modeling_modern_protein.py

"""
ModernProteinLM: A next-generation protein encoder combining:
- ModernBERT architectural improvements (RoPE, Pre-LN, GeGLU, FlashAttention-compatible)
- ELECTRA-style discriminative pre-training
- Deep & narrow design optimal for protein sequences
- Curriculum masking (30% -> 5%)
- Span masking for protein structural motifs

Architecture goals (~150M params):
- 28 layers, hidden 576, heads 9, intermediate 2304 (GeGLU)
- RoPE position embeddings (no absolute PE)
- Pre-LayerNorm with extra LN after embedding
- No dropout (following ESM-2)
- Tied input/output embeddings
"""

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, PretrainedConfig
from transformers.modeling_outputs import MaskedLMOutput, SequenceClassifierOutput, TokenClassifierOutput


class ModernProteinLMConfig(PretrainedConfig):
    model_type = "modern_protein_lm"
    
    def __init__(
        self,
        vocab_size=33,
        hidden_size=576,
        num_hidden_layers=28,
        num_attention_heads=9,
        intermediate_size=2304,
        hidden_act="gelu",
        hidden_dropout_prob=0.0,
        attention_probs_dropout_prob=0.0,
        max_position_embeddings=1026,
        initializer_range=0.02,
        layer_norm_eps=1e-12,
        position_embedding_type="rotary",
        rope_theta=10000.0,
        use_geglu=True,
        tie_word_embeddings=True,
        pad_token_id=1,
        mask_token_id=32,
        cls_token_id=0,
        eos_token_id=2,
        **kwargs,
    ):
        super().__init__(
            pad_token_id=pad_token_id,
            mask_token_id=mask_token_id,
            cls_token_id=cls_token_id,
            eos_token_id=eos_token_id,
            tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.intermediate_size = intermediate_size
        self.hidden_act = hidden_act
        self.hidden_dropout_prob = hidden_dropout_prob
        self.attention_probs_dropout_prob = attention_probs_dropout_prob
        self.max_position_embeddings = max_position_embeddings
        self.initializer_range = initializer_range
        self.layer_norm_eps = layer_norm_eps
        self.position_embedding_type = position_embedding_type
        self.rope_theta = rope_theta
        self.use_geglu = use_geglu


class RotaryEmbedding(nn.Module):
    """RoPE (Rotary Position Embedding) for protein sequences."""
    
    def __init__(self, dim, max_seq_len=1026, base=10000.0, device=None):
        super().__init__()
        self.dim = dim
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.max_seq_len = max_seq_len
        
    def forward(self, seq_len, device):
        t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype)
        freqs = torch.outer(t, self.inv_freq)
        emb = torch.cat((freqs, freqs), dim=-1)
        return emb.cos().to(torch.float32), emb.sin().to(torch.float32)


def rotate_half(x):
    x1, x2 = x.chunk(2, dim=-1)
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(q, k, cos, sin):
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


class ModernProteinAttention(nn.Module):
    """Multi-head attention with RoPE and optional FlashAttention."""
    
    def __init__(self, config: ModernProteinLMConfig):
        super().__init__()
        self.num_heads = config.num_attention_heads
        self.head_dim = config.hidden_size // config.num_attention_heads
        self.scale = self.head_dim ** -0.5
        
        self.qkv_proj = nn.Linear(config.hidden_size, 3 * config.hidden_size, bias=False)
        self.out_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
        
        self.rotary_emb = RotaryEmbedding(self.head_dim, max_seq_len=config.max_position_embeddings, base=config.rope_theta)
        
        self.dropout = nn.Dropout(config.attention_probs_dropout_prob) if config.attention_probs_dropout_prob > 0 else None
        
    def forward(self, hidden_states, attention_mask=None, output_attentions=False):
        batch_size, seq_len, _ = hidden_states.shape
        
        qkv = self.qkv_proj(hidden_states)
        qkv = qkv.reshape(batch_size, seq_len, 3, self.num_heads, self.head_dim)
        qkv = qkv.permute(2, 0, 3, 1, 4)  # (3, B, H, T, D)
        q, k, v = qkv[0], qkv[1], qkv[2]
        
