ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

CUDA: fix q_nope_absorbed prec for DS 2 Lite f16 (#13137)

Browse source
This commit is contained in:
Johannes Gäßler 2025-04-28 09:29:26 +02:00 • committed by GitHub
parent 85f36e5e71
commit 69699be48a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 5 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

4
ggml/src/ggml-cuda/ggml-cuda.cu
View file

@ -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) {