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

feat: SGLang-style paged attention kernels replace page-table path

Browse Source 
- PagedAttentionParams uses flat KV pool + req_to_token + kv_indptr/qo_indptr instead of page_table
- MMA split-KV decode and split-Q prefill kernels with indirect ragged-batch addressing
- Prefill kernel accepts 4D mask (causal-aware); decode kernel supports 2D mask
- CudaBackend is inference-only: kv_cache=None raises, no torch fallback
- benchmark.py: required --ckpt, --backend/--compare options
- Parallel build isolates build-temp/build-lib per subprocess
- Standalone test covers decode/prefill with mask, 27 cases pass
This commit is contained in:
ViperEkura
2026-08-01 15:41:25 +08:00
parent 9960f79920
commit 41dcf0feb9
17 changed files with 1683 additions and 535 deletions
Download Patch File Download Diff File Expand all files Collapse all files
csrc/kernels/attn_dispatchers.cuh
+35 -7
View File
@@ -12,6 +12,7 @@
#include "attn_prefill_split_q_mma.cuh"
#include "attn_decode_split_kv_mma.cuh"
#include "attn_paged_decode_split_kv_mma.cuh"
#include "attn_paged_prefill_split_q_mma.cuh"
#endif
// Cached SM count — cudaDeviceGetAttribute is a host-side call that was
@@ -145,18 +146,18 @@ static inline void dispatch_decode(AttentionParams<bf16>& p) {
}
// ======================================================================
// Paged Decode
// Paged Decode (SGLang-style: flat pool + req_to_token + kv_indptr)
// ======================================================================
#ifndef ASTRAI_NO_MMA
template <int HEAD_DIM, bool IsCausal, bool HasMask>
static inline void launch_paged_decode_mma(PagedAttentionParams<bf16>& p, int group_size) {
static inline void launch_paged_decode_mma(PagedAttentionParams<bf16>& p, int) {
int G = p.q_head / p.kv_head;
constexpr int MAX_G = 16;
constexpr int BC = 16;
int num_passes = (G + MAX_G - 1) / MAX_G;
int tiles_total = (p.kv_len + BC - 1) / BC;
p.num_splits = compute_num_splits(p.batch * p.kv_head, tiles_total, 2);
int tiles_total = (p.max_seq_len + BC - 1) / BC;
p.num_splits = compute_num_splits(p.batch * p.kv_head * num_passes, tiles_total, 2);
constexpr int STAGES = 2;
using Traits = KernelTraits<HEAD_DIM, BC, 1, STAGES>;
dim3 grid(p.kv_head * num_passes, p.batch, p.num_splits);
@@ -166,10 +167,10 @@ static inline void launch_paged_decode_mma(PagedAttentionParams<bf16>& p, int gr
template <int HEAD_DIM, bool IsCausal, bool HasMask>
static inline void launch_paged_decode_scalar(PagedAttentionParams<bf16>& p, int group_size) {
int chunks_total = (p.kv_len + PDC_CHUNK - 1) / PDC_CHUNK;
int chunks_total = (p.max_seq_len + PDC_CHUNK - 1) / PDC_CHUNK;
p.num_splits = compute_num_splits(p.batch * p.kv_head, chunks_total);
size_t smem = PDC_CHUNK * p.head_dim * sizeof(bf16);
int g = min(group_size, 32); // cap at 32 to respect 1024-thread limit
int g = min(group_size, 32);
dim3 grid(p.batch * p.kv_head, 1, p.num_splits);
dim3 block(32, g);
paged_attn_decode_split_kv_kernel<HEAD_DIM, IsCausal, HasMask><<<grid, block, smem>>>(p);
@@ -182,10 +183,37 @@ static inline void dispatch_paged_decode(PagedAttentionParams<bf16>& p) {
int group_size = p.q_head / p.kv_head;
#ifndef ASTRAI_NO_MMA
DISPATCH_CAUSAL_MASK(is_causal, has_mask, launch_paged_decode_mma, HEAD_DIM, p, group_size);
DISPATCH_CAUSAL_MASK(is_causal, has_mask, launch_paged_decode_mma, HEAD_DIM, p, 0);
#else
DISPATCH_CAUSAL_MASK(is_causal, has_mask, launch_paged_decode_scalar, HEAD_DIM, p, group_size);
#endif
paged_attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
}
// ======================================================================
// Paged Prefill (SGLang-style: flat pool + ragged batch)
// ======================================================================
#ifndef ASTRAI_NO_MMA
template <int HEAD_DIM, bool IsCausal, bool HasMask>
static inline void launch_paged_prefill_mma(PagedAttentionParams<bf16>& p) {
constexpr int WARPS = 4;
constexpr int BC = (HEAD_DIM <= 128) ? 32 : 16;
using Traits = KernelTraits<HEAD_DIM, BC, WARPS, 2>;
int max_q_tiles = (p.max_q_len + Traits::BR * WARPS - 1) / (Traits::BR * WARPS);
dim3 grid(max_q_tiles, p.q_head, p.batch);
dim3 block(Traits::NUM_THREADS);
paged_attn_prefill_split_q_mma_kernel<Traits, IsCausal, HasMask><<<grid, block>>>(p);
}
#endif
template <int HEAD_DIM>
static inline void dispatch_paged_prefill(PagedAttentionParams<bf16>& p) {
bool is_causal = (p.causal_offset >= 0);
bool has_mask = (p.use_mask && p.mask);
#ifndef ASTRAI_NO_MMA
DISPATCH_CAUSAL_MASK(is_causal, has_mask, launch_paged_prefill_mma, HEAD_DIM, p);
#endif
}