str_bamba/bamba.py

from .generation import GenerationMixin
from mamba_ssm.modules.mamba2 import Mamba2
from mamba_ssm.modules.mha import MHA
from mamba_ssm.modules.mlp import GatedMLP
from mamba_ssm.modules.block import Block
from mamba_ssm.models.mixer_seq_simple import _init_weights
from mamba_ssm.utils.hf import load_config_hf, load_state_dict_hf
from .bamba_modules import BertEmbeddings, BertPooler, BertPreTrainingHeads, BlockCrossAttention
from .bamba_config import BambaConfig, BambaEncoderDecoderConfig

try:
    from mamba_ssm.ops.triton.layer_norm import RMSNorm, layer_norm_fn, rms_norm_fn
except ImportError:
    RMSNorm, layer_norm_fn, rms_norm_fn = None, None, None

from typing import List, Optional, Tuple, Union
from collections import namedtuple
import torch.backends.cudnn as cudnn
import math
import random
from functools import partial
import json
import os
import copy
import torch
import torch.nn as nn
import pandas as pd
import numpy as np
import gc
from tqdm import tqdm


def create_block(
    d_model,
    d_intermediate,
    block_class,
    ssm_cfg=None,
    attn_layer_idx=None,
    attn_cfg=None,
    norm_epsilon=1e-5,
    rms_norm=False,
    residual_in_fp32=False,
    fused_add_norm=False,
    layer_idx=None,
    device=None,
    dtype=None,
):
    if ssm_cfg is None:
        ssm_cfg = {}
    if attn_layer_idx is None:
        attn_layer_idx = []
    if attn_cfg is None:
        attn_cfg = {}
    factory_kwargs = {"device": device, "dtype": dtype}
    if layer_idx not in attn_layer_idx:
        # Create a copy of the config to modify
        ssm_cfg = copy.deepcopy(ssm_cfg) if ssm_cfg is not None else {}
        ssm_layer = ssm_cfg.pop("layer", "Mamba1")
        if ssm_layer not in ["Mamba1", "Mamba2"]:
            raise ValueError(f"Invalid ssm_layer: {ssm_layer}, only support Mamba1 and Mamba2")
        mixer_cls = partial(
            Mamba2 if ssm_layer == "Mamba2" else Mamba,
            layer_idx=layer_idx,
            **ssm_cfg,
            **factory_kwargs
        )
    else:
        mixer_cls = partial(MHA, layer_idx=layer_idx, **attn_cfg, **factory_kwargs)
    norm_cls = partial(
        nn.LayerNorm if not rms_norm else RMSNorm, eps=norm_epsilon, **factory_kwargs
    )
    if d_intermediate == 0:
        mlp_cls = nn.Identity
    else:
        mlp_cls = partial(
            GatedMLP, hidden_features=d_intermediate, out_features=d_model, **factory_kwargs
        )
    block = block_class(
        d_model,
        mixer_cls,
        mlp_cls,
        norm_cls=norm_cls,
        fused_add_norm=fused_add_norm,
        residual_in_fp32=residual_in_fp32,
    )
    if isinstance(block, BlockCrossAttention) and factory_kwargs["dtype"] is not None:
        block.encoder_attn.type(factory_kwargs["dtype"]).to(factory_kwargs["device"])
    block.layer_idx = layer_idx
    return block


class BambaMixerModel(nn.Module):
    def __init__(
        self,
        d_model: int,
        n_layer: int,
        d_intermediate: int,
        vocab_size: int,
        max_position_embeddings: int,
        is_decoder: bool = False,
        ssm_cfg=None,
        attn_layer_idx=None,
        attn_cfg=None,
        norm_epsilon: float = 1e-5,
        rms_norm: bool = False,
        initializer_cfg=None,
        fused_add_norm=False,
        residual_in_fp32=False,
        device=None,
        dtype=None,
    ) -> None:
        factory_kwargs = {"device": device, "dtype": dtype}
        super().__init__()
        self.residual_in_fp32 = residual_in_fp32

        self.is_decoder = is_decoder

        if is_decoder:
            self.embedding = nn.Embedding(vocab_size, d_model, **factory_kwargs)
        else:
            self.embedding = BertEmbeddings(vocab_size, d_model, max_position_embeddings, **factory_kwargs)

