ViperEkura/AstrAI
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

refactor: adopt FA2-style KernelTraits + compile-time causal/mask dispatch

Browse Source 
- Introduce KernelTraits<HEAD_DIM, BC, WARPS, STAGES> compile-time config bundle, replacing scattered <KD, NC8, KT2, ...> template params
- Template all MMA and scalar kernels on IsCausal/HasMask bools to eliminate inner-loop runtime branches
- Dispatch to 4-path IsCausal/HasMask kernel variants at entry points based on p.causal_offset and p.use_mask
- Update standalone test files with new kernel signatures, add causal test cases
- Fix duplicate using bf16 in MMA kernels that include attn_mma_utils.cuh
This commit is contained in:
ViperEkura
2026-07-21 21:52:46 +08:00
parent ccf728a1b7
commit a01e8bbe98
13 changed files with 682 additions and 611 deletions
Download Patch File Download Diff File Expand all files Collapse all files
csrc/kernels/attn_decode.cu
+39 -22
View File
@@ -3,9 +3,26 @@
#ifndef ASTRAI_NO_MMA
#include "attn_decode_split_kv_mma.cuh"
template <int HEAD_DIM, int BC, int STAGES, bool IsCausal, bool HasMask>
static void launch_mma_decode_impl(AttentionParams<bf16>& p) {
using Traits = KernelTraits<HEAD_DIM, BC, 1, STAGES>;
int tiles_total = (p.kv_len + BC - 1) / BC;
p.num_splits = compute_num_splits(p.batch * p.kv_head, tiles_total);
alloc_split_partials(p);
attn_decode_split_kv_mma_kernel<Traits, IsCausal, HasMask><<<dim3(p.kv_head, p.batch, p.num_splits), 32>>>(p);
attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
}
template <int HEAD_DIM, int BC, bool IsCausal, bool HasMask>
static void launch_mma_decode(AttentionParams<bf16>& p) {
constexpr int STAGES = (HEAD_DIM <= 128) ? 2 : 1;
launch_mma_decode_impl<HEAD_DIM, BC, STAGES, IsCausal, HasMask>(p);
}
#endif
// Scalar fallback: one warp per query head, split-KV across grid.z.
template <int HEAD_DIM, bool IsCausal, bool HasMask>
static void launch_scalar_decode(AttentionParams<bf16>& p) {
int group_size = p.q_head / p.kv_head;
int chunks_total = (p.kv_len + DC_CHUNK - 1) / DC_CHUNK;
@@ -13,37 +30,38 @@ static void launch_scalar_decode(AttentionParams<bf16>& p) {
alloc_split_partials(p);
size_t smem = DC_CHUNK * p.head_dim * sizeof(bf16);
attn_decode_split_kv_kernel<<<dim3(p.batch * p.kv_head, 1, p.num_splits), dim3(32, group_size), smem>>>(p);
dim3 grid(p.batch * p.kv_head, 1, p.num_splits);
dim3 block(32, group_size);
attn_decode_split_kv_kernel<HEAD_DIM, IsCausal, HasMask><<<grid, block, smem>>>(p);
attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
}
#ifndef ASTRAI_NO_MMA
// MMA head-packing requires G <= 16 (BR=16 rows). sm_80+ tensor-core
// + cp.async wins even at G=1 (decode is memory-bound, not compute-bound).
// STAGES=2 (double-buffer) for D<=128 (smem 16 KB); STAGES=1 for D=256
// (double-buffer would be 32 KB, near the 48 KB static cap — keep single
// to preserve occupancy).
template <int HEAD_DIM, int BC, int STAGES = (HEAD_DIM <= 128) ? 2 : 1>
static void launch_mma_decode(AttentionParams<bf16>& p) {
int tiles_total = (p.kv_len + BC - 1) / BC;
p.num_splits = compute_num_splits(p.batch * p.kv_head, tiles_total);
alloc_split_partials(p);
attn_decode_split_kv_mma_kernel<HEAD_DIM, BC, STAGES><<<dim3(p.kv_head, p.batch, p.num_splits), 32>>>(p);
attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
}
#endif
template <int HEAD_DIM>
static void dispatch_decode(AttentionParams<bf16>& p) {
bool is_causal = (p.causal_offset >= 0);
bool has_mask = (p.use_mask && p.mask);
#ifndef ASTRAI_NO_MMA
int G = p.q_head / p.kv_head;
if (G >= 1 && G <= 16) {
launch_mma_decode<HEAD_DIM, 32>(p);
if (is_causal) {
if (has_mask) launch_mma_decode<HEAD_DIM, 32, true, true>(p);
else launch_mma_decode<HEAD_DIM, 32, true, false>(p);
} else {
if (has_mask) launch_mma_decode<HEAD_DIM, 32, false, true>(p);
else launch_mma_decode<HEAD_DIM, 32, false, false>(p);
}
return;
}
#endif
launch_scalar_decode(p);
if (is_causal) {
if (has_mask) launch_scalar_decode<HEAD_DIM, true, true>(p);
else launch_scalar_decode<HEAD_DIM, true, false>(p);
} else {
if (has_mask) launch_scalar_decode<HEAD_DIM, false, true>(p);
else launch_scalar_decode<HEAD_DIM, false, false>(p);
}
}
torch::Tensor attn_decode(
@@ -60,7 +78,6 @@ torch::Tensor attn_decode(
TORCH_CHECK(p.q_len == 1, "Q seq_len must be 1");
TORCH_CHECK(p.head_dim % 32 == 0, "head_dim must be multiple of 32");
// O matches Q's original layout
auto O = torch::empty_strided(q.sizes(), q.strides(), q.options());
auto O_view = (layout == 1) ? O.transpose(1, 2) : O;
p.o = (bf16*)O_view.data_ptr();