"""v31: Shared-weight recurrent depth.

Physically N distinct blocks; each block is applied K times sequentially.
Effective depth = N·K with only N params. Tests whether discrete forward
passes benefit disproportionately from depth at the same param cost.
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F

from model import sign_ste, BinaryEmbedding
from model_v18 import BitBlockV18, IntBinaryAttention


class BitLMv31(nn.Module):
    def __init__(self, vocab_size=128, d_model=256, n_unique_blocks=4, block_repeat=2,
                 n_heads=8, d_ff=512, max_seq_len=256):
        super().__init__()
        self.vocab_size = vocab_size
        self.d_model = d_model
        self.n_unique_blocks = n_unique_blocks
        self.block_repeat = block_repeat
        self.n_layers = n_unique_blocks * block_repeat  # effective depth
        self.max_seq_len = max_seq_len
        self.embed = BinaryEmbedding(vocab_size, d_model)
        # Only N unique block parameter sets, but we apply each K times.
        self.unique_blocks = nn.ModuleList([
            BitBlockV18(d_model, n_heads, d_ff) for _ in range(n_unique_blocks)
        ])
        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.unique_blocks:
            for _ in range(self.block_repeat):
                x = blk(x)
        W_out = sign_ste(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__':
    from model_v16 import set_gumbel_tau
    set_gumbel_tau(0.5)
    # Match v17's 5M with 4 unique blocks × 2 repeat = 8 effective layers
    m = BitLMv31(vocab_size=128, d_model=256, n_unique_blocks=4, block_repeat=2,
                 n_heads=8, d_ff=512)
    n = sum(p.numel() for p in m.parameters())
    print(f'v31 (4 blocks × 2 repeat): {n:,} params ({n/1e6:.2f}M)')
    x = torch.randint(0, 128, (2, 64))
    y = torch.randint(0, 128, (2, 64))
    logits, loss = m(x, y)
    loss.backward()
    print(f'  loss={loss.item():.3f}, backward OK')