modeling_c2llm.py

import os
import gc
import inspect
import math
import multiprocessing as mp
import queue
from multiprocessing import Queue
import warnings
from typing import Any, Union, List, Dict, Literal, Optional
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from transformers import PretrainedConfig

from transformers import Qwen2Config
from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa, _prepare_4d_attention_mask, _prepare_4d_attention_mask_for_sdpa
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    is_flash_attn_2_available,
    is_flash_attn_greater_or_equal_2_10,
    logging,
    replace_return_docstrings,
)
import numpy as np
from transformers import Qwen2Config
from transformers import Qwen2ForCausalLM
import inspect
import math
import os
import warnings
from typing import List, Optional, Tuple, Union
from tqdm import tqdm, trange
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa, _prepare_4d_attention_mask, _prepare_4d_attention_mask_for_sdpa
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    is_flash_attn_2_available,
    is_flash_attn_greater_or_equal_2_10,
    logging,
    replace_return_docstrings,
)
import numpy as np
import torch
import os
import argparse
import json
from tqdm import tqdm
from typing import cast, List, Union, Tuple
from transformers import AutoTokenizer, AutoModel  # pylint: disable=C0413
from peft import LoraConfig, get_peft_model, TaskType
import time
import torch.nn.functional as F
import sys
import time
import torch 
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from tqdm import tqdm, trange
from collections import defaultdict
from transformers import AutoTokenizer, AutoModel, AutoModelForCausalLM, AutoConfig
import torch.distributed as dist
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
import sys
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
import re
import logging
logging.getLogger().setLevel(logging.INFO)
from .configuration_c2llm import C2LLMConfig
from transformers.models.qwen2.modeling_qwen2 import Qwen2DecoderLayer, Qwen2Attention

class MAB_POST(nn.Module):
    def __init__(self, dim_Q, dim_K, dim_V, num_heads, ln=False):
        super(MAB_POST, self).__init__()
        self.dim_V = dim_V
        self.num_heads = num_heads
        self.fc_q = nn.Linear(dim_Q, dim_V)
        self.fc_k = nn.Linear(dim_K, dim_V)
        self.fc_v = nn.Linear(dim_K, dim_V)
        if ln:
            self.ln0 = nn.LayerNorm(dim_V)
            self.ln1 = nn.LayerNorm(dim_V)
        self.fc_o = nn.Linear(dim_V, dim_V)
        nn.init.xavier_uniform_(self.fc_q.weight)
        nn.init.xavier_uniform_(self.fc_k.weight)
        nn.init.xavier_uniform_(self.fc_v.weight)
        nn.init.xavier_uniform_(self.fc_o.weight)

    def forward(self, Q, K, pad_mask=None):

        Q_ = self.fc_q(Q)
        K_, V_ = self.fc_k(K), self.fc_v(K)

        dim_split = self.dim_V // self.num_heads
        Q_ = torch.cat(Q_.split(dim_split, 2), 0)
        K_ = torch.cat(K_.split(dim_split, 2), 0)
        V_ = torch.cat(V_.split(dim_split, 2), 0)

        pad_mask = pad_mask.unsqueeze(1).repeat(self.num_heads, Q.size(1), 1)
        score = Q_.bmm(K_.transpose(1,2))/math.sqrt(self.dim_V)
        score = score.masked_fill(pad_mask == 0, -1e12)
        A = torch.softmax(score, 2)
        A = A * pad_mask
        O = torch.cat(A.bmm(V_).split(Q.size(0), 0), 2) 
        O = Q + O
        O = O if getattr(self, 'ln0', None) is None else self.ln0(O)
        O = O + F.relu(self.fc_o(O))
        O = O if getattr(self, 'ln1', None) is None else self.ln1(O)
        return O


class PMA(nn.Module):
    def __init__(self, dim, compressed_dim, num_heads, num_seeds, ln=False, pma_mode=None):
        super(PMA, self).__init__()
        self.S = nn.Parameter(torch.Tensor(1, num_seeds, compressed_dim))
        nn.init.xavier_uniform_(self.S)
        if pma_mode == 'post_normal':
            self.mab = MAB_POST(compressed_dim, dim, compressed_dim, num_heads, ln=ln)
        else:
            raise ValueError(f"Error, the pma_mode {pma_mode} is not implemented !")

