import torch
import torch.nn as nn
import math


class Model(nn.Module):
    """
    Rotary Position Embedding (RoPE) applied to protein residue sequences.
    Used in ESM-2 and similar protein language models.  Encodes absolute
    position via rotation of query/key pairs in 2-D subspaces, preserving
    relative-position information in dot-product attention.
    """

    def __init__(self, dim, max_seq_len=2048, base=10000.0):
        super().__init__()
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq)
        self.max_seq_len = max_seq_len
        self.dim = dim

    def _build_cos_sin(self, seq_len: int, device: torch.device):
        t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype)
        freqs = torch.outer(t, self.inv_freq)
        emb = torch.cat([freqs, freqs], dim=-1)
        return emb.cos(), emb.sin()

    @staticmethod
    def _rotate_half(x: torch.Tensor) -> torch.Tensor:
        x1, x2 = x.chunk(2, dim=-1)
        return torch.cat([-x2, x1], dim=-1)

    def forward(self, q: torch.Tensor, k: torch.Tensor) -> tuple:
        """
        Args:
            q: (B, H, N, D)  – queries
            k: (B, H, N, D)  – keys
        Returns:
            (q_rot, k_rot) each (B, H, N, D) with RoPE applied
        """
        N = q.shape[2]
        cos, sin = self._build_cos_sin(N, q.device)  # (N, D)
        cos = cos.unsqueeze(0).unsqueeze(0)
        sin = sin.unsqueeze(0).unsqueeze(0)
        q_rot = q * cos + self._rotate_half(q) * sin
        k_rot = k * cos + self._rotate_half(k) * sin
        return q_rot, k_rot


dim = 64
num_heads = 8
seq_len = 256
batch_size = 4


def get_inputs():
    p = popcorn_pri
    bs = p.jitter_int(batch_size)
    nh = p.jitter_int(num_heads)
    sl = p.jitter_int(seq_len)
    q = torch.randn(bs, nh, sl, dim)
    k = torch.randn(bs, nh, sl, dim)
    return [q, k]


def get_init_inputs():
    return [dim]

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 rope_kernel(
    const float* __restrict__ q,
    const float* __restrict__ k,
    float* __restrict__ q_rot,
    float* __restrict__ k_rot,
    const float* __restrict__ cos,
    const float* __restrict__ sin,
    int B, int H, int N, int D
) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    int total = B * H * N * D;
    if (idx >= total) return;

    int d = idx % D;
    int n = (idx / D) % N;
    int h = (idx / (D * N)) % H;
    int b = idx / (D * N * H);

    int half_d = D / 2;
    float c = cos[n * D + d];
    float s = sin[n * D + d];

    float val_q = q[idx];
    float val_k = k[idx];

    if (d < half_d) {
        int idx2 = idx + half_d;
        float val_q2 = q[idx2];
        float val_k2 = k[idx2];
        q_rot[idx] = val_q * c - val_q2 * s;
        k_rot[idx] = val_k * c - val_k2 * s;
    } else {
        int idx1 = idx - half_d;
        float val_q1 = q[idx1];
        float val_k1 = k[idx1];
        q_rot[idx] = val_q1 * s + val_q * c;
        k_rot[idx] = val_k1 * s + val_k * c;
    }
}

std::tuple<torch::Tensor, torch::Tensor> apply_rope(
    torch::Tensor q,
    torch::Tensor k,
    torch::Tensor cos,
    torch::Tensor sin
) {
    int B = q.size(0);
    int H = q.size(1);
    int N = q.size(2);
    int D = q.size(3);

    auto q_rot = torch::empty_like(q);
    auto k_rot = torch::empty_like(k);

    int total = B * H * N * D;
    int threads = 256;
    int blocks = (total + threads - 1) / threads;

    rope_kernel<<<blocks, threads>>>(
        q.data_ptr<float>(),
        k.data_ptr<float>(),
        q_rot.data_ptr<float>(),
        k_rot.data_ptr<float>(),
        cos.data_ptr<float>(),
        sin.data_ptr<float>(),
        B, H, N, D
    );

    return std::make_tuple(q_rot, k_rot);
}
'''

cpp_source = '''
std::tuple<torch::Tensor, torch::Tensor> apply_rope(
    torch::Tensor q,
    torch::Tensor k,
    torch::Tensor cos,
    torch::Tensor sin
);
'''

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

class ModelNew(nn.Module):
    def __init__(self, dim, max_seq_len=2048, base=10000.0):
        super().__init__()
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq)
        self.max_seq_len = max_seq_len
        self.dim = dim

    def forward(self, q: torch.Tensor, k: torch.Tensor) -> tuple:
        N = q.shape[2]
        device = q.device
        dtype = self.inv_freq.dtype
        t = torch.arange(N, device=device, dtype=dtype)
        freqs = torch.outer(t, self.inv_freq)
        emb = torch.cat([freqs, freqs], dim=-1)
        cos = emb.cos().contiguous()
        sin = emb.sin().contiguous()
        return module.apply_rope(q, k, cos, sin)