        # We change the order of residual and layer norm:
        # Instead of LN -> Attn / MLP -> Add, we do:
        # Add -> LN -> Attn / MLP / Mixer, returning both the residual branch (output of Add) and
        # the main branch (output of MLP / Mixer). The model definition is unchanged.
        # This is for performance reason: we can fuse add + layer_norm.
        self.fused_add_norm = fused_add_norm
        if self.fused_add_norm:
            if layer_norm_fn is None or rms_norm_fn is None:
                raise ImportError("Failed to import Triton LayerNorm / RMSNorm kernels")

        if is_decoder:
            block_class = BlockCrossAttention
        else:
            block_class = Block

        self.layers = nn.ModuleList(
            [
                create_block(
                    d_model,
                    d_intermediate=d_intermediate,
                    block_class=block_class,
                    ssm_cfg=ssm_cfg,
                    attn_layer_idx=attn_layer_idx,
                    attn_cfg=attn_cfg,
                    norm_epsilon=norm_epsilon,
                    rms_norm=rms_norm,
                    residual_in_fp32=residual_in_fp32,
                    fused_add_norm=fused_add_norm,
                    layer_idx=i,
                    **factory_kwargs,
                )
                for i in range(n_layer)
            ]
        )

        self.norm_f = (nn.LayerNorm if not rms_norm else RMSNorm)(
            d_model, eps=norm_epsilon, **factory_kwargs
        )

        if not is_decoder:
            self.pooler = BertPooler(d_model, **factory_kwargs)

        self.apply(
            partial(
                _init_weights,
                n_layer=n_layer,
                **(initializer_cfg if initializer_cfg is not None else {}),
                n_residuals_per_layer=1 if d_intermediate == 0 else 2,  # 2 if we have MLP
            )
        )

    def allocate_inference_cache(self, batch_size, max_seqlen, dtype=None, **kwargs):
        return {
            i: layer.allocate_inference_cache(batch_size, max_seqlen, dtype=dtype, **kwargs)
            for i, layer in enumerate(self.layers)
        }

    def forward(self, input_ids, token_type_ids=None, inference_params=None, encoder_hidden_states=None, attention_mask=None, **mixer_kwargs):
        if self.is_decoder:
            hidden_states = self.embedding(input_ids)
        else:
            hidden_states = self.embedding(input_ids, token_type_ids)
        residual = None
        for layer in self.layers:
            if self.is_decoder:
                hidden_states, residual = layer(
                    hidden_states, residual, inference_params=inference_params, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask, **mixer_kwargs
                )
            else:
                hidden_states, residual = layer(
                    hidden_states, residual, inference_params=inference_params, **mixer_kwargs
                )
        if not self.fused_add_norm:
            residual = (hidden_states + residual) if residual is not None else hidden_states
            hidden_states = self.norm_f(residual.to(dtype=self.norm_f.weight.dtype))
        else:
            # Set prenorm=False here since we don't need the residual
            hidden_states = layer_norm_fn(
                hidden_states,
                self.norm_f.weight,
                self.norm_f.bias,
                eps=self.norm_f.eps,
                residual=residual,
                prenorm=False,
                residual_in_fp32=self.residual_in_fp32,
                is_rms_norm=isinstance(self.norm_f, RMSNorm)
            )
        if not self.is_decoder:
            pooled_output = self.pooler(hidden_states)
            return hidden_states, pooled_output
        return hidden_states


class BambaEncoder(nn.Module):

    def __init__(
        self,
        config: BambaConfig,
        initializer_cfg=None,
        device=None,
        dtype=None,
    ) -> None:
        self.config = config
        d_model = config.d_model
        n_layer = config.n_layer
        d_intermediate = config.d_intermediate
        vocab_size = config.vocab_size
        max_position_embeddings = config.max_position_embeddings
        ssm_cfg = config.ssm_cfg
        attn_layer_idx = config.attn_layer_idx
        attn_cfg = config.attn_cfg
        rms_norm = config.rms_norm
        residual_in_fp32 = config.residual_in_fp32
        fused_add_norm = config.fused_add_norm
        pad_vocab_size_multiple = config.pad_vocab_size_multiple
        factory_kwargs = {"device": device, "dtype": dtype}

