tensorrt_llm 中 layer norm 插件的 USE_DIFF_OF_SQUARES 实现可能导致结果出现 nan

[xiatwhu]

• Environment
  • TensorRT 9.0.0.2
  • Versions of CUDA(12.1), CUBLAS(12.1.3.1)
  • Container used (registry.cn-hangzhou.aliyuncs.com/trt-hackathon/trt-hackathon:final_v1)
  • NVIDIA driver version (535.86.05)
• Reproduction Steps USE_DIFF_OF_SQUARES 采用了 Var[x] = E[x²] - E[x]² 的公式来计算方差，但由于 float 能表示的精度有限，在某些情况下会导致计算出来的 Var[x] 为负数，此时通过 rsqrtf 计算得到的标准差的倒数为 nan，从而导致经过 layer norm 之后输出为 nan。

使用 torch 进行简单的验证

import torch

x = torch.full((1, 10240), 256., device='cuda:0', dtype=torch.float32)
x[0, -1] = 255

x_square_mean = (x * x).mean()
x_mean = x.mean()
x_var = x_square_mean - x_mean * x_mean
x_inv_std = torch.rsqrt(x_var + 1e-5)

print(x_var, x_inv_std)

• Expected Behavior 输出的方差大于 0，标准差的倒数大于 0

• Actual Behavior 输出的方差小于 0，标准差的倒数为 nan

  tensor(-0.0039, device='cuda:0') tensor(nan, device='cuda:0')

• Additional Notes 编写 C++ 代码来测试 tensorrt_llm 中的 kernel，确认经过 layer norm 计算后结果为 nan

  
  #include "cuda_runtime.h"
  #include "tensorrt_llm/kernels/layernormKernels.h"

include

include

float malloc_device(std::vector& host) { float dev = nullptr; cudaMalloc(&dev, host.size() sizeof(float)); cudaMemcpy(dev, host.data(), host.size() sizeof(float), cudaMemcpyDefault); return dev; }

int main() { const int size = 10240;

auto host_input = std::vector<float>(size, 256.f);
host_input[size - 1] = 255.f;
auto host_gamma = std::vector<float>(size, 1.f);
auto host_beta = std::vector<float>(size, 0.f);
auto host_output = std::vector<float>(size, 0.f);

float* dev_input = malloc_device(host_input);
float* dev_gamma = malloc_device(host_gamma);
float* dev_beta = malloc_device(host_beta);
float* dev_output = malloc_device(host_output);

tensorrt_llm::kernels::invokeGeneralLayerNorm(
        dev_output, dev_input, dev_gamma, dev_beta, 1e-5f, 1, size);

cudaMemcpy(host_output.data(), dev_output, size * sizeof(float), cudaMemcpyDefault);

std::cout << host_output[0] << std::endl;

return 0;

}

修复方法：
- 方案一：只需要添加一个 variance = max(variance, 0.f) 的保护即可修复该 bug，在代码 tensorrt_llm/kernels/layernormKernels.cu 中 118 行
```cpp
    if (threadIdx.x == 0)
    {
        mean = mean / hidden_dim;
        s_mean = mean;
        if (USE_DIFF_OF_SQUARES)
        {
            variance = (variance / hidden_dim) - (mean * mean); // Var[x] = E[x²] - E[x]²
            // 此处添加一行 variance = max(variance, 0.f)
            s_variance = rsqrtf(variance + eps);
        }
    }
    __syncthreads();

• 方案二：可以采用 torch 源码中的 Welford 算法来计算方差