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

Merge branch 'master' into gg/flash-attn

Browse source
This commit is contained in:
Georgi Gerganov 2024-02-01 19:51:41 +02:00
commit 41d136b602
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
9 changed files with 400 additions and 17 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml-cuda.cu
View file

@ -9213,9 +9213,9 @@ static void ggml_cuda_op_dequantize_mul_mat_vec(
if (src1_convert_f16) {
src1_dfloat = src1_dfloat_a.alloc(ne00);
ggml_cpy_f32_f16_cuda((const char *) src1_ddf_i, (char *) src1_dfloat, ne00,
ne00, 1, sizeof(float), 0, 0,
ne00, 1, sizeof(half), 0, 0, 0, 0, 0, 0, stream);
const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
GGML_ASSERT(to_fp16_cuda != nullptr);
to_fp16_cuda(src1_ddf_i, src1_dfloat, ne00, stream);
}
#else
const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion