EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

CUDA: add roll (#14919)

Browse Source 
* CUDA: add roll

* Make everything const, use __restrict__
This commit is contained in:
Aman Gupta
2025-07-29 14:45:18 +08:00
committed by GitHub
parent 8ad7b3e65b
commit 0a5036bee9
3 changed files with 81 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@@ -31,6 +31,7 @@
#include "ggml-cuda/pool2d.cuh"
#include "ggml-cuda/quantize.cuh"
#include "ggml-cuda/rope.cuh"
#include "ggml-cuda/roll.cuh"
#include "ggml-cuda/scale.cuh"
#include "ggml-cuda/softmax.cuh"
#include "ggml-cuda/ssm-conv.cuh"
@@ -2419,6 +2420,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
case GGML_OP_ROPE_BACK:
ggml_cuda_op_rope_back(ctx, dst);
break;
case GGML_OP_ROLL:
ggml_cuda_op_roll(ctx, dst);
break;
case GGML_OP_IM2COL:
ggml_cuda_op_im2col(ctx, dst);
break;
@@ -3411,6 +3415,11 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
memcpy(&max_bias, (const float *) op->op_params + 1, sizeof(float));
return max_bias == 0.0f;
}
case GGML_OP_ROLL:
if(op->src[0]->type == GGML_TYPE_F32) {
return true;
}
return false;
case GGML_OP_ROPE:
case GGML_OP_ROPE_BACK: {
return op->src[0]->nb[0] == ggml_type_size(op->src[0]->type) && ggml_is_contiguous_2(op->src[0]);