    def forward(self, X, pad_mask):
        if self.S.dtype != torch.bfloat16:
            X = X.float()
        return self.mab(self.S.repeat(X.size(0), 1, 1), X, pad_mask)



class MAB_POST_v2(nn.Module):
    def __init__(self, dim_Q, dim_K, dim_V, num_heads, ln=False):
        super(MAB_POST_v2, self).__init__()
        self.dim_V = dim_V
        self.num_heads = num_heads
        self.fc_q = nn.Linear(dim_Q, dim_V)
        self.fc_k = nn.Linear(dim_K, dim_V)
        self.fc_v = nn.Linear(dim_K, dim_V)

        if ln:
            self.ln0 = nn.LayerNorm(dim_V)
            self.ln1 = nn.LayerNorm(dim_V)
        self.fc_o = nn.Linear(dim_V, dim_V)
        nn.init.xavier_uniform_(self.fc_q.weight)
        nn.init.xavier_uniform_(self.fc_k.weight)
        nn.init.xavier_uniform_(self.fc_v.weight)
        nn.init.xavier_uniform_(self.fc_o.weight)



    # Q(B, num_seed, D), pad_mask (bs, seq) Post-LN
    def forward(self, Q, K, pad_mask=None):

        Q_tmp = self.fc_q(Q) # B, num_seed, C
        K_, V_ = self.fc_k(K), self.fc_v(K) # B, L, C

        dim_split = self.dim_V // self.num_heads
        Q_ = torch.cat(Q_tmp.split(dim_split, 2), 0) # (B* num_head, num_seed, C)
        K_ = torch.cat(K_.split(dim_split, 2), 0) # (B* num_head, L, C)
        V_ = torch.cat(V_.split(dim_split, 2), 0) # (B* num_head,L, C)

        pad_mask = pad_mask.unsqueeze(1).repeat(self.num_heads, Q.size(1), 1) # (B*num_head, num_seed, L)
        score = Q_.bmm(K_.transpose(1,2))/math.sqrt(self.dim_V) # (B*num_head, num_seed, L) 
        score = score.masked_fill(pad_mask == 0, -1e12) # B,num_seed,L
        A = torch.softmax(score, 2)  # (B*num_head, num_seed, L)
        A = A * pad_mask
        O = torch.cat(A.bmm(V_).split(Q.size(0), 0), 2) # (B, num_seed, D)
        O = Q_tmp + O
        # O = torch.cat((Q_ + A.bmm(V_)).split(Q.size(0), 0), 2)
        O = O if getattr(self, 'ln0', None) is None else self.ln0(O)
        O = O + F.relu(self.fc_o(O))
        O = O if getattr(self, 'ln1', None) is None else self.ln1(O)
        return O




class PMA_v2(nn.Module):
    def __init__(self, dim, compressed_dim, num_heads, num_seeds, ln=False):
        super(PMA_v2, self).__init__()
        self.S = nn.Parameter(torch.Tensor(1, num_seeds, dim))
        nn.init.xavier_uniform_(self.S)
        # if pma_mode == 'post_normal':
        self.mab = MAB_POST_v2(dim, dim, compressed_dim, num_heads, ln=ln)
        # elif pma_mode == 'pre_normal':
            # self.mab = MAB_PRE_NORMAL(dim, dim, compressed_dim, num_heads, ln=ln)
        # elif pma_mode == 'pre_gptj':
            # self.mab = MAB_PRE_GPTJ(dim, dim, compressed_dim, num_heads, ln=ln)
        # else:
            # raise ValueError(f"Error, the pma_mode {pma_mode} is not implemented !")
    # X: (bs, seq, emb), pad_mask: (bs, seq)
    def forward(self, X, pad_mask):
        if self.S.dtype != torch.bfloat16:
            X = X.float()
        return self.mab(self.S.expand(X.size(0), -1, -1), X, pad_mask)


class C2LLMModel(PreTrainedModel):
    config_class = C2LLMConfig 
    config: C2LLMConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Qwen2DecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": Qwen2DecoderLayer,
        "attentions": Qwen2Attention,
    }


class C2LLMForEmbedding(C2LLMModel):

    config_class = C2LLMConfig
    model_type   = "c2llm"

    def __init__(self, config):
        super().__init__(config)
        qwen_cfg = Qwen2Config.from_dict(config.to_dict())
        self.plm_model = AutoModelForCausalLM.from_config(qwen_cfg) 
        self.embedding_method = config.embedding_method
        self.inf_seq_length = 2048
        self.padding_side = config.padding_side

        self.emb_dim = self.plm_model.model.embed_tokens.weight.size(1)
        self.keep_max_layer = self.plm_model.config.num_hidden_layers
        self.num_heads = config.pma_num_heads
        self.ln = config.pma_ln
        self.norm = config.pma_norm
        self.pma_mode = config.pma_norm_mode
        self.compressed_dim = config.compressed_dim
        
