed
Jun 25, 2025
Collaborator
Yikun
left a comment
Yikun
left a comment
There was a problem hiding this comment.
Considering that we have been digg this problem for a long long long time, it would be great if there is a regression test (ut or e2e test) to check and prevent the break again.
Feel free to add test in a separate PR.
wangxiyuan
approved these changes
Jun 26, 2025
Collaborator
|
Let's add the ut later once the test framwork is strong enough. |
Collaborator
Author
Maybe we should re-evaluate the code that generates the various attention masks to identify any conflicts with the padding logic, and add some tests for masks, as these do not exist in vLLM. |
Yikun
added
long-term-test
weijinqian0
pushed a commit
to weijinqian0/vllm-ascend
that referenced
this pull request
Jun 30, 2025
…rV3 (vllm-project#1416) ### What this PR does / why we need it? Reset all unused positions in `NPUModelRunner` to prevent out-of-bounds asserts in the `GatherV3` operator. Currently, in [`get_splitfuse_attn_mask`](https://github.com/vllm-project/vllm-ascend/blob/main/vllm_ascend/attention/attention.py#L124), the `position` tensor may contain values that exceed the dimensions of the attention mask, triggering a `GatherV3` boundary check failure. These invalid indices originate from stale “dirty” entries left over in `position` due to padding logic in the ACL graph. Specifically, in [`_process_reqs`](https://github.com/vllm-project/vllm-ascend/blob/main/vllm_ascend/worker/model_runner_v1.py#L989), the variable `num_input_tokens` is always greater than or equal to `total_num_scheduled_tokens`, so any positions not explicitly cleared from a previous batch will persist and cause this sporadic error. BTW, in the original vLLM implementation, masks are constructed internally using other args, so these lingering values do not surface. However, on the Ascend platform—where split-fuse attention requires externally supplied masks—these residual indices become critical and lead to this elusive, hard-to-reproduce failure. The fix is to explicitly reset or zero out all unused entries in the `position` tensor before passing it to `GatherV3`, ensuring that every index lies within the valid range of the attention mask. Closes: vllm-project#1038 ### Does this PR introduce _any_ user-facing change? No Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
venus-taibai
pushed a commit
to venus-taibai/vllm-ascend
that referenced
this pull request
Sep 17, 2025
…f-bounds in GatherV3 (vllm-project#1416) Merge branch wengang/cherry-pick-1416 of git@code.alipay.com:Theta/vllm-ascend.git into dev-v0.9.1.0622 https://code.alipay.com/Theta/vllm-ascend/pull_requests/222 Reviewed-by: 子宏 <tanzhiqiang.tzq@antgroup.com> * [Bugfix] Reset all unused positions to prevent out-of-bounds in GatherV3 (vllm-project#1416)
chopper0126
pushed a commit
to chopper0126/vllm-ascend
that referenced
this pull request
Oct 16, 2025
…rV3 (vllm-project#1416) ### What this PR does / why we need it? Reset all unused positions in `NPUModelRunner` to prevent out-of-bounds asserts in the `GatherV3` operator. Currently, in [`get_splitfuse_attn_mask`](https://github.com/vllm-project/vllm-ascend/blob/main/vllm_ascend/attention/attention.py#L124), the `position` tensor may contain values that exceed the dimensions of the attention mask, triggering a `GatherV3` boundary check failure. These invalid indices originate from stale “dirty” entries left over in `position` due to padding logic in the ACL graph. Specifically, in [`_process_reqs`](https://github.com/vllm-project/vllm-ascend/blob/main/vllm_ascend/worker/model_runner_v1.py#L989), the variable `num_input_tokens` is always greater than or equal to `total_num_scheduled_tokens`, so any positions not explicitly cleared from a previous batch will persist and cause this sporadic error. BTW, in the original vLLM implementation, masks are constructed internally using other args, so these lingering values do not surface. However, on the Ascend platform—where split-fuse attention requires externally supplied masks—these residual indices become critical and lead to this elusive, hard-to-reproduce failure. The fix is to explicitly reset or zero out all unused entries in the `position` tensor before passing it to `GatherV3`, ensuring that every index lies within the valid range of the attention mask. Closes: vllm-project#1038 ### Does this PR introduce _any_ user-facing change? No Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
Angazenn
pushed a commit
to Angazenn/vllm-ascend
that referenced
this pull request
Oct 21, 2025
…rV3 (vllm-project#1416) ### What this PR does / why we need it? Reset all unused positions in `NPUModelRunner` to prevent out-of-bounds asserts in the `GatherV3` operator. Currently, in [`get_splitfuse_attn_mask`](https://github.com/vllm-project/vllm-ascend/blob/main/vllm_ascend/attention/attention.py#L124), the `position` tensor may contain values that exceed the dimensions of the attention mask, triggering a `GatherV3` boundary check failure. These invalid indices originate from stale “dirty” entries left over in `position` due to padding logic in the ACL graph. Specifically, in [`_process_reqs`](https://github.com/vllm-project/vllm-ascend/blob/main/vllm_ascend/worker/model_runner_v1.py#L989), the variable `num_input_tokens` is always greater than or equal to `total_num_scheduled_tokens`, so any positions not explicitly cleared from a previous batch will persist and cause this sporadic error. BTW, in the original vLLM implementation, masks are constructed internally using other args, so these lingering values do not surface. However, on the Ascend platform—where split-fuse attention requires externally supplied masks—these residual indices become critical and lead to this elusive, hard-to-reproduce failure. The fix is to explicitly reset or zero out all unused entries in the `position` tensor before passing it to `GatherV3`, ensuring that every index lies within the valid range of the attention mask. Closes: vllm-project#1038 ### Does this PR introduce _any_ user-facing change? No Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
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?
Reset all unused positions in
NPUModelRunnerto prevent out-of-bounds asserts in theGatherV3operator.Currently, in
get_splitfuse_attn_mask, thepositiontensor may contain values that exceed the dimensions of the attention mask, triggering aGatherV3boundary check failure. These invalid indices originate from stale “dirty” entries left over inpositiondue to padding logic in the ACL graph. Specifically, in_process_reqs, the variablenum_input_tokensis always greater than or equal tototal_num_scheduled_tokens, so any positions not explicitly cleared from a previous batch will persist and cause this sporadic error.BTW, in the original vLLM implementation, masks are constructed internally using other args, so these lingering values do not surface. However, on the Ascend platform—where split-fuse attention requires externally supplied masks—these residual indices become critical and lead to this elusive, hard-to-reproduce failure.
The fix is to explicitly reset or zero out all unused entries in the
positiontensor before passing it toGatherV3, ensuring that every index lies within the valid range of the attention mask.Closes: #1038
Does this PR introduce any user-facing change?
No
How was this patch tested?