Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[]

[]

[]

[ "I think you first need to initialize the pipeline, and then call it with any parameters you like. Example:\r\n\r\n```\r\nfrom transformers import pipeline\r\nnlp = pipeline(\"question-answering\")\r\ncontext = r\"\"\"\r\nExtractive Question Answering is the task of extracting an answer from a text given a question. An example of a\r\nquestion answering dataset is the SQuAD dataset, which is entirely based on that task. If you would like to fine-tune\r\na model on a SQuAD task, you may leverage the examples/question-answering/run_squad.py script.\r\n\r\nresult = nlp(question=\"What is extractive question answering?\", context=context, max_answer_len=100)\r\n```\r\n\r\nFor the documentation of the `__call__` method of `QuestionAnsweringPipeline`, see [here](https://huggingface.co/transformers/main_classes/pipelines.html#transformers.QuestionAnsweringPipeline.__call__).", "This solved my issue. Thank you very much." ]

[]

[]

[ "Hi @varna9000 \r\n\r\nThe script expects jsonline format where each line is a `json` object. for ex:\r\n```\r\n{\"title\": \"summarised text\",\"body\": \"text to be summarized\"}\r\n{\"title\": \"summarised text\",\"body\": \"text to be summarized\"}\r\n```\r\n\r\nYou can find more info in the `datasets` doc here https://huggingface.co/docs/datasets/loading_datasets.html#json-files", "Thank you! :)", "Hi, I somehow encountered the same problem. Can I ask you how did you solve it?", "> Hi, I somehow encountered the same problem. Can I ask you how did you solve it?\r\n\r\nEncountered too.", "@judy-jii @hihihihiwsf it has been answered. You have to pass jsonlines format, not json.", "For anyone else who gets here from google, make sure your JSON lines are not wrapped in quote marks or something similar. Each line should be a valid JSON object, not a string. " ]

[]

[]

[ "You need to run `make style` to update the table deps ;-) " ]

[]

[]

[ "Hey @alexvaca0, \r\n\r\nCould you please provide us the command you used to train `t5-large` with `run_seq2seq.py` so that we can reproduce? ", "Also note that the timing is just a tqdm timing, that takes into account how long each iteration takes. Since the model saving takes a long time, it's normal to see it jump around a saving step and never go back to the best score of the first iterations.", "I don't intend to see it go back to the best score of the first iterations, but long after the last time it has been saved, it still takes twice as much time per iteration. Let me share with you the command used:\r\n\r\nFirst, I pre-trained t5 using the concepts explained here (taken from https://huggingface.co/transformers/master/model_doc/t5.html):\r\n\r\n```{python}\r\ninput_ids = tokenizer('The <extra_id_0> walks in <extra_id_1> park', return_tensors='pt').input_ids\r\nlabels = tokenizer('<extra_id_0> cute dog <extra_id_1> the <extra_id_2>', return_tensors='pt').input_ids\r\n# the forward function automatically creates the correct decoder_input_ids\r\nloss = model(input_ids=input_ids, labels=labels).loss\r\n```\r\n\r\nFor that I used the following command:\r\n\r\n```{bash}\r\nnohup python transformers/examples/seq2seq/run_seq2seq.py \\\r\n --model_name_or_path t5-large \\\r\n --do_eval --do_train \\\r\n --train_file /perturbed_data/tr.csv \\\r\n --validation_file /perturbed_data/val.csv \\\r\n --output_dir t5_lm \\\r\n --overwrite_output_dir \\\r\n --per_device_train_batch_size=2 \\\r\n --per_device_eval_batch_size=16 \\\r\n --eval_accumulation_steps=10 \\\r\n --max_source_length 346 \\\r\n --max_target_length 60 \\\r\n --val_max_target_length 60 --evaluation_strategy steps \\\r\n --gradient_accumulation_steps 128 --num_train_epochs=20 --eval_beams=1 \\\r\n --load_best_model_at_end --save_steps 100 --logging_steps 100 --learning_rate 7e-5 > bart_basic.txt &\r\n```\r\n\r\nThen, with that model trained, I tried to fine-tune it on a summarization task with the following command:\r\n\r\n```{bash}\r\npython transformers/examples/seq2seq/run_seq2seq.py \\\r\n --model_name_or_path t5-lm \\\r\n --do_eval --do_train \\\r\n --train_file summary_train_df.csv \\\r\n --validation_file summary_val_df.csv \\\r\n --output_dir t5_0802 \\\r\n --overwrite_output_dir \\\r\n --per_device_train_batch_size=2 \\\r\n --per_device_eval_batch_size=16 \\\r\n --eval_accumulation_steps=10 \\\r\n --max_source_length 346 \\\r\n --max_target_length 60 \\\r\n --val_max_target_length 60 --evaluation_strategy steps \\\r\n --gradient_accumulation_steps 32 --num_train_epochs=20 --eval_beams=1 \\\r\n --load_best_model_at_end --save_steps 250 --logging_steps 250 --learning_rate 3e-5\r\n```\r\n\r\nThis is for the second part of the issue, that is, the fact that when trying to load my lm-trained t5 from disk and re-train it on a summarization corpus the batches it processes suddenly occupy much more than when they're processed by t5-large (without re-training on the LM task). \r\n\r\nFor the other part of the issue, the only command that needs to be run is:\r\n\r\n```{bash}\r\npython transformers/examples/seq2seq/run_seq2seq.py \\\r\n --model_name_or_path t5-large \\\r\n --do_eval --do_train \\\r\n --train_file summary_train_df.csv \\\r\n --validation_file summary_val_df.csv \\\r\n --output_dir t5_0802 \\\r\n --overwrite_output_dir \\\r\n --per_device_train_batch_size=2 \\\r\n --per_device_eval_batch_size=16 \\\r\n --eval_accumulation_steps=10 \\\r\n --max_source_length 346 \\\r\n --max_target_length 60 \\\r\n --val_max_target_length 60 --evaluation_strategy steps \\\r\n --gradient_accumulation_steps 32 --num_train_epochs=20 --eval_beams=1 \\\r\n --load_best_model_at_end --save_steps 250 --logging_steps 250 --learning_rate 3e-5\r\n```\r\n\r\n@patrickvonplaten @sgugger ", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "That's the moment to create that list :)", "Really like the idea of improving our ONNX compatibility in a more reliable way. \r\n\r\nIn this sense, I'm not sure this is the easiest way for approaching the problem. \r\nMay be it would be more suitable to just attempt to export the model through `keras2onnx` and report errors. This can also allow us to more easily test compatibility with various ONNX opset (10 is the minimum required).\r\n\r\nWe already have the `keras2onnx` dependency when using the optional requirements `onnxruntime`\r\n\r\nAlso, regarding the list of model to be supported I think we would like to have: \r\n- BERT\r\n- GPT2\r\n- BART (cc @Narsil wdyt?)", "> In this sense, I'm not sure this is the easiest way for approaching the problem. \r\nMay be it would be more suitable to just attempt to export the model through `keras2onnx` and report errors. This can also allow us to more easily test compatibility with various ONNX opset (10 is the minimum required).\r\n\r\nI'm not in favour of running the export directly in the tests, as it is less flexible and not a compatible solution with other frameworks/SDKs. We can add any other opsets without problems with the proposed approach, but I don't think that going below the 12 is good idea. Only a few of the current models are compliant with opset < 12. Also, the default in the convert script is 11, not 10, so maybe we can propose opset 11 to be aligned. I think the proposed models below are compliant but we will largely reduce the number of compliant models with ONNX.\r\n\r\n> Also, regarding the list of model to be supported I think we would like to have:\r\nBERT\r\nGPT2\r\nBART (cc @Narsil wdyt?)\r\n\r\nThese three are ok on my side!", "Add BART and GPT2 as a mandatory green test.", "Testing ONNX operator support might be more complicated than this. \r\n\r\nEach operator in itself supports a set of different shape(s) as input/output combined with different data types and various dynamic axis support ... \r\n\r\nI would go for the easiest solution, well tested of using the official converter to report incompatibilities.", "The problem with the solution to use the converter is that we cannot have the full list of incompatible operators, it will stop at the first encounter one, which would be too much annoying IMO. I think we can also assume that as long as the operator belongs to this list https://github.com/onnx/tensorflow-onnx/blob/master/support_status.md it is compliant. Until now, this assumption is true for all of our models.\r\n\r\nUnless you know a case for which it is not true?\r\n\r\nAlso, I'm afraid to add a dependency to the onnxruntime would switch the quick test into a slow test, which reduces the traceability of a potential change that will break it.\r\n\r\nIf @LysandreJik, @sgugger and @patrickvonplaten agree on making the TF tests dependent on the two keras2onnx and onnxruntime packages, I can add a slow test that will run the following pipeline:\r\n\r\n1. Create a SavedModel\r\n2. Convert this SavedModel into ONNX with keras2onnx\r\n3. Run the converted model with onnxruntime", "We can add a test depending on `keras2onnx` or `onnxruntime` with a `@require_onnx` decorator. If you decide to go down this road, according to the time spent doing those tests, we'll probably put them in the slow suite (which is okay, no need to test that the model opsets on each PR)", "I like the idea to add a decorator. I will add a slow test doing this in addition to the quick test.", "I have reworked the quick test. Now, we can easily specify against which opset we want to test a model to be compliant. In the `onnx.json` file, all the operators are split in multiple opset, where each of them corresponds to the list of operators implemented in it. This should be way much easier to maintain and more flexible to use.\r\n\r\nIn addition to this I have added slow test that runs a complete pipeline of \"keras model -> ONNX model -> optimized ONNX model -> quantized ONNX model\".", "As proposed by @mfuntowicz I switched the min required opset version from 12 to 10 for BERT, GPT2 and BART.", "> Do you have an idea of how long the slow tests take ?\r\n\r\nDepending of the model between 1 and 5min.\r\n\r\n> According to the information gathered, would it be possible (in a next PR) to have a doc referencing the opset compliancy/onnx support for each model?\r\n\r\nDo you mean to have an entire page about ONNX? Or just to add a paragraph in the doc of every model about it?\r\n\r\nI think it is also important to mention that the model `TFGPT2ForSequenceClassification` cannot be converted into ONNX for now. The reason is because of the `tf.map_fn` function, that internally creates a `tf.while` with an iterator of type `tf.variant` which is not allowed in ONNX.", "LGTM on my side!\r\n\r\n@LysandreJik I have fixed the issue with `TFGPT2ForSequenceClassification`, so now it is compliant with ONNX.\r\n\r\n@mfuntowicz I should have addressed your comments, please double check ^^", "LGTM 👍🏻 ", "@LysandreJik Feel free to merge if the recent changes look ok for you!", "@LysandreJik Yes, this is exactly that :) I plan to apply this update to the other causal models one by one 😉 " ]

