claude-opus-4-1-20250805 / cudae7a6dc

#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,
    torch::Tensor top_p
) {
    // Input validation
    TORCH_CHECK(probs.dim() == 2, "probs must be 2D tensor");
    TORCH_CHECK(top_k.dim() == 1, "top_k must be 1D tensor");
    TORCH_CHECK(top_p.dim() == 1, "top_p must be 1D tensor");
    
    const int64_t batch_size = probs.size(0);
    const int64_t vocab_size = probs.size(1);
    
    TORCH_CHECK(vocab_size == 128256, "vocab_size must be 128256, got ", vocab_size);
    TORCH_CHECK(top_k.size(0) == batch_size, "top_k size must match batch_size");
    TORCH_CHECK(top_p.size(0) == batch_size, "top_p size must match batch_size");
    
    // Ensure tensors are on CUDA
    TORCH_CHECK(probs.is_cuda(), "probs must be on CUDA");
    TORCH_CHECK(top_k.is_cuda(), "top_k must be on CUDA");
    TORCH_CHECK(top_p.is_cuda(), "top_p must be on CUDA");
    
    // Convert to correct dtypes 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);
    }
    if (top_p.scalar_type() != torch::kFloat32) {
        top_p = top_p.to(torch::kFloat32);
    }
    
    // Ensure contiguous memory layout
    probs = probs.contiguous();
    top_k = top_k.contiguous();
    top_p = top_p.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();
    
    // Launch kernel
    launch_top_k_top_p_sampling(
        probs.data_ptr<float>(),
        top_k.data_ptr<int32_t>(),
        top_p.data_ptr<float>(),
        samples.data_ptr<int64_t>(),
        static_cast<int>(batch_size),
        stream
    );
    
    // Ensure kernel completion
    cudaError_t error = cudaStreamSynchronize(stream);
    if (error != cudaSuccess) {
        TORCH_CHECK(false, "CUDA kernel execution failed: ", cudaGetErrorString(error));
    }
    
    return samples;
}

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