beeper_model.py

"""
beeper_model.py - Core model module for Beeper
Extracted from the training code for use in inference/apps
"""

import os
import re
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional
from safetensors.torch import load_file as load_safetensors

# =========================================================================================
# Model Components
# =========================================================================================

class CausalSelfAttention(nn.Module):
    def __init__(self, dim: int, n_heads: int, attn_dropout: float = 0.0):
        super().__init__()
        assert dim % n_heads == 0
        self.nh = n_heads
        self.hd = dim // n_heads
        self.qkv = nn.Linear(dim, 3 * dim, bias=False)
        self.proj = nn.Linear(dim, dim, bias=False)
        self.attn_dropout = attn_dropout

    def forward(self, x):
        B, T, C = x.shape
        qkv = self.qkv(x)
        q, k, v = qkv.chunk(3, dim=-1)
        q = q.view(B, T, self.nh, self.hd).transpose(1, 2)
        k = k.view(B, T, self.nh, self.hd).transpose(1, 2)
        v = v.view(B, T, self.nh, self.hd).transpose(1, 2)

        # Use scaled_dot_product_attention when available
        y = F.scaled_dot_product_attention(
            q, k, v,
            is_causal=True,
            dropout_p=self.attn_dropout if self.training else 0.0,
        )
        
        y = y.transpose(1, 2).contiguous().view(B, T, C)
        return self.proj(y)


class MLP(nn.Module):
    def __init__(self, dim, mlp_ratio=4.0, dropout=0.1):
        super().__init__()
        hidden = int(dim * mlp_ratio)
        self.fc1 = nn.Linear(dim, hidden)
        self.fc2 = nn.Linear(hidden, dim)
        self.drop = nn.Dropout(dropout)
    
    def forward(self, x):
        x = self.fc1(x)
        x = F.gelu(x, approximate="tanh")
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x


class BeeperRoseGPT(nn.Module):
    def __init__(self, cfg: dict):
        super().__init__()
        V = cfg.get("vocab_size", 8192)
        D = cfg.get("dim", 512)
        Ctx = cfg.get("context", 512)
        H = cfg.get("n_heads", 8)
        L = cfg.get("n_layers", 6)
        MR = cfg.get("mlp_ratio", 4.0)
        RD = cfg.get("resid_dropout", 0.1)
        AD = cfg.get("dropout", 0.0)

        self.vocab_size = V
        self.context = Ctx
        
        # Core transformer components
        self.token_emb = nn.Embedding(V, D)
        self.pos_emb = nn.Parameter(torch.zeros(1, Ctx, D))
        self.drop = nn.Dropout(RD)

        self.blocks = nn.ModuleList([
            nn.ModuleDict({
                "norm1": nn.LayerNorm(D),
                "attn": CausalSelfAttention(D, H, attn_dropout=AD),
                "norm2": nn.LayerNorm(D),
                "mlp": MLP(D, mlp_ratio=MR, dropout=RD),
            }) for _ in range(L)
        ])
        
        self.norm = nn.LayerNorm(D)
        self.lm_head = nn.Linear(D, V, bias=False)
        
        # Weight tying
        self.lm_head.weight = self.token_emb.weight

        # Rose components (for compatibility, may not be used in inference)
        self.rose_proj = nn.Linear(D, D, bias=False)
        self.rose_anchors = nn.Parameter(torch.randn(3, D) / (D**0.5))

        # Pentachora placeholders (not needed for inference but for weight compatibility)
        self.register_buffer("pent_inited", torch.tensor(0, dtype=torch.uint8), persistent=False)
        self.penta_coarse = None
        self.penta_medium = None
        self.penta_fine = None

        self.apply(self._init)
        self.grad_checkpoint = False

    @staticmethod
    def _init(m):
        if isinstance(m, nn.Linear):
            nn.init.normal_(m.weight, mean=0.0, std=0.02)
            if m.bias is not None:
                nn.init.zeros_(m.bias)
        elif isinstance(m, nn.Embedding):
            nn.init.normal_(m.weight, mean=0.0, std=0.02)

    def _block_forward(self, blk, x):
        x = x + blk["attn"](blk["norm1"](x))
        x = x + blk["mlp"](blk["norm2"](x))
        return x

    def backbone(self, idx):
        B, T = idx.shape
        x = self.token_emb(idx) + self.pos_emb[:, :T, :]
        x = self.drop(x)
        
        for blk in self.blocks:
            x = self._block_forward(blk, x)
        
        return self.norm(x)

    def forward(self, idx):
        h = self.backbone(idx)
        return self.lm_head(h)

    def hidden_states(self, idx):
        return self.backbone(idx)

    def load_state_dict(self, state_dict, strict=False):
        """Custom load that handles pentachora bank initialization gracefully"""
        # Clean state dict keys
        cleaned = {}
        for k, v in state_dict.items():
            if k.startswith("_orig_mod."):
                k = k[10:]
            if k.startswith("module."):
                k = k[7:]
            cleaned[k] = v
        
