Repository navigation
Conversation
McZyWu
changed the title
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Withdrawn after a direct comparison with the existing Ascend kernel
The original premise was incorrect for the installed Ascend Triton compiler: the existing kernel already compiles and produces correct output for 3, 5, 6, 10 and 20 heads. Ten original-kernel NPU configurations, including non-power-of-two query/KV heads, sliced strides, full/partial RoPE and dynamic graph inputs, pass an independent reference check.
A direct stacked-kernel A/B also shows the proposed masked specialization is slightly slower. Seven alternating trials, 100 operations captured per graph, 3 replays per measurement, on A3 with another model resident:
Outputs match bitwise. The earlier comparison of five per-layer calls versus one stacked call measured the benefit of stacking; it did not establish any benefit over the existing stacked kernel. It must not be used to justify this replacement.
This PR is closed and its merge is reverted from the local performance integration. Framework PR31 keeps the separately validated 2D-to-3D input adaptation and does not require this PR. No TPOT benefit is attributed to kernel PR13.