        super().__init__()
        if vocab_size % pad_vocab_size_multiple != 0:
            vocab_size += pad_vocab_size_multiple - (vocab_size % pad_vocab_size_multiple)
        self.backbone = BambaMixerModel(
            d_model=d_model,
            n_layer=n_layer,
            d_intermediate=d_intermediate,
            vocab_size=vocab_size,
            max_position_embeddings=max_position_embeddings,
            is_decoder=False,
            ssm_cfg=ssm_cfg,
            attn_layer_idx=attn_layer_idx,
            attn_cfg=attn_cfg,
            rms_norm=rms_norm,
            initializer_cfg=initializer_cfg,
            fused_add_norm=fused_add_norm,
            residual_in_fp32=residual_in_fp32,
            **factory_kwargs,
        )
        self.cls = BertPreTrainingHeads(vocab_size, d_model, **factory_kwargs)

        # Initialize weights and apply final processing
        self.apply(
            partial(
                _init_weights,
                n_layer=n_layer,
                **(initializer_cfg if initializer_cfg is not None else {}),
            )
        )
        self.tie_weights()

    def tie_weights(self):
        if self.config.tie_embeddings:
            self.lm_head.weight = self.backbone.embedding.weight

    def allocate_inference_cache(self, batch_size, max_seqlen, dtype=None, **kwargs):
        return self.backbone.allocate_inference_cache(batch_size, max_seqlen, dtype=dtype, **kwargs)

    def forward(self, input_ids, token_type_ids=None, position_ids=None, inference_params=None, num_last_tokens=0, **mixer_kwargs):
        """
        "position_ids" is just to be compatible with Transformer generation. We don't use it.
        num_last_tokens: if > 0, only return the logits for the last n tokens
        """
        hidden_states, pooled_output = self.backbone(input_ids, token_type_ids, inference_params=inference_params, **mixer_kwargs)
        if num_last_tokens > 0:
            hidden_states = hidden_states[:, -num_last_tokens:]
        lm_logits, seq_relationship_score = self.cls(hidden_states, pooled_output)
        CausalLMOutput = namedtuple("CausalLMOutput", ["logits", "seq_relationship_logits", "hidden_states"])
        return CausalLMOutput(logits=lm_logits, seq_relationship_logits=seq_relationship_score, hidden_states=hidden_states)

    @classmethod
    def from_pretrained(cls, pretrained_model_name, device=None, dtype=None, **kwargs):
        config_data = load_config_hf(pretrained_model_name)
        config = MambaConfig(**config_data)
        model = cls(config, device=device, dtype=dtype, **kwargs)
        model.load_state_dict(load_state_dict_hf(pretrained_model_name, device=device, dtype=dtype))
        return model

    def save_pretrained(self, save_directory):
        """
        Minimal implementation of save_pretrained for MambaLMHeadModel.
        Save the model and its configuration file to a directory.
        """
        # Ensure save_directory exists
        os.makedirs(save_directory, exist_ok=True)

        # Save the model's state_dict
        model_path = os.path.join(save_directory, 'pytorch_model.bin')
        torch.save(self.state_dict(), model_path)

        # Save the configuration of the model
        config_path = os.path.join(save_directory, 'config.json')
        with open(config_path, 'w') as f:
            json.dump(self.config.__dict__, f, indent=4)


class BambaDecoder(nn.Module, GenerationMixin):

    def __init__(
        self,
        config: BambaConfig,
        initializer_cfg=None,
        device=None,
        dtype=None,
    ) -> None:
        self.config = config
        d_model = config.d_model
        n_layer = config.n_layer
        d_intermediate = config.d_intermediate
        vocab_size = config.vocab_size
        max_position_embeddings = config.max_position_embeddings
        ssm_cfg = config.ssm_cfg
        attn_layer_idx = config.attn_layer_idx
        attn_cfg = config.attn_cfg
        rms_norm = config.rms_norm
        residual_in_fp32 = config.residual_in_fp32
        fused_add_norm = config.fused_add_norm
        pad_vocab_size_multiple = config.pad_vocab_size_multiple
        factory_kwargs = {"device": device, "dtype": dtype}

