"""v52: v48 BitNet + standard softmax attention (not Gumbel-argmax).

Our v16→v48 chain has always used Gumbel hard-attention where each query
attends to exactly ONE position. That's a severe restriction — every real
1-bit LLM paper uses vanilla softmax attention on float scores derived
from ±1 QK. Weights stay ±1; attention matrix A is a float softmax over
the ±1-derived integer scores.

Change from v48: gumbel_hard_attention(…) → softmax(scores, dim=-1).
Everything else identical.
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F

from model import sign_ste, sign_ste_clipped, BinaryEmbedding
from model_v47 import RMSNorm, BitLinearScaled, BitLinearScaledRaw, BitFFNScaled


class SoftmaxBinaryAttention(nn.Module):
    """±1 Q/K/V/O weights; vanilla softmax attention over scaled integer scores."""
    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 = BitLinearScaled(d_model, d_model)
        self.k_proj = BitLinearScaled(d_model, d_model)
        self.v_proj = BitLinearScaled(d_model, d_model)
        self.o_proj = BitLinearScaledRaw(d_model, d_model)
        slopes = torch.tensor([1 << i for i in range(n_heads)], dtype=torch.long)
        self.register_buffer('alibi_slopes_int', slopes)
        self.scale = 1.0 / math.sqrt(self.head_dim)

    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)   # ±1
        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)) * self.scale   # float
        pos = torch.arange(T, device=x.device)
        dist = (pos.unsqueeze(0) - pos.unsqueeze(1)).abs()
        alibi = self.alibi_slopes_int.view(1, H, 1, 1).to(scores.dtype) \
                * dist.view(1, 1, T, T).to(scores.dtype) * self.scale
        scores = scores - alibi

        mask = torch.triu(torch.ones(T, T, device=x.device, dtype=torch.bool), diagonal=1)
        scores = scores.masked_fill(mask, float('-inf'))
        A = F.softmax(scores, dim=-1)                                 # float softmax
        O = torch.matmul(A, V)                                        # float
        O = O.transpose(1, 2).contiguous().view(B, T, D)
        return self.o_proj(O)


class BitBlockV52(nn.Module):
    def __init__(self, d_model, n_heads, d_ff):
        super().__init__()
        self.norm1 = RMSNorm(d_model)
        self.attn = SoftmaxBinaryAttention(d_model, n_heads)
        self.norm2 = RMSNorm(d_model)
        self.ffn = BitFFNScaled(d_model, d_ff)

    def forward(self, x):
        x = x + self.attn(self.norm1(x))
        x = x + self.ffn(self.norm2(x))
        return x


class BitLMv52(nn.Module):
    def __init__(self, vocab_size=128, d_model=512, n_layers=4, n_heads=8,
                 d_ff=192, 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 = BinaryEmbedding(vocab_size, d_model)
        self.blocks = nn.ModuleList([
            BitBlockV52(d_model, n_heads, d_ff) for _ in range(n_layers)
        ])
        self.norm_out = RMSNorm(d_model)
        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)
        x = self.norm_out(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


if __name__ == '__main__':
    m = BitLMv52(d_model=512, n_layers=4, d_ff=192)
    n = sum(p.numel() for p in m.parameters())
    print(f'v52 softmax-attn: {n:,} ({n/1e6:.3f}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')