claude-opus-4-1-20250805 / cudaef57df

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

torch::Tensor run(
    torch::Tensor probs,
    torch::Tensor top_k,
    torch::Tensor top_p
) {
    // Validate input dimensions
    TORCH_CHECK(probs.dim() == 2, "probs must be a 2D tensor");
    TORCH_CHECK(top_k.dim() == 1, "top_k must be a 1D tensor");
    TORCH_CHECK(top_p.dim() == 1, "top_p must be a 1D tensor");
    
    const int batch_size = probs.size(0);
    const int vocab_size = probs.size(1);
    
    // Check vocabulary size
    TORCH_CHECK(vocab_size == VOCAB_SIZE, 
                "Vocabulary size must be exactly 151936, got ", vocab_size);
    
    // Check batch size consistency
    TORCH_CHECK(top_k.size(0) == batch_size, 
                "top_k batch size mismatch: expected ", batch_size, ", got ", top_k.size(0));
    TORCH_CHECK(top_p.size(0) == batch_size, 
                "top_p batch size mismatch: expected ", batch_size, ", got ", top_p.size(0));
    
    // Check data types
    TORCH_CHECK(probs.scalar_type() == torch::ScalarType::Float, 
                "probs must be float32");
    TORCH_CHECK(top_k.scalar_type() == torch::ScalarType::Int, 
                "top_k must be int32");
    TORCH_CHECK(top_p.scalar_type() == torch::ScalarType::Float, 
                "top_p must be float32");
    
    // Ensure tensors are on CUDA device
    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(top_p.is_cuda(), "top_p must be on CUDA device");
    
    // Make tensors contiguous if needed
    probs = probs.contiguous();
    top_k = top_k.contiguous();
    top_p = top_p.contiguous();
    
    // Create output tensor for samples
    auto options = torch::TensorOptions()
        .dtype(torch::kInt64)
        .device(probs.device())
        .requires_grad(false);
    torch::Tensor samples = torch::empty({batch_size}, options);
    
    // Get current CUDA stream
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
    
    // Launch the kernel
    launch_top_k_top_p_sampling(
        probs.data_ptr<float>(),
        top_k.data_ptr<int>(),
        top_p.data_ptr<float>(),
        samples.data_ptr<int64_t>(),
        batch_size,
        stream
    );
    
    // Ensure kernel completion before returning
    cudaError_t err = cudaStreamSynchronize(stream);
    if (err != cudaSuccess) {
        TORCH_CHECK(false, "CUDA kernel execution failed: ", cudaGetErrorString(err));
    }
    
    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"));
}