feat: scalar paged prefill fallback and decode causal fix
- Add scalar paged prefill kernel mirroring split-Q MMA indexing for sm<80 - Wire scalar path into dispatch_paged_prefill under ASTRAI_NO_MMA - Fix paged decode scalar causal mask dropping all kv>0 for decode
This commit is contained in:
@@ -8,6 +8,7 @@
|
||||
#include "attn_prefill_split_q.cuh"
|
||||
#include "attn_decode_split_kv.cuh"
|
||||
#include "attn_paged_decode_split_kv.cuh"
|
||||
#include "attn_paged_prefill_split_q.cuh"
|
||||
#ifndef ASTRAI_NO_MMA
|
||||
#include "attn_prefill_split_q_mma.cuh"
|
||||
#include "attn_decode_split_kv_mma.cuh"
|
||||
@@ -208,6 +209,16 @@ static inline void launch_paged_prefill_mma(PagedAttentionParams<bf16>& p) {
|
||||
}
|
||||
#endif
|
||||
|
||||
template <int HEAD_DIM, bool IsCausal, bool HasMask>
|
||||
static inline void launch_paged_prefill_scalar(PagedAttentionParams<bf16>& p) {
|
||||
constexpr int G = 8, ROWS = 32, P_BC = 32;
|
||||
int max_q_tiles = (p.max_q_len + ROWS - 1) / ROWS;
|
||||
dim3 grid(max_q_tiles, p.q_head, p.batch);
|
||||
dim3 block(G, ROWS);
|
||||
paged_attn_prefill_split_q_kernel<HEAD_DIM, G, ROWS, P_BC, IsCausal, HasMask>
|
||||
<<<grid, block>>>(p);
|
||||
}
|
||||
|
||||
template <int HEAD_DIM>
|
||||
static inline void dispatch_paged_prefill(PagedAttentionParams<bf16>& p) {
|
||||
bool is_causal = (p.causal_offset >= 0);
|
||||
@@ -215,5 +226,7 @@ static inline void dispatch_paged_prefill(PagedAttentionParams<bf16>& p) {
|
||||
|
||||
#ifndef ASTRAI_NO_MMA
|
||||
DISPATCH_CAUSAL_MASK(is_causal, has_mask, launch_paged_prefill_mma, HEAD_DIM, p);
|
||||
#else
|
||||
DISPATCH_CAUSAL_MASK(is_causal, has_mask, launch_paged_prefill_scalar, HEAD_DIM, p);
|
||||
#endif
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user