finish fisherrao_rmsnorm #124

S1/uucoco_#124/fisherrao_rmsnorm_cuda.py

@@ -0,0 +1,75 @@
+import os
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+
+cuda_source = """
+#include <torch/extension.h>
+#include <cuda_runtime.h>
+
+__global__ void fisherrao_dist_kernel(const float* __restrict__ x, 
+                                      const float* __restrict__ y, 
+                                      float* __restrict__ out, 
+                                      int size) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < size) {
+        float eps = 1e-6f;
+        float val_x = fabsf(x[idx]) + eps;
+        float val_y = fabsf(y[idx]) + eps;
+        // Metric: |log(x) - log(y)|
+        out[idx] = fabsf(logf(val_x) - logf(val_y));
+    }
+}
+
+torch::Tensor fisherrao_dist_cuda(torch::Tensor x, torch::Tensor y) {
+    auto size = x.numel();
+    auto out = torch::empty_like(x);
+
+    const int block_size = 256;
+    int grid_size = (size + block_size - 1) / block_size;
+
+    fisherrao_dist_kernel<<<grid_size, block_size>>>(
+        x.data_ptr<float>(),
+        y.data_ptr<float>(),
+        out.data_ptr<float>(),
+        size
+    );
+
+    return out;
+}
+"""
+
+cpp_source = "torch::Tensor fisherrao_dist_cuda(torch::Tensor x, torch::Tensor y);"
+
+module = load_inline(
+    name="fisherrao_rmsnorm_ext",
+    cpp_sources=cpp_source,
+    cuda_sources=cuda_source,
+    functions=["fisherrao_dist_cuda"],
+    verbose=False,
+    with_cuda=True
+)
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim, eps=1e-6):
+        super().__init__()
+        self.scale = nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x):
+        var = x.pow(2).mean(-1, keepdim=True)
+        norm_x = x * torch.rsqrt(var + self.eps)
+        return norm_x * self.scale
+
+
+class ModelNew(nn.Module):
+    def __init__(self, dim):
+        super(ModelNew, self).__init__()
+        self.rmsnorm = RMSNorm(dim)
+        self.op = module
+
+    def forward(self, x, y):
+        dist = self.op.fisherrao_dist_cuda(x.contiguous(), y.contiguous())
+        out = self.rmsnorm(dist)
+        return out.mean()

S1/uucoco_#124/fisherrao_rmsnorm_torch.py

@@ -0,0 +1,37 @@
+import torch
+import torch.nn as nn
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim, eps=1e-6):
+        super().__init__()
+        self.scale = nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x):
+        var = x.pow(2).mean(-1, keepdim=True)
+        norm_x = x * torch.rsqrt(var + self.eps)
+        return norm_x * self.scale
+
+class Model(nn.Module):
+    def __init__(self, dim):
+        super(Model, self).__init__()
+        self.rmsnorm = RMSNorm(dim)
+
+    def forward(self, x, y):
+        eps = 1e-6
+        val_x = torch.abs(x) + eps
+        val_y = torch.abs(y) + eps
+        dist = torch.abs(torch.log(val_x) - torch.log(val_y))
+        out = self.rmsnorm(dist)
+        return out.mean()
+
+batch_size = 16
+input_dim = 1024
+
+def get_inputs():
+    x = torch.randn(batch_size, input_dim)
+    y = torch.randn(batch_size, input_dim)
+    return [x, y]
+
+def get_init_inputs():
+    return [input_dim]