[]

[]

[ "Yes, it does save the tokenization on disk to avoid loading everything in RAM.\r\n\r\n> It's expected that the tokenized texts occupy less space than pure texts\r\n\r\nYou want the script to destroy your dataset?\r\n\r\nThe previous version of the scripts were loading every text in RAM so I would argue using Datasets makes it actually more possible to train with big datasets.\r\nThe examples are also just examples for quick fine-tuning/pre-training. If you are at the stage were your datasets don't even fit in disk space, some tweaks inside them are expected.", "@sgugger The thing is that it uses soooo much disk. When it had processed 18.7GB of texts it was using 2.1TB of disk... My dataset fits in the disk, and I'm sure the tokenized dataset should fit too, as it's actually lighter than text (it's only a list of integers per text), what doesn't fit in disk are the pyarrow objects created by datasets. \r\nSorry if I didn't explain myself clearly before, but it's not a problem with RAM or Memory, the main problem is that even when trying to pre-tokenize the whole dataset, saving it to disk for further use, it's not possible because the objects stored in disk by datasets library use 2 orders of magnitude more disk space than the original texts.", "Mmm, pinging @lhoestq as this seems indeed a huge bump in memory now that I see the numbers (I understood twice the space, not 100 times more, sorry!)", "Same discussion on the `datasets` repo: https://github.com/huggingface/datasets/issues/1825\r\n\r\n> tokenizing a dataset using map takes a lot of space since it can store input_ids but also token_type_ids, attention_mask and special_tokens_mask. Moreover if your tokenization function returns python integers then by default they'll be stored as int64 which can take a lot of space. Padding can also increase the size of the tokenized dataset.\r\n\r\nTo go forward it would be nice to optimize what we actually need to be stored on disk. If some columns are not useful maybe they can be dropped (and possibly recreated on-the-fly if needed). We should also tweak the tensors precisions.\r\n\r\nAnother approach would be to tokenize on-the-fly for pretraining.", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "I'm trying to reproduce your error but not managing. Could you indicate the checkpoint you are using and the command you used to launch your training? Also, what is the exact content of your `checkpoint-1000` folder? It should have a `config.json` alongside a `pytorch_model.bin`.", "Thanks @sgugger for the quick reply. I was testing using the checkpoint-1000 but all other folder checkpoints have the similar file contents. They are \r\n1) scheduler.pt\r\n2) tokenizer_config.json\r\n3) optimizer.pt\r\n4) sentencepiece.bpe.model\r\n5) trainer_state.json\r\n6) pytorch_model.bin\r\n7) special_tokens_map.json\r\n8) training_args.bin.\r\n\r\nThe checkpoints all had the tokenizer_config.json instead of config.json and the contents of the tokenizer_config.json had the following.\r\n\r\n`{\"bos_token\": \"<s>\", \"eos_token\": \"</s>\", \"sep_token\": \"</s>\", \"cls_token\": \"<s>\", \"unk_token\": \"<unk>\", \"pad_token\": \"<pad>\", \"mask_token\": \"<mask>\", \"model_max_length\": 512, \"name_or_path\": \"xlm-roberta-base\"}`", "I meant the checkpoint from the model hub (your model identifier). And the command that you use to run the script please. The model config should be saved along the rest, I'm trying to find out why that is not the case.", "@sgugger The checkpoint folder was created by language modeling run_mlm.py script. The run_mlm.py script which I ran is with the following args.\r\n\r\n```\r\npython run_mlm.py \\\r\n --model_name_or_path xlm-roberta-base \\\r\n --train_file train_file \\\r\n --validation_file valid_file \\\r\n --do_train \\\r\n --do_eval \\\r\n --output_dir output_path \\\r\n --logging_dir log_path \\\r\n --logging_steps 100o \\\r\n --max_seq_length 512 \\\r\n --pad_to_max_length \\\r\n --learning_rate 2e-5 \\\r\n --per_device_train_batch_size 2 \\\r\n --per_device_eval_batch_size 2 \\\r\n --overwrite_output_dir \\\r\n --num_train_epochs 2 \\\r\n --eval_steps 100o \\\r\n --mlm_probability 0.15 \\\r\n --evaluation_strategy \"steps\"\r\n```\r\n\r\nAnd to access the model and tokenizer from the checkpoint folder I used the following commands:\r\n```\r\nfrom transformers import AutoTokenizer, AutoModelForSequenceClassification\r\ntokenizer = AutoTokenizer.from_pretrained(ckpt_path)\r\nmodel = AutoModelForSequenceClassification.from_pretrained(\r\n ckpt_path, \r\n num_labels = 3, \r\n output_attentions = False,\r\n output_hidden_states = False,\r\n )\r\n```", "Mmm, I ran the same command as you did (replacing the 100o by 100, I think it's a typo) on an env similar to yours (transformers master and PyTorch 1.7.1, Python 3.7.9). The problem does not occur, everything is properly saved and your second snippet runs without problem.", "@sgugger Yes it is a typo in the code snippet above, sorry for that. Maybe I will retry the whole process in a different environment and test it out.", "@sgugger I tried in in a Redhat server and the code generated the config.json without a hitch but when I ran the same code in AWS Sagemaker. The config.json file was not generated. And the log information during the AWS Sagemaker run is shared below. But not sure if this is related to the Sagemaker environment or not.\r\n\r\n```\r\n02/10/2021 19:34:51 - INFO - logger - Trainer.model is not a `PreTrainedModel`, only saving its state dict.\r\n/usr/local/lib/python3.6/dist-packages/torch/nn/parallel/_functions.py:64: UserWarning: Was asked to gather along dimension 0, but all input tensors were scalars; will instead unsqueeze and return a vector.\r\n warnings.warn('Was asked to gather along dimension 0, but all '\r\n\r\n```\r\n", "Ah thanks, this is helpful! How are you launching the script in sagemaker exactly?", "@sgugger I zipped all my transformers file into a zip file and ran the run_mlm.py script in a docker image using the configuration given in the github repo. All the hyperparameters were passed in a json format to the estimator and the logging dir was changed into the tensorboard directory in sagemaker. And may I know what is causing the above information in pretraining process?", "@sgugger Something like this\r\n\r\n```\r\ntensorboard_logs = log_s3_path\r\n\r\nfrom sagemaker.debugger import TensorBoardOutputConfig\r\ntensorboard_output_config = TensorBoardOutputConfig(\r\n s3_output_path=tensorboard_logs,\r\n container_local_output_path='/opt/tensorboard/'\r\n)\r\n\r\nimport json\r\n# JSON encode hyperparameters.