菜单
Optimize FlashAttention-style fused attention kernels in Triton for NVIDIA and AMD GPUs. Covers online softmax, tiled QK/AV GEMM, causal masking, and memory-efficient attention. Use when writing or optimizing self-attention, cross-attention, or any QKV attention kernel.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | flash-attention-kernel |
| description | Optimize FlashAttention-style fused attention kernels in Triton for NVIDIA and AMD GPUs. Covers online softmax, tiled QK/AV GEMM, causal masking, and memory-efficient attention. Use when writing or optimizing self-attention, cross-attention, or any QKV attention kernel. |
FlashAttention fuses the entire attention computation (Q@K^T -> softmax -> @V) into a single kernel, avoiding materialization of the full N x N attention matrix in HBM.
For typical LLM shapes, attention is IO-bound (dominated by HBM reads of Q/K/V). The key metrics are TFLOPS and memory savings (O(N) instead of O(N^2)).
The key insight: compute softmax incrementally as K/V tiles stream through:
m_i (row max) and l_i (exp sum) per query rowqk = Q @ K^T, update max, rescale accumulatorl_iThis avoids two passes over the attention matrix.
[BLOCK_M, D][BLOCK_N, D]QK = Q @ K^T and acc += softmax(QK) @ VFor causal (autoregressive) attention:
kv_end = min(N, (pid_m + 1) * BLOCK_M) for causalwhere(causal_mask, qk, -inf) for partial tiles| Parameter | Typical Values |
|---|---|
| BLOCK_M | 64, 128 |
| BLOCK_N | 32, 64 |
| D (head dim) | 64, 128 (constexpr) |
| num_warps | 4 (D<=64), 8 (D>64) |
| num_stages | 2 |
AMD: same structure but num_warps maps to wavefronts of 64.
python skills/kernels/flash-attention/test_flash_attention.py
torch.nn.functional.scaled_dot_product_attention4 * B * H * N * N * D / latency / 1e12 (2 GEMMs, fwd only)acc *= alpha[:, None] when max updatesoffs_m >= offs_n not offs_m > offs_ntl.trans on AMD: works but may need explicit .contiguous() on input KOptimize fused cross-entropy loss kernels in Triton for NVIDIA and AMD GPUs. Covers fused log-softmax + NLL, online log-sum-exp, and large vocabulary handling. Use when writing or optimizing cross-entropy, label-smoothed CE, or similar loss kernels.
Optimize dense matrix multiplication (GEMM) kernels in Triton for NVIDIA and AMD GPUs. Covers tiled blocking, L2 cache grouping, tensor core utilization, and dual-platform autotune. Use when writing or optimizing matmul, linear layers, or any GEMM-based kernel.
Optimize RMS Normalization kernels in Triton for NVIDIA and AMD GPUs. Covers row-parallel reduction, vectorized loads, warp shuffle patterns, and rsqrt usage. Use when writing or optimizing RMSNorm, LayerNorm, or similar normalization kernels.
Optimize Rotary Position Embedding (RoPE) kernels in Triton for NVIDIA and AMD GPUs. Covers interleaved/sequential layouts, sincos precomputation, and multi-row processing. Use when writing or optimizing RoPE, position embeddings, or rotary transformations.
Optimize fused softmax kernels in Triton for NVIDIA and AMD GPUs. Covers online max/sum reduction, multi-row processing, numerical stability, and warp-level patterns. Use when writing or optimizing softmax, log-softmax, or similar row-wise normalization kernels.
Chooses block/tile dimensions, warp counts, pipeline stages, and L2-aware block grouping for GPU kernels (Triton/CUDA/HIP). Use when autotuning matmul-like kernels, tuning occupancy, or aligning tile geometry with M/N/K and NVIDIA vs AMD thread models.