finish TanhGLU #55

S1/uucoco_#55/TanhGLU_cuda.py

@@ -0,0 +1,89 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+
+
+class ModelNew(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self._compile_cuda_kernel()
+
+    def _compile_cuda_kernel(self):
+        cpp_source = """
+        #include <torch/extension.h>
+        torch::Tensor tanhglu_cuda(torch::Tensor input);
+        """
+
+        cuda_source = """
+        #include <cuda_runtime.h>
+
+        __global__ void tanhglu_vec4_kernel(
+            const float4* __restrict__ x,
+            float4* __restrict__ y,
+            int vec_dim_out,
+            int n_vec_out) 
+        {
+            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 = tanhf(g.x) * a.x;
+                out.y = tanhf(g.y) * a.y;
+                out.z = tanhf(g.z) * a.z;
+                out.w = tanhf(g.w) * a.w;
+
+                y[i] = out;
+            }
+        }
+
+        torch::Tensor tanhglu_cuda(torch::Tensor input) {
+            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;
+
+            tanhglu_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
+            );
+
+            return output;
+        }
+        """
+
+        self.op = load_inline(
+            name="tanhglu_opt_vec4",
+            cpp_sources=cpp_source,
+            cuda_sources=cuda_source,
+            functions=["tanhglu_cuda"],
+            extra_cuda_cflags=["-O3", "--use_fast_math"],
+            verbose=False
+        )
+
+    def forward(self, x):
+        return self.op.tanhglu_cuda(x)

S1/uucoco_#55/TanhGLU_torch.py

@@ -0,0 +1,21 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class Model(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        gate, act = x.chunk(2, dim=-1)
+        return torch.tanh(gate) * act
+
+batch_size = 1024
+feature_dim = 4096
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+def get_init_inputs():
+    return []

S1/uucoco_#55/prompt.txt

@@ -0,0 +1,63 @@
+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 CUDA kernel implements optimized Tanh Gated Linear Unit (GLU) with:
+
+Memory Optimization:
+
+Vectorized memory access using float4 for 4x bandwidth
+
+Contiguous tensor inputs for coalesced memory access
+
+Direct element-wise computation without temporary storage
+
+Parallelization Strategy:
+
+Grid-stride loop for efficient workload distribution
+
+256 threads per block optimal configuration
+
+Automatic grid size calculation with 65535 block limit
+
+Computational Optimization:
+
+Tanh GLU: tanh(gate) * activation
+
+Fast math compilation flags for optimized tanhf()
+
+Efficient indexing for gate and activation components
+
+Work Distribution:
+
+Each thread processes 4 elements via float4
+
+Automatic indexing calculation for gate and activation vectors
+
+Direct multiplication of tanh-activated gate with activation
+
+The implementation provides maximum throughput through vectorization and fast math optimizations for the tanh function, requiring input feature dimension to be divisible by 8 for optimal performance.
+
+
+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
+import torch.nn.functional as F
+
+class Model(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        gate, act = x.chunk(2, dim=-1)
+        return torch.tanh(gate) * act
+
+batch_size = 1024
+feature_dim = 4096
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+def get_init_inputs():
+    return []

S1/uucoco_#55/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 TanhGLU_torch import Model, get_inputs, get_init_inputs
+from TanhGLU_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()