forked from ccf-ai-infra/GPUCodeForces
101 lines
3.3 KiB
Python
101 lines
3.3 KiB
Python
import torch
|
|
import torch.nn as nn
|
|
from torch.utils.cpp_extension import load_inline
|
|
|
|
|
|
class ModelNew(nn.Module):
|
|
def __init__(self, beta=1.0, threshold=20.0):
|
|
super().__init__()
|
|
self.beta = beta
|
|
self.threshold = threshold
|
|
self._compile_cuda_kernel()
|
|
|
|
def _compile_cuda_kernel(self):
|
|
cpp_source = """
|
|
#include <torch/extension.h>
|
|
torch::Tensor softplus_glu_cuda(torch::Tensor input, float beta, float threshold);
|
|
"""
|
|
|
|
cuda_source = """
|
|
#include <cuda_runtime.h>
|
|
|
|
__device__ __forceinline__ float softplus_f(float x, float beta, float threshold) {
|
|
float bx = beta * x;
|
|
if (bx > threshold) return x;
|
|
return (1.0f / beta) * log1pf(expf(bx));
|
|
}
|
|
|
|
__global__ void softplus_glu_vec4_kernel(
|
|
const float4* __restrict__ x,
|
|
float4* __restrict__ y,
|
|
int vec_dim_out,
|
|
int n_vec_out,
|
|
float beta,
|
|
float threshold)
|
|
{
|
|
int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
|
int stride = gridDim.x * blockDim.x;
|
|
|
|
for (int i = idx; i < n_vec_out; i += stride) {
|
|
int row = i / vec_dim_out;
|
|
int col = i % vec_dim_out;
|
|
|
|
int gate_idx = row * (2 * vec_dim_out) + col;
|
|
int act_idx = gate_idx + vec_dim_out;
|
|
|
|
float4 g = x[gate_idx];
|
|
float4 a = x[act_idx];
|
|
float4 out;
|
|
|
|
out.x = softplus_f(g.x, beta, threshold) * a.x;
|
|
out.y = softplus_f(g.y, beta, threshold) * a.y;
|
|
out.z = softplus_f(g.z, beta, threshold) * a.z;
|
|
out.w = softplus_f(g.w, beta, threshold) * a.w;
|
|
|
|
y[i] = out;
|
|
}
|
|
}
|
|
|
|
torch::Tensor softplus_glu_cuda(torch::Tensor input, float beta, float threshold) {
|
|
auto x_c = input.contiguous();
|
|
|
|
int last_dim = x_c.size(-1);
|
|
TORCH_CHECK(last_dim % 8 == 0, "Feature dim must be divisible by 8 for float4 optimization");
|
|
|
|
auto out_sizes = x_c.sizes().vec();
|
|
out_sizes.back() /= 2;
|
|
auto output = torch::empty(out_sizes, x_c.options());
|
|
|
|
int numel_out = output.numel();
|
|
int n_vec_out = numel_out / 4;
|
|
int vec_dim_out = out_sizes.back() / 4;
|
|
|
|
int threads = 256;
|
|
int blocks = (n_vec_out + threads - 1) / threads;
|
|
if (blocks > 65535) blocks = 65535;
|
|
if (blocks == 0) blocks = 1;
|
|
|
|
softplus_glu_vec4_kernel<<<blocks, threads>>>(
|
|
reinterpret_cast<const float4*>(x_c.data_ptr<float>()),
|
|
reinterpret_cast<float4*>(output.data_ptr<float>()),
|
|
vec_dim_out,
|
|
n_vec_out,
|
|
beta,
|
|
threshold
|
|
);
|
|
|
|
return output;
|
|
}
|
|
"""
|
|
|
|
self.op = load_inline(
|
|
name="softplus_glu_opt_vec4",
|
|
cpp_sources=cpp_source,
|
|
cuda_sources=cuda_source,
|
|
functions=["softplus_glu_cuda"],
|
|
extra_cuda_cflags=["-O3", "--use_fast_math"],
|
|
verbose=False
|
|
)
|
|
|
|
def forward(self, x):
|
|
return self.op.softplus_glu_cuda(x, self.beta, self.threshold) |