"""v23: Track IV.B — multi-prototype output head.

Standard v18 head: logit[c] = popcount(h ⊕ embed[c])  # single ±1 prototype per char
v23 head: logit[c] = max_k popcount(h ⊕ proto[c, k])  # K prototypes per char

The max-over-k captures multi-modal character distributions that a single ±1
prototype cannot represent. Still pure-integer: each popcount is a standard
XNOR-popcount, max is an integer compare tree. Inference cost: K× more
popcounts at the head, negligible because head is ~1% of FLOPs.

Training: use log-sum-exp as a soft-max at train time (collapses to max at τ→0
with the annealed Gumbel temperature we already use).
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F

from model import sign_ste, sign_ste_clipped, BitLinear, BitFFN, BinaryEmbedding
from model_v18 import IntBinaryAttention
from model_v16 import set_gumbel_tau


class BitBlockV23(nn.Module):
    def __init__(self, d_model, n_heads, d_ff):
        super().__init__()
        self.attn = IntBinaryAttention(d_model, n_heads)
        self.ffn = BitFFN(d_model, d_ff)  # standard v18 FFN

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


class BitLMv23(nn.Module):
    def __init__(self, vocab_size=128, d_model=256, n_layers=8, n_heads=8, d_ff=512,
                 max_seq_len=256, K_proto=4):
        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.K = K_proto
        self.embed = BinaryEmbedding(vocab_size, d_model)
        self.blocks = nn.ModuleList([
            BitBlockV23(d_model, n_heads, d_ff) for _ in range(n_layers)
        ])
        # Multi-prototype output: (vocab_size, K, d_model) ±1 via sign_ste
        self.out_codebook = nn.Parameter(torch.randn(vocab_size, K_proto, 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(self.out_codebook)  # (V, K, D)
        # Scores: (B, T, D) × (V, K, D) -> (B, T, V, K)
        scores = torch.einsum('btd,vkd->btvk', x, W_out)
        # Soft-max at train (smooth over K), hard-max at inference-eval.
        # Using logsumexp with a learned inverse-temperature eases training.
        # Collapses to max as the network matures (the learned scale grows).
        scaled = scores * self.logit_scale  # (B, T, V, K)
        # Use logsumexp over K dim: logsumexp ≈ max when scaled values are peaked.
        logits = torch.logsumexp(scaled, dim=-1) + self.out_bias  # (B, T, V)
        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 forward_eval_argmax(self, idx):
        """Hard-max variant for inference — pure integer."""
        x = self.embed(idx)
        for blk in self.blocks:
            x = blk(x)
        W_out = sign_ste(self.out_codebook)
        scores = torch.einsum('btd,vkd->btvk', x, W_out)  # integer popcount
        best_over_k, _ = scores.max(dim=-1)  # (B, T, V)
        return best_over_k

    @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_gumbel_tau(0.5)
    for K in [2, 4, 8]:
        m = BitLMv23(K_proto=K)
        n = sum(p.numel() for p in m.parameters())
        print(f'v23 K={K}: {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')