\r\ndef json_encode_hyperparameters(hyperparameters):\r\n return {str(k): json.dumps(v) for (k, v) in hyperparameters.items()}\r\n\r\nhyperparameters = json_encode_hyperparameters({\r\n \"sagemaker_program\": \"run_mlm.py\",\r\n \"sagemaker_submit_directory\": code_path_s3,\r\n \"model_name_or_path\": \"xlm-roberta-base\",\r\n \"num_train_epochs\": 1,\r\n \"learning_rate\" : 2e-5,\r\n \"max_seq_length\" : 512,\r\n \"eval_steps\": 1000,\r\n \"mlm_probability\": 0.15,\r\n \"logging_steps\": 1000,\r\n \"save_steps\": 1000,\r\n \"per_device_train_batch_size\": 2,\r\n \"per_device_eval_batch_size\": 2,\r\n \"output_dir\": \"/opt/ml/output\",\r\n \"logging_dir\": \"/opt/tensorboard/\",\r\n \"train_file\": train_file,\r\n \"validation_file\": validation_file,\r\n }) \r\n\r\nfrom sagemaker.estimator import Estimator\r\nestimator = Estimator(\r\n image_uri=<docker_image_path>,\r\n output_path=output_path,\r\n output_kms_key=kms_key,\r\n role=role,\r\n tensorboard_output_config = tensorboard_output_config,\r\n instance_count = 1,\r\n instance_type=<instance_type>,\r\n hyperparameters=hyperparameters,\r\n)\r\nestimator.fit(job_name=training_job_name, inputs ={\"training\":f'{training_data}',\r\n \"validation\":validation_file_s3_path, })\r\n```", "Mmmm, Could you try with the latest version of Transformers (a source install of the released v4.3.2?) It seems the model that the `Trainer` saved has been wrapped by something (since it doesn't find it's a `PreTrainedModel`) but I'm not finding what. When trying on SageMaker on my side, I get a regular `PreTrainedModel` and it saves properly.\r\n\r\nJust for my information, what kind of instance are you using (1 or several GPUs?)", "@sgugger Currently I am using version 4.2.2, and I am running a multi-gpu instance (p3.8xlarge - 4 GPUs). And one more quick question, does the model checkpoint saved not possible to use for finetuning. I already know the config.json file for the XLM-Roberta architecture so it is possible to using it along with the model variables stored in every checkpoint even though the Trainer class is not able to find it as a PreTrainedModel as a backup option or the model saved in each checkpoint would not be updated from the pretrained model even though it is being pretrained.", "You can load the model weights manually with something like:\r\n```\r\nconfig = AutoConfig.from_pretrained(\"xlm-roberta-base\")\r\nmodel = AutoModel.from_config(config)\r\nmodel.load_state_dict(torch.load(checkpoint_file))\r\n```\r\n", "@sgugger I tried the running with v4.3.2. But I am getting tokenizer errors, which did not occur before.\r\nthread '<unnamed>' panicked at 'index out of bounds: the len is 300 but the index is 300', /__w/tokenizers/tokenizers/tokenizers/src/tokenizer/normalizer.rs:382:21\r\npyo3_runtime.PanicException: index out of bounds: the len is 300 but the index is 300\r\n\r\nAnd the saved checkpoint folder has multiple files for pytorch, which one should I load using torch.load()", "I can't help you if you don't show me the content of that folder. I'm also unsure of the error for your tokenizer since there is no `tokenzier` in the snippet of code I pasted above.", "@sgugger I think I did not convey the error correctly. I was having confusion on what file I should you as the checkpoint file in the code which you have shared.\r\n\r\n```\r\nconfig = AutoConfig.from_pretrained(\"xlm-roberta-base\")\r\nmodel = AutoModel.from_config(config)\r\nmodel.load_state_dict(torch.load(checkpoint_file))\r\n```\r\nAs in the checkpoint directory I have these files. \r\n\r\n1. scheduler.pt\r\n2. tokenizer_config.json\r\n3. optimizer.pt\r\n4. sentencepiece.bpe.model\r\n5. trainer_state.json\r\n6. pytorch_model.bin\r\n7. special_tokens_map.json\r\n8. training_args.bin.\r\n\r\nAnd the tokenizer error which I have pasted is when trying to trigger the run_mlm.py script of transformers v4.3.2 on the training dataset as you had told me to try with the latest version of transformers. The new script fails in the tokenization step itself.", "I will try to reproduce your error for the tokenizer. For the checkpoint file, you have to use `pytorch_model.bin`, this is where your model weights are.", "@sgugger I tried to load the model from the ```pytorch_model.bin``` But I am getting the following error.\r\n```\r\n--------------------------------------------------------------------------\r\nRuntimeError Traceback (most recent call last)\r\n<ipython-input-4-3bd59185f8cc> in <module>\r\n 2 model = AutoModel.from_config(config)\r\n 3 model_path = \"/home/ec2-user/SageMaker/samples/new_checkpoint_test/checkpoint-1000/pytorch_model.bin\"\r\n----> 4 model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))\r\n\r\n~/anaconda3/envs/pytorch_latest_p36/lib/python3.6/site-packages/torch/nn/modules/module.py in load_state_dict(self, state_dict, strict)\r\n 1050 if len(error_msgs) > 0:\r\n 1051 raise RuntimeError('Error(s) in loading state_dict for {}:\\n\\t{}'.format(\r\n-> 1052 self.__class__.__name__, \"\\n\\t\".join(error_msgs)))\r\n 1053 return _IncompatibleKeys(missing_keys, unexpected_keys)\r\n 1054 \r\n\r\nRuntimeError: Error(s) in loading state_dict for RobertaModel:\r\n\tMissing key(s) in state_dict: \"embeddings.position_ids\", \"embeddings.word_embeddings.weight\", \"embeddings.position_embeddings.weight\", \"embeddings.token_type_embeddings.weight\", \"embeddings.LayerNorm.weight\", \"embeddings.LayerNorm.bias\", \"encoder.layer.0.attention.self.query.weight\", \"encoder.layer.0.attention.self.query.bias\", \"encoder.layer.0.attention.self.key.weight\", \"encoder.layer.0.attention.self.key.bias\", \"encoder.layer.0.attention.self.value.weight\", \"encoder.layer.0.attention.self.value.bias\", \"encoder.layer.0.attention.output.dense.weight\", \"encoder.layer.0.attention.output.dense.bias\", \"encoder.layer.0.attention.output.LayerNorm.weight\", \"encoder.layer.0.attention.output.LayerNorm.bias\", \"encoder.layer.0.intermediate.dense.weight\", \"encoder.layer.0.intermediate.dense.bias\", \"encoder.layer.0.output.dense.weight\", \"encoder.layer.0.output.dense.bias\", \"encoder.layer.0.output.LayerNorm.weight\", \"encoder.layer.0.output.LayerNorm.bias\", \"encoder.layer.1.attention.self.query.weight\", \"encoder.layer.1.attention.self.query.bias\", \"encoder.layer.1.attention.self.key.weight\", \"encoder.layer.1.attention.self.key.bias\", \"encoder.layer.1.attention.self.value.weight\", \"encoder.layer.1.attention.self.value.bias\", \"encoder.layer.1.attention.output.dense.weight\", \"encoder.layer.1.attention.output.dense.bias\", \"encoder.layer.1.attention.output.LayerNorm.weight\", \"encoder.layer.1.attention.output.LayerNorm.bias\", \"encoder.layer.1.intermediate.dense.weight\", \"encoder.layer.1.intermediate.dense.bias\", \"encoder.layer.1.output.dense.weight\", \"encoder.layer.1.output.dense.bias\", \"encoder.layer.1.output.LayerNorm.weight\", \"encoder.layer.1.output.LayerNorm.bias\", \"encoder.layer.2.attention.self.query.weight\", \"encoder.layer.2.attention.self.query.bias\", \"encoder.layer.2.attention.self.key.weight\", \"encoder.layer.2.attention.self.key.bias\", \"encoder.layer.2.attention.self.value.weight\", \"encoder.layer.2.attention.self.value.bias\", \"encoder.layer.2.attention.output.dense.weight\", \"encoder.layer.2.attention.output.dense.bias\", \"encoder.layer.2.attention.output.LayerNorm.weight\", \"encoder.layer.2.attention.output.LayerNorm.bias\", \"encoder.layer.2.intermediate.dense.weight\", \"encoder.layer.2.intermediate.dense.bias\", \"encoder.layer.2.output.dense.weight\", \"encoder.layer.2.output.dense.bias\", \"encoder.layer.2.output.LayerNorm.weight\", \"encoder.layer.2.output.LayerNorm.bias\", \"encoder.layer.3.attention.self.query.weight\", \"encoder.layer.3.attention.self.query.bias\", \"encoder.layer.3.attention.self.key.weight\", \"encoder.layer.3.attention.self.key.bias\", \"encoder.layer.3.attention.self.value.weight\", \"encoder.layer.3.attention.self.value.bias\", \"encoder.layer.3.attention.output