"""v12: noise-annealed STE on v3 architecture (Issue 2 isolation).

Self-contained reimplementation of v3 where every sign-STE gets annealed
additive Gaussian noise during training: sign(x + N(0, σ²)).
σ anneals 1.0 -> 0.05 over training, injected via module-level holder.
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F

# Module-level noise sigma; training script calls v12_set_sigma().
_NOISE = {'sigma': 0.0}


def set_noise_sigma(sigma: float):
    _NOISE['sigma'] = float(sigma)


def _sigma():
    return _NOISE['sigma']


def sign_ste_noisy(x):
    sigma = _sigma()
    if sigma > 1e-8 and x.requires_grad:
        x_n = x + torch.randn_like(x) * sigma
    else:
        x_n = x
    out = torch.where(x_n >= 0, torch.ones_like(x), -torch.ones_like(x))
    return x + (out - x).detach()


def sign_ste_clipped_noisy(x):
    sigma = _sigma()
    if sigma > 1e-8 and x.requires_grad:
        x_n = x + torch.randn_like(x) * sigma
    else:
        x_n = x
    out = torch.where(x_n >= 0, torch.ones_like(x), -torch.ones_like(x))
    x_clip = torch.clamp(x, -1.0, 1.0)
    return x_clip + (out - x_clip).detach()


def sign_ste_clean(x):
    """Non-noisy sign STE for activations."""
    out = torch.where(x >= 0, torch.ones_like(x), -torch.ones_like(x))
    return x + (out - x).detach()


def sign_ste_clipped_clean(x):
    out = torch.where(x >= 0, torch.ones_like(x), -torch.ones_like(x))
    x_clip = torch.clamp(x, -1.0, 1.0)
    return x_clip + (out - x_clip).detach()


class BitLinearRawN(nn.Module):
    """Weights use noisy STE (for exploration); activations use clean STE."""
    def __init__(self, in_features, out_features, binarize_input=True):
        super().__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.binarize_input = binarize_input
        self.weight = nn.Parameter(torch.randn(out_features, in_features) * 0.02)

    def forward(self, x):
        W = sign_ste_noisy(self.weight)  # noise on weight
        if self.binarize_input:
            x = sign_ste_clipped_clean(x)  # clean STE on activations
        return F.linear(x, W)


class BitLinearN(nn.Module):
    def __init__(self, in_features, out_features, binarize_input=True):
        super().__init__()
        self.raw = BitLinearRawN(in_features, out_features, binarize_input=binarize_input)
        self.threshold = nn.Parameter(torch.zeros(out_features))
        self.scale = 1.0 / math.sqrt(in_features)

    def forward(self, x):
        s = self.raw(x) * self.scale - self.threshold
        return sign_ste_clipped_clean(s)


class BiAttentionN(nn.Module):
    def __init__(self, d_model, n_heads):
        super().__init__()
        assert d_model % n_heads == 0
        self.d_model = d_model
        self.n_heads = n_heads
        self.head_dim = d_model // n_heads
        self.q_proj = BitLinearN(d_model, d_model)
        self.k_proj = BitLinearN(d_model, d_model)
        self.v_proj = BitLinearN(d_model, d_model)
        self.o_proj = BitLinearN(d_model, d_model)
        self.attn_threshold = nn.Parameter(torch.zeros(n_heads))
        slopes = torch.tensor([2.0 ** (i - 2) for i in range(n_heads)])
        self.register_buffer('alibi_slopes', slopes)
        self.register_buffer('_causal_mask', torch.empty(0), persistent=False)

    def _get_mask(self, T, device):
        if self._causal_mask.shape[-1] < T or self._causal_mask.device != device:
            m = torch.triu(torch.ones(T, T, device=device, dtype=torch.bool), diagonal=1)
            self._causal_mask = m
        return self._causal_mask[:T, :T]

