gpu
/
ChatRWKV-gradio
mirror of https://huggingface.co/spaces/BlinkDL/ChatRWKV-gradio


  
1

	
2

	
3

	
4

	
5

	
6

	
7

	
8

	
9

	
10

	
11

	
12

	
13

	
14

	
15

	
16

	
17

	
18

	
19

	
20

	
21

	
22

	
23

	
24

	
25

	
26

	
27

	
28

	
29

	
30

	
31

	
32

	
33

	
34

	
            #include <torch/extension.h>
#include "ATen/ATen.h"
#include <c10/cuda/CUDAGuard.h>
typedef at::BFloat16 bf16;
typedef at::Half fp16;
typedef float fp32;

void cuda_forward_bf16(int B, int T, int C, int H, float *state, bf16 *r, bf16 *k, bf16 *v, float *w, bf16 *u, bf16 *y);
void cuda_forward_fp16(int B, int T, int C, int H, float *state, fp16 *r, fp16 *k, fp16 *v, float *w, fp16 *u, fp16 *y);
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>());
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward_bf16", &forward_bf16, "rwkv5 forward_bf16");
    m.def("forward_fp16", &forward_fp16, "rwkv5 forward_fp16");
    m.def("forward_fp32", &forward_fp32, "rwkv5 forward_fp32");
}
TORCH_LIBRARY(rwkv5, m) {
    m.def("forward_bf16", forward_bf16);
    m.def("forward_fp16", forward_fp16);
    m.def("forward_fp32", forward_fp32);
}