        # Apply RoPE
        cos, sin = self.rotary_emb(seq_len, device=hidden_states.device)
        cos = cos[None, None, :, :]  # (1, 1, T, D)
        sin = sin[None, None, :, :]
        q, k = apply_rotary_pos_emb(q, k, cos, sin)
        
        # Try FlashAttention if available
        try:
            from flash_attn import flash_attn_func
            if attention_mask is None and q.dtype in [torch.float16, torch.bfloat16]:
                attn_output = flash_attn_func(q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), 
                                              dropout_p=self.dropout.p if self.dropout else 0.0,
                                              causal=False)
                attn_output = attn_output.transpose(1, 2)
            else:
                raise ImportError("Fallback to standard attention")
        except (ImportError, AttributeError):
            # Standard scaled dot-product attention
            attn_scores = torch.matmul(q, k.transpose(-2, -1)) * self.scale
            
            if attention_mask is not None:
                attn_scores = attn_scores + attention_mask
            
            attn_probs = F.softmax(attn_scores, dim=-1, dtype=torch.float32).to(q.dtype)
            if self.dropout is not None:
                attn_probs = self.dropout(attn_probs)
            
            attn_output = torch.matmul(attn_probs, v)
        
        attn_output = attn_output.transpose(1, 2).reshape(batch_size, seq_len, -1)
        attn_output = self.out_proj(attn_output)
        
        if output_attentions:
            return attn_output, attn_probs
        return attn_output, None


class GeGLU(nn.Module):
    """GeGLU activation: GELU(gate) * value. More expressive than GELU alone."""
    
    def __init__(self, config: ModernProteinLMConfig):
        super().__init__()
        self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
        self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
        self.act = nn.GELU()
        
    def forward(self, x):
        gate = self.act(self.gate_proj(x))
        up = self.up_proj(x)
        return self.down_proj(gate * up)


class ModernProteinMLP(nn.Module):
    def __init__(self, config: ModernProteinLMConfig):
        super().__init__()
        if config.use_geglu:
            self.mlp = GeGLU(config)
        else:
            self.mlp = nn.Sequential(
                nn.Linear(config.hidden_size, config.intermediate_size, bias=False),
                nn.GELU(),
                nn.Linear(config.intermediate_size, config.hidden_size, bias=False),
            )
    
    def forward(self, x):
        return self.mlp(x)


class ModernProteinLayer(nn.Module):
    """Pre-LN transformer layer with optional parallel formulation."""
    
    def __init__(self, config: ModernProteinLMConfig):
        super().__init__()
        self.ln1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.attn = ModernProteinAttention(config)
        self.ln2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.mlp = ModernProteinMLP(config)
        
    def forward(self, hidden_states, attention_mask=None, output_attentions=False):
        # Pre-LN: LN -> Attn -> Residual
        attn_out, attn_weights = self.attn(self.ln1(hidden_states), attention_mask, output_attentions)
        hidden_states = hidden_states + attn_out
        
        # Pre-LN: LN -> MLP -> Residual
        mlp_out = self.mlp(self.ln2(hidden_states))
        hidden_states = hidden_states + mlp_out
        
        return hidden_states, attn_weights


class ModernProteinLM(PreTrainedModel):
    config_class = ModernProteinLMConfig
    
    def __init__(self, config: ModernProteinLMConfig):
        super().__init__(config)
        self.config = config
        
        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
        self.embed_ln = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        
        self.layers = nn.ModuleList([
            ModernProteinLayer(config) for _ in range(config.num_hidden_layers)
        ])
        
        self.final_ln = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        
        # Initialize weights
        self._init_weights()
        
