modeling_nvembed.py

from typing import List, Union, Dict, Mapping, Optional, Tuple, TypedDict
import torch
import os
import json
import numpy as np
from functools import partial
from contextlib import nullcontext
from transformers import AutoModel, PreTrainedTokenizerFast, BatchEncoding, DataCollatorWithPadding
from transformers.modeling_utils import PreTrainedModel
from transformers.models.auto import AutoTokenizer
from transformers.models.mistral.modeling_mistral import MISTRAL_INPUTS_DOCSTRING
from transformers.modeling_outputs import BaseModelOutputWithPast
from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_attention_mask_for_sdpa
from transformers import MistralModel, MistralConfig
from transformers.cache_utils import Cache, DynamicCache
from transformers.utils import (
    add_start_docstrings_to_model_forward,
    logging,
)
from einops import rearrange, repeat
from tqdm.auto import tqdm
from datasets import Dataset
from torch.utils.data import DataLoader
from .configuration_nvembed import NVEmbedConfig, LatentAttentionConfig, BidirectionalMistralConfig

logger = logging.get_logger(__name__)

class NVEmbedFeatures(TypedDict):
    input_dict: torch.Tensor
    attention_mask: torch.Tensor
    pool_mask: torch.Tensor

class BidirectionalMistralModel(MistralModel):
    config_class = BidirectionalMistralConfig
    
    def __init__(self, config: MistralConfig):
        super().__init__(config)
        for layer in self.layers:
            layer.self_attn.is_causal = False
        self._attn_implementation = "eager"

    @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
        elif input_ids is not None:
            batch_size, seq_length = input_ids.shape
        elif inputs_embeds is not None:
            batch_size, seq_length, _ = inputs_embeds.shape
        else:
            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        past_key_values_length = 0

        if use_cache:
            use_legacy_cache = not isinstance(past_key_values, Cache)
            if use_legacy_cache:
                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
            past_key_values_length = past_key_values.get_usable_length(seq_length)

        if position_ids is None:
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(
                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
            )
            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
        else:
            position_ids = position_ids.view(-1, seq_length).long()

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
            is_padding_right = attention_mask[:, -1].sum().item() != batch_size
            if is_padding_right:
                raise ValueError(
                    "You are attempting to perform batched generation with padding_side='right'"
                    " this may lead to unexpected behaviour for Flash Attention version of Mistral. Make sure to "
                    " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
                )

        if self._attn_implementation == "flash_attention_2":
            # 2d mask is passed through the layers
            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
        elif self._attn_implementation == "sdpa" and not output_attentions:
            # output_attentions=True can not be supported when using SDPA, and we fall back on
            # the manual implementation that requires a 4D causal mask in all cases.
            attention_mask = _prepare_4d_attention_mask_for_sdpa(
                attention_mask, inputs_embeds.dtype
            )
        else:
            # 4d mask is passed through the layers
            attention_mask = _prepare_4d_attention_mask(
                attention_mask, inputs_embeds.dtype,
            )

        hidden_states = inputs_embeds

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        next_decoder_cache = None

        for decoder_layer in self.layers:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    decoder_layer.__call__,
                    hidden_states,
                    attention_mask,
                    position_ids,
                    past_key_values,
                    output_attentions,
                    use_cache,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                )

            hidden_states = layer_outputs[0]

            if use_cache:
                next_decoder_cache = layer_outputs[2 if output_attentions else 1]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

        hidden_states = self.norm(hidden_states)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        next_cache = None
        if use_cache:
            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache

        if not return_dict:
            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )
    
def _move_to_device(maybe_tensor, device: torch.device):
    if torch.is_tensor(maybe_tensor):
        return maybe_tensor.to(device, non_blocking=device.type == "cuda")
    elif isinstance(maybe_tensor, dict):
        return {key: _move_to_device(value, device) for key, value in maybe_tensor.items()}
    elif isinstance(maybe_tensor, list):
        return [_move_to_device(x, device) for x in maybe_tensor]
    elif isinstance(maybe_tensor, tuple):
        return tuple([_move_to_device(x, device) for x in maybe_tensor])
    elif isinstance(maybe_tensor, Mapping):
        return type(maybe_tensor)({k: _move_to_device(v, device) for k, v in maybe_tensor.items()})
    else:
        return maybe_tensor

def move_to_device(sample, device: torch.device):
    if device.type == "cpu":
        return sample
    
    if len(sample) == 0:
        return {}
    return _move_to_device(sample, device)


