flash-attn-res 0.1.2

Attention Residuals (AttnRes) kernels

Flash Attention Residuals

1.4x faster inference/training vs. an optimized torch.compile impl. of the paper’s two-phase batched attention with online softmax

20% reduction in training memory (without activation checkpointing)*

*Benchmarked on H100. Dependent on problem size and setup.

Credits:

Thanks to Mohamed Osman (https://github.com/spaghettiSystems) and Cartesia for advising on and supporting the development of this kernel.

Roadmap:

• Proper backward eval
• Implement in CuTE and CUDA
• Tune precision
• Mixed FP16 and BF16 and store quantization scale
• Stochastic rounding
• Make into Python package

Insights:

• Normalizing in phase 1 keeps outputs bounded (convex combination of values) so bf16 error doesn't scale with softmax flatness. Phase 2 computes in fp32, and the reduction algebra matches split-KV Flash Attention.
• Certain dimensions, especially NUM_QUERIES_PER_BLOCK, are small so semi-elementwise (B, T) kernel with static_range is better than doing tl.dot
• Kernel is memory bound and doing semi-elementwise allows for kernel fusion
• NUM_SOURCE_BLOCKS and NUM_QUERIES_PER_BLOCK should be autotuning keys, unlike with torch.compile, which allows for faster kernels
• Small NUM_QUERIES_PER_BLOCK so eviction_policy should be "evict_last"