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

_cpp_src = r"""
#include <torch/extension.h>
#include <vector>

std::vector<torch::Tensor> rope_forward_cuda(torch::Tensor q, torch::Tensor k, torch::Tensor cos, torch::Tensor sin);

std::vector<torch::Tensor> rope_forward(torch::Tensor q, torch::Tensor k, torch::Tensor cos, torch::Tensor sin) {
    return rope_forward_cuda(q, k, cos, sin);
}
"""

_cuda_src = r"""
#include <torch/extension.h>
#include <ATen/cuda/CUDAContext.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <vector>

__global__ void rope_kernel_64(const float* __restrict__ q,
                               const float* __restrict__ k,
                               const float* __restrict__ cos,
                               const float* __restrict__ sin,
                               float* __restrict__ qo,
                               float* __restrict__ ko,
                               int total_pairs,
                               int N) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx >= total_pairs) return;
    int j = idx & 31;
    int tmp = idx >> 5;
    int n = tmp % N;
    int bh = tmp / N;
    int base = ((bh * N + n) << 6) + j;
    int base2 = base + 32;
    int cs = n * 32 + j;
    float c = cos[cs];
    float s = sin[cs];
    float q1 = q[base];
    float q2 = q[base2];
    qo[base]  = q1 * c - q2 * s;
    qo[base2] = q2 * c + q1 * s;
    float k1 = k[base];
    float k2 = k[base2];
    ko[base]  = k1 * c - k2 * s;
    ko[base2] = k2 * c + k1 * s;
}

__global__ void rope_kernel_generic(const float* __restrict__ q,
                                    const float* __restrict__ k,
                                    const float* __restrict__ cos,
                                    const float* __restrict__ sin,
                                    float* __restrict__ qo,
                                    float* __restrict__ ko,
                                    int total_pairs,
                                    int N,
                                    int D,
                                    int half) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx >= total_pairs) return;
    int j = idx % half;
    int tmp = idx / half;
    int n = tmp % N;
    int bh = tmp / N;
    int base = (bh * N + n) * D + j;
    int base2 = base + half;
    int cs = n * half + j;
    float c = cos[cs];
    float s = sin[cs];
    float q1 = q[base];
    float q2 = q[base2];
    qo[base]  = q1 * c - q2 * s;
    qo[base2] = q2 * c + q1 * s;
    float k1 = k[base];
    float k2 = k[base2];
    ko[base]  = k1 * c - k2 * s;
    ko[base2] = k2 * c + k1 * s;
}

std::vector<torch::Tensor> rope_forward_cuda(torch::Tensor q, torch::Tensor k, torch::Tensor cos, torch::Tensor sin) {
    auto qo = torch::empty_like(q);
    auto ko = torch::empty_like(k);
    int B = (int)q.size(0);
    int H = (int)q.size(1);
    int N = (int)q.size(2);
    int D = (int)q.size(3);
    int half = D / 2;
    int total_pairs = B * H * N * half;
    const int threads = 256;
    int blocks = (total_pairs + threads - 1) / threads;
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
    if (D == 64) {
        rope_kernel_64<<<blocks, threads, 0, stream>>>(q.data_ptr<float>(), k.data_ptr<float>(), cos.data_ptr<float>(), sin.data_ptr<float>(), qo.data_ptr<float>(), ko.data_ptr<float>(), total_pairs, N);
    } else {
        rope_kernel_generic<<<blocks, threads, 0, stream>>>(q.data_ptr<float>(), k.data_ptr<float>(), cos.data_ptr<float>(), sin.data_ptr<float>(), qo.data_ptr<float>(), ko.data_ptr<float>(), total_pairs, N, D, half);
    }
    return {qo, ko};
}
"""

_rope_ext = load_inline(
    name="rope_ext_opt2",
    cpp_sources=[_cpp_src],
    cuda_sources=[_cuda_src],
    functions=["rope_forward"],
    extra_cflags=["-O3"],
    extra_cuda_cflags=["-O3"],
    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))
        t = torch.arange(max_seq_len, dtype=inv_freq.dtype)
        freqs = torch.outer(t, inv_freq)
        self.register_buffer("cos_cached", freqs.cos(), persistent=False)
        self.register_buffer("sin_cached", freqs.sin(), persistent=False)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.max_seq_len = max_seq_len
        self.dim = dim
        self.base = base

    def forward(self, q: torch.Tensor, k: torch.Tensor) -> tuple:
        N = q.shape[2]
        return tuple(_rope_ext.rope_forward(q, k, self.cos_cached[:N], self.sin_cached[:N]))