def input_transform_func(
    tokenizer: PreTrainedTokenizerFast,
    examples: Dict[str, List],
    always_add_eos: bool,
    max_length: int,
    instruction: str,
) -> BatchEncoding:
    if always_add_eos:
        examples['input_texts'] = [instruction + input_example + tokenizer.eos_token for input_example in examples['input_texts']]
    batch_dict = tokenizer(
        examples['input_texts'],
        max_length=max_length,
        padding=True,
        return_token_type_ids=False,
        return_tensors="pt",
        truncation=True)
    return batch_dict


class PreNorm(torch.nn.Module):
    def __init__(self, dim, fn, context_dim = None):
        super().__init__()
        self.fn = fn
        self.norm = torch.nn.LayerNorm(dim)
        self.norm_context = torch.nn.LayerNorm(context_dim) if exists(context_dim) else None

    def forward(self, x, **kwargs):
        x = self.norm(x)
        if exists(self.norm_context):
            context = kwargs['context']
            normed_context = self.norm_context(context)
            kwargs.update(context = normed_context)
        return self.fn(x, **kwargs)

class GEGLU(torch.nn.Module):
    def forward(self, x):
        x, gates = x.chunk(2, dim = -1)
        return x * torch.nn.functional.gelu(gates)

class FeedForward(torch.nn.Module):
    def __init__(self, dim, mult = 4):
        super().__init__()
        self.net = torch.nn.Sequential(torch.nn.Linear(dim, dim * mult * 2),
            GEGLU(),
            torch.nn.Linear(dim * mult, dim))

    def forward(self, x):
        return self.net(x)

def exists(val):
    return val is not None

def default(val, d):
    return val if exists(val) else d


class Attention(torch.nn.Module):
    def __init__(self, query_dim, context_dim = None, heads = 8, dim_head = 64):
        super().__init__()
        inner_dim = dim_head * heads
        context_dim = default(context_dim, query_dim)
        self.scale = dim_head ** -0.5
        self.heads = heads

        self.to_q = torch.nn.Linear(query_dim, inner_dim, bias = False)
        self.to_kv = torch.nn.Linear(context_dim, inner_dim * 2, bias = False)
        self.to_out = torch.nn.Linear(inner_dim, query_dim, bias = False)

    def forward(self, x, context = None, mask = None):
        h = self.heads
        q = self.to_q(x)
        context = default(context, x)
        k, v = self.to_kv(context).chunk(2, dim = -1)
        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h = h), (q, k, v))
        with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_mem_efficient=True):
            out = torch.nn.functional.scaled_dot_product_attention(q, k, v)
        out = rearrange(out, '(b h) n d -> b n (h d)', h = h)
        return self.to_out(out)


class LatentAttentionModel(PreTrainedModel):
    config_class = LatentAttentionConfig

    def __init__(self, config: LatentAttentionConfig):
        super().__init__(config)
        ## cross-attention block
        num_latents, latent_dim, cross_heads, cross_dim_head = config.num_latents_value, config.latent_dim, config.num_cross_heads, config.cross_dim_head
        dim = config.hidden_dim
        # init latent_attention and latents
        self.cross_attend_blocks = torch.nn.ModuleList([
            PreNorm(latent_dim, Attention(latent_dim, dim, heads = cross_heads, dim_head = cross_dim_head),
                    context_dim = dim),
            PreNorm(latent_dim, FeedForward(latent_dim)),
        ])
        self.output_normalize = config.output_normalize
        self.register_parameter("latents", torch.nn.Parameter(torch.randn(num_latents, latent_dim)))

    def forward(self, hiddens, attention_mask: torch.Tensor=None):
        ## cross-attention block
        cross_attn, cross_ff = self.cross_attend_blocks
        b, *_, device = *hiddens.shape, hiddens.device
        x = repeat(self.latents, 'n d -> b n d', b = b)
        hiddens = cross_attn(hiddens, context = x, mask = None) + hiddens
        hiddens = cross_ff(hiddens) + hiddens
        if attention_mask !=None:
            s = torch.sum(hiddens * attention_mask.unsqueeze(-1).float(), dim=1)
            d = attention_mask.sum(dim=1, keepdim=True).float()
            hiddens = s / d
            if self.output_normalize:
                hiddens = torch.nn.functional.normalize(hiddens, p=2, dim=-1)
        return hiddens
    
