forked from ccf-ai-infra/GPUCodeForces
104 lines
3.6 KiB
Python
104 lines
3.6 KiB
Python
# swiglu_cuda.py
|
|
import torch
|
|
import torch.nn as nn
|
|
from torch.utils.cpp_extension import load_inline
|
|
|
|
from swiglu_torch import feature_dim
|
|
|
|
|
|
assert (feature_dim / 2) % 2 == 0, "feature_dim/2 must be a multiple of 2 for float2 vectorization"
|
|
|
|
class ModelNew(nn.Module):
|
|
|
|
def __init__(self):
|
|
super().__init__()
|
|
self._compile_cuda_kernel()
|
|
|
|
def _compile_cuda_kernel(self):
|
|
cpp_source = """
|
|
#include <torch/extension.h>
|
|
#include <ATen/cuda/CUDAContext.h>
|
|
torch::Tensor swiglu_forward_cuda(torch::Tensor input);
|
|
"""
|
|
cuda_source = """
|
|
#include <cuda_runtime.h>
|
|
#include <cmath> // For expf
|
|
|
|
|
|
__global__ void swiglu_fused_vectorized_kernel(
|
|
const float* __restrict__ x,
|
|
float* __restrict__ y,
|
|
int feature_dim, // 原始输入的特征维度
|
|
int n_elements_out // 输出张量的元素总数
|
|
) {
|
|
const float2* x2 = reinterpret_cast<const float2*>(x);
|
|
float2* y2 = reinterpret_cast<float2*>(y);
|
|
|
|
int n_work_items = n_elements_out / 2;
|
|
int grid_stride = gridDim.x * blockDim.x;
|
|
|
|
for (int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
|
idx < n_work_items;
|
|
idx += grid_stride)
|
|
{
|
|
|
|
int feature_dim_out_f2 = (feature_dim / 2) / 2;
|
|
int feature_dim_in_f2 = feature_dim / 2;
|
|
|
|
int row = idx / feature_dim_out_f2;
|
|
int col_f2 = idx % feature_dim_out_f2;
|
|
|
|
int gate_idx_f2 = row * feature_dim_in_f2 + col_f2;
|
|
int act_idx_f2 = gate_idx_f2 + feature_dim_out_f2;
|
|
|
|
float2 gate_vec = x2[gate_idx_f2];
|
|
float2 act_vec = x2[act_idx_f2];
|
|
|
|
float sigmoid_gate_x = 1.0f / (1.0f + expf(-gate_vec.x));
|
|
float silu_out_x = gate_vec.x * sigmoid_gate_x;
|
|
|
|
float sigmoid_gate_y = 1.0f / (1.0f + expf(-gate_vec.y));
|
|
float silu_out_y = gate_vec.y * sigmoid_gate_y;
|
|
|
|
y2[idx] = make_float2(silu_out_x * act_vec.x, silu_out_y * act_vec.y);
|
|
}
|
|
}
|
|
|
|
torch::Tensor swiglu_forward_cuda(torch::Tensor input) {
|
|
input = input.contiguous();
|
|
TORCH_CHECK(input.size(-1) % 2 == 0, "Last dimension must be even for SwiGLU");
|
|
|
|
auto original_sizes = input.sizes().vec();
|
|
int feature_dim = original_sizes.back();
|
|
original_sizes.back() /= 2;
|
|
auto output = torch::empty(original_sizes, input.options());
|
|
|
|
int n_elements_out = output.numel();
|
|
TORCH_CHECK(n_elements_out % 2 == 0, "Output elements must be even for float2 vectorization");
|
|
|
|
const int block_size = 256;
|
|
const int n_work_items = n_elements_out / 2;
|
|
|
|
const int grid_size = (n_work_items + block_size - 1) / block_size;
|
|
|
|
swiglu_fused_vectorized_kernel<<<grid_size, block_size>>>(
|
|
input.data_ptr<float>(),
|
|
output.data_ptr<float>(),
|
|
feature_dim,
|
|
n_elements_out
|
|
);
|
|
return output;
|
|
}
|
|
"""
|
|
|
|
self.swiglu_op = load_inline(
|
|
name="swiglu_fused_vectorized_op_fixed",
|
|
cpp_sources=cpp_source,
|
|
cuda_sources=cuda_source,
|
|
functions=["swiglu_forward_cuda"],
|
|
extra_cuda_cflags=["-O3", "--use_fast_math"],
|
|
verbose=True
|
|
)
|
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
return self.swiglu_op.swiglu_forward_cuda(x) |