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

perf: optimize small-head causal prefill

Browse Source 
- map D=32 and D=64 causal prefill to BC=64 tiles

- add small-head correctness and benchmark coverage
This commit is contained in:
ViperEkura
2026-08-14 23:25:49 +08:00
parent 69f35c46e0
commit b5afe3d7a4
2 changed files with 34 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
csrc/kernels/attn_dispatchers.cuh
+24 -4
View File
@@ -59,14 +59,34 @@ inline int compute_num_splits(int base_blocks, int tiles_total,
// ======================================================================
#ifndef ASTRAI_NO_MMA
template <int BC_>
struct PrefillKernelConfig {
static constexpr int BC = BC_;
static constexpr int WARPS = 4;
static constexpr int STAGES = 2;
};
// Compile-time configuration map shared by contiguous and paged prefill.
// Unsupported head dimensions intentionally have no mapping.
template <int HEAD_DIM, bool IsCausal>
struct PrefillConfigMap;
template <> struct PrefillConfigMap<32, false> : PrefillKernelConfig<32> {};
template <> struct PrefillConfigMap<32, true> : PrefillKernelConfig<64> {};
template <> struct PrefillConfigMap<64, false> : PrefillKernelConfig<32> {};
template <> struct PrefillConfigMap<64, true> : PrefillKernelConfig<64> {};
template <> struct PrefillConfigMap<128, false> : PrefillKernelConfig<32> {};
template <> struct PrefillConfigMap<128, true> : PrefillKernelConfig<32> {};
template <> struct PrefillConfigMap<256, false> : PrefillKernelConfig<16> {};
template <> struct PrefillConfigMap<256, true> : PrefillKernelConfig<16> {};
template <typename KV>
struct PrefillLauncherMMA {
template <int HEAD_DIM, bool IsCausal, bool HasMask>
static void launch(AttentionParams<bf16>& p, cudaStream_t stream) {
constexpr int WARPS = 4;
constexpr int BC = (HEAD_DIM <= 128) ? 32 : 16;
using Traits = KernelTraits<HEAD_DIM, BC, WARPS, 2>;
constexpr int ROWS = Traits::BR * WARPS;
using Config = PrefillConfigMap<HEAD_DIM, IsCausal>;
using Traits = KernelTraits<HEAD_DIM, Config::BC, Config::WARPS, Config::STAGES>;
constexpr int ROWS = Traits::BR * Config::WARPS;
dim3 grid(KV::host_q_blocks(p, ROWS), p.q_head,
KV::kPaged ? 1 : p.batch);
dim3 block(Traits::NUM_THREADS);
csrc/tests/attn_test.cu
+10 -1
View File
@@ -118,7 +118,8 @@ static void bench_decode() {
printf("\n===== DECODE BENCH (warmup=%d iters=%d) =====\n", WARMUP, ITERS);
print_bench_header();
for (int ci = 0; ci < 6; ci++) {
int n = sizeof(cfgs) / sizeof(cfgs[0]);
for (int ci = 0; ci < n; ci++) {
int B = cfgs[ci][0], Hq = cfgs[ci][1], Hk = cfgs[ci][2];
int sl = cfgs[ci][3], D = cfgs[ci][4];
size_t nQ = (size_t)B * Hq * D;
@@ -229,6 +230,12 @@ static int run_prefill_test(int B, int Hq, int Hk, int ql, int kl, int D, int ca
static void bench_prefill() {
const int cfgs[][7] = {
{1,32,4,1024,1024,32,0},
{1,32,4,1024,1024,32,1},
{1,32,4,4096,4096,32,1},
{1,32,4,1024,1024,64,0},
{1,32,4,1024,1024,64,1},
{1,32,4,4096,4096,64,1},
{1,32,4,512,512,128,0},
{1,32,4,1024,1024,128,0},
{1,32,4,2048,2048,128,0},
@@ -324,7 +331,9 @@ int main() {
{
const int configs[][7] = {
{1,2,1,64,128,32,0}, // scalar fallback D=32
{1,4,2,256,256,32,1}, // causal D=32 dispatch
{1,2,1,64,128,64,0}, // tiny: B,Hq,Hk,q,kv,D,causal
{1,4,2,256,256,64,1}, // causal D=64 dispatch
{1,32,4,512,512,128,0}, // standard
{1,32,4,128,256,128,0}, // medium
{1,4,2,256,256,128,1}, // causal