.dense.weight\", \"encoder.layer.3.attention.output.dense.bias\", \"encoder.layer.3.attention.output.LayerNorm.weight\", \"encoder.layer.3.attention.output.LayerNorm.bias\", \"encoder.layer.3.intermediate.dense.weight\", \"encoder.layer.3.intermediate.dense.bias\", \"encoder.layer.3.output.dense.weight\", \"encoder.layer.3.output.dense.bias\", \"encoder.layer.3.output.LayerNorm.weight\", \"encoder.layer.3.output.LayerNorm.bias\", \"encoder.layer.4.attention.self.query.weight\", \"encoder.layer.4.attention.self.query.bias\", \"encoder.layer.4.attention.self.key.weight\", \"encoder.layer.4.attention.self.key.bias\", \"encoder.layer.4.attention.self.value.weight\", \"encoder.layer.4.attention.self.value.bias\", \"encoder.layer.4.attention.output.dense.weight\", \"encoder.layer.4.attention.output.dense.bias\", \"encoder.layer.4.attention.output.LayerNorm.weight\", \"encoder.layer.4.attention.output.LayerNorm.bias\", \"encoder.layer.4.intermediate.dense.weight\", \"encoder.layer.4.intermediate.dense.bias\", \"encoder.layer.4.output.dense.weight\", \"encoder.layer.4.output.dense.bias\", \"encoder.layer.4.output.LayerNorm.weight\", \"encoder.layer.4.output.LayerNorm.bias\", \"encoder.layer.5.attention.self.query.weight\", \"encoder.layer.5.attention.self.query.bias\", \"encoder.layer.5.attention.self.key.weight\", \"encoder.layer.5.attention.self.key.bias\", \"encoder.layer.5.attention.self.value.weight\", \"encoder.layer.5.attention.self.value.bias\", \"encoder.layer.5.attention.output.dense.weight\", \"encoder.layer.5.attention.output.dense.bias\", \"encoder.layer.5.attention.output.LayerNorm.weight\", \"encoder.layer.5.attention.output.LayerNorm.bias\", \"encoder.layer.5.intermediate.dense.weight\", \"encoder.layer.5.intermediate.dense.bias\", \"encoder.layer.5.output.dense.weight\", \"encoder.layer.5.output.dense.bias\", \"encoder.layer.5.output.LayerNorm.weight\", \"encoder.layer.5.output.LayerNorm.bias\", \"encoder.layer.6.attention.self.query.weight\", \"encoder.layer.6.attention.self.query.bias\", \"encoder.layer.6.attention.self.key.weight\", \"encoder.layer.6.attention.self.key.bias\", \"encoder.layer.6.attention.self.value.weight\", \"encoder.layer.6.attention.self.value.bias\", \"encoder.layer.6.attention.output.dense.weight\", \"encoder.layer.6.attention.output.dense.bias\", \"encoder.layer.6.attention.output.LayerNorm.weight\", \"encoder.layer.6.attention.output.LayerNorm.bias\", \"encoder.layer.6.intermediate.dense.weight\", \"encoder.layer.6.intermediate.dense.bias\", \"encoder.layer.6.output.dense.weight\", \"encoder.layer.6.output.dense.bias\", \"encoder.layer.6.output.LayerNorm.weight\", \"encoder.layer.6.output.LayerNorm.bias\", \"encoder.layer.7.attention.self.query.weight\", \"encoder.layer.7.attention.self.query.bias\", \"encoder.layer.7.attention.self.key.weight\", \"encoder.layer.7.attention.self.key.bias\", \"encoder.layer.7.attention.self.value.weight\", \"encoder.layer.7.attention.self.value.bias\", \"encoder.layer.7.attention.output.dense.weight\", \"encoder.layer.7.attention.output.dense.bias\", \"encoder.layer.7.attention.output.LayerNorm.weight\", \"encoder.layer.7.attention.output.LayerNorm.bias\", \"encoder.layer.7.intermediate.dense.weight\", \"encoder.layer.7.intermediate.dense.bias\", \"encoder.layer.7.output.dense.weight\", \"encoder.layer.7.output.dense.bias\", \"encoder.layer.7.output.LayerNorm.weight\", \"encoder.layer.7.output.LayerNorm.bias\", \"encoder.layer.8.attention.self.query.weight\", \"encoder.layer.8.attention.self.query.bias\", \"encoder.layer.8.attention.self.key.weight\", \"encoder.layer.8.attention.self.key.bias\", \"encoder.layer.8.attention.self.value.weight\", \"encoder.layer.8.attention.self.value.bias\", \"encoder.layer.8.attention.output.dense.weight\", \"encoder.layer.8.attention.output.dense.bias\", \"encoder.layer.8.attention.output.LayerNorm.weight\", \"encoder.layer.8.attention.output.LayerNorm.bias\", \"encoder.layer.8.intermediate.dense.weight\", \"encoder.layer.8.intermediate.dense.bias\", \"encoder.layer.8.output.dense.weight\", \"encoder.layer.8.output.dense.bias\", \"encoder.layer.8.output.LayerNorm.weight\", \"encoder.layer.8.output.LayerNorm.bias\", \"encoder.layer.9.attention.self.query.weight\", \"encoder.layer.9.attention.self.query.bias\", \"encoder.layer.9.attention.self.key.weight\", \"encoder.layer.9.attention.self.key.bias\", \"encoder.layer.9.attention.self.value.weight\", \"encoder.layer.9.attention.self.value.bias\", \"encoder.layer.9.attention.output.dense.weight\", \"encoder.layer.9.attention.output.dense.bias\", \"encoder.layer.9.attention.output.LayerNorm.weight\", \"encoder.layer.9.attention.output.LayerNorm.bias\", \"encoder.layer.9.intermediate.dense.weight\", \"encoder.layer.9.intermediate.dense.bias\", \"encoder.layer.9.output.dense.weight\", \"encoder.layer.9.output.dense.bias\", \"encoder.layer.9.output.LayerNorm.weight\", \"encoder.layer.9.output.LayerNorm.bias\", \"encoder.layer.10.attention.self.query.weight\", \"encoder.layer.10.attention.self.query.bias\", \"encoder.layer.10.attention.self.key.weight\", \"encoder.layer.10.attention.self.key.bias\", \"encoder.layer.10.attention.self.value.weight\", \"encoder.layer.10.attention.self.value.bias\", \"encoder.layer.10.attention.output.dense.weight\", \"encoder.layer.10.attention.output.dense.bias\", \"encoder.layer.10.attention.output.LayerNorm.weight\", \"encoder.layer.10.attention.output.LayerNorm.bias\", \"encoder.layer.10.intermediate.dense.weight\", \"encoder.layer.10.intermediate.dense.bias\", \"encoder.layer.10.output.dense.weight\", \"encoder.layer.10.output.dense.bias\", \"encoder.layer.10.output.LayerNorm.weight\", \"encoder.layer.10.output.LayerNorm.bias\", \"encoder.layer.11.attention.self.query.weight\", \"encoder.layer.11.attention.self.query.bias\", \"encoder.layer.11.attention.self.key.weight\", \"encoder.layer.11.attention.self.key.bias\", \"encoder.layer.11.attention.self.value.weight\", \"encoder.layer.11.attention.self.value.bias\", \"encoder.layer.11.attention.output.dense.weight\", \"encoder.layer.11.attention.output.dense.bias\", \"encoder.layer.11.attention.output.LayerNorm.weight\", \"encoder.layer.11.attention.output.LayerNorm.bias\", \"encoder.layer.11.intermediate.dense.weight\", \"encoder.layer.11.intermediate.dense.bias\", \"encoder.layer.11.output.dense.weight\", \"encoder.layer.11.output.dense.bias\", \"encoder.layer.11.output.LayerNorm.weight\", \"encoder.layer.11.output.LayerNorm.bias\", \"pooler.dense.weight\", \"pooler.dense.bias\". \r\n\tUnexpected key(s) in state_dict: \"roberta.embeddings.position_ids\", \"roberta.embeddings.word_embeddings.weight\", \"roberta.embeddings.position_embeddings.weight\", \"roberta.embeddings.token_type_embeddings.weight\", \"roberta.embeddings.LayerNorm.weight\", \"roberta.embeddings.LayerNorm.bias\", \"roberta.encoder.layer.0.attention.self.query.weight\", \"roberta.encoder.layer.0.attention.self.query.bias\", \"roberta.encoder.layer.0.attention.self.key.weight\", \"roberta.encoder.layer.0.attention.self.key.bias\", \"roberta.encoder.layer.0.attention.self.value.weight\", \"roberta.encoder.layer.0.attention.self.value.bias\", \"roberta.encoder.layer.0.attention.output.dense.weight\", \"roberta.encoder.layer.0.attention.output.dense.bias\", \"roberta.encoder.layer.0.attention.output.LayerNorm.weight\", \"roberta.encoder.layer.0.attention.output.LayerNorm.bias\", \"roberta.encoder.layer.0.intermediate.dense.weight\", \"roberta.encoder.layer.0.intermediate.dense.bias\", \"roberta.encoder.layer.0.output.dense.weight\", \"roberta.encoder.layer.0.output.dense.bias\", \"roberta.encoder.layer.0.output.LayerNorm.weight\", \"roberta.encoder.layer.0.output.LayerNorm.bias\", \"roberta.encoder.layer.1.attention.self.query.weight\", \"roberta.encoder.layer.1.attention.self.query.bias\", \"roberta.encoder.layer.1.attention.self.key.weight\", \"roberta.encoder.layer.1.attention.self.key.bias\", \"roberta.encoder.layer.1.attention.self.value.weight\", \"roberta.encoder.layer.1.attention.self.value.bias\", \"roberta.encoder.layer.1.attention.output.dense.weight\", \"roberta.encoder.layer.1.attention.output.dense.bias\", \"roberta.encoder.layer.1.attention.output.LayerNorm.weight\", \"roberta.encoder.layer.1.attention.output.LayerNorm.bias\", \"roberta.encoder.layer.1.intermediate.dense.weight\", \"roberta.encoder.layer.1.intermediate.dense.bias\", \"roberta.encoder.layer.1.output.dense.weight\", \"roberta.encoder.layer.1.output.dense.bias\", \"roberta.encoder.layer.1.output.LayerNorm.weight\", \"roberta.encoder.layer.1.output.LayerNorm.bias\", \"roberta.encoder.layer.2.attention.self.query.weight\", \"roberta.encoder.layer.2.attention.self.query.bias\", \"roberta.encoder.layer.2.attention.self.key.weight\", \"roberta.encoder.layer.2.attention.self.key.bias\", \"roberta.encoder.layer.2.attention.self.value.weight\", \"roberta.encoder.layer.2.attention.self.value.bias\", \"roberta.encoder.layer.2.attention.output.dense.weight\", \"roberta.encoder.layer.2.attention.output.dense.bias\", \"roberta.encoder.layer.2.attention.output.LayerNorm.weight\", \"roberta.encoder.layer.2.attention.output.LayerNorm.bias\", \"roberta.encoder.layer.2.intermediate.dense.weight\", \"roberta.encoder.layer.2.intermediate.dense.bias\", \"roberta.encoder.layer.2.output.dense.weight\", \"roberta.encoder.layer.2.output.dense.bias\", \"roberta.encoder.layer.2.output.LayerNorm.weight\", \"roberta.encoder.layer.2.output.LayerNorm.bias\", \"roberta.encoder.layer.3.attention.self.query.weight\", \"roberta.encoder.layer.3.attention.self.query.bias\", \"roberta.encoder.layer.3.attention.self.key.weight\", \"roberta.encoder.layer.3.attention.self.key.bias\", \"roberta.encoder.layer.3.attention.self.value.weight\", \"roberta.encoder.layer.3.attention.self.value.bias\", \"roberta.encoder.layer.3.attention.output.dense.weight\", \"roberta.encoder.layer.3.attention.output.dense.bias\", \"roberta.encoder.layer.3.attention.output.LayerNorm.weight\", \"roberta.encoder.layer.3.attention.output.LayerNorm.bias\", \"roberta.encoder.layer.3.intermediate.dense.weight\", \"roberta.encoder.layer.3.intermediate.dense.bias\", \"roberta.encoder.layer.3.output.dense.weight\", \"roberta.encoder.layer.3.output.dense.bias\", \"roberta.encoder.layer.3.output.LayerNorm.weight\", \"roberta.encoder.layer.3.output.LayerNorm.bias\", \"roberta.encoder.layer.4.attention.self.query.weight\", \"roberta.encoder.layer.4.attention.self.query.bias\", \"roberta.encoder.layer.4.attention.self.key.weight\", \"roberta.encoder.layer.4.attention.self.key.bias\", \"roberta.encoder.layer.4.attention.self.value.weight\", \"roberta.encoder.layer.4.attention.self.value.bias\", \"roberta.encoder.layer.4.attention.output.dense.weight\", \"roberta.encoder.layer.4.attention.output.dense.bias\", \"roberta.encoder.layer.4.attention.output.LayerNorm.weight\", \"roberta.encoder.layer.4.attention.output.LayerNorm.bias\", \"roberta.encoder.layer.4.intermediate.dense.weight\", \"roberta.encoder.layer.4.intermediate.dense.bias\", \"roberta.encoder.layer.4.output.dense.weight\", \"roberta.encoder.layer.4.output.dense.bias\", \"roberta.encoder.layer.4.output.LayerNorm.weight\", \"roberta.encoder.layer.4.output.LayerNorm.bias\", \"roberta.encoder.layer.5.attention.self.query.weight\", \"roberta.encoder.layer.5.attention.self.query.bias\", \"roberta.encoder.layer.5.attention.self.key.weight\", \"roberta.encoder.layer.5.attention.self.key.bias\", \"roberta.encoder.layer.5.attention.self.value.weight\", \"roberta.encoder.layer.5.attention.self.value.bias\", \"roberta.encoder.layer.5.attention.output.dense.weight\", \"roberta.encoder.layer.5.attention.output.dense.bias\", \"roberta.encoder.layer.5.attention.output.LayerNorm.weight\", \"roberta.encoder.layer.5.attention.output.LayerNorm.bias\", \"roberta.encoder.layer.5.intermediate.dense.weight\", \"roberta.encoder.layer.5.intermediate.dense.bias\", \"roberta.encoder.layer.5.output.dense.weight\", \"roberta.encoder.layer.5.output.dense.bias\", \"roberta.encoder.layer.5.output.LayerNorm.weight\", \"roberta.encoder.layer.5.output.LayerNorm.bias\", \"roberta.encoder.layer.6.attention.self.query.weight\", \"roberta.encoder.layer.6.attention.self.query.bias\", \"roberta.encoder.layer.6.attention.self.key.weight\", \"roberta.encoder.layer.6.attention.self.key.bias\", \"roberta.encoder.layer.6.attention.self.value.weight\", \"roberta.encoder.layer.6.attention.self.value.bias\", \"roberta.encoder.layer.6.attention.output.dense.weight\", \"roberta.encoder.layer.6.attention.output.dense.bias\", \"roberta.encoder.layer.6.attention.output.LayerNorm.weight\", \"roberta.encoder.layer.6.attention.output.LayerNorm.bias\", \"roberta.encoder.layer.6.intermediate.dense.weight\", \"roberta.encoder.layer.6.intermediate.dense.bias\", \"roberta.encoder.layer.6.output.dense.weight\", \"roberta.encoder.layer.6.output.dense.bias\", \"roberta.encoder.layer.6.output.LayerNorm.weight\", \"roberta.encoder.layer.6.output.LayerNorm.bias\", \"roberta.encoder.layer.7.attention.self.query.weight\", \"roberta.encoder.layer.7.attention.self.query.bias\", \"roberta.encoder.layer.7.attention.self.key.weight\", \"roberta.encoder.layer.7.attention.self.key.bias\", \"roberta.encoder.layer.7.attention.self.value.weight\", \"roberta.encoder.layer.7.attention.self.value.bias\", \"roberta.encoder.layer.7.attention.output.dense.weight\", \"roberta.encoder.layer.7.attention.output.dense.bias\", \"roberta.encoder.layer.7.attention.output.LayerNorm.weight\", \"roberta.encoder.layer.7.attention.output.LayerNorm.bias\", \"roberta.encoder.layer.7.intermediate.dense.weight\", \"roberta.encoder.layer.7.intermediate.dense.bias\", \"roberta.encoder.layer.7.output.dense.weight\", \"roberta.encoder.layer.7.output.dense.bias\", \"roberta.encoder.layer.7.output.LayerNorm.weight\", \"roberta.encoder.layer.7.output.LayerNorm.bias\", \"roberta.encoder.layer.8.attention.self.query.weight\", \"roberta.encoder.layer.8.attention.self.query.bias\", \"roberta.encoder.layer.8.attention.self.key.weight\", \"roberta.encoder.layer.8.attention.self.key.bias\", \"roberta.encoder.layer.8.attention.self.value.weight\", \"roberta.encoder.layer.8.attention.self.value.bias\", \"roberta.encoder.layer.8.attention.output.dense.weight\", \"roberta.encoder.layer.8.attention.output.dense.bias\", \"roberta.encoder.layer.8.attention.output.LayerNorm.weight\", \"roberta.encoder.layer.8.attention.output.LayerNorm.bias\", \"roberta.encoder.layer.8.intermediate.dense.weight\", \"roberta.encoder.layer.8.intermediate.dense.bias\", \"roberta.encoder.layer.8.output.dense.weight\", \"roberta.encoder.layer.8.output.dense.bias