import math

import torch
import torch.nn as nn


class Model(nn.Module):
    """
    Scaled low-precision attention with explicit dequantization semantics.

    The inputs are quantized tensors plus per-head scales. This mirrors the data
    path of FP8-style scaled matmul attention even though the reference computes
    everything through explicit dequantization in PyTorch.
    """

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

    def forward(
        self,
        q_q: torch.Tensor,
        k_q: torch.Tensor,
        v_q: torch.Tensor,
        q_scale: torch.Tensor,
        k_scale: torch.Tensor,
        v_scale: torch.Tensor,
    ) -> torch.Tensor:
        q = q_q.float() * q_scale.view(1, 1, -1, 1)
        k = k_q.float() * k_scale.view(1, 1, -1, 1)
        v = v_q.float() * v_scale.view(1, 1, -1, 1)

        batch_size, seq_len, num_heads, head_dim = q.shape
        out = torch.zeros_like(q)
        scale = 1.0 / math.sqrt(head_dim)

        for b in range(batch_size):
            for h in range(num_heads):
                for t in range(seq_len):
                    q_t = q[b, t, h]
                    k_hist = k[b, : t + 1, h]
                    v_hist = v[b, : t + 1, h]
                    scores = (k_hist * q_t.unsqueeze(0)).sum(dim=-1) * scale
                    attn = torch.softmax(scores, dim=0)
                    out[b, t, h] = (attn.unsqueeze(-1) * v_hist).sum(dim=0)

        return out


batch_size = 2
seq_len = 72
num_heads = 4
head_dim = 32


def _quantize_per_head(x: torch.Tensor):
    scale = x.abs().amax(dim=(0, 1, 3)).clamp(min=1e-4) / 127.0
    q = torch.clamp(torch.round(x / scale.view(1, 1, -1, 1)), -127, 127).to(torch.int8)
    return q, scale.to(torch.float32)


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)
    q_q, q_scale = _quantize_per_head(q)
    k_q, k_scale = _quantize_per_head(k)
    v_q, v_scale = _quantize_per_head(v)
    return [q_q, k_q, v_q, q_scale, k_scale, v_scale]


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>
#include <math.h>

__global__ void attention_kernel(
    const int8_t* __restrict__ q_q,
    const int8_t* __restrict__ k_q,
    const int8_t* __restrict__ v_q,
    const float* __restrict__ q_scale,
    const float* __restrict__ k_scale,
    const float* __restrict__ v_scale,
    float* __restrict__ out,
    int batch_size,
    int seq_len,
    int num_heads,
    int head_dim,
    float scale
) {
    int bh = blockIdx.x;
    int b = bh / num_heads;
    int h = bh % num_heads;

    if (b >= batch_size) return;

    float q_s = q_scale[h];
    float k_s = k_scale[h];
    float v_s = v_scale[h];

    for (int t = threadIdx.x; t < seq_len; t += blockDim.x) {
        float q_t[32];
        for (int d=0; d<head_dim; d++) {
            int idx = ((b * seq_len + t) * num_heads + h) * head_dim + d;
            q_t[d] = static_cast<float>(q_q[idx]) * q_s;
        }

        float scores[72];
        float max_score = -1e9f;
        for (int s=0; s<=t; s++) {
            float dot = 0.0f;
            for (int d=0; d<head_dim; d++) {
                int idx = ((b * seq_len + s) * num_heads + h) * head_dim + d;
                float k_val = static_cast<float>(k_q[idx]) * k_s;
                dot += k_val * q_t[d];
            }
            scores[s] = dot * scale;
            if (scores[s] > max_score) max_score = scores[s];
        }

        float sum_exp = 0.0f;
        for (int s=0; s<=t; s++) {
            scores[s] = expf(scores[s] - max_score);
            sum_exp += scores[s];
        }
        float inv_sum = 1.0f / sum_exp;

        float out_t[32] = {0};
        for (int s=0; s<=t; s++) {
            float attn = scores[s] * inv_sum;
            for (int d=0; d<head_dim; d++) {
                int idx = ((b * seq_len + s) * num_heads + h) * head_dim + d;
                float v_val = static_cast<float>(v_q[idx]) * v_s;
                out_t[d] += attn * v_val;
            }
        }

        for (int d=0; d<head_dim; d++) {
            int idx = ((b * seq_len + t) * num_heads + h) * head_dim + d;
            out[idx] = out_t[d];
        }
    }
}

torch::Tensor attention_cuda(
    torch::Tensor q_q,
    torch::Tensor k_q,
    torch::Tensor v_q,
    torch::Tensor q_scale,
    torch::Tensor k_scale,
    torch::Tensor v_scale
) {
    int batch_size = q_q.size(0);
    int seq_len = q_q.size(1);
    int num_heads = q_q.size(2);
    int head_dim = q_q.size(3);

    auto options = torch::TensorOptions().dtype(torch::kFloat32).device(q_q.device());
    auto out = torch::zeros({batch_size, seq_len, num_heads, head_dim}, options);

    float attn_scale = 1.0f / sqrtf(static_cast<float>(head_dim));

    int threads = 32;
    int blocks = batch_size * num_heads;
    attention_kernel<<<blocks, threads>>>(
        q_q.data_ptr<int8_t>(),
        k_q.data_ptr<int8_t>(),
        v_q.data_ptr<int8_t>(),
        q_scale.data_ptr<float>(),
        k_scale.data_ptr<float>(),
        v_scale.data_ptr<float>(),
        out.data_ptr<float>(),
        batch_size, seq_len, num_heads, head_dim, attn_scale
    );

    return out;
}
"""

cpp_source = """
torch::Tensor attention_cuda(torch::Tensor q_q, torch::Tensor k_q, torch::Tensor v_q, torch::Tensor q_scale, torch::Tensor k_scale, torch::Tensor v_scale);
"""

module = load_inline(
    name="attention_module",
    cpp_sources=cpp_source,
    cuda_sources=cuda_source,
    functions=["attention_cuda"],
    verbose=False,
    extra_cuda_cflags=['-O3']
)

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

    def forward(self, q_q, k_q, v_q, q_scale, k_scale, v_scale):
        q_q = q_q.cuda().contiguous()
        k_q = k_q.cuda().contiguous()
        v_q = v_q.cuda().contiguous()
        q_scale = q_scale.cuda().contiguous()
        k_scale = k_scale.cuda().contiguous()
        v_scale = v_scale.cuda().contiguous()
        out = self.module.attention_cuda(q_q, k_q, v_q, q_scale, k_scale, v_scale)
        return out.cpu()