        # Initialize pentachora if present in checkpoint
        if "penta_coarse" in cleaned:
            self.penta_coarse = nn.Parameter(cleaned["penta_coarse"])
        if "penta_medium" in cleaned:
            self.penta_medium = nn.Parameter(cleaned["penta_medium"])
        if "penta_fine" in cleaned:
            self.penta_fine = nn.Parameter(cleaned["penta_fine"])
        
        return super().load_state_dict(cleaned, strict=strict)


# =========================================================================================
# Generation
# =========================================================================================

def _detokenize(text: str) -> str:
    """Clean up tokenization artifacts"""
    text = re.sub(r"\s+([,.;:!?%])", r"\1", text)
    text = re.sub(r"\s+([\)\]\}])", r"\1", text)
    text = re.sub(r"([\(\[\{])\s+", r"\1", text)
    return text


@torch.no_grad()
def generate(
    model: BeeperRoseGPT,
    tok,  # Tokenizer
    cfg: dict,
    prompt: str,
    max_new_tokens: int = 120,
    temperature: float = None,
    top_k: int = None,
    top_p: float = None,
    repetition_penalty: float = None,
    presence_penalty: float = None,
    frequency_penalty: float = None,
    device: Optional[torch.device] = None,
    detokenize: bool = True
) -> str:
    """
    Generate text from Beeper model with various sampling strategies.
    """
    # Use defaults from config if not specified
    temperature = temperature if temperature is not None else cfg.get("temperature", 0.9)
    top_k = top_k if top_k is not None else cfg.get("top_k", 40)
    top_p = top_p if top_p is not None else cfg.get("top_p", 0.9)
    repetition_penalty = repetition_penalty if repetition_penalty is not None else cfg.get("repetition_penalty", 1.1)
    presence_penalty = presence_penalty if presence_penalty is not None else cfg.get("presence_penalty", 0.6)
    frequency_penalty = frequency_penalty if frequency_penalty is not None else cfg.get("frequency_penalty", 0.0)

    device = device or next(model.parameters()).device
    model.eval()
    
    # Encode prompt
    ids = tok.encode(prompt).ids
    x = torch.tensor([ids], dtype=torch.long, device=device)
    
    # Track token frequencies for penalties
    vocab_size = cfg.get("vocab_size", 8192)
    counts = torch.zeros(vocab_size, dtype=torch.int32, device=device)
    for t in ids:
        if 0 <= t < vocab_size:
            counts[t] += 1

    # Generate tokens
    for _ in range(max_new_tokens):
        # Get logits for next token
        context_window = cfg.get("context", 512)
        logits = model(x[:, -context_window:])
        logits = logits[:, -1, :]

        # Apply repetition penalty
        if repetition_penalty and repetition_penalty != 1.0:
            mask = counts > 0
            if mask.any():
                pos = logits[:, mask] > 0
                logits[:, mask][pos] /= repetition_penalty
                logits[:, mask][~pos] *= repetition_penalty

        # Apply presence and frequency penalties
        if presence_penalty or frequency_penalty:
            pen = counts.float() * (frequency_penalty or 0.0) + (counts > 0).float() * (presence_penalty or 0.0)
            logits = logits - pen.unsqueeze(0)

        # Temperature scaling
        logits = logits / max(1e-8, temperature)

        # Top-k filtering
        if top_k and top_k > 0:
            k = min(top_k, logits.size(-1))
            v, ix = torch.topk(logits, k, dim=-1)
            filt = torch.full_like(logits, float("-inf"))
            logits = filt.scatter_(-1, ix, v)

        # Top-p (nucleus) filtering
        if top_p and top_p < 1.0:
            sorted_logits, sorted_indices = torch.sort(logits, descending=True)
            probs = F.softmax(sorted_logits, dim=-1)
            cumulative_probs = torch.cumsum(probs, dim=-1)
            
            # Find cutoff
            cutoff_idx = (cumulative_probs > top_p).float().argmax(dim=-1)
            mask = torch.arange(logits.size(-1), device=device).unsqueeze(0) > cutoff_idx.unsqueeze(-1)
            sorted_logits = sorted_logits.masked_fill(mask, float("-inf"))
            logits = torch.full_like(logits, float("-inf")).scatter(-1, sorted_indices, sorted_logits)

        # Sample next token
        probs = F.softmax(logits, dim=-1)
        next_id = torch.multinomial(probs, num_samples=1)
        
        # Append to sequence
        x = torch.cat([x, next_id], dim=1)
        counts[next_id.item()] += 1

    # Decode output
    output = tok.decode(x[0].tolist())
    return _detokenize(output) if detokenize else output