\", \"roberta.encoder.layer.8.output.LayerNorm.weight\", \"roberta.encoder.layer.8.output.LayerNorm.bias\", \"roberta.encoder.layer.9.attention.self.query.weight\", \"roberta.encoder.layer.9.attention.self.query.bias\", \"roberta.encoder.layer.9.attention.self.key.weight\", \"roberta.encoder.layer.9.attention.self.key.bias\", \"roberta.encoder.layer.9.attention.self.value.weight\", \"roberta.encoder.layer.9.attention.self.value.bias\", \"roberta.encoder.layer.9.attention.output.dense.weight\", \"roberta.encoder.layer.9.attention.output.dense.bias\", \"roberta.encoder.layer.9.attention.output.LayerNorm.weight\", \"roberta.encoder.layer.9.attention.output.LayerNorm.bias\", \"roberta.encoder.layer.9.intermediate.dense.weight\", \"roberta.encoder.layer.9.intermediate.dense.bias\", \"roberta.encoder.layer.9.output.dense.weight\", \"roberta.encoder.layer.9.output.dense.bias\", \"roberta.encoder.layer.9.output.LayerNorm.weight\", \"roberta.encoder.layer.9.output.LayerNorm.bias\", \"roberta.encoder.layer.10.attention.self.query.weight\", \"roberta.encoder.layer.10.attention.self.query.bias\", \"roberta.encoder.layer.10.attention.self.key.weight\", \"roberta.encoder.layer.10.attention.self.key.bias\", \"roberta.encoder.layer.10.attention.self.value.weight\", \"roberta.encoder.layer.10.attention.self.value.bias\", \"roberta.encoder.layer.10.attention.output.dense.weight\", \"roberta.encoder.layer.10.attention.output.dense.bias\", \"roberta.encoder.layer.10.attention.output.LayerNorm.weight\", \"roberta.encoder.layer.10.attention.output.LayerNorm.bias\", \"roberta.encoder.layer.10.intermediate.dense.weight\", \"roberta.encoder.layer.10.intermediate.dense.bias\", \"roberta.encoder.layer.10.output.dense.weight\", \"roberta.encoder.layer.10.output.dense.bias\", \"roberta.encoder.layer.10.output.LayerNorm.weight\", \"roberta.encoder.layer.10.output.LayerNorm.bias\", \"roberta.encoder.layer.11.attention.self.query.weight\", \"roberta.encoder.layer.11.attention.self.query.bias\", \"roberta.encoder.layer.11.attention.self.key.weight\", \"roberta.encoder.layer.11.attention.self.key.bias\", \"roberta.encoder.layer.11.attention.self.value.weight\", \"roberta.encoder.layer.11.attention.self.value.bias\", \"roberta.encoder.layer.11.attention.output.dense.weight\", \"roberta.encoder.layer.11.attention.output.dense.bias\", \"roberta.encoder.layer.11.attention.output.LayerNorm.weight\", \"roberta.encoder.layer.11.attention.output.LayerNorm.bias\", \"roberta.encoder.layer.11.intermediate.dense.weight\", \"roberta.encoder.layer.11.intermediate.dense.bias\", \"roberta.encoder.layer.11.output.dense.weight\", \"roberta.encoder.layer.11.output.dense.bias\", \"roberta.encoder.layer.11.output.LayerNorm.weight\", \"roberta.encoder.layer.11.output.LayerNorm.bias\", \"lm_head.bias\", \"lm_head.dense.weight\", \"lm_head.dense.bias\", \"lm_head.layer_norm.weight\", \"lm_head.layer_norm.bias\", \"lm_head.decoder.weight\", \"lm_head.decoder.bias\". \r\n```", "Oh sorry, the proper class is `AutoModelForMaskedLM` (since this is your current task), not `AutoModel`.", "@sgugger Thanks, I am able to read the model checkpoint now. But the v4.3.2 tokenizer issue still persists, I have tried it in various environments.", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Just to make sure, can you try installing sentencepiece? `pip install sentencepiece`", "Pip says\r\n`Requirement already satisfied: sentencepiece in c:\\users\\chrs\\.virtualenvs\\pythonproject-wdxdk-rq\\lib\\site-packages (0.1.95)`\r\nPipenv \"installs it\" (I guess it just links it) and writes it to the lock-file. Running the example again I get the same error about Protobuf.", "Okay, thank you for trying. Could you show me the steps you did to get this error, seeing as you get the errors on both your cloud instance and your windows machine? I'll try it on my Windows machine and try to reproduce the issue to find out what's happening.", "Yeah the steps are as follows:\r\n1. Create a new pipenv environment\r\n2. Install sentence-transformers\r\n3. Create a python file with the following content\r\n`from sentence-transformers import SentenceTransformer`\r\n`SentenceTransformer('T-Systems-onsite/cross-en-de-roberta-sentence-transformer')`\r\n4. Run the python file => Error", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Facing the same issue with T5. Following demo code:\r\n\r\n```\r\nfrom transformers import AutoTokenizer, T5ForConditionalGeneration\r\n\r\nmodel_name = \"allenai/unifiedqa-t5-small\" \r\ntokenizer = AutoTokenizer.from_pretrained(model_name)\r\n```\r\n", "I had the same problem. I had tried many things like here [Link](https://stackoverflow.com/questions/31308812/no-module-named-google-protobuf) but nothing fixed the problem.\r\n\r\nWith the same environment I worked with the fastai library, which installs quite a few packages. So I created a new environment without fastai and now it works.\r\n\r\nname: [NAME]\r\nchannels:\r\n - conda-forge\r\n - pytorch\r\ndependencies:\r\n - python=3.8\r\n - pandas\r\n - numpy\r\n - scikit-learn\r\n - seaborn\r\n - pytest\r\n - twine\r\n - pip\r\n - ftfy\r\n - xlrd\r\n - ipykernel\r\n - notebook\r\n - pip\r\n - pip:\r\n - azureml-core==1.0.*\r\n - azureml-sdk==1.0.*\r\n - pandas==1.0.5\r\n - numpy==1.17.*\r\n - fastavro==0.22.*\r\n - pandavro==1.5.*\r\n - sentencepiece==0.1.95\r\n\t- datasets==1.8.0\r\n\t- transformers==4.7.0\r\n\t- seqeval==1.2.2\r\n\t- tensorflow==2.5.0\r\n\t- ipywidgets==7.6.3", "As mentioned over [here](https://github.com/huggingface/transformers/issues/9515#issuecomment-869188308), `pip install protobuf` could help. ", "This is still a problem.\r\n\r\nOn an ubuntu cloud instance, I installed in a venv:\r\n```\r\ntorch\r\ntransformers\r\npandas\r\nseaborn\r\njupyter\r\nsentencepiece\r\nprotobuf==3.20.1\r\n```\r\n\r\nI had to downgrade protobuf to 3.20.x for it to work.\r\n\r\nExpected behaviour would be that it works without the need to search the internet to land at this fix.", "Thanks @raoulg. I had the same issue working with the pegasus model, actually from an example in huggingface's new book. Downgrading to 3.20.x was the solution.", "I didn't have to downgrade, just install a missing `protobuf` (latest version). This can be reproduced in e.g. a `Hugging Face` example for e.g. DONUT document classifier using our latest CUDA 11.8 containers: `mirekphd/cuda-11.8-cudnn8-devel-ubuntu22.04:20230928`. Note that the official `nvidia/cuda/11.8.0-cudnn8-devel-ubuntu22.04` containers seem to come with `protobuf` already preinstalled, so you won't reproduce the bug there).\r\n\r\nPerhaps `protobuf` should be added explicitly as a dependency of `transformers`?", "I'm still facing the same error. I have fine tuned mistral model, but I'm trying to inference it, it's still giving me:\r\n\r\nCould not complete request to HuggingFace API, Status Code: 500, Error: \\nLlamaConverter requires the protobuf library but it was not found in your environment. Checkout the instructions on the\\ninstallation page of its repo: https://github.com/protocolbuffers/protobuf/tree/master/python#installation and follow the ones\\nthat match your environment. Please note that you may need to restart your runtime after installation.\\n\r\n\r\nI've done: pip install protobuf, in both env (fine tuning and inferencing)" ]

