RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.LongTensor [1, 128]] is at version 3; expected version 2 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

[jordane95]

System Info

• transformers version: 4.25.1
• Platform: Linux-5.13.0-27-generic-x86_64-with-debian-bullseye-sid
• Python version: 3.7.15
• Huggingface_hub version: 0.11.1
• PyTorch version (GPU?): 1.10.1+cu113 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

Who can help?

@sgugger @ArthurZucker

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

I want to train a embedding-based retrieval qa system by minimizing the contrastive loss of correct (q,a) pairs against in-batch negatives. I also want it to be run on multiple gpus. But I run into the problem of backward propagation in position embedding layer of BERT (which I infer from the error log) when runing in distributed manner. I don't know where is broken (trainer? BertModel? pytorch?)

btw, the code works in single gpu setting

Command that I ran:

torchrun --nproc_per_node 2 retrieval_qa.py \
    --model_name_or_path bert-base-uncased \
    --output_dir debug \
    --max_steps 10000 \
    --remove_unused_columns False \
    --learning_rate 5e-5 \
    --logging_steps 10 \
    --save_steps 500 \
    --warmup_ratio 0.0 \
    --per_device_train_batch_size 16 \
    --normalize True

Error details:

***** Running training *****
  Num examples = 20360
  Num Epochs = 16
  Instantaneous batch size per device = 16
  Total train batch size (w. parallel, distributed & accumulation) = 32
  Gradient Accumulation steps = 1
  Total optimization steps = 10000
  Number of trainable parameters = 109482240
  0%|                                                                                                                     | 0/10000 [00:00<?, ?it/s][W python_anomaly_mode.cpp:104] Warning: Error detected in EmbeddingBackward0. Traceback of forward call that caused the error:
  File "retrieval_qa.py", line 213, in <module>
    main()
  File "retrieval_qa.py", line 209, in main
    trainer.train()
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/transformers/trainer.py", line 1531, in train
    ignore_keys_for_eval=ignore_keys_for_eval,
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/transformers/trainer.py", line 1775, in _inner_training_loop
    tr_loss_step = self.training_step(model, inputs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/transformers/trainer.py", line 2523, in training_step
    loss = self.compute_loss(model, inputs)
  File "retrieval_qa.py", line 142, in compute_loss
    token_type_ids=inputs[k]['token_type_ids'],
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/parallel/distributed.py", line 886, in forward
    output = self.module(*inputs[0], **kwargs[0])
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "retrieval_qa.py", line 103, in forward
    model_output = self.model(**kwargs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/transformers/models/bert/modeling_bert.py", line 1019, in forward
    past_key_values_length=past_key_values_length,
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/transformers/models/bert/modeling_bert.py", line 236, in forward
    position_embeddings = self.position_embeddings(position_ids)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/modules/sparse.py", line 160, in forward
    self.norm_type, self.scale_grad_by_freq, self.sparse)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/functional.py", line 2044, in embedding
    return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
 (function _print_stack)
[W python_anomaly_mode.cpp:104] Warning: Error detected in EmbeddingBackward0. Traceback of forward call that caused the error:
  File "retrieval_qa.py", line 213, in <module>
    main()
  File "retrieval_qa.py", line 209, in main
    trainer.train()
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/transformers/trainer.py", line 1531, in train
    ignore_keys_for_eval=ignore_keys_for_eval,
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/transformers/trainer.py", line 1775, in _inner_training_loop
    tr_loss_step = self.training_step(model, inputs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/transformers/trainer.py", line 2523, in training_step
    loss = self.compute_loss(model, inputs)
  File "retrieval_qa.py", line 142, in compute_loss
    token_type_ids=inputs[k]['token_type_ids'],
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/parallel/distributed.py", line 886, in forward
    output = self.module(*inputs[0], **kwargs[0])
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "retrieval_qa.py", line 103, in forward
    model_output = self.model(**kwargs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/transformers/models/bert/modeling_bert.py", line 1019, in forward
    past_key_values_length=past_key_values_length,
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/transformers/models/bert/modeling_bert.py", line 236, in forward
    position_embeddings = self.position_embeddings(position_ids)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/modules/sparse.py", line 160, in forward
    self.norm_type, self.scale_grad_by_freq, self.sparse)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/nn/functional.py", line 2044, in embedding
    return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
 (function _print_stack)
Traceback (most recent call last):
  File "retrieval_qa.py", line 213, in <module>
    main()
  File "retrieval_qa.py", line 209, in main
    trainer.train()
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/transformers/trainer.py", line 1531, in train
    ignore_keys_for_eval=ignore_keys_for_eval,
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/transformers/trainer.py", line 1775, in _inner_training_loop
    tr_loss_step = self.training_step(model, inputs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/transformers/trainer.py", line 2541, in training_step
Traceback (most recent call last):
  File "retrieval_qa.py", line 213, in <module>
    loss.backward()
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/_tensor.py", line 307, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
  File "/data01/lizehan/anaconda3/envs/beir/lib/python3.7/site-packages/torch/autograd/__init__.py", line 156, in backward
    allow_unreachable=True, accumulate_grad=True)  # allow_unreachable flag
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.LongTensor [1, 128]] is at version 3; expected version 2 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

