mla ps support paged 64 and 3buffer layout for ds3.2 #1917
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
Contributor
There was a problem hiding this comment.
Pull request overview
This pull request adds support for paged 64 and 3-buffer layout for MLA (Multi-Head Latent Attention) operations in DeepSpeed 3.2. The PR introduces new parameters (page_size and nhead_kv) across the entire call stack to enable flexible paging strategies and support for a specialized 3-buffer KV cache layout with FP8 quantization.
Changes:
- Added
page_sizeandnhead_kvparameters throughout the MLA API stack (Python, C++, and CUDA) - Introduced support for "byte" datatype to handle 3-buffer layout with separate nope/scale/rope buffers
- Added new kernel assembly files (
mla.co,mla_page64.co) for optimized page size 64 support - Updated metadata generation logic to properly calculate KV offsets for paged layouts
- Added test infrastructure for 3-buffer layout with helper functions for initialization and validation
Reviewed changes
Copilot reviewed 13 out of 15 changed files in this pull request and generated 11 comments.
Show a summary per file
| File | Description |
|---|---|
| op_tests/test_mla_sparse.py | Added kv_last_page_lens and page_size/nhead_kv parameters to function calls |
| op_tests/test_mla_persistent.py | Major additions: 3-buffer KV cache helper functions, new test cases for 3-buffer layout |
| op_tests/test_mla.py | Updated function signatures to include new page_size and nhead_kv parameters |
| hsa/gfx942/mla/mla_page64.co | New assembly kernel binary for page size 64 support |
| hsa/gfx942/mla/mla_asm.csv | Added kernel configuration entry for "byte" datatype |
| hsa/gfx942/mla/mla.co | New assembly kernel binary for 3-buffer layout support |
| csrc/py_itfs_cu/asm_mla.cu | Extended to handle "byte" datatype and removed page_size extraction from KV tensor |
| csrc/kernels/mla/metadata/v1_comm.cuh | Code formatting improvements and copyright update |
| csrc/kernels/mla/metadata/v1_2_device.cuh | Added paged layout support with proper kv_offset calculation |
| csrc/kernels/mla/metadata.cu | Added kv_last_page_lens and page_size parameters to metadata generation |
| csrc/include/rocm_ops.hpp | Updated Python bindings with new parameters |
| csrc/include/mla.h | Updated function signatures with new parameters |
| csrc/include/attention_asm_mla.h | Updated function signatures with new parameters |
| aiter/ops/attention.py | Added default values for backward compatibility |
| aiter/mla.py | Added page_size and nhead_kv parameters with default values |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
minmengdie
force-pushed
the
mmd/dev/mla_ps_paged3buffer
branch
from
2dc00f6 to
1c7550d
Compare
ruanjm
reviewed
Jan 29, 2026
shengnxu
approved these changes
Jan 29, 2026
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
support for paged 64 and 3-buffer layout for MLA (Multi-Head Latent Attention) operations in DeepSpeed 3.2
Technical Details
Added page_size and nhead_kv parameters throughout the MLA API stack (Python, C++, and CUDA)
Introduced support for "byte" datatype to handle 3-buffer layout with separate nope/scale/rope buffers
Added new kernel assembly files (mla_a16w8_qh16_m16x4_n16x1_coex0_mask1_ps_page64_ds32.co) for optimized page size 64 support
Updated metadata generation logic to properly calculate KV offsets for paged layouts
Added test infrastructure for 3-buffer layout with helper functions for initialization and validation
Test Plan
python3 op_tests/test_mla_persistent.py -blk=64 -d=bf16 -kvd=fp8 -pl=3BUFFER
Test Result
Submission Checklist