菜单
5-stage kernel correctness verification protocol for Triton and CUDA kernels. Covers numerical correctness, dtype sensitivity, edge cases, determinism, and stress testing. Use when verifying kernel correctness, running validation, or setting up test harnesses.
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
Optimize 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 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.
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.
| name | kernel-verification |
| description | 5-stage kernel correctness verification protocol for Triton and CUDA kernels. Covers numerical correctness, dtype sensitivity, edge cases, determinism, and stress testing. Use when verifying kernel correctness, running validation, or setting up test harnesses. |
Every kernel optimization must pass correctness verification before performance is measured. This skill defines a 5-stage protocol adapted from AutoKernel.
Compare kernel output against PyTorch reference on standard shapes.
ref = pytorch_reference(inputs)
out = triton_kernel(inputs)
assert torch.allclose(out.float(), ref.float(), atol=atol, rtol=rtol)
| Dtype | atol | rtol | Notes |
|---|---|---|---|
| float32 | 1e-5 | 1e-5 | Strict |
| float16 | 1e-2 | 1e-2 | FP16 has limited precision |
| bfloat16 | 1e-1 | 1e-1 | BF16 has even less mantissa |
| fp8 | 5e-1 | 5e-1 | Very coarse |
Test all target dtypes (fp16, bf16, fp32) and mixed combinations. Many bugs only appear in specific precision modes.
Test boundary conditions:
Run the same input 10 times and verify bitwise identical output. Non-determinism indicates race conditions or uninitialized memory.
results = [triton_kernel(inputs) for _ in range(10)]
for r in results[1:]:
assert torch.equal(r, results[0]), "Non-deterministic!"
Run with large inputs and many iterations to catch:
When verifying a kernel:
def verify_kernel(kernel_fn, ref_fn, input_gen, n_shapes=5, n_dtypes=2, n_repeats=3):
for shape in generate_shapes(n_shapes):
for dtype in [torch.float16, torch.bfloat16][:n_dtypes]:
inputs = input_gen(shape, dtype)
ref = ref_fn(*inputs)
for _ in range(n_repeats):
out = kernel_fn(*inputs)
assert torch.allclose(out.float(), ref.float(), atol=get_atol(dtype), rtol=get_rtol(dtype))
return True