import torch
import torch.nn as nn


class Model(nn.Module):
    """
    Scalar-gated delta-rule linear attention recurrence.

    A recurrent state matrix is decayed by a scalar gate per head, updated by a
    rank-1 key/value outer product, then read out by a query vector.
    """

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

    def forward(
        self,
        q: torch.Tensor,
        k: torch.Tensor,
        v: torch.Tensor,
        gate: torch.Tensor,
        beta: torch.Tensor,
    ) -> torch.Tensor:
        batch_size, seq_len, num_heads, head_dim = q.shape
        state = torch.zeros(batch_size, num_heads, head_dim, head_dim, dtype=q.dtype, device=q.device)
        outputs = []

        for t in range(seq_len):
            gate_t = gate[:, t].unsqueeze(-1).unsqueeze(-1)
            beta_t = beta[:, t].unsqueeze(-1).unsqueeze(-1)
            kv_outer = k[:, t].unsqueeze(-1) * v[:, t].unsqueeze(-2)
            state = gate_t * state + beta_t * kv_outer
            y_t = (q[:, t].unsqueeze(-1) * state).sum(dim=-2)
            outputs.append(y_t)

        return torch.stack(outputs, dim=1)


batch_size = 2
seq_len = 128
num_heads = 4
head_dim = 24


def get_inputs():
    p = popcorn_pri
    q = torch.randn(p.jitter_int(batch_size), p.jitter_int(seq_len), p.jitter_int(num_heads), p.jitter_int(head_dim, align=8), dtype=torch.float32)
    k = torch.randn(p.jitter_int(batch_size), p.jitter_int(seq_len), p.jitter_int(num_heads), p.jitter_int(head_dim, align=8), dtype=torch.float32)
    v = torch.randn(p.jitter_int(batch_size), p.jitter_int(seq_len), p.jitter_int(num_heads), p.jitter_int(head_dim, align=8), dtype=torch.float32)
    gate = torch.sigmoid(torch.randn(p.jitter_int(batch_size), p.jitter_int(seq_len), p.jitter_int(num_heads), dtype=torch.float32))
    beta = torch.sigmoid(torch.randn(p.jitter_int(batch_size), p.jitter_int(seq_len), p.jitter_int(num_heads), dtype=torch.float32))
    return [q, k, v, gate, beta]


def get_init_inputs():
    return []

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

_cpp = r'''
#include <torch/extension.h>
torch::Tensor delta_attn_forward(torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor gate, torch::Tensor beta);
'''

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

__global__ __launch_bounds__(32, 8)
void delta_cols_kernel(const float* __restrict__ q,
                       const float* __restrict__ k,
                       const float* __restrict__ v,
                       const float* __restrict__ gate,
                       const float* __restrict__ beta,
                       float* __restrict__ out,
                       int B, int T, int H, int D) {
    int col = blockIdx.x;
    int bh = blockIdx.y;
    int b = bh / H;
    int h = bh - b * H;
    int lane = threadIdx.x;
    float s = 0.0f;

    #pragma unroll 1
    for (int t = 0; t < T; ++t) {
        int gh = (b * T + t) * H + h;
        int base = gh * D;
        float g = gate[gh];
        float be = beta[gh];
        float vj = v[base + col];
        float prod = 0.0f;

        if (lane < D) {
            s = fmaf(g, s, be * k[base + lane] * vj);
            prod = q[base + lane] * s;
        }

        prod += __shfl_down_sync(0xffffffff, prod, 16);
        prod += __shfl_down_sync(0xffffffff, prod, 8);
        prod += __shfl_down_sync(0xffffffff, prod, 4);
        prod += __shfl_down_sync(0xffffffff, prod, 2);
        prod += __shfl_down_sync(0xffffffff, prod, 1);

        if (lane == 0) out[base + col] = prod;
    }
}

torch::Tensor delta_attn_forward(torch::Tensor q, torch::Tensor k, torch::Tensor v, torch::Tensor gate, torch::Tensor beta) {
    int B = q.size(0);
    int T = q.size(1);
    int H = q.size(2);
    int D = q.size(3);
    auto out = torch::empty({B, T, H, D}, q.options());
    dim3 grid(D, B * H);
    delta_cols_kernel<<<grid, 32>>>(q.data_ptr<float>(), k.data_ptr<float>(), v.data_ptr<float>(), gate.data_ptr<float>(), beta.data_ptr<float>(), out.data_ptr<float>(), B, T, H, D);
    return out;
}
'''

_ext = load_inline(
    name='delta_attn_final_ext',
    cpp_sources=_cpp,
    cuda_sources=_cuda,
    functions=['delta_attn_forward'],
    verbose=False,
    extra_cuda_cflags=['-O3', '--use_fast_math'],
)

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

    def forward(self, q, k, v, gate, beta):
        return _ext.delta_attn_forward(q, k, v, gate, beta)