flash-attn-res 0.1.8

Attention Residuals (AttnRes) kernels

Flash Attention Residuals

1.4x faster inference/training vs. torch.compile impl. of the paper’s two-phase batched attn. + online softmax

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

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

Reference: https://arxiv.org/abs/2603.15031 (Kimi Team, MoonshotAI, 2026)

Credits:

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

Install

pip install flash-attn-res

Usage

This package contains Triton kernels, triton_op wrappers compatible with torch.compile, and an experimental high-performance Block AttenRes autograd implementation. See src and examples folders.

Roadmap:

• More robust autograd impl.
• Precision tuning
• Mixed FP16 and BF16 and store quantization scale
• Stochastic rounding
• CuTE, CUDA, and other DSLs implementation

Development Notes:

• 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"