        self.mha_pma_disc = PMA_v2(self.emb_dim, self.compressed_dim, self.num_heads, 1, ln=self.ln)
        self.pool = None
        self.target_devices = self.get_target_devices(None)
        self.tokenizer = AutoTokenizer.from_pretrained(config.tokenizer_name_or_path, padding_side=config.padding_side) if config.tokenizer_name_or_path is not None else None
        self.config_class = C2LLMConfig


    def pma_embedding(self, mha_pma, A, mask):
        res = mha_pma(A, mask).squeeze(1)
        return res

    def get_hidden_states(self, **inputs):
        outputs = self.plm_model(inputs['input_ids'], inputs['attention_mask'], output_hidden_states=True)
        return outputs.hidden_states[self.keep_max_layer]

    def get_sentence_embedding(self, embedding_method, hidden_states, emb_type, attention_mask):

        
        if embedding_method == 'pma':
            
            if emb_type == 'disc':
                res_embedding = self.pma_embedding(self.mha_pma_disc, hidden_states, attention_mask)
                if self.norm:
                    res_embedding = torch.nn.functional.normalize(res_embedding, p=2.0, dim=-1, eps=1e-12, out=None)
                return res_embedding
            else:
                raise NotImplementedError(f"emb type {emb_type} hasn't been implemented")
        else:
            raise NotImplementedError(f"embedding method {embedding_method} hasn't been implemented")


    @staticmethod
    def get_target_devices(devices: Union[str, int, List[str], List[int]]) -> List[str]:
        """

        Args:
            devices (Union[str, int, List[str], List[int]]): specified devices, can be `str`, `int`, list of `str`, or list of `int`.

        Raises:
            ValueError: Devices should be a string or an integer or a list of strings or a list of integers.

        Returns:
            List[str]: A list of target devices in format.
        """
        if devices is None:
            if torch.cuda.is_available():
                return [f"cuda:{i}" for i in range(torch.cuda.device_count())]
            elif is_torch_npu_available():
                return [f"npu:{i}" for i in range(torch.npu.device_count())]
            elif hasattr(torch, "musa") and torch.musa.is_available():
                return [f"musa:{i}" for i in range(torch.musa.device_count())]
            elif torch.backends.mps.is_available():
                try:
                    return [f"mps:{i}" for i in range(torch.mps.device_count())]
                except:
                    return ["mps"]
            else:
                return ["cpu"]
        elif isinstance(devices, str):
            return [devices]
        elif isinstance(devices, int):
            if hasattr(torch, "musa") and torch.musa.is_available():
                return [f"musa:{devices}"]
            else:
                return [f"cuda:{devices}"]
        elif isinstance(devices, list):
            if isinstance(devices[0], str):
                return devices
            elif isinstance(devices[0], int):
                if hasattr(torch, "musa") and torch.musa.is_available():
                    return [f"musa:{device}" for device in devices]
                else:
                    return [f"cuda:{device}" for device in devices]
            else:
                raise ValueError("devices should be a string or an integer or a list of strings or a list of integers.")
        else:
            raise ValueError("devices should be a string or an integer or a list of strings or a list of integers.")

        # adapted from https://github.com/UKPLab/sentence-transformers/blob/1802076d4eae42ff0a5629e1b04e75785d4e193b/sentence_transformers/SentenceTransformer.py#L807
    
    
    def start_multi_process_pool(
        self,
        process_target_func: Any,
    ) -> Dict[Literal["input", "output", "processes"], Any]:
        """
        Starts a multi-process pool to process the encoding with several independent processes
        via :meth:`SentenceTransformer.encode_multi_process <sentence_transformers.SentenceTransformer.encode_multi_process>`.

        This method is recommended if you want to encode on multiple GPUs or CPUs. It is advised
        to start only one process per GPU. This method works together with encode_multi_process
        and stop_multi_process_pool.

        Returns:
            Dict[str, Any]: A dictionary with the target processes, an input queue, and an output queue.
        """
        if self.plm_model is None or self.mha_pma_disc is None:
            raise ValueError("Model is not initialized.")

