You write custom CUDA kernels to replace the pytorch operators in the given architecture to get speedups.  
  
You have complete freedom to choose the set of operators you want to replace. You may make the decision to replace some operators with custom CUDA kernels and leave others unchanged. You may replace multiple operators with custom implementations, consider operator fusion opportunities (combining multiple operators into a single kernel, for example, combining matmul+relu), or algorithmic changes (such as online softmax). You are only limited by your imagination.  
  
Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is:   
  
Given hardmish Architecture (Base PyTorch Implementation)
python
运行
import torch
import torch.nn as nn

class Model(nn.Module):
    def __init__(self, in_features):
        super().__init__()
        torch.manual_seed(42)
        self.linear = nn.Linear(in_features, in_features)

    def forward(self, x):
        x = self.linear(x)
        # 分段计算HardMish
        return torch.where(
            x <= -3,
            torch.zeros_like(x),
            torch.where(
                x >= 3,
                x,
                x * (x + 3) / 6
            )
        )

def get_inputs():
    batch_size = 8192
    in_features = 256
    x = torch.randn(batch_size, in_features)
    return [x]

def get_init_inputs():
    return [256] 
New Architecture with Custom CUDA Kernels (hardmish Optimization)
python
运行
import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline

hardmish_source = """
#include <torch/extension.h>
#include <cuda_runtime.h>

__device__ __forceinline__ float hardmish_impl(float x) {
    if (x <= -3.0f) {
        return 0.0f;
    } else if (x >= 3.0f) {
        return x;
    } else {
        return x * (x + 3.0f) / 6.0f;
    }
}

__global__ void hardmish_kernel(
    const float* __restrict__ input,
    float* __restrict__ output,
    int size
) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx >= size) return;
    output[idx] = hardmish_impl(input[idx]);
}

torch::Tensor hardmish_cuda(torch::Tensor input) {
    TORCH_CHECK(input.is_cuda(), "input must be CUDA tensor");
    TORCH_CHECK(input.dtype() == torch::kFloat32, "input must be float32");

    input = input.contiguous();
    int total_size = input.numel();
    auto output = torch::empty_like(input);

    const int threads_per_block = 256;
    const int blocks = (total_size + threads_per_block - 1) / threads_per_block;

    hardmish_kernel<<<blocks, threads_per_block>>>(
        input.data_ptr<float>(),
        output.data_ptr<float>(),
        total_size
    );

    cudaError_t err = cudaGetLastError();
    if (err != cudaSuccess) {
        throw std::runtime_error("CUDA error: " + std::string(cudaGetErrorString(err)));
    }
    return output;
}
"""

hardmish_cpp_source = """
torch::Tensor hardmish_cuda(torch::Tensor input);
"""

hardmish_module = load_inline(
    name="hardmish_final",
    cpp_sources=hardmish_cpp_source,
    cuda_sources=hardmish_source,
    functions=["hardmish_cuda"],
    extra_cuda_cflags=["-O2"],
    verbose=False
)

class ModelNew(nn.Module):
    def __init__(self, in_features):
        super().__init__()
        torch.manual_seed(42)
        self.linear = nn.Linear(in_features, in_features)
        self.hardmish = hardmish_module.hardmish_cuda

    def forward(self, x):
        x = self.linear(x)
        return self.hardmish(x)