S1/uucoco_#124/prompt.txt

@@ -0,0 +1,68 @@
+You write custom CUDA kernels to replace the pytorch operators in the given GeGLU 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 chunk+gelu+elementwise_mul), or algorithmic changes (such as optimized memory access patterns). You are only limited by your imagination.
+
+This code implements Fisher-Rao distance + RMSNorm (Root Mean Square Normalization) with CUDA optimizations:
+
+Element-wise parallelism - Each thread computes distance between x[i] and y[i] independently.
+
+Fisher-Rao metric - Computes |log(x) - log(y)| as distance on probability simplex.
+
+Numerical stability - Adds ε=1e-6 to absolute values to avoid log(0).
+
+Fused distance calculation - Computes absolute log difference in single kernel.
+
+Memory coalescing - Contiguous memory access patterns.
+
+CUDA math functions - Uses fabsf() and logf() for hardware acceleration.
+
+Simple grid-stride mapping - Standard 1D grid/block for element-wise operations.
+
+Custom RMSNorm implementation - PyTorch module for root mean square normalization.
+
+Post-processing - Applies RMSNorm to Fisher-Rao distance elements.
+
+Batch processing - Handles all elements in parallel regardless of shape.
+
+
+
+
+
+Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is:
+import torch
+import torch.nn as nn
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim, eps=1e-6):
+        super().__init__()
+        self.scale = nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x):
+        var = x.pow(2).mean(-1, keepdim=True)
+        norm_x = x * torch.rsqrt(var + self.eps)
+        return norm_x * self.scale
+
+class Model(nn.Module):
+    def __init__(self, dim):
+        super(Model, self).__init__()
+        self.rmsnorm = RMSNorm(dim)
+
+    def forward(self, x, y):
+        eps = 1e-6
+        val_x = torch.abs(x) + eps
+        val_y = torch.abs(y) + eps
+        dist = torch.abs(torch.log(val_x) - torch.log(val_y))
+        out = self.rmsnorm(dist)
+        return out.mean()
+
+batch_size = 16
+input_dim = 1024
+
+def get_inputs():
+    x = torch.randn(batch_size, input_dim)
+    y = torch.randn(batch_size, input_dim)
+    return [x, y]
+
+def get_init_inputs():
+    return [input_dim]

S1/uucoco_#124/run_code.py

@@ -0,0 +1,77 @@
+###########################################################
+# 性能和精度验证程序
+###########################################################
+import torch
+import torch.nn as nn
+import time
+from fisherrao_rmsnorm_torch import Model, get_inputs, get_init_inputs
+from fisherrao_rmsnorm_cuda  import ModelNew
+
+
+def run_benchmark():
+    # 检查 CUDA 是否可用
+    if not torch.cuda.is_available():
+        print("CUDA 不可用，请确保您有可用的 NVIDIA GPU 并已正确安装 PyTorch CUDA 版本。")
+        return
+    else:
+        device = torch.device("cuda")
+
+    # 初始化模型
+    init_inputs = get_init_inputs()
+    init_inputs = [
+        x.cuda(device=device) if isinstance(x, torch.Tensor) else x for x in init_inputs
+    ]
+    inputs = get_inputs()
+    inputs = [
+        x.cuda(device=device) if isinstance(x, torch.Tensor) else x for x in inputs
+    ]
+
+    torch_model = Model(*init_inputs).cuda()
+    cuda_model = ModelNew(*init_inputs).cuda()
+
+    torch_model.eval()
+    cuda_model.eval()
+
+    print("-------------------- 精度对齐验证 --------------------")
+    with torch.no_grad():
+        output_torch = torch_model(*inputs)
+        output_cuda = cuda_model(*inputs)
+
+    precision_flag = torch.allclose(output_torch, output_cuda, rtol=1e-03)
+    if precision_flag:
+        print("✅ 精度对齐：两个模型的输出结果非常接近。")
+    else:
+        print("❌ 精度不一致！")
+
+    print("\n-------------------- 性能加速比测试 --------------------")
+    num_iterations = 100
+
+    # PyTorch 模型计时
+    torch.cuda.synchronize()
+    start_time = time.time()
+    for _ in range(num_iterations):
+        _ = torch_model(*inputs)
+    torch.cuda.synchronize()
+    torch_time = (time.time() - start_time) / num_iterations
+
+    # 自定义 CUDA 内核计时
+    torch.cuda.synchronize()
+    start_time = time.time()
+    for _ in range(num_iterations):
+        _ = cuda_model(*inputs)
+    torch.cuda.synchronize()
+    cuda_time = (time.time() - start_time) / num_iterations
+
+    print(f"PyTorch torch.relu 平均执行时间: {torch_time:.6f} 秒")
+    print(f"自定义 CUDA 内核 平均执行时间: {cuda_time:.6f} 秒")
+    speedup = 0
+    if cuda_time > 0:
+        speedup = torch_time / cuda_time
+        print(f"加速比 (Speedup): {speedup:.2f}x")
+    else:
+        print("CUDA 内核执行时间为0，无法计算加速比。")
+    return precision_flag, speedup
+
+
+if __name__ == "__main__":
+    precision_flag, speedup = run_benchmark()