Source code of retrieval_qa.py

import logging
import os
import sys
from typing import Dict, List, Tuple, Optional, Any, Union

import torch
from torch import nn
from torch.nn import functional as F

from transformers import AutoConfig, AutoModel, AutoTokenizer
from transformers import (
    HfArgumentParser,
    set_seed,
)

import os
from dataclasses import dataclass, field
from typing import Optional, List
from transformers import TrainingArguments
from transformers import DataCollatorWithPadding
from transformers.trainer import Trainer

import logging
logger = logging.getLogger(__name__)


# Name of the files used for checkpointing
TRAINING_ARGS_NAME = "training_args.bin"
TRAINER_STATE_NAME = "trainer_state.json"
OPTIMIZER_NAME = "optimizer.pt"
SCHEDULER_NAME = "scheduler.pt"
SCALER_NAME = "scaler.pt"


@dataclass
class ModelArguments:
    model_name_or_path: str = field(
        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
    )
    config_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
    )
    tokenizer_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
    )
    cache_dir: Optional[str] = field(
        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
    )

    normalize: bool = field(default=False)
    pooling: str = field(default='mean')


@dataclass
class QPCollator(DataCollatorWithPadding):
    """
    Wrapper that does conversion from List[Tuple[encode_qry, encode_psg]] to List[qry], List[psg]
    and pass batch separately to the actual collator.
    Abstract out data detail for the model.
    """
    max_q_len: int = 32
    max_p_len: int = 128

    def __call__(self, features):

        keys = list(features[0].keys())
        collated_batch = {}

        for key in keys:
            if not isinstance(features[0][key], str):
                continue
            text = [f[key] for f in features]
            # print(text)
            text_batch = self.tokenizer(
                text,
                padding='max_length',
                truncation=True,
                max_length=self.max_p_len,
                return_tensors="pt",
            )
            collated_batch[key] = text_batch

        return collated_batch


class AutoModelForSentenceEmbedding(nn.Module):
    def __init__(
        self,
        model_name_or_path,
        tokenizer=None,
        pooling='cls',
        normalize=True,
    ):
        super(AutoModelForSentenceEmbedding, self).__init__()

        self.model = AutoModel.from_pretrained(model_name_or_path)
        self.tokenizer = tokenizer if tokenizer else AutoTokenizer.from_pretrained(model_name_or_path)
        self.pooling = pooling
        self.normalize = normalize


    def forward(self, **kwargs):
        model_output = self.model(**kwargs)
        embeddings = self.mean_pooling(model_output, kwargs['attention_mask'])
        if self.normalize:
            embeddings = F.normalize(embeddings, p=2, dim=1)

        return embeddings

    def mean_pooling(self, model_output, attention_mask):
        token_embeddings = model_output[0]  # First element of model_output contains all token embeddings
        input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
        return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

    def save_pretrained(self, output_path):
        self.model.save_pretrained(output_path)



class EmbeddingTrainer(Trainer):

    def _save(self, output_dir: Optional[str] = None, state_dict=None):
        # If we are executing this function, we are the process zero, so we don't check for that.
        output_dir = output_dir if output_dir is not None else self.args.output_dir
        os.makedirs(output_dir, exist_ok=True)
        logger.info(f"Saving model checkpoint to {output_dir}")
        self.model.save_pretrained(output_dir)

        if self.tokenizer is not None:
            self.tokenizer.save_pretrained(output_dir)

