src/domain_tokenizer/models/modeling.py

"""
DomainTransformer Model — GPT-style causal decoder for domain token sequences.

Architecture follows:
  - NoPE (no positional encoding) — Kazemnejad et al. 2023 (arXiv:2305.19466)
  - Pre-norm (LayerNorm before attention and FFN) — GPT-2 style
  - F.scaled_dot_product_attention with is_causal=True — auto FlashAttention
  - Weight tying between token embedding and LM head
  - Scaled residual initialization: 1/sqrt(2*N_layers)

Reference sizes (Nubank nuFormer, arXiv:2507.23267):
  - 24M:  6 layers, d=512, 8 heads
  - 330M: 24 layers, d=1024, 16 heads
"""

import math
from typing import Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import PreTrainedModel
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast

from .configuration import DomainTransformerConfig


class DomainTransformerAttention(nn.Module):
    """Multi-head self-attention with NoPE.

    Uses F.scaled_dot_product_attention for automatic FlashAttention/SDPA dispatch.
    No positional encoding — causal masking via is_causal=True.
    """

    def __init__(self, config: DomainTransformerConfig):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.scaling = self.head_dim ** -0.5

        self.q_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
        self.out_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
        self.attn_dropout = config.attention_probs_dropout_prob

    def forward(self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        B, T, C = hidden_states.shape
        q = self.q_proj(hidden_states).view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
        k = self.k_proj(hidden_states).view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
        v = self.v_proj(hidden_states).view(B, T, self.num_heads, self.head_dim).transpose(1, 2)

        # Convert HF-style attention_mask (1=attend, 0=ignore, long) to SDPA format
        sdpa_mask = None
        use_causal = True
        if attention_mask is not None:
            sdpa_mask = attention_mask[:, None, None, :].to(dtype=q.dtype)
            sdpa_mask = (1.0 - sdpa_mask) * torch.finfo(q.dtype).min
            use_causal = False

        attn_out = F.scaled_dot_product_attention(
            q, k, v, attn_mask=sdpa_mask,
            dropout_p=self.attn_dropout if self.training else 0.0,
            is_causal=use_causal, scale=self.scaling,
        )
        attn_out = attn_out.transpose(1, 2).contiguous().reshape(B, T, C)
        return self.out_proj(attn_out)


class DomainTransformerMLP(nn.Module):
    """Two-layer FFN with GELU activation (GPT-2 style)."""

    def __init__(self, config: DomainTransformerConfig):
        super().__init__()
        self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=True)
        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=True)
        self.act = nn.GELU(approximate="tanh")
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        return self.dropout(self.down_proj(self.act(self.up_proj(hidden_states))))


class DomainTransformerBlock(nn.Module):
    """Single transformer block with pre-norm architecture."""

    def __init__(self, config: DomainTransformerConfig):
        super().__init__()
        self.ln_1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.attn = DomainTransformerAttention(config)
        self.ln_2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.mlp = DomainTransformerMLP(config)

    def forward(self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.attn(self.ln_1(hidden_states), attention_mask)
        hidden_states = residual + hidden_states
        residual = hidden_states
        hidden_states = self.mlp(self.ln_2(hidden_states))
        hidden_states = residual + hidden_states
        return hidden_states


class DomainTransformerPreTrainedModel(PreTrainedModel):
    """Base class with weight initialization."""
    config_class = DomainTransformerConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True

    def _init_weights(self, module: nn.Module):
        std = self.config.initializer_range
        if isinstance(module, nn.Linear):
            nn.init.normal_(module.weight, mean=0.0, std=std)
            if module.bias is not None:
                nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            nn.init.normal_(module.weight, mean=0.0, std=std)
            if module.padding_idx is not None:
                nn.init.zeros_(module.weight[module.padding_idx])
        elif isinstance(module, nn.LayerNorm):
            nn.init.zeros_(module.bias)
            nn.init.ones_(module.weight)


class DomainTransformerModel(DomainTransformerPreTrainedModel):
    """The bare DomainTransformer: embeddings + blocks + final layernorm."""

    def __init__(self, config: DomainTransformerConfig):
        super().__init__(config)
        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
        self.embed_dropout = nn.Dropout(config.hidden_dropout_prob)
        self.blocks = nn.ModuleList([DomainTransformerBlock(config) for _ in range(config.num_hidden_layers)])
        self.ln_f = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.gradient_checkpointing = False
        self.post_init()

    def get_input_embeddings(self):
        return self.embed_tokens

    def set_input_embeddings(self, value):
        self.embed_tokens = value

    def forward(self, input_ids=None, attention_mask=None, inputs_embeds=None, **kwargs):
        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)
        hidden_states = self.embed_dropout(inputs_embeds)
        for block in self.blocks:
            if self.gradient_checkpointing and self.training:
                hidden_states = torch.utils.checkpoint.checkpoint(block, hidden_states, attention_mask, use_reentrant=False)
            else:
                hidden_states = block(hidden_states, attention_mask)
        hidden_states = self.ln_f(hidden_states)
        return BaseModelOutputWithPast(last_hidden_state=hidden_states)


class DomainTransformerForCausalLM(DomainTransformerPreTrainedModel):
    """DomainTransformer with a causal language modeling head.

    The LM head is weight-tied with the token embedding layer.
    Loss is computed via standard shifted cross-entropy.
    """
    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}

    def __init__(self, config: DomainTransformerConfig):
        super().__init__(config)
        self.model = DomainTransformerModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.post_init()

    def get_input_embeddings(self):
        return self.model.embed_tokens

    def set_input_embeddings(self, value):
        self.model.embed_tokens = value

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def forward(self, input_ids=None, attention_mask=None, labels=None, inputs_embeds=None, **kwargs):
        outputs = self.model(input_ids=input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds)
        hidden_states = outputs.last_hidden_state
        logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss = F.cross_entropy(shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1), ignore_index=-100)

        return CausalLMOutputWithPast(loss=loss, logits=logits)

    def get_user_embedding(self, input_ids, attention_mask=None):
        """Extract user-level embedding from the last non-padding token."""
        outputs = self.model(input_ids=input_ids, attention_mask=attention_mask)
        hidden_states = outputs.last_hidden_state
        if attention_mask is not None:
            seq_lengths = attention_mask.sum(dim=1) - 1
            batch_idx = torch.arange(hidden_states.size(0), device=hidden_states.device)
            return hidden_states[batch_idx, seq_lengths]
        else:
            return hidden_states[:, -1, :]


DomainTransformerConfig.register_for_auto_class()
DomainTransformerForCausalLM.register_for_auto_class("AutoModelForCausalLM")