Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

cuda : concat implementation for quantized types (#25303)

Browse Source 
* cuda : concat implementation for quantized types

* chore : apply am17an clever suggestion to shorten the code

---------

Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
This commit is contained in:
fairydreaming
2026-07-05 23:26:24 +08:00
committed by GitHub
co-authored by Stanisław Szymczyk
parent a4107133a6
commit 78d2f52468
2 changed files with 50 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/ggml-cuda.cu
+18 -6
View File
@@ -5387,12 +5387,24 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
ggml_type src1_type = op->src[1]->type;
return src0_type == src1_type &&
src0_type == op->type &&
!ggml_is_quantized(src0_type) &&
ggml_blck_size(src0_type) == 1 &&
(ggml_type_size(src0_type) == 1 ||
ggml_type_size(src0_type) == 2 ||
ggml_type_size(src0_type) == 4 ||
ggml_type_size(src0_type) == 8);
(
(
ggml_is_quantized(src0_type) &&
ggml_is_contiguous(op->src[0]) &&
ggml_is_contiguous(op->src[1]) &&
op->src[0]->ne[0] % ggml_blck_size(src0_type) == 0 &&
op->src[1]->ne[0] % ggml_blck_size(src0_type) == 0
) || (
!ggml_is_quantized(src0_type) &&
ggml_blck_size(src0_type) == 1 &&
(
ggml_type_size(src0_type) == 1 ||
ggml_type_size(src0_type) == 2 ||
ggml_type_size(src0_type) == 4 ||
ggml_type_size(src0_type) == 8
)
)
);
} break;
case GGML_OP_CONV_TRANSPOSE_1D:
{