public_git
/
whisper.cpp
mirror of https://github.com/ggml-org/whisper.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

ggml : require contiguous src for ROLL on CUDA and Metal (llama/25928) 

ggml_roll only asserts nb[0] == ggml_type_size, so a permuted src is a
valid input, but the CUDA and Metal roll kernels index by ne alone and
never read the nb strides. A non-contiguous src therefore produced
silently wrong results. Neither backend declared a contiguity
requirement in supports_op, so the scheduler did not fall back to the
CPU implementation, which does handle strides correctly.

Add the requirement to both backends, matching the existing
GGML_OP_ROPE guard, and add a permuted test_roll case.
This commit is contained in:
Yash Raj Pandey 2026-08-10 08:01:44 -04:00 committed by Georgi Gerganov
parent f09a97cf6d
commit b3bc904638
2 changed files with 3 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@ -5185,7 +5185,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
return max_bias == 0.0f; return max_bias == 0.0f;
} }
case GGML_OP_ROLL: case GGML_OP_ROLL:
if(op->src[0]->type == GGML_TYPE_F32) { if(op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0])) {
return true; return true;
} }
return false; return false;

3
ggml/src/ggml-metal/ggml-metal-device.m
View File

@ -1268,8 +1268,9 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
case GGML_OP_ARGSORT: case GGML_OP_ARGSORT:
case GGML_OP_TOP_K: case GGML_OP_TOP_K:
case GGML_OP_ARANGE: case GGML_OP_ARANGE:
case GGML_OP_ROLL:
return true; return true;
case GGML_OP_ROLL:
return ggml_is_contiguous(op->src[0]);
case GGML_OP_FLASH_ATTN_EXT: case GGML_OP_FLASH_ATTN_EXT:
// for new head sizes, add checks here // for new head sizes, add checks here
if (op->src[0]->ne[0] != 32 && if (op->src[0]->ne[0] != 32 &&