    def forward(self, x):
        B, T, D = x.shape
        H, Dh = self.n_heads, self.head_dim
        Q = self.q_proj(x).view(B, T, H, Dh).transpose(1, 2)
        K = self.k_proj(x).view(B, T, H, Dh).transpose(1, 2)
        V = self.v_proj(x).view(B, T, H, Dh).transpose(1, 2)
        scores = torch.matmul(Q, K.transpose(-2, -1)) / math.sqrt(Dh)
        pos = torch.arange(T, device=x.device).float()
        dist = (pos.unsqueeze(0) - pos.unsqueeze(1)).abs()
        alibi_bias = self.alibi_slopes.view(1, H, 1, 1) * dist.view(1, 1, T, T) / math.sqrt(Dh)
        scores = scores - alibi_bias
        mask = self._get_mask(T, x.device)
        scores = scores.masked_fill(mask, -1e9)
        tau = self.attn_threshold.view(1, H, 1, 1)
        A = sign_ste_clipped_clean(scores - tau)
        A = A.masked_fill(mask, -1.0)
        O = torch.matmul(A, V)
        O = O.transpose(1, 2).contiguous().view(B, T, D)
        return self.o_proj(O)


class BitFFNN(nn.Module):
    def __init__(self, d_model, d_ff):
        super().__init__()
        self.gate = BitLinearN(d_model, d_ff)
        self.up = BitLinearN(d_model, d_ff)
        self.down = BitLinearN(d_ff, d_model)

    def forward(self, x):
        return self.down(self.gate(x) * self.up(x))


class BitBlockN(nn.Module):
    def __init__(self, d_model, n_heads, d_ff):
        super().__init__()
        self.attn = BiAttentionN(d_model, n_heads)
        self.ffn = BitFFNN(d_model, d_ff)

    def forward(self, x):
        a = self.attn(x)
        f = self.ffn(x)
        return sign_ste_clean(x + a + f)


class BinaryEmbeddingN(nn.Module):
    def __init__(self, vocab_size, d_model):
        super().__init__()
        self.vocab_size = vocab_size
        self.d_model = d_model
        self.weight = nn.Parameter(torch.randn(vocab_size, d_model) * 0.02)

    def forward(self, idx):
        W = sign_ste_noisy(self.weight)
        return F.embedding(idx, W)


class BitLMv12(nn.Module):
    def __init__(self, vocab_size=128, d_model=256, n_layers=8, n_heads=8, d_ff=512, max_seq_len=256):
        super().__init__()
        self.vocab_size = vocab_size
        self.d_model = d_model
        self.n_layers = n_layers
        self.max_seq_len = max_seq_len
        self.embed = BinaryEmbeddingN(vocab_size, d_model)
        self.blocks = nn.ModuleList([BitBlockN(d_model, n_heads, d_ff) for _ in range(n_layers)])
        self.out_codebook = nn.Parameter(torch.randn(vocab_size, d_model) * 0.02)
        self.logit_scale = nn.Parameter(torch.tensor(1.0 / math.sqrt(d_model)))
        self.out_bias = nn.Parameter(torch.zeros(vocab_size))

    def forward(self, idx, targets=None):
        x = self.embed(idx)
        for blk in self.blocks:
            x = blk(x)
        W_out = sign_ste_noisy(self.out_codebook)
        scores = torch.matmul(x, W_out.t())
        logits = scores * self.logit_scale + self.out_bias
        loss = None
        if targets is not None:
            loss = F.cross_entropy(logits.view(-1, self.vocab_size), targets.view(-1))
        return logits, loss

    @torch.no_grad()
    def generate(self, idx, max_new_tokens=200, temperature=1.0, top_k=None):
        self.eval()
        for _ in range(max_new_tokens):
            idx_cond = idx[:, -self.max_seq_len:]
            logits, _ = self(idx_cond)
            logits = logits[:, -1, :] / max(temperature, 1e-5)
            if top_k is not None:
                v, _ = torch.topk(logits, top_k)
                logits[logits < v[:, [-1]]] = -float('inf')
            probs = F.softmax(logits, dim=-1)
            nxt = torch.multinomial(probs, num_samples=1)
            idx = torch.cat([idx, nxt], dim=1)
        return idx


if __name__ == '__main__':
    set_noise_sigma(0.5)
    m = BitLMv12()
    n = sum(p.numel() for p in m.parameters())
    print(f"v12 params: {n:,} ({n/1e6:.2f}M)")
    x = torch.randint(0, 128, (2, 64))
    y = torch.randint(0, 128, (2, 64))
    logits, loss = m(x, y)
    print("logits:", logits.shape, "loss:", loss.item())
    loss.backward()
    print("backward OK")