Rafail Giavrimis
687e778927
* CUDA: fuse rms_norm + mul + rope (+ view + set_rows) * tests: add broadcast weight case to rms_norm_mul_rope * CUDA: check memory ranges before rms_norm rope fusion * CUDA: check memory ranges in rope set_rows fusion
12 lines
486 B
Plaintext
12 lines
486 B
Plaintext
#include "common.cuh"
|
|
|
|
#define CUDA_ROPE_BLOCK_SIZE 256
|
|
|
|
void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
|
|
|
|
void ggml_cuda_op_rope_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
|
|
|
|
void ggml_cuda_op_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * set_rows);
|
|
|
|
void ggml_cuda_op_rms_norm_mul_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * rms_norm, ggml_tensor * mul, ggml_tensor * rope, ggml_tensor * set_rows);
|