Skip to content

cuda : fix multi-seq, quantized FA #14820

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
wants to merge 1 commit into from
Closed

cuda : fix multi-seq, quantized FA #14820

Show file tree
Hide file tree
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: 14 additions & 6 deletions ggml/src/ggml-cuda/fattn-common.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -745,10 +745,14 @@ void launch_fattn(
size_t nb23 = V ? V->nb[3] : nb13;

if (need_f16_K && K->type != GGML_TYPE_F16) {
GGML_ASSERT(ggml_is_contiguously_allocated(K));
K_f16.alloc(ggml_nelements(K));
const int64_t n_seq = K->ne[3];
const int64_t n_eps = (K->nb[3]/ggml_type_size(K->type))*ggml_blck_size(K->type); // elements per sequence

K_f16.alloc(n_seq*n_eps);
to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(K->type);
to_fp16(K_data, K_f16.ptr, ggml_nelements(K), main_stream);
for (int s = 0; s < n_seq; ++s) {
to_fp16(K_data + s*K->nb[3], K_f16.ptr + s*n_eps, n_eps, main_stream);
}
K_data = (char *) K_f16.ptr;

const size_t bs = ggml_blck_size(K->type);
Expand All @@ -760,10 +764,14 @@ void launch_fattn(
}

if (V && need_f16_V && V->type != GGML_TYPE_F16) {
GGML_ASSERT(ggml_is_contiguously_allocated(V));
V_f16.alloc(ggml_nelements(V));
const int64_t n_seq = V->ne[3];
const int64_t n_eps = (V->nb[3]/ggml_type_size(V->type))*ggml_blck_size(V->type);

V_f16.alloc(n_seq*n_eps);
to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(V->type);
to_fp16(V_data, V_f16.ptr, ggml_nelements(V), main_stream);
for (int s = 0; s < n_seq; ++s) {
to_fp16(V_data + s*V->nb[3], V_f16.ptr + s*n_eps, n_eps, main_stream);
}
V_data = (char *) V_f16.ptr;

const size_t bs = ggml_blck_size(V->type);
Expand Down
2 changes: 2 additions & 0 deletions tests/test-backend-ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5525,6 +5525,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_timestep_embedding());
test_cases.emplace_back(new test_leaky_relu());

test_cases.emplace_back(new test_flash_attn_ext(128, 128, 4, {1, 3}, 512, 128, true, 0.0f, 0.0f, GGML_PREC_DEFAULT, GGML_TYPE_Q8_0));

for (int hsk : { 64, 80, 128, 192, 256, 576 }) {
for (int hsv : { 64, 80, 128, 192, 256, 512 }) {
if (hsk != 192 && hsk != 576 && hsk != hsv) continue;
Expand Down
Loading