theluyuan/RWKV-Runner
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

upgrade to rwkv 0.8.20

Browse Source
This commit is contained in:
josc146
2023-11-03 23:27:14 +08:00
parent 35e92d2aef
commit 1f81a1e5a8
6 changed files with 188 additions and 359 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
backend-python/rwkv_pip/cuda/rwkv5_op.cpp vendored
View File

@@ -1,5 +1,6 @@
#include <torch/extension.h>
#include "ATen/ATen.h"
#include <c10/cuda/CUDAGuard.h>
typedef at::BFloat16 bf16;
typedef at::Half fp16;
typedef float fp32;
@@ -9,12 +10,15 @@ void cuda_forward_fp16(int B, int T, int C, int H, float *state, fp16 *r, fp16 *
void cuda_forward_fp32(int B, int T, int C, int H, float *state, fp32 *r, fp32 *k, fp32 *v, float *w, fp32 *u, fp32 *y);
void forward_bf16(int64_t B, int64_t T, int64_t C, int64_t H, torch::Tensor &state, torch::Tensor &r, torch::Tensor &k, torch::Tensor &v, torch::Tensor &w, torch::Tensor &u, torch::Tensor &y) {
const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
cuda_forward_bf16(B, T, C, H, state.data_ptr<float>(), r.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), w.data_ptr<float>(), u.data_ptr<bf16>(), y.data_ptr<bf16>());
}
void forward_fp16(int64_t B, int64_t T, int64_t C, int64_t H, torch::Tensor &state, torch::Tensor &r, torch::Tensor &k, torch::Tensor &v, torch::Tensor &w, torch::Tensor &u, torch::Tensor &y) {
const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
cuda_forward_fp16(B, T, C, H, state.data_ptr<float>(), r.data_ptr<fp16>(), k.data_ptr<fp16>(), v.data_ptr<fp16>(), w.data_ptr<float>(), u.data_ptr<fp16>(), y.data_ptr<fp16>());
}
void forward_fp32(int64_t B, int64_t T, int64_t C, int64_t H, torch::Tensor &state, torch::Tensor &r, torch::Tensor &k, torch::Tensor &v, torch::Tensor &w, torch::Tensor &u, torch::Tensor &y) {
const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
cuda_forward_fp32(B, T, C, H, state.data_ptr<float>(), r.data_ptr<fp32>(), k.data_ptr<fp32>(), v.data_ptr<fp32>(), w.data_ptr<float>(), u.data_ptr<fp32>(), y.data_ptr<fp32>());
}