Enable gptoss_sink #1753
Open
Enable gptoss_sink #1753
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
Signed-off-by: Linjun-AMD <Jun.Lin@amd.com>
Contributor
There was a problem hiding this comment.
Pull request overview
This PR adds support for a sink_ptr parameter across the multi-head attention (MHA) forward pass implementations. The sink_ptr enables the "gptoss_sink" feature, which is a mechanism for attention sink tokens in transformer models.
Key changes:
- Added
sink_ptrparameter to all MHA forward function signatures (regular, varlen, and batch prefill variants) - Added validation logic to ensure
sink_ptrmatches device and shape requirements, with automatic dtype conversion to float32 - Propagated
sink_ptrthrough the call stack from Python interfaces to C++/CUDA implementations
Reviewed changes
Copilot reviewed 10 out of 10 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
| csrc/py_itfs_cu/asm_mha_varlen_fwd.cu | Added nullptr sink_ptr argument to mha_fwd_args constructor |
| csrc/py_itfs_cu/asm_mha_fwd.cu | Added nullptr sink_ptr argument to mha_fwd_args constructor |
| csrc/py_itfs_ck/mha_varlen_fwd_kernels.cu | Added sink_ptr parameter handling and validation in varlen forward kernels |
| csrc/py_itfs_ck/mha_fwd_kernels.cu | Added sink_ptr parameter to standard MHA forward kernels |
| csrc/py_itfs_ck/mha_batch_prefill_kernels.cu | Added sink_ptr parameter to batch prefill kernels |
| csrc/include/torch/mha_varlen_fwd.h | Updated header to include sink_ptr parameter in function signature |
| csrc/include/torch/mha_fwd.h | Updated header to include sink_ptr parameter in function signature |
| csrc/include/torch/mha_batch_prefill.h | Updated header to include sink_ptr parameter in function signature |
| csrc/include/rocm_ops.hpp | Added sink_ptr pybind argument definitions for all MHA variants |
| aiter/ops/mha.py | Added sink_ptr parameter to all Python MHA functions with device/shape validation |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Signed-off-by: Linjun-AMD <Jun.Lin@amd.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Motivation
This PR adds support for a sink_ptr parameter across the multi-head attention (MHA) forward pass implementations. The sink_ptr enables the "gptoss_sink" feature, which is a mechanism for attention sink tokens in transformer models.
Technical Details
Added sink_ptr parameter to all MHA forward function signatures (regular, varlen, and batch prefill variants)
Added validation logic to ensure sink_ptr matches device and shape requirements, with automatic dtype conversion to float32
Propagated sink_ptr through the call stack from Python interfaces to C++/CUDA implementations
Test Plan
Test in ck repo
Test Result
local test passd
Submission Checklist