        # Tie embeddings if requested
        if config.tie_word_embeddings:
            self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
            self.lm_head.weight = self.embeddings.weight
        else:
            self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
    
    def _init_weights(self):
        for module in self.modules():
            if isinstance(module, nn.Linear):
                nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
                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=self.config.initializer_range)
            elif isinstance(module, nn.LayerNorm):
                nn.init.ones_(module.weight)
                nn.init.zeros_(module.bias)
    
    def get_input_embeddings(self):
        return self.embeddings
    
    def set_input_embeddings(self, value):
        self.embeddings = value
    
    def forward(
        self,
        input_ids,
        attention_mask=None,
        position_ids=None,
        labels=None,
        output_attentions=False,
        output_hidden_states=False,
        return_dict=True,
    ):
        batch_size, seq_len = input_ids.shape
        
        # Embedding
        hidden_states = self.embeddings(input_ids)
        hidden_states = self.embed_ln(hidden_states)
        
        # Attention mask for padding
        if attention_mask is not None:
            # (B, T) -> (B, 1, 1, T) for broadcasting
            attention_mask = (1.0 - attention_mask[:, None, None, :]) * -10000.0
        
        all_hidden_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None
        
        # Transformer layers
        for layer in self.layers:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
            
            hidden_states, attn_weights = layer(hidden_states, attention_mask, output_attentions)
            
            if output_attentions:
                all_attentions += (attn_weights,)
        
        hidden_states = self.final_ln(hidden_states)
        
        if output_hidden_states:
            all_hidden_states += (hidden_states,)
        
        # LM head
        logits = self.lm_head(hidden_states)
        
        loss = None
        if labels is not None:
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
        
        if not return_dict:
            output = (logits,)
            if output_hidden_states:
                output += (all_hidden_states,)
            if output_attentions:
                output += (all_attentions,)
            return ((loss,) + output) if loss is not None else output
        
        return MaskedLMOutput(
            loss=loss,
            logits=logits,
            hidden_states=all_hidden_states,
            attentions=all_attentions,
        )
    
    def get_sequence_embedding(self, input_ids, attention_mask=None):
        """Extract CLS or mean-pooled embedding for downstream tasks."""
        outputs = self.forward(
            input_ids=input_ids,
            attention_mask=attention_mask,
            output_hidden_states=True,
            return_dict=True,
        )
        hidden = outputs.hidden_states[-1]
        
        if attention_mask is not None:
            # Mean pool over non-padded positions
            mask_expanded = attention_mask.unsqueeze(-1).float()
            sum_hidden = (hidden * mask_expanded).sum(dim=1)
            pooled = sum_hidden / mask_expanded.sum(dim=1).clamp(min=1e-9)
        else:
            pooled = hidden[:, 0]  # CLS token
        
        return pooled


class ModernProteinLMForMaskedLM(ModernProteinLM):
    """Masked Language Model wrapper."""
    pass


class ModernProteinLMForSequenceClassification(PreTrainedModel):
    config_class = ModernProteinLMConfig
    
    def __init__(self, config: ModernProteinLMConfig):
        super().__init__(config)
        self.modern_protein = ModernProteinLM(config)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
        
    def forward(self, input_ids, attention_mask=None, labels=None):
        pooled = self.modern_protein.get_sequence_embedding(input_ids, attention_mask)
        logits = self.classifier(pooled)
        
        loss = None
        if labels is not None:
            if self.config.num_labels == 1:
                loss_fct = nn.MSELoss()
                loss = loss_fct(logits.squeeze(-1), labels.float())
            else:
                loss_fct = nn.CrossEntropyLoss()
                loss = loss_fct(logits, labels)
        
        return SequenceClassifierOutput(loss=loss, logits=logits)


class ModernProteinLMForTokenClassification(PreTrainedModel):
    config_class = ModernProteinLMConfig
    
    def __init__(self, config: ModernProteinLMConfig):
        super().__init__(config)
        self.modern_protein = ModernProteinLM(config)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
        
    def forward(self, input_ids, attention_mask=None, labels=None):
        outputs = self.modern_protein(
            input_ids=input_ids,
            attention_mask=attention_mask,
            return_dict=True,
        )
        logits = self.classifier(outputs.hidden_states[-1])
        
        loss = None
        if labels is not None:
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
        
        return TokenClassifierOutput(loss=loss, logits=logits)