Skip to content

cuda: reserve the f16 K/V scratch for the short-query MMA route (#346) #350

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
professorpalmer wants to merge 1 commit into from
Closed

cuda: reserve the f16 K/V scratch for the short-query MMA route (#346) #350

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 18 additions & 2 deletions ggml/src/ggml-cuda/fattn.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,13 +5,22 @@
#include "fattn-vec.cuh"
#include "fattn.cuh"

// 3-4 queries with a quantized V take the 64-column f16 tiles (#189): launch_fattn converts K and V to f16 in the
// op's extra buffer, also when an in-place quantized K/V kernel exists for the head size. ggml_cuda_flash_attn_ext_get_alloc_size
// must reserve that buffer for these ops.
static bool ggml_cuda_fattn_mma_short_query_f16(const int cc, const ggml_tensor * dst) {
const ggml_tensor * Q = dst->src[0];
const ggml_tensor * V = dst->src[2];
return turing_mma_available(cc) && ggml_is_quantized(V->type) && Q->ne[1] > 2 && Q->ne[1] <= 4;
}

template <int DKQ, int DV, int ncols2, ggml_type type_KV = GGML_TYPE_F16>
static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
const ggml_tensor * Q = dst->src[0];
const ggml_tensor * V = dst->src[2];

if (turing_mma_available(cc) && ggml_is_quantized(V->type) && Q->ne[1] > 2 && Q->ne[1] <= 4) {
if (ggml_cuda_fattn_mma_short_query_f16(cc, dst)) {
ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 64/ncols2, ncols2>(ctx, dst);
return;
}
Expand Down Expand Up @@ -579,7 +588,8 @@ size_t ggml_cuda_flash_attn_ext_get_alloc_size(int device, const ggml_tensor * d

switch (kernel) {
case BEST_FATTN_KERNEL_MMA_F16:
if (ggml_cuda_fattn_mma_kv_native_supported(dst)) {
if (ggml_cuda_fattn_mma_kv_native_supported(dst) &&
!ggml_cuda_fattn_mma_short_query_f16(ggml_cuda_info().devices[device].cc, dst)) {
// In-place quantized K/V kernel: nothing to reserve beyond dst.
break;
}
Expand Down Expand Up @@ -617,6 +627,12 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
ggml_cuda_flash_attn_ext_vec(ctx, dst);
break;
case BEST_FATTN_KERNEL_MMA_F16:
if (dst->buffer && ggml_cuda_fattn_mma_short_query_f16(ggml_cuda_info().devices[ggml_cuda_get_device()].cc, dst)) {
// the f16 K/V this route writes after dst must fit in what the buffer reserved for dst
const ggml_cuda_flash_attn_ext_f16_extra_data need = ggml_cuda_flash_attn_ext_get_f16_extra_data(dst, true, true);
GGML_ASSERT(need.end - (uintptr_t) dst->data <= ggml_backend_buffer_get_alloc_size(dst->buffer, dst) &&
"flash attention: f16 K/V scratch not reserved for the short-query route");
}
ggml_cuda_flash_attn_ext_mma_f16(ctx, dst);
break;
}
Expand Down