"""v24: Track III.C (integer FFN) + Track IV.B (multi-prototype head) combined.

Stacks the two best non-attention improvements:
  - FFN intermediate = integer clipped to [-B, +B] (from v22)
  - Output head = K ±1 prototypes per char, max over K (from v23)

Weights remain 1-bit ±1. Intermediate FFN activation is 4-bit signed integer.
Attention still Gumbel one-hot. All still integer-only at inference.
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F

from model import sign_ste, sign_ste_clipped, BitLinear, BinaryEmbedding
from model_v18 import IntBinaryAttention
from model_v22 import IntFFN
from model_v16 import set_gumbel_tau


class BitBlockV24(nn.Module):
    def __init__(self, d_model, n_heads, d_ff, B=7):
        super().__init__()
        self.attn = IntBinaryAttention(d_model, n_heads)
        self.ffn = IntFFN(d_model, d_ff, B=B)

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


class BitLMv24(nn.Module):
    def __init__(self, vocab_size=128, d_model=256, n_layers=8, n_heads=8, d_ff=512,
                 max_seq_len=256, B=7, K_proto=8):
        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.B = B
        self.embed = BinaryEmbedding(vocab_size, d_model)
        self.blocks = nn.ModuleList([
            BitBlockV24(d_model, n_heads, d_ff, B=B) for _ in range(n_layers)
        ])
        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)
        scores = torch.einsum('btd,vkd->btvk', x, W_out)
        logits = torch.logsumexp(scores * self.logit_scale, dim=-1) + 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_gumbel_tau(0.5)
    m = BitLMv24(B=7, K_proto=8)
    n = sum(p.numel() for p in m.parameters())
    print(f'v24 B=7 K=8: {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')