        logging.info("Start multi-process pool on devices: {}".format(", ".join(map(str, self.target_devices))))

        self.to("cpu")
        self.share_memory()
        ctx = mp.get_context("spawn")
        input_queue = ctx.Queue()
        output_queue = ctx.Queue()
        processes = []

        for device_id in tqdm(self.target_devices, desc='initial target device'):
            p = ctx.Process(
                target=process_target_func,
                args=(device_id, self, input_queue, output_queue),
                daemon=True,
            )
            p.start()
            processes.append(p)

        return {"input": input_queue, "output": output_queue, "processes": processes}



    @staticmethod
    def _encode_multi_process_worker(
        target_device: str, model: 'C2LLMForEmbedding', input_queue: Queue, results_queue: Queue
    ) -> None:
        model = model.to(target_device)
        while True:
            try:
                chunk_id, sentences, kwargs = (
                    input_queue.get()
                )
                embeddings = model.encode_single_device(
                    sentences,
                    device=target_device,
                    **kwargs
                )

                results_queue.put([chunk_id, embeddings])
            except queue.Empty:
                break
    
    def encode_multi_process(
        self,
        sentences: List[str],
        pool: Dict[Literal["input", "output", "processes"], Any],
        **kwargs
    ):

        chunk_size = math.ceil(len(sentences) / len(pool["processes"]))

        input_queue = pool["input"]
        last_chunk_id = 0
        chunk = []

        for sentence in sentences:
            chunk.append(sentence)
            if len(chunk) >= chunk_size:
                input_queue.put(
                    [last_chunk_id, chunk, kwargs]
                )
                last_chunk_id += 1
                chunk = []

        if len(chunk) > 0:
            input_queue.put([last_chunk_id, chunk, kwargs])
            last_chunk_id += 1

        output_queue = pool["output"]
        results_list = sorted(
            [output_queue.get() for _ in trange(last_chunk_id, desc="")],
            key=lambda x: x[0],
        )
        embeddings = self._concatenate_results_from_multi_process([result[1] for result in results_list])
        return embeddings

    def _concatenate_results_from_multi_process(self, results_list: List[Union[torch.Tensor, np.ndarray, Any]]):
        """concatenate and return the results from all the processes

        Args:
            results_list (List[Union[torch.Tensor, np.ndarray, Any]]): A list of results from all the processes.

        Raises:
            NotImplementedError: Unsupported type for results_list

        Returns:
            Union[torch.Tensor, np.ndarray]: return the embedding vectors in a numpy array or tensor.
        """
        if isinstance(results_list[0], torch.Tensor):
            # move all tensors to the same device
            results_list = [res.to(self.target_devices[0]) for res in results_list]
            return torch.cat(results_list, dim=0)
        elif isinstance(results_list[0], np.ndarray):
            return np.concatenate(results_list, axis=0)
        else:
            raise NotImplementedError("Unsupported type for results_list")


    def forward(self, input_ids: torch.Tensor, attention_mask: torch.Tensor, return_dict: bool=True, **kwargs):
        outputs = self.plm_model(input_ids, attention_mask, output_hidden_states=True)
        hidden_states = outputs.hidden_states[self.keep_max_layer]
        embeddings = self.get_sentence_embedding(self.embedding_method, hidden_states, 'disc', attention_mask)
        if not return_dict:
            return (embeddings,)
        return {"sentence_embedding": embeddings}

    def encode_single_device(
        self, 
        sentences: Union[List[str], str],
        batch_size: int = 16, 
        convert_to_numpy: bool = False,        
        convert_to_tensor: bool = True, 
        show_progress_bar: bool = True, 
        max_seq_length: int = 2048, 
        device: Optional[str] = None,
        **kwargs: Any
    ):
        if max_seq_length is None:
            max_seq_length = self.inf_seq_length

        input_is_string = False        
        if isinstance(sentences, str) or not hasattr(sentences, "__len__"):
            sentences = [sentences]
            input_is_string = True
        all_embeddings = []
        length_sorted_idx = np.argsort([-len(s) for s in sentences])
        sentences_sorted = [sentences[idx] for idx in length_sorted_idx] # 大到小重排
        with torch.no_grad():
            for start_index in trange(0, len(sentences), batch_size, desc="Batches", disable=not show_progress_bar):
                sentences_batch = sentences_sorted[start_index: start_index + batch_size]
                inputs = self.tokenizer(sentences_batch, padding=True, truncation=True, max_length=max_seq_length, return_tensors='pt').to(self.plm_model.device)
                hidden_states = self.get_hidden_states(**inputs)
                embeddings = self.get_sentence_embedding(self.embedding_method, hidden_states, 'disc', inputs['attention_mask'])
                embeddings = embeddings.detach()
                if convert_to_numpy:
                    if embeddings.dtype == torch.bfloat16:
                        embeddings = embeddings.cpu().to(torch.float32)
                    else:
                        embeddings = embeddings.cpu()
                all_embeddings.extend(embeddings)
        all_embeddings = [all_embeddings[idx] for idx in np.argsort(length_sorted_idx)]
        if convert_to_tensor:
            all_embeddings = torch.stack(all_embeddings)
        elif convert_to_numpy:
            all_embeddings = np.asarray([emb.numpy() for emb in all_embeddings])

