import torch
import torch.nn as nn


class Model(nn.Module):
    """
    Apply rotary position embedding to incoming keys and update the KV cache.

    This isolates a common decode-time kernel: rotate the fresh key vectors for
    the current positions and store both keys and values into the running cache.
    """

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

    def forward(
        self,
        k_new: torch.Tensor,
        v_new: torch.Tensor,
        cos: torch.Tensor,
        sin: torch.Tensor,
        cache_k: torch.Tensor,
        cache_v: torch.Tensor,
        positions: torch.Tensor,
    ) -> torch.Tensor:
        updated_k = cache_k.clone()
        updated_v = cache_v.clone()
        rot_dim = k_new.shape[-1]

        for b in range(k_new.shape[0]):
            for t in range(k_new.shape[1]):
                pos = int(positions[b, t].item())
                c = cos[pos]
                s = sin[pos]
                x = k_new[b, t]
                x_even = x[..., 0::2]
                x_odd = x[..., 1::2]
                rot_even = x_even * c - x_odd * s
                rot_odd = x_even * s + x_odd * c
                rotated = torch.empty_like(x)
                rotated[..., 0::2] = rot_even
                rotated[..., 1::2] = rot_odd
                updated_k[b, pos] = rotated
                updated_v[b, pos] = v_new[b, t]

        return torch.stack((updated_k, updated_v), dim=0)


batch_size = 2
cache_len = 128
update_len = 16
num_heads = 4
head_dim = 32


def _rope_tables(cache_len: int, half_dim: int):
    positions = torch.arange(cache_len, dtype=torch.float32).unsqueeze(1)
    freqs = torch.pow(10000.0, -torch.arange(half_dim, dtype=torch.float32) / half_dim).unsqueeze(0)
    angles = positions * freqs
    return torch.cos(angles), torch.sin(angles)


def get_inputs():
    p = popcorn_pri
    k_new = torch.randn(p.jitter_int(batch_size), p.jitter_int(update_len), p.jitter_int(num_heads), p.jitter_int(head_dim, align=8), dtype=torch.float32)
    v_new = torch.randn(p.jitter_int(batch_size), p.jitter_int(update_len), p.jitter_int(num_heads), p.jitter_int(head_dim, align=8), dtype=torch.float32)
    cos, sin = _rope_tables(p.jitter_int(cache_len), p.jitter_int(head_dim, align=8) // 2)
    cache_k = torch.randn(p.jitter_int(batch_size), p.jitter_int(cache_len), p.jitter_int(num_heads), p.jitter_int(head_dim, align=8), dtype=torch.float32)
    cache_v = torch.randn(p.jitter_int(batch_size), p.jitter_int(cache_len), p.jitter_int(num_heads), p.jitter_int(head_dim, align=8), dtype=torch.float32)
    base = torch.arange(p.jitter_int(update_len), dtype=torch.int32).unsqueeze(0).repeat(p.jitter_int(batch_size), 1)
    positions = (base + torch.tensor([[7], [31]], dtype=torch.int32)) % p.jitter_int(cache_len)
    return [k_new, v_new, cos, sin, cache_k, cache_v, positions]


def get_init_inputs():
    return []

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

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

__global__ void rope_update_kernel(
    const float* __restrict__ k_new,
    const float* __restrict__ v_new,
    const float* __restrict__ cos,
    const float* __restrict__ sin,
    const float* __restrict__ cache_k,
    const float* __restrict__ cache_v,
    const int* __restrict__ positions,
    float* __restrict__ updated_k,
    float* __restrict__ updated_v,
    int batch_size,
    int update_len,
    int num_heads,
    int head_dim,
    int cache_len
) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    int total = batch_size * update_len;
    if (idx >= total) return;

    int b = idx / update_len;
    int t = idx % update_len;

    int pos = positions[b * update_len + t];
    if (pos < 0 || pos >= cache_len) return;

    int rot_dim = head_dim;
    int half_dim = head_dim / 2;

    const float* c = cos + pos * half_dim;
    const float* s = sin + pos * half_dim;

    for (int h = 0; h < num_heads; ++h) {
        int base = ((b * update_len + t) * num_heads + h) * head_dim;
        int out_base = (b * cache_len + pos) * num_heads * head_dim + h * head_dim;

        for (int i = 0; i < half_dim; ++i) {
            float x_even = k_new[base + 2 * i];
            float x_odd = k_new[base + 2 * i + 1];
            float cc = c[i];
            float ss = s[i];
            updated_k[out_base + 2 * i] = x_even * cc - x_odd * ss;
            updated_k[out_base + 2 * i + 1] = x_even * ss + x_odd * cc;
        }

        for (int d = 0; d < head_dim; ++d) {
            updated_v[out_base + d] = v_new[base + d];
        }
    }
}

torch::Tensor rope_kv_update(
    torch::Tensor k_new,
    torch::Tensor v_new,
    torch::Tensor cos,
    torch::Tensor sin,
    torch::Tensor cache_k,
    torch::Tensor cache_v,
    torch::Tensor positions
) {
    auto updated_k = cache_k.clone();
    auto updated_v = cache_v.clone();

    int batch_size = k_new.size(0);
    int update_len = k_new.size(1);
    int num_heads = k_new.size(2);
    int head_dim = k_new.size(3);
    int cache_len = cache_k.size(1);
    int half_dim = head_dim / 2;

    int total = batch_size * update_len;
    int threads = 256;
    int blocks = (total + threads - 1) / threads;

    rope_update_kernel<<<blocks, threads>>>(
        k_new.data_ptr<float>(),
        v_new.data_ptr<float>(),
        cos.data_ptr<float>(),
        sin.data_ptr<float>(),
        cache_k.data_ptr<float>(),
        cache_v.data_ptr<float>(),
        positions.data_ptr<int>(),
        updated_k.data_ptr<float>(),
        updated_v.data_ptr<float>(),
        batch_size,
        update_len,
        num_heads,
        head_dim,
        cache_len
    );

    cudaDeviceSynchronize();

    return torch::stack({updated_k, updated_v}, 0);
}
'''

cpp_source = r'''
torch::Tensor rope_kv_update(
    torch::Tensor k_new,
    torch::Tensor v_new,
    torch::Tensor cos,
    torch::Tensor sin,
    torch::Tensor cache_k,
    torch::Tensor cache_v,
    torch::Tensor positions
);
'''

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

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

    def forward(
        self,
        k_new: torch.Tensor,
        v_new: torch.Tensor,
        cos: torch.Tensor,
        sin: torch.Tensor,
        cache_k: torch.Tensor,
        cache_v: torch.Tensor,
        positions: torch.Tensor,
    ) -> torch.Tensor:
        return self.module.rope_kv_update(k_new, v_new, cos, sin, cache_k, cache_v, positions)
```