fix(rpa-v3): add sliding window mask to h64 kernel and attention_sink to h128 #1185
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
… to h128 Fixes vllm-project#1169 This PR fixes two issues in the ragged paged attention v3 kernels: 1. **kernel_hd64.py (h64)**: Added missing sliding window mask in the kernel. The original code only skipped fetching KV blocks outside the window but didn't apply token-level masking within partially-covered blocks. 2. **kernel.py (h128)**: Added attention_sink support following the same pattern as the h64 kernel. Attention sinks allow the model to "dump" attention to a virtual token that doesn't contribute to the output. Uses LEFT concatenation semantics where sink logits are prepended before softmax, then removed after. Changes: - kernel_hd64.py: Added `if sliding_window is not None` mask in flash_attention - kernel.py: Added attention_sink parameter to all functions (ref impl, kernel, prepare_inputs, validation, main function) - kernel.py: Initialize m_prev with sink values and l_prev with 1.0 for proper online softmax tracking across blocks when using attention_sink
erfanzar
requested review from
bythew3i,
kyuyeunk and
yaochengji
as code owners
Collaborator
|
Hi @erfanzar, thanks for the quick PR. Here are few comments:
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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Outdated suggestions cannot be applied.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
Fixes #1169
This PR fixes two issues in the ragged-paged attention v3 kernels:
kernel_hd64.py (h64): Added missing sliding window mask in the kernel. The
original code only skipped fetching KV blocks outside the window, but didn't apply
token-level masking within partially-covered blocks.
kernel.py (h128): Added attention_sink support following the same pattern as the
h64 kernel. Attention sinks allow the model to "dump" attention to a virtual token that
doesn't contribute to the output.
Changes