claude-opus-4-1-20250805 / cudab62d75

#include <torch/extension.h>
#include <cuda_runtime.h>
#include <vector>
#include <stdexcept>
#include "kernel.h"

torch::Tensor run(torch::Tensor probs, torch::Tensor top_k) {
    // Input validation
    TORCH_CHECK(probs.dim() == 2, "probs must be 2-dimensional, got ", probs.dim());
    TORCH_CHECK(top_k.dim() == 1, "top_k must be 1-dimensional, got ", top_k.dim());
    TORCH_CHECK(probs.size(1) == 151936, 
                "vocab_size must be 151936, got ", probs.size(1));
    TORCH_CHECK(probs.size(0) == top_k.size(0), 
                "batch dimensions must match: probs has ", probs.size(0), 
                ", top_k has ", top_k.size(0));
    
    // Device checks
    TORCH_CHECK(probs.is_cuda(), "probs must be on CUDA device");
    TORCH_CHECK(top_k.is_cuda(), "top_k must be on CUDA device");
    TORCH_CHECK(probs.device() == top_k.device(), 
                "All inputs must be on the same device");
    
    // Type conversion if needed
    if (probs.scalar_type() != torch::kFloat32) {
        probs = probs.to(torch::kFloat32);
    }
    if (top_k.scalar_type() != torch::kInt32) {
        top_k = top_k.to(torch::kInt32);
    }
    
    // Get batch size
    const int batch_size = probs.size(0);
    
    // Ensure contiguous memory layout
    probs = probs.contiguous();
    top_k = top_k.contiguous();
    
    // Create output tensor
    auto options = torch::TensorOptions()
        .dtype(torch::kInt64)
        .device(probs.device());
    torch::Tensor samples = torch::empty({batch_size}, options);
    
    // Get current CUDA stream
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
    
    // Get raw data pointers
    const float* probs_ptr = probs.data_ptr<float>();
    const int32_t* top_k_ptr = top_k.data_ptr<int32_t>();
    int64_t* samples_ptr = samples.data_ptr<int64_t>();
    
    // Launch kernel
    launchTopKSampling(
        probs_ptr,
        top_k_ptr,
        samples_ptr,
        batch_size,
        stream
    );
    
    // Synchronize to ensure kernel completion
    cudaStreamSynchronize(stream);
    
    // Check for kernel execution errors
    cudaError_t err = cudaGetLastError();
    TORCH_CHECK(err == cudaSuccess, 
                "CUDA kernel execution failed: ", cudaGetErrorString(err));
    
    return samples;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("run", &run, "Top-k sampling from probability distributions",
          py::arg("probs"), py::arg("top_k"));
}