        super().__init__()
        if vocab_size % pad_vocab_size_multiple != 0:
            vocab_size += pad_vocab_size_multiple - (vocab_size % pad_vocab_size_multiple)
        self.backbone = BambaMixerModel(
            d_model=d_model,
            n_layer=n_layer,
            d_intermediate=d_intermediate,
            vocab_size=vocab_size,
            max_position_embeddings=max_position_embeddings,
            is_decoder=True,
            ssm_cfg=ssm_cfg,
            attn_layer_idx=attn_layer_idx,
            attn_cfg=attn_cfg,
            rms_norm=rms_norm,
            initializer_cfg=initializer_cfg,
            fused_add_norm=fused_add_norm,
            residual_in_fp32=residual_in_fp32,
            **factory_kwargs,
        )
        self.lm_head = nn.Linear(d_model, vocab_size, bias=False, **factory_kwargs)

        # Initialize weights and apply final processing
        self.apply(
            partial(
                _init_weights,
                n_layer=n_layer,
                **(initializer_cfg if initializer_cfg is not None else {}),
            )
        )
        self.tie_weights()

    def tie_weights(self):
        if self.config.tie_embeddings:
            self.lm_head.weight = self.backbone.embedding.weight

    def allocate_inference_cache(self, batch_size, max_seqlen, dtype=None, **kwargs):
        return self.backbone.allocate_inference_cache(batch_size, max_seqlen, dtype=dtype, **kwargs)

    def forward(self, input_ids, token_type_ids=None, position_ids=None, inference_params=None, num_last_tokens=0, encoder_hidden_states=None, attention_mask=None, **mixer_kwargs):
        """
        "position_ids" is just to be compatible with Transformer generation. We don't use it.
        num_last_tokens: if > 0, only return the logits for the last n tokens
        """
        hidden_states = self.backbone(
            input_ids, token_type_ids, inference_params=inference_params, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask, **mixer_kwargs
        )
        if num_last_tokens > 0:
            hidden_states = hidden_states[:, -num_last_tokens:]
        lm_logits = self.lm_head(hidden_states)
        CausalLMOutput = namedtuple("CausalLMOutput", ["logits"])
        return CausalLMOutput(logits=lm_logits)

    @classmethod
    def from_pretrained(cls, pretrained_model_name, device=None, dtype=None, **kwargs):
        config_data = load_config_hf(pretrained_model_name)
        config = MambaConfig(**config_data)
        model = cls(config, device=device, dtype=dtype, **kwargs)
        model.load_state_dict(load_state_dict_hf(pretrained_model_name, device=device, dtype=dtype))
        return model

    def save_pretrained(self, save_directory):
        """
        Minimal implementation of save_pretrained for MambaLMHeadModel.
        Save the model and its configuration file to a directory.
        """
        # Ensure save_directory exists
        os.makedirs(save_directory, exist_ok=True)

        # Save the model's state_dict
        model_path = os.path.join(save_directory, 'pytorch_model.bin')
        torch.save(self.state_dict(), model_path)

        # Save the configuration of the model
        config_path = os.path.join(save_directory, 'config.json')
        with open(config_path, 'w') as f:
            json.dump(self.config.__dict__, f, indent=4)


class BambaEncoderDecoder(nn.Module, GenerationMixin):

    def __init__(
        self,
        config: BambaEncoderDecoderConfig,
        tokenizer=None,
        initializer_cfg=None,
        device=None,
        dtype=None,
    ) -> None:
        self.config = config
        self.encoder_config = config.encoder_config
        self.decoder_config = config.decoder_config
        factory_kwargs = {"device": device, "dtype": dtype}
        self.tokenizer = tokenizer

        super().__init__()
        self.encoder = BambaEncoder(self.encoder_config, **factory_kwargs)
        self.decoder = BambaDecoder(self.decoder_config, **factory_kwargs)

        self.device = device

        self.tie_weights()
        self._set_seed(config.seed)