class NVEmbedModel(PreTrainedModel):
    config_class = NVEmbedConfig

    def __init__(self, config: NVEmbedConfig):
        super().__init__(config)
        self.latent_attention_model = AutoModel.from_config(config.latent_attention_config)
        self.embedding_model = AutoModel.from_config(
            config.text_config,
        ) if config.text_config is not None else None
        self.tokenizer = AutoTokenizer.from_pretrained(config.text_config._name_or_path, trust_remote_code=True) if config.text_config is not None else None
        self.padding_side = config.padding_side
        self.is_mask_instruction = config.is_mask_instruction
        self.add_eos = config.add_eos
        self.mask_type = config.mask_type
        if config.add_pad_token and self.tokenizer is not None:
            self.add_pad_token()

    def add_pad_token(self):
        self.tokenizer.pad_token = self.tokenizer.eos_token
        self.tokenizer.padding_side = self.padding_side
    
    def prepare_kwargs_from_batch(self, batch_dict: dict, instruction_lens: int, device: torch.device):
        batch_dict = move_to_device(batch_dict, device)
        attention_mask = batch_dict['attention_mask'].clone() if 'attention_mask' in batch_dict else None
        if (attention_mask is not None and
            self.padding_side == "right" and
            self.is_mask_instruction == True and
            instruction_lens > 0):
            # Mask out the instruction tokens for mean-pooling
            attention_mask[:, :instruction_lens] = 0
        features: NVEmbedFeatures = {
            'input_ids': torch.tensor(batch_dict.get('input_ids').to(batch_dict.get('input_ids')).long()),
            'attention_mask': batch_dict['attention_mask'],
            'pool_mask': attention_mask,
        }
        return features

    @torch.no_grad()
    def _do_encode(self,
        prompts: List[str],
        batch_size,
        instruction: str,
        max_length: int=4096,
        num_workers: int=32,
    ) -> np.ndarray:
        dataset: Dataset = Dataset.from_dict({'input_texts': prompts})
        dataset.set_transform(partial(input_transform_func,
                                      self.tokenizer,
                                      always_add_eos=True,
                                      max_length=max_length,
                                      instruction=instruction))

        data_collator = DataCollatorWithPadding(self.tokenizer)
        data_loader = DataLoader(
            dataset,
            batch_size=batch_size,
            shuffle=False,
            drop_last=False,
            num_workers=num_workers,
            collate_fn=data_collator,
            pin_memory=True)

        if self.padding_side == "right" and self.is_mask_instruction == True and len(instruction) > 0:
            instruction_lens = len(self.tokenizer.tokenize(instruction))
        else:
            instruction_lens = 0

        encoded_embeds = []
        device = next(self.embedding_model.parameters()).device
        for batch_dict in tqdm(data_loader, desc='encoding', mininterval=10):
            features = self.prepare_kwargs_from_batch(batch_dict, instruction_lens, device=device)
            embeds=self(**features)["sentence_embeddings"].squeeze(1)
            encoded_embeds.append(embeds.cpu().numpy())
        return np.concatenate(encoded_embeds, axis=0)

    def forward(self, input_ids: torch.Tensor, attention_mask: torch.Tensor, pool_mask: Optional[torch.Tensor]=None, return_dict: bool=True):
        autocast_ctx = torch.cuda.amp.autocast if torch.cuda.is_available() else nullcontext
        with autocast_ctx():
            ## decoder only layer
            outputs = self.embedding_model(
                input_ids=input_ids,
                attention_mask=attention_mask,
            )
            ## latent attention layer
            embeds = self.latent_attention_model(
                outputs.last_hidden_state,
                pool_mask,
            )
        if not return_dict:
            return (embeds,)
        return {"sentence_embeddings": embeds}
        
    
    @torch.no_grad()
    def encode(self, prompts: List[str], instruction: str="", max_length: int=4096):
        if self.padding_side == "right" and self.is_mask_instruction == True and len(instruction) > 0:
            instruction_lens = len(self.tokenizer.tokenize(instruction))
        else:
            instruction_lens = 0
        
        device = next(self.embedding_model.parameters()).device
        batch_dict = input_transform_func(self.tokenizer,
                                          {"input_texts": [prompt for prompt in prompts]},
                                          always_add_eos=True,
                                          max_length=max_length,
                                          instruction=instruction)

        features: NVEmbedFeatures = self.prepare_kwargs_from_batch(batch_dict, instruction_lens, device=device)
        return self(**features)["sentence_embeddings"].squeeze(1)


## AutoModel Register
AutoModel.register(NVEmbedConfig, NVEmbedModel)
AutoModel.register(LatentAttentionConfig, LatentAttentionModel)
AutoModel.register(BidirectionalMistralConfig, BidirectionalMistralModel)

## Register for auto class
NVEmbedModel.register_for_auto_class("AutoModel")
LatentAttentionModel.register_for_auto_class("AutoModel")
BidirectionalMistralModel.register_for_auto_class("AutoModel")