[]

[]

[ "There is `batch_decode`, yes, the docs are [here](https://huggingface.co/transformers/internal/tokenization_utils.html?highlight=batch_decode#transformers.tokenization_utils_base.PreTrainedTokenizerBase.batch_decode).\r\n\r\n@sgugger I wonder if we shouldn't make the docs of this method more prominent? The \"Utilities for tokenizer\" page mentions: \"Most of those are only useful if you are studying the code of the tokenizers in the library.\", but `batch_decode` and `decode` are only found here, and are very important methods of the tokenization pipeline.", "We should add them to the `PreTrainedTokenizer` and `PreTrainedTokenizerFast` documentation. Or did you want to add them to all models?\r\n", "@LysandreJik `tokenizer.batch_decode` and `tokenizer.decode` in loop, both the functions take almost the same time. can you suggest something, how can I speed up the decoding in T5? why is batch_decode not as fast as batch_encode_plus? Is there a way to make decoding even faster?", "Unfortunately we have no way to go faster than that.", "@LysandreJik this function is used in compute_metrics. and it seems it is limited to the number of GPUs ( it uses the same number of `--nproc_per_node` when doing ddp training, how is it possible to extend that to the maximum number of cores) any guide on how to fix and maybe do a PR? " ]

[]

[]

[ "Hi @BigBird01, thank you for opening the PR! Can you let me know once you're satisfied with your changes so that we can take a look? Thank you!", "hi, Lysandre\r\n\r\nI already tested the code and model. I think it’s good to go. Hope we can merge it into master soon, as there are a lot of people in the community waiting for a try with it.\r\n\r\nThanks!\r\nPengcheng\r\n\r\nGet Outlook for iOS<https://aka.ms/o0ukef>\r\n________________________________\r\nFrom: Lysandre Debut <[email protected]>\r\nSent: Thursday, February 4, 2021 10:21:00 PM\r\nTo: huggingface/transformers <[email protected]>\r\nCc: Pengcheng He <[email protected]>; Mention <[email protected]>\r\nSubject: Re: [huggingface/transformers] Integrate DeBERTa v2(the 1.5B model surpassed human performance on Su… (#10018)\r\n\r\n\r\nHi @BigBird01<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FBigBird01&data=04%7C01%7CPengcheng.H%40microsoft.com%7C4e5ccdeff6dc4cda999b08d8c99e36f9%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637481028648975065%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=fqbkcIs%2FwVkKIqCXBn0NjbZUx2ws7CO6wPKIg8CbQUU%3D&reserved=0>, thank you for opening the PR! Can you let me know once you're satisfied with your changes so that we can take a look? Thank you!\r\n\r\n—\r\nYou are receiving this because you were mentioned.\r\nReply to this email directly, view it on GitHub<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fhuggingface%2Ftransformers%2Fpull%2F10018%23issuecomment-773820084&data=04%7C01%7CPengcheng.H%40microsoft.com%7C4e5ccdeff6dc4cda999b08d8c99e36f9%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637481028648975065%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=UaaB2syxP0gd3AKFP%2BRkxK0TU%2Fh8kEwS8vhu%2FCFvPfQ%3D&reserved=0>, or unsubscribe<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAJDNDRRDJHJGIDNKSUGZ63TS5OE4ZANCNFSM4XDYVO7A&data=04%7C01%7CPengcheng.H%40microsoft.com%7C4e5ccdeff6dc4cda999b08d8c99e36f9%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637481028648985020%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=zYLDqVBRXr3TFu%2FNXABOll%2BBPvT54RvQt1cOHRG4dT4%3D&reserved=0>.\r\n", "I see, thanks. As mentioned by e-mail, I think the correct approach here is to create a `deberta-v2` folder that contains all of the changes, rather than implementing changes in the original `deberta` folder.\r\n\r\nCan I handle that for you?", "> I see, thanks. As mentioned by e-mail, I think the correct approach here is to create a `deberta-v2` folder that contains all of the changes, rather than implementing changes in the original `deberta` folder.\r\n> \r\n> Can I handle that for you?\r\n\r\nBut I think the current implementation is better. First ,the current changes not only contain the new features of v2 but also some improvements to v1. Second, the change between v2 and v1 is small. I also tested all the models with current implementation, and I didn't find any regression. Third and the most important, by creating another folder for deberta-v2 we need to add redundant code and tests to cover v2. This may introduce additional maintain effort in the future. \r\n\r\nLet me know what's your thought.\r\n", "In that case, just feel free to take over the change and follow the rule to merge it to master. Please let me know when you finish it and I will take a test over it. Thanks in advance @[email protected]<mailto:[email protected]>\n\nGet Outlook for iOS<https://aka.ms/o0ukef>\n________________________________\nFrom: Lysandre Debut <[email protected]>\nSent: Thursday, February 4, 2021 10:55:24 PM\nTo: huggingface/transformers <[email protected]>\nCc: Pengcheng He <[email protected]>; Mention <[email protected]>\nSubject: Re: [huggingface/transformers] Integrate DeBERTa v2(the 1.5B model surpassed human performance on Su… (#10018)\n\n\n@LysandreJik commented on this pull request.\n\nThe issues with modifying the code of the first version are:\n\n * We might inadvertently modify some of the behavior of the past model\n * We don't know what is the difference between the first and second version\n\nFor example here the DisentangledSelfAttention layer gets radically changed, with some layer name changes, which makes me dubious that you can load first version checkpoints inside.\n\nFinally, you make a good point regarding maintainability. However, we can still enforce this by building some tools which ensure that the code does not diverge. We have this setup for a multitude of models, for example BART is very similar to mBART, Pegasus, Marian.\n\nPlease take a look at the mBART code and look for the \"# Copied from ...\" comments, such as the following:\n\nhttps://github.com/huggingface/transformers/blob/3be965c5dbee794a7a3606df6a1ae36a0d65904d/src/transformers/models/mbart/modeling_mbart.py#L96-L108<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fhuggingface%2Ftransformers%2Fblob%2F3be965c5dbee794a7a3606df6a1ae36a0d65904d%2Fsrc%2Ftransformers%2Fmodels%2Fmbart%2Fmodeling_mbart.py%23L96-L108&data=04%7C01%7CPengcheng.H%40microsoft.com%7Ce6bf8235a2654c89b0bc08d8c9a30491%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637481049290703674%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=pvA7y0lA326x2LKdGrnEaFbrjl7AI5tVvk3fotCeQM0%3D&reserved=0>\n\nThis ensures that the two implementations do not diverge, it helps identify where the code is different, and it is what we've chosen to go through in order to keep readability to a maximum.\n\n—\nYou are receiving this because you were mentioned.\nReply to this email directly, view it on GitHub<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fhuggingface%2Ftransformers%2Fpull%2F10018%23pullrequestreview-584064416&data=04%7C01%7CPengcheng.H%40microsoft.com%7Ce6bf8235a2654c89b0bc08d8c9a30491%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637481049290703674%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=6taj4jBCjpauA8sHZETQSZgUWxe7IaPYLDzorhvY4EE%3D&reserved=0>, or unsubscribe<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAJDNDRUB2E3QCL4LSA6B34TS5OI5ZANCNFSM4XDYVO7A&data=04%7C01%7CPengcheng.H%40microsoft.com%7Ce6bf8235a2654c89b0bc08d8c9a30491%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637481049290713636%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=q2G2HWxOOd1Hz2KJzEAeGDm8QTVolDlkkoLFrCUyrsE%3D&reserved=0>.\n", "This works for me, thank you for your understanding. I'll ping you once the PR can be reviewed.", "Great! Thanks!\n\nGet Outlook for iOS<https://aka.ms/o0ukef>\n________________________________\nFrom: Lysandre Debut <[email protected]>\nSent: Thursday, February 4, 2021 11:03:05 PM\nTo: huggingface/transformers <[email protected]>\nCc: Pengcheng He <[email protected]>; Mention <[email protected]>\nSubject: Re: [huggingface/transformers] Integrate DeBERTa v2(the 1.5B model surpassed human performance on Su… (#10018)\n\n\nThis works for me, thank you for your understanding. I'll ping you once the PR can be reviewed.\n\n—\nYou are receiving this because you were mentioned.\nReply to this email directly, view it on GitHub<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fhuggingface%2Ftransformers%2Fpull%2F10018%23issuecomment-773838474&data=04%7C01%7CPengcheng.H%40microsoft.com%7C3092474ab5094c50311108d8c9a417aa%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637481053901879595%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=N6LqcZcPYCxNOm%2BBvLmPt3ZvG%2FmegyaNZFn5AEqLz10%3D&reserved=0>, or unsubscribe<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAJDNDRQD247WMEFTE7SRXNTS5OJ2TANCNFSM4XDYVO7A&data=04%7C01%7CPengcheng.H%40microsoft.com%7C3092474ab5094c50311108d8c9a417aa%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637481053901879595%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=lL68lmcS4iGlBfZvNttTSTsXpCq8hNTWhJUDzbFMtsA%3D&reserved=0>.\n", "PR to split the two models is here: https://github.com/BigBird01/transformers/pull/1", "@BigBird01 just wanted to ask if the new additions involve the base and large versions of v2 as well, because i saw that new base and large deberta models were added as well, or will they be just v1?", "> @BigBird01 just wanted to ask if the new additions involve the base and large versions of v2 as well, because i saw that new base and large deberta models were added as well, or will they be just v1?\r\n\r\nFor v2 we don't have base and large yet. But we will add them in the future.", "Are there any bottlenecks preventing this from being merged?", "> @BigBird01 just wanted to ask if the new additions involve the base and large versions of v2 as well, because i saw that new base and large deberta models were added as well, or will they be just v1?\r\n\r\nI think @LysandreJik will merge the changes to master soon.\r\n\r\n> PR to split the two models is here: [BigBird01#1](https://github.com/BigBird01/transformers/pull/1)\r\n\r\nThanks @LysandreJik. I just reviewed the PR and I'm good with it\r\n\r\n> Are there any bottlenecks preventing this from being merged?\r\n\r\n", "After playing around with the model, I don't think we need pre-load hooks after all. In order to load the MNLI checkpoints, you just need to specify to the model that it needs three labels. It can be done as follows:\r\n\r\n```py\r\nfrom transformers import DebertaV2ForSequenceClassification\r\n\r\nmodel = DebertaV2ForSequenceClassification.from_pretrained(\"microsoft/deberta-v2-xlarge-mnli\", num_labels=3)\r\n```\r\n\r\nBut this should be taken care of in the configuration. I believe all your MNLI model configurations should have the `num_labels` field set to `3` in order to be loadable.\r\n\r\n---\r\n\r\nFollowing this, I found a few issues with the XLARGE MNLI checkpoint. When loading it in the `DebertaForSequenceClassification` model, I get the following messages:\r\n\r\n```\r\nSome weights of the model checkpoint at microsoft/deberta-xlarge-mnli were not used when initializing DebertaForSequenceClassification: ['deberta.encoder.layer.0.attention.self.query_proj.weight', 'deberta.encoder.layer.0.attention.self.query_proj.bias', 'deberta.encoder.layer.0.attention.self.key_proj.weight', 'deberta.encoder.layer.0.attention.self.key_proj.bias', 'deberta.encoder.layer.0.attention.self.value_proj.weight', 'deberta.encoder.layer.0.attention.self.value_proj.bias', 'deberta.encoder.layer.0.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.0.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.0.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.0.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.1.attention.self.query_proj.weight', 'deberta.encoder.layer.1.attention.self.query_proj.bias', 'deberta.encoder.layer.1.attention.self.key_proj.weight', 'deberta.encoder.layer.1.attention.self.key_proj.bias', 'deberta.encoder.layer.1.attention.self.value_proj.weight', 'deberta.encoder.layer.1.attention.self.value_proj.bias', 'deberta.encoder.layer.1.