[Metal] Add Flash Attention VJP for training #2995
Open
+3,613
−67
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
Author
|
Notes: I'm working on https://github.com/mlx-node/mlx-node and trying to port some features in trl. |
mlx/backend/metal/kernels/steel/attn/kernels/steel_attention_vjp_dkv.h
Outdated
Show resolved
Hide resolved
Brooooooklyn
force-pushed
the
flash-attn
branch
5 times, most recently
from
568ff36 to
26b5857
Compare
Brooooooklyn
force-pushed
the
flash-attn
branch
7 times, most recently
from
d9089ef to
dd8daf1
Compare
Implement fused backward pass (VJP) for scaled_dot_product_attention on Metal GPU, enabling efficient training without falling back to unfused attention. - **dQ Kernel** (steel_attention_vjp_dq.h): Computes query gradients - Outer loop over KV blocks, inner accumulation for dQ - Uses log2 domain for numerical stability - **dK/dV Kernel** (steel_attention_vjp_dkv.h): Computes key/value gradients - K-row ownership model eliminates atomic operations - Each simdgroup owns exclusive K rows to prevent races - Optimized path for short sequences (L ≤ 8) - Uses shared memory for efficient reduction - Float32 accumulators for half/bfloat16 precision - Logsumexp caching from forward pass - Proper GQA (grouped query attention) support - Causal mask support - Comprehensive test coverage for all code paths - No gradient support for mask or attention sinks (falls back to unfused) - Requires logsumexp from forward pass (training mode only) - Head dimension D=256 not supported in vector VJP (threadgroup memory) Co-Authored-By: Claude <[email protected]>
Brooooooklyn
force-pushed
the
flash-attn
branch
from
dd8daf1 to
5c78507
Compare
Author
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.
Summary
Implements fused backward pass (VJP) for
scaled_dot_product_attentionon Metal GPU. This enables efficient gradient computation during training without falling back to unfused (decomposed) attention operations.Changes
New Files
mlx/backend/metal/kernels/sdpa_vector_vjp.h- Vector VJP kernel for short sequences (L ≤ 8)mlx/backend/metal/kernels/steel/attn/kernels/steel_attention_vjp_dq.h- STEEL dQ gradient kernelmlx/backend/metal/kernels/steel/attn/kernels/steel_attention_vjp_dkv.h- STEEL dK/dV gradient kernelModified Files
mlx/backend/metal/scaled_dot_product_attention.cpp- VJP dispatch logic (+840 lines)mlx/fast.cpp/mlx/fast_primitives.h- Logsumexp caching, VJP routingpython/tests/test_fast_sdpa.py- Comprehensive VJP tests (+220 lines)Implementation Notes
Uses a two-kernel approach to avoid atomic operations:
dQ kernel (
steel_attention_vjp_dq.h):dK/dV kernel (
steel_attention_vjp_dkv.h):Vector VJP (
sdpa_vector_vjp.h):Key Features
Limitations
Test Plan
test_sdpa_gradpassestest_sdpa_grad_vector_path- short sequences (L=1,4,7,8)test_sdpa_grad_steel_path- longer sequences (L=16,32,128,256)test_sdpa_grad_head_dims- head dimensions (D=32,64,96,128)test_sdpa_grad_gqa- GQA configurations (4:1, 8:1, 16:1, MHA)test_sdpa_grad_dtypes- float16, bfloat16, float32test_sdpa_grad_edge_cases- L=1, non-power-of-2, large batch, qL≠kvLAll 21 SDPA tests pass (1 skipped for unrelated disabled feature).