        if input_is_string:
            all_embeddings = all_embeddings[0]
        return all_embeddings


    def encode(self, sentences, batch_size=16, convert_to_numpy=False,
            convert_to_tensor=True, show_progress_bar=True, max_seq_length=None, **kwargs):

        if max_seq_length is None:
            max_seq_length = self.inf_seq_length
        
        if convert_to_tensor == convert_to_numpy:
            convert_to_tensor=True
            convert_to_numpy=False
            
        if isinstance(sentences, str) or len(self.target_devices) == 1:
            return self.encode_single_device(
                sentences,
                batch_size=batch_size,
                convert_to_numpy=convert_to_numpy,
                convert_to_tensor=convert_to_tensor,
                show_progress_bar=show_progress_bar,
                max_seq_length=max_seq_length,
                device=self.target_devices[0],
                **kwargs
            )
        if self.pool is None:
            self.pool = self.start_multi_process_pool(C2LLMForEmbedding._encode_multi_process_worker)
        

        all_embeddings = []
        length_sorted_idx = np.argsort([-len(s) for s in sentences])
        sentences_sorted = [sentences[idx] for idx in length_sorted_idx] # 大到小重排
        with torch.no_grad():
            for start_index in trange(0, len(sentences), batch_size, desc="Batches", disable=not show_progress_bar):
                sentences_batch = sentences_sorted[start_index: start_index + batch_size]
                embeddings_batch = self.encode_multi_process(
                    sentences_batch,
                    self.pool,
                    convert_to_numpy=convert_to_numpy,
                    convert_to_tensor=convert_to_tensor,
                    show_progress_bar=show_progress_bar,
                    max_seq_length=max_seq_length,
                    **kwargs
                )
                embeddings_batch = embeddings_batch.detach()
                if convert_to_numpy:
                    if embeddings_batch.dtype == torch.bfloat16:
                        embeddings_batch = embeddings_batch.cpu().to(torch.float32)
                    else:
                        embeddings_batch = embeddings_batch.cpu()
                all_embeddings.extend(embeddings_batch)
        all_embeddings = [all_embeddings[idx] for idx in np.argsort(length_sorted_idx)]
        if convert_to_tensor:
            all_embeddings = torch.stack(all_embeddings)
        elif convert_to_numpy:
            all_embeddings = np.asarray([emb.numpy() for emb in all_embeddings])


        return all_embeddings


    def encode_queries(self, sentences, batch_size=16, convert_to_numpy=False,
            convert_to_tensor=True, show_progress_bar=True, max_seq_length=None, **kwargs):
        if max_seq_length is None:
            max_seq_length = self.inf_seq_length
        
        if convert_to_tensor == convert_to_numpy:
            convert_to_tensor=True
            convert_to_numpy=False

        return self.encode(
            sentences=sentences,
            batch_size=batch_size,
            convert_to_numpy=convert_to_numpy,
            convert_to_tensor=convert_to_tensor,
            show_progress_bar=show_progress_bar,
            max_seq_length=max_seq_length,
            **kwargs
        )

    
    def encode_corpus(self, sentences, batch_size=16, convert_to_numpy=False,
            convert_to_tensor=True, show_progress_bar=True, max_seq_length=None, **kwargs):

        if max_seq_length is None:
            max_seq_length = self.inf_seq_length
        
        if convert_to_tensor == convert_to_numpy:
            convert_to_tensor=True
            convert_to_numpy=False
        sentences = [sentence['title']+' '+sentence['text'] for sentence in sentences]

        return self.encode(
            sentences=sentences,
            batch_size=batch_size,
            convert_to_numpy=convert_to_numpy,
            convert_to_tensor=convert_to_tensor,
            show_progress_bar=show_progress_bar,
            max_seq_length=max_seq_length,
            **kwargs
    
        )

    @staticmethod
    def stop_multi_process_pool(pool: Dict[Literal["input", "output", "processes"], Any]) -> None:
        """
        Stops all processes started with start_multi_process_pool.

        Args:
            pool (Dict[str, object]): A dictionary containing the input queue, output queue, and process list.

        Returns:
            None
        """
        for p in pool["processes"]:
            p.terminate()

        for p in pool["processes"]:
            p.join()
            p.close()

        pool["input"].close()
        pool["output"].close()
        pool = None

    def stop_self_pool(self):
        if self.pool is not None:
            self.stop_multi_process_pool(self.pool)
            self.pool = None
        try:
            self.model.to('cpu')
            torch.cuda.empty_cache()
        except:
            pass
        if gc is not None and callable(gc.collect):
            gc.collect()

    def __del__(self):
        self.stop_self_pool()