char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Rafail GiavrimisCUDA: fuse rms_norm + mul + rope (+ view + set_rows) (#26767)687e77892

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
486 B11 linesraw
1#include "common.cuh"
2
3#define CUDA_ROPE_BLOCK_SIZE 256
4
5void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
6
7void ggml_cuda_op_rope_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
8
9void ggml_cuda_op_rope_fused(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * set_rows);
10
11void 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);