        # Good practice: save your training arguments together with the trained model
        torch.save(self.args, os.path.join(output_dir, TRAINING_ARGS_NAME))
    
    def compute_loss(self, model, inputs, return_outputs=False):

        all_embeddings = {}
        for k in ['question', 'answer']:
            all_embeddings[k] = model(
                input_ids=inputs[k]['input_ids'],
                attention_mask=inputs[k]['attention_mask'],
                token_type_ids=inputs[k]['token_type_ids'],
            )
        embeddings_query = all_embeddings['question']
        embeddings_pos = all_embeddings['answer']

        scores = embeddings_query @ embeddings_pos.T
        labels = torch.arange(0, embeddings_query.shape[0], dtype=torch.long, device=embeddings_query.device)

        self.cross_entropy = torch.nn.CrossEntropyLoss(reduction='mean')
        loss = self.cross_entropy(scores, labels)
        return loss


def main():
    parser = HfArgumentParser((ModelArguments, TrainingArguments))

    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
        model_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
    else:
        model_args, training_args = parser.parse_args_into_dataclasses()
        model_args: ModelArguments
        training_args: TrainingArguments

    if (
            os.path.exists(training_args.output_dir)
            and os.listdir(training_args.output_dir)
            and training_args.do_train
            and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
    )

    set_seed(training_args.seed)

    tokenizer = AutoTokenizer.from_pretrained(
        model_args.model_name_or_path,
        cache_dir=model_args.cache_dir
    )
    model = AutoModelForSentenceEmbedding(
        model_args.model_name_or_path,
        pooling=model_args.pooling,
        normalize=model_args.normalize,
    )

    from datasets import load_dataset
    wq = load_dataset('wiki_qa', split='train')
    train_dataset = wq.remove_columns('label')

    data_collator = QPCollator(tokenizer=tokenizer)

    torch.autograd.set_detect_anomaly(True)
    trainer = EmbeddingTrainer(
        model=model,
        args=training_args,
        train_dataset=train_dataset,
        data_collator=data_collator,
        tokenizer=tokenizer,
    )

    trainer.train()


if __name__ == "__main__":
    main()

Expected behavior

Currently there is no problem on single gpu.
I want this code to run normally on multi-gpus. But it seems somewhere is broken...
It's hard to find where the problem is cause I'm not super familar with how pytorch/trainer/bertmodel works in distributed manner...
Could you help me? Thanks!

[ArthurZucker]

Hey!
Given how big the reproduction script is, I'm gonna say this is probably related to the way you are wrapping the use of transformers models, and would recommend you to ask on the forum to see if anyone in the community can help you with this!
I won't have time to dive into this, maybe @younesbelkada

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[younesbelkada]

Hi @jordane95 @ArthurZucker
Sadly I won"t have time to dig into that :/
@jordane95 do you still face the issue on the main branch of transformers?

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[jordane95]

Hi @jordane95 @ArthurZucker Sadly I won"t have time to dig into that :/ @jordane95 do you still face the issue on the main branch of transformers?

Yeah, this seems to be a problem involved with the siamese architecture? Althogh I can avoid this error by moving loss computation operations in compute_loss function of the trainer class to the forward function of model class, I'm still curious why this error occurs.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[anaivebird]

@jordane95 any idea of what happened in this error? Thanks

[silencio94]

@anaivebird If you are using BertModel class, try to put position_ids and token_type_ids in model input manually. (typically torch.arange(batch_size).unsqueeze(0).expand_as(input_ids) and torch.zeros_like(input_ids)).

[Ray-ui]

Hi @jordane95 @ArthurZucker Sadly I won"t have time to dig into that :/ @jordane95 do you still face the issue on the main branch of transformers?

Yeah, this seems to be a problem involved with the siamese architecture? Althogh I can avoid this error by moving loss computation operations in compute_loss function of the trainer class to the forward function of model class, I'm still curious why this error occurs.

Thank you for your suggestion. I encountered the same issue as well. I’m also trying to use SimCLR or another contrastive learning framework based on the BERT class. However, I was able to resolve the problem using the method you provided.