    def tie_weights(self):
        if self.config.tie_word_embeddings:
            self.decoder.backbone.embedding.weight = self.encoder.backbone.embedding.word_embeddings.weight

    def allocate_inference_cache(self, batch_size, max_seqlen, dtype=None, **kwargs):
        return self.decoder.backbone.allocate_inference_cache(batch_size, max_seqlen, dtype=dtype, **kwargs)

    def forward(self, encoder_input_ids, decoder_input_ids, token_type_ids=None, attention_mask=None, position_ids=None, inference_params=None, num_last_tokens=0, **mixer_kwargs):
        """
        "position_ids" is just to be compatible with Transformer generation. We don't use it.
        num_last_tokens: if > 0, only return the logits for the last n tokens
        """
        encoder_hidden_states = self.encoder(encoder_input_ids, inference_params=inference_params, **mixer_kwargs).hidden_states
        lm_logits = self.decoder(decoder_input_ids, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask, inference_params=inference_params, **mixer_kwargs).logits
        if num_last_tokens > 0:
            hidden_states = hidden_states[:, -num_last_tokens:]
        CausalLMOutput = namedtuple("CausalLMOutput", ["logits"])
        return CausalLMOutput(logits=lm_logits)

    @classmethod
    def from_pretrained(cls, pretrained_model_name, device=None, dtype=None, **kwargs):
        config_data = load_config_hf(pretrained_model_name)
        config = MambaConfig(**config_data)
        model = cls(config, device=device, dtype=dtype, **kwargs)
        model.load_state_dict(load_state_dict_hf(pretrained_model_name, device=device, dtype=dtype))
        return model

    def save_pretrained(self, save_directory):
        """
        Minimal implementation of save_pretrained for MambaLMHeadModel.
        Save the model and its configuration file to a directory.
        """
        # Ensure save_directory exists
        os.makedirs(save_directory, exist_ok=True)

        # Save the model's state_dict
        model_path = os.path.join(save_directory, 'pytorch_model.bin')
        torch.save(self.state_dict(), model_path)

        # Save the configuration of the model
        config_path = os.path.join(save_directory, 'config.json')
        with open(config_path, 'w') as f:
            json.dump(self.config.__dict__, f, indent=4)

    def _set_seed(self, value):
        print('Random Seed:', value)
        random.seed(value)
        torch.manual_seed(value)
        torch.cuda.manual_seed(value)
        torch.cuda.manual_seed_all(value)
        np.random.seed(value)
        cudnn.deterministic = True
        cudnn.benchmark = False

    def extract_embeddings(self, smiles):
        tokens = self.tokenizer(smiles, padding=True, truncation=True, return_tensors='pt')
        
        idx = tokens['input_ids'].to(self.device)
        mask = tokens['attention_mask'].to(self.device)
        outputs = self.encoder(input_ids=idx)
        hidden_states = outputs.hidden_states

        token_embeddings = hidden_states
        input_mask_expanded = mask.unsqueeze(-1).expand(token_embeddings.size()).float()
        sum_embeddings = torch.sum(token_embeddings * input_mask_expanded, 1)
        sum_mask = torch.clamp(input_mask_expanded.sum(1), min=1e-9)
        embeddings = sum_embeddings / sum_mask

        return embeddings

    def encode(self, smiles, useCuda=False, batch_size=100, return_torch=False):
        """Extract efficiently SMILES embeddings per batches."""
        # TODO: remove useCuda argument

        # handle single str or a list of str
        smiles = pd.Series(smiles) if isinstance(smiles, str) else pd.Series(list(smiles))

        # process in batches
        n_split = smiles.shape[0] // batch_size if smiles.shape[0] >= batch_size else smiles.shape[0]
        embeddings = [
            self.extract_embeddings(list(batch)).cpu().detach().numpy() 
                for batch in tqdm(np.array_split(smiles, n_split))
        ]
        flat_list = [item for sublist in embeddings for item in sublist]

        # clear GPU memory
        torch.cuda.empty_cache()
        gc.collect()

        if return_torch:
            return torch.tensor(flat_list)
        return pd.DataFrame(flat_list)