Fix cutlass_blackwell_fmha_custom_op and add comprehensive FMHA tests #5108
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
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify project configuration. |
Contributor
|
@jsisometa has exported this pull request. If you are a Meta employee, you can view the originating Diff in D86583157. |
…pytorch#5108) Summary: X-link: facebookresearch/FBGEMM#2113 This diff fixes the cutlass_blackwell_fmha_custom_op.py to be fully functional and adds comprehensive testing for Blackwell FMHA (Fused Multi-Head Attention). ## Changes Made: ### 1. Fixed `cutlass_blackwell_fmha_custom_op.py` - Added missing parameters to `fmha_fwd`: `page_table`, `seqlen_k`, `window_size_left`, `window_size_right`, `bottom_right` - Added missing parameters to `fmha_bwd`: `softmax_scale`, `window_size_left`, `window_size_right`, `bottom_right`, `deterministic` - Fixed parameter type issues: `torch.ops.fbgemm.fmha_fwd/bwd` expect `int` and `bool` types, not `Optional[int]` or `Optional[bool]` - Added proper default value handling: - `window_size_left = -1` (default for no left window) - `window_size_right = -1` (default for no right window) - `bottom_right = True` (default) - `deterministic = False` (default) - Updated `_backward`, `_setup_context`, and wrapper functions to properly pass all parameters - The custom op now correctly wraps `torch.ops.fbgemm.fmha_fwd` and `torch.ops.fbgemm.fmha_bwd` ### 2. Created `blackwell_fmha.py` Test File - Structured following `blackwell_gdpa.py` as reference - Uses `cutlass_blackwell_fmha_custom_op` (Cutlass implementation) for forward and backward passes - Compares against `jagged_flash_attention_v2` (Triton JFA v2 implementation) - Tests BF16 dtype only (as specified) - Tests both forward outputs and backward gradients (dq, dk, dv) - Runs 10 random test configurations with varying batch sizes, sequence lengths, and number of heads - Uses `generate_jagged_data` utility for proper test data generation ### 3. Updated BUCK Dependencies - Changed from `//ads_mkl/ops:jfa` to `//ads_mkl/ops/triton:triton_jfa_v2` - Added `//ads_mkl/ops/utils:jfa_utils` for data generation utilities - Changed from `blackwell_attention_ops_gpu` to `blackwell_attention` to include Python bindings --- > Generated by [Confucius Code Assist (CCA)](https://www.internalfb.com/wiki/Confucius/Analect/Shared_Analects/Confucius_Code_Assist_(CCA)/) [Session](https://www.internalfb.com/confucius?session_id=96622022-bc27-11f0-bdba-7c8c09f29af2&tab=Chat), [Trace](https://www.internalfb.com/confucius?session_id=96622022-bc27-11f0-bdba-7c8c09f29af2&tab=Trace) Differential Revision: D86583157
jsisometa
force-pushed
the
export-D86583157
branch
from
d7b1d11 to
bed170c
Compare
Contributor
|
This pull request has been merged in 6350109. |
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:
This diff fixes the cutlass_blackwell_fmha_custom_op.py to be fully functional and adds comprehensive testing for Blackwell FMHA (Fused Multi-Head Attention).
Changes Made:
1. Fixed
cutlass_blackwell_fmha_custom_op.pyfmha_fwd:page_table,seqlen_k,window_size_left,window_size_right,bottom_rightfmha_bwd:softmax_scale,window_size_left,window_size_right,bottom_right,deterministictorch.ops.fbgemm.fmha_fwd/bwdexpectintandbooltypes, notOptional[int]orOptional[bool]window_size_left = -1(default for no left window)window_size_right = -1(default for no right window)bottom_right = True(default)deterministic = False(default)_backward,_setup_context, and wrapper functions to properly pass all parameterstorch.ops.fbgemm.fmha_fwdandtorch.ops.fbgemm.fmha_bwd2. Created
blackwell_fmha.pyTest Fileblackwell_gdpa.pyas referencecutlass_blackwell_fmha_custom_op(Cutlass implementation) for forward and backward passesjagged_flash_attention_v2(Triton JFA v2 implementation)generate_jagged_datautility for proper test data generation3. Updated BUCK Dependencies
//ads_mkl/ops:jfato//ads_mkl/ops/triton:triton_jfa_v2//ads_mkl/ops/utils:jfa_utilsfor data generation utilitiesblackwell_attention_ops_gputoblackwell_attentionto include Python bindingsDifferential Revision: D86583157