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

__global__ void rope_kernel(
    const float* __restrict__ q,
    const float* __restrict__ k,
    float* __restrict__ q_out,
    float* __restrict__ k_out,
    const float* __restrict__ cos,
    const float* __restrict__ sin,
    const int B,
    const int H,
    const int N,
    const 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);

    float c = cos[n * D + d];
    float s = sin[n * D + d];

    int half = D / 2;
    float val_q, val_k, rot_q, rot_k;

    if (d < half) {
        val_q = q[idx];
        val_k = k[idx];
        int d2 = d + half;
        rot_q = -q[b * H * N * D + h * N * D + n * D + d2];
        rot_k = -k[b * H * N * D + h * N * D + n * D + d2];
        q_out[idx] = val_q * c + rot_q * s;
        k_out[idx] = val_k * c + rot_k * s;
    } else {
        val_q = q[idx];
        val_k = k[idx];
        int d2 = d - half;
        rot_q = q[b * H * N * D + h * N * D + n * D + d2];
        rot_k = k[b * H * N * D + h * N * D + n * D + d2];
        q_out[idx] = val_q * c + rot_q * s;
        k_out[idx] = val_k * c + rot_k * s;
    }
}

std::vector<torch::Tensor> rope_cuda(
    torch::Tensor q,
    torch::Tensor k,
    torch::Tensor inv_freq,
    int max_seq_len
) {
    auto B = q.size(0);
    auto H = q.size(1);
    auto N = q.size(2);
    auto D = q.size(3);

    auto options = torch::TensorOptions().dtype(q.dtype()).device(q.device());
    auto cos = torch::empty({N, D}, options);
    auto sin = torch::empty({N, D}, options);

    // Build freqs and cos/sin on GPU
    auto t = torch::arange(N, options);
    auto freqs = t.unsqueeze(1) * inv_freq.unsqueeze(0);  // (N, D/2)
    auto emb = torch::cat({freqs, freqs}, 1);  // (N, D)
    cos = torch::cos(emb);
    sin = torch::sin(emb);

    auto q_out = torch::empty_like(q);
    auto k_out = 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_out.data_ptr<float>(),
        k_out.data_ptr<float>(),
        cos.data_ptr<float>(),
        sin.data_ptr<float>(),
        B, H, N, D
    );

    return {q_out, k_out};
}
'''

cpp_source = r'''
std::vector<torch::Tensor> rope_cuda(
    torch::Tensor q,
    torch::Tensor k,
    torch::Tensor inv_freq,
    int max_seq_len
);
'''

rope_module = load_inline(
    name='rope_module',
    cpp_sources=cpp_source,
    cuda_sources=cuda_source,
    functions=['rope_cuda'],
    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:
        outs = rope_module.rope_cuda(q, k, self.inv_freq, self.max_seq_len)
        return outs[0], outs[1]