Skip to content

CUDA: fix q_nope_absorbed precision for Deepseek 2 Lite f16 #13137

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

Merged
merged 1 commit into from
Apr 28, 2025
Merged

CUDA: fix q_nope_absorbed precision for Deepseek 2 Lite f16 #13137

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
4 changes: 2 additions & 2 deletions ggml/include/ggml.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -393,8 +393,8 @@ extern "C" {

// precision
enum ggml_prec {
GGML_PREC_DEFAULT,
GGML_PREC_F32,
GGML_PREC_DEFAULT = 0, // stored as ggml_tensor.op_params, 0 by default
GGML_PREC_F32 = 10,
};

// model file types
Expand Down
4 changes: 2 additions & 2 deletions ggml/src/ggml-cuda/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1935,8 +1935,8 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
ggml_cuda_mul_mat_vec(ctx, src0, src1, nullptr, dst);
} else if (!split && use_mul_mat_vec_q) {
ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
} else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
&& !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
} else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
dst->op_params[0] == GGML_PREC_DEFAULT && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
// general KQ + KQV multi-batch without FlashAttention
ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
} else if (use_mul_mat_vec) {
Expand Down
1 change: 1 addition & 0 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10149,6 +10149,7 @@ struct llm_build_deepseek2 : public llm_graph_context {

// {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
ggml_mul_mat_set_prec(q_nope_absorbed, GGML_PREC_F32);
cb(q_nope_absorbed, "q_nope_absorbed", il);

// {kv_lora_rank, n_head, n_tokens}
Expand Down
Loading