import torch
import torch.nn as nn


class Model(nn.Module):
    """
    Weighted expert-output combine back into token-major order.

    Expert outputs are scattered back to token positions and scaled by the
    routing weight associated with each expert contribution.
    """

    def __init__(self):
        super().__init__()

    def forward(
        self,
        expert_hidden: torch.Tensor,
        token_idx: torch.Tensor,
        gates: torch.Tensor,
        num_tokens: int,
    ) -> torch.Tensor:
        out = torch.zeros(num_tokens, expert_hidden.shape[1], dtype=expert_hidden.dtype, device=expert_hidden.device)
        for row in range(expert_hidden.shape[0]):
            token = int(token_idx[row].item())
            out[token] += gates[row] * expert_hidden[row]
        return out


num_tokens = 1536
hidden_dim = 128
fanout = 2


def get_inputs():
    return popcorn_pri.moe_combine_inputs(num_tokens, hidden_dim, fanout)




def get_init_inputs():
    return []

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline

cuda_source = '''
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>

__global__ void scatter_add_kernel(const float* __restrict__ expert_hidden,
                                    const int64_t* __restrict__ token_idx,
                                    const float* __restrict__ gates,
                                    float* __restrict__ out,
                                    int N, int D) {
    int row = blockIdx.x * blockDim.x + threadIdx.x;
    if (row < N) {
        int64_t token = token_idx[row];
        float gate = gates[row];
        const float* src = expert_hidden + row * D;
        float* dst = out + token * D;
        for (int d = 0; d < D; d++) {
            atomicAdd(&dst[d], gate * src[d]);
        }
    }
}

torch::Tensor scatter_add_cuda(torch::Tensor expert_hidden,
                               torch::Tensor token_idx,
                               torch::Tensor gates,
                               int num_tokens) {
    int N = expert_hidden.size(0);
    int D = expert_hidden.size(1);
    auto out = torch::zeros({num_tokens, D}, expert_hidden.options());
    const int block_size = 256;
    const int num_blocks = (N + block_size - 1) / block_size;
    scatter_add_kernel<<<num_blocks, block_size>>>(
        expert_hidden.data_ptr<float>(),
        token_idx.data_ptr<int64_t>(),
        gates.data_ptr<float>(),
        out.data_ptr<float>(),
        N, D
    );
    return out;
}
'''

cpp_source = '''
torch::Tensor scatter_add_cuda(torch::Tensor expert_hidden,
                               torch::Tensor token_idx,
                               torch::Tensor gates,
                               int num_tokens);
'''

module = load_inline(
    name='scatter_add',
    cpp_sources=[cpp_source],
    cuda_sources=[cuda_source],
    functions=['scatter_add_cuda'],
    verbose=False
)

class ModelNew(nn.Module):
    def __init__(self):
        super().__init__()
        self.module = module

    def forward(self, expert_hidden, token_idx, gates, num_tokens):
        token_idx = token_idx.to(torch.int64)
        return self.module.scatter_add_cuda(expert_hidden, token_idx, gates, num_tokens)