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.1.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.1.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.1.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.2.attention.self.query_proj.weight', 'deberta.encoder.layer.2.attention.self.query_proj.bias', 'deberta.encoder.layer.2.attention.self.key_proj.weight', 'deberta.encoder.layer.2.attention.self.key_proj.bias', 'deberta.encoder.layer.2.attention.self.value_proj.weight', 'deberta.encoder.layer.2.attention.self.value_proj.bias', 'deberta.encoder.layer.2.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.2.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.2.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.2.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.3.attention.self.query_proj.weight', 'deberta.encoder.layer.3.attention.self.query_proj.bias', 'deberta.encoder.layer.3.attention.self.key_proj.weight', 'deberta.encoder.layer.3.attention.self.key_proj.bias', 'deberta.encoder.layer.3.attention.self.value_proj.weight', 'deberta.encoder.layer.3.attention.self.value_proj.bias', 'deberta.encoder.layer.3.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.3.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.3.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.3.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.4.attention.self.query_proj.weight', 'deberta.encoder.layer.4.attention.self.query_proj.bias', 'deberta.encoder.layer.4.attention.self.key_proj.weight', 'deberta.encoder.layer.4.attention.self.key_proj.bias', 'deberta.encoder.layer.4.attention.self.value_proj.weight', 'deberta.encoder.layer.4.attention.self.value_proj.bias', 'deberta.encoder.layer.4.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.4.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.4.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.4.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.5.attention.self.query_proj.weight', 'deberta.encoder.layer.5.attention.self.query_proj.bias', 'deberta.encoder.layer.5.attention.self.key_proj.weight', 'deberta.encoder.layer.5.attention.self.key_proj.bias', 'deberta.encoder.layer.5.attention.self.value_proj.weight', 'deberta.encoder.layer.5.attention.self.value_proj.bias', 'deberta.encoder.layer.5.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.5.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.5.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.5.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.6.attention.self.query_proj.weight', 'deberta.encoder.layer.6.attention.self.query_proj.bias', 'deberta.encoder.layer.6.attention.self.key_proj.weight', 'deberta.encoder.layer.6.attention.self.key_proj.bias', 'deberta.encoder.layer.6.attention.self.value_proj.weight', 'deberta.encoder.layer.6.attention.self.value_proj.bias', 'deberta.encoder.layer.6.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.6.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.6.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.6.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.7.attention.self.query_proj.weight', 'deberta.encoder.layer.7.attention.self.query_proj.bias', 'deberta.encoder.layer.7.attention.self.key_proj.weight', 'deberta.encoder.layer.7.attention.self.key_proj.bias', 'deberta.encoder.layer.7.attention.self.value_proj.weight', 'deberta.encoder.layer.7.attention.self.value_proj.bias', 'deberta.encoder.layer.7.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.7.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.7.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.7.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.8.attention.self.query_proj.weight', 'deberta.encoder.layer.8.attention.self.query_proj.bias', 'deberta.encoder.layer.8.attention.self.key_proj.weight', 'deberta.encoder.layer.8.attention.self.key_proj.bias', 'deberta.encoder.layer.8.attention.self.value_proj.weight', 'deberta.encoder.layer.8.attention.self.value_proj.bias', 'deberta.encoder.layer.8.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.8.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.8.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.8.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.9.attention.self.query_proj.weight', 'deberta.encoder.layer.9.attention.self.query_proj.bias', 'deberta.encoder.layer.9.attention.self.key_proj.weight', 'deberta.encoder.layer.9.attention.self.key_proj.bias', 'deberta.encoder.layer.9.attention.self.value_proj.weight', 'deberta.encoder.layer.9.attention.self.value_proj.bias', 'deberta.encoder.layer.9.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.9.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.9.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.9.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.10.attention.self.query_proj.weight', 'deberta.encoder.layer.10.attention.self.query_proj.bias', 'deberta.encoder.layer.10.attention.self.key_proj.weight', 'deberta.encoder.layer.10.attention.self.key_proj.bias', 'deberta.encoder.layer.10.attention.self.value_proj.weight', 'deberta.encoder.layer.10.attention.self.value_proj.bias', 'deberta.encoder.layer.10.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.10.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.10.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.10.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.11.attention.self.query_proj.weight', 'deberta.encoder.layer.11.attention.self.query_proj.bias', 'deberta.encoder.layer.11.attention.self.key_proj.weight', 'deberta.encoder.layer.11.attention.self.key_proj.bias', 'deberta.encoder.layer.11.attention.self.value_proj.weight', 'deberta.encoder.layer.11.attention.self.value_proj.bias', 'deberta.encoder.layer.11.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.11.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.11.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.11.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.12.attention.self.query_proj.weight', 'deberta.encoder.layer.12.attention.self.query_proj.bias', 'deberta.encoder.layer.12.attention.self.key_proj.weight', 'deberta.encoder.layer.12.attention.self.key_proj.bias', 'deberta.encoder.layer.12.attention.self.value_proj.weight', 'deberta.encoder.layer.12.attention.self.value_proj.bias', 'deberta.encoder.layer.12.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.12.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.12.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.12.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.13.attention.self.query_proj.weight', 'deberta.encoder.layer.13.attention.self.query_proj.bias', 'deberta.encoder.layer.13.attention.self.key_proj.weight', 'deberta.encoder.layer.13.attention.self.key_proj.bias', 'deberta.encoder.layer.13.attention.self.value_proj.weight', 'deberta.encoder.layer.13.attention.self.value_proj.bias', 'deberta.encoder.layer.13.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.13.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.13.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.13.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.14.attention.self.query_proj.weight', 'deberta.encoder.layer.14.attention.self.query_proj.bias', 'deberta.encoder.layer.14.attention.self.key_proj.weight', 'deberta.encoder.layer.14.attention.self.key_proj.bias', 'deberta.encoder.layer.14.attention.self.value_proj.weight', 'deberta.encoder.layer.14.attention.self.value_proj.bias', 'deberta.encoder.layer.14.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.14.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.14.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.14.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.15.attention.self.query_proj.weight', 'deberta.encoder.layer.15.attention.self.query_proj.bias', 'deberta.encoder.layer.15.attention.self.key_proj.weight', 'deberta.encoder.layer.15.attention.self.key_proj.bias', 'deberta.encoder.layer.15.attention.self.value_proj.weight', 'deberta.encoder.layer.15.attention.self.value_proj.bias', 'deberta.encoder.layer.15.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.15.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.15.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.15.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.16.attention.self.query_proj.weight', 'deberta.encoder.layer.16.attention.self.query_proj.bias', 'deberta.encoder.layer.16.attention.self.key_proj.weight', 'deberta.encoder.layer.16.attention.self.key_proj.bias', 'deberta.encoder.layer.16.attention.self.value_proj.weight', 'deberta.encoder.layer.16.attention.self.value_proj.bias', 'deberta.encoder.layer.16.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.16.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.16.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.16.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.17.attention.self.query_proj.weight', 'deberta.encoder.layer.17.attention.self.query_proj.bias', 'deberta.encoder.layer.17.attention.self.key_proj.weight', 'deberta.encoder.layer.17.attention.self.key_proj.bias', 'deberta.encoder.layer.17.attention.self.value_proj.weight', 'deberta.encoder.layer.17.attention.self.value_proj.bias', 'deberta.encoder.layer.17.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.17.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.17.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.17.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.18.attention.self.query_proj.weight', 'deberta.encoder.layer.18.attention.self.query_proj.bias', 'deberta.encoder.layer.18.attention.self.key_proj.weight', 'deberta.encoder.layer.18.attention.self.key_proj.bias', 'deberta.encoder.layer.18.attention.self.value_proj.weight', 'deberta.encoder.layer.18.attention.self.value_proj.bias', 'deberta.encoder.layer.18.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.18.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.18.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.18.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.19.attention.self.query_proj.weight', 'deberta.encoder.layer.19.attention.self.query_proj.bias', 'deberta.encoder.layer.19.attention.self.key_proj.weight', 'deberta.encoder.layer.19.attention.self.key_proj.bias', 'deberta.encoder.layer.19.attention.self.value_proj.weight', 'deberta.encoder.layer.19.attention.self.value_proj.bias', 'deberta.encoder.layer.19.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.19.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.19.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.19.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.20.attention.self.query_proj.weight', 'deberta.encoder.layer.20.attention.self.query_proj.bias', 'deberta.encoder.layer.20.attention.self.key_proj.weight', 'deberta.encoder.layer.20.attention.self.key_proj.bias', 'deberta.encoder.layer.20.attention.self.value_proj.weight', 'deberta.encoder.layer.20.attention.self.value_proj.bias', 'deberta.encoder.layer.20.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.20.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.20.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.20.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.21.attention.self.query_proj.weight', 'deberta.encoder.layer.21.attention.self.query_proj.bias', 'deberta.encoder.layer.21.attention.self.key_proj.weight', 'deberta.encoder.layer.21.attention.self.key_proj.bias', 'deberta.encoder.layer.21.attention.self.value_proj.weight', 'deberta.encoder.layer.21.attention.self.value_proj.bias', 'deberta.encoder.layer.21.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.21.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.21.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.21.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.22.attention.self.query_proj.weight', 'deberta.encoder.layer.22.attention.self.query_proj.bias', 'deberta.encoder.layer.22.attention.self.key_proj.weight', 'deberta.encoder.layer.22.attention.self.key_proj.bias', 'deberta.encoder.layer.22.attention.self.value_proj.weight', 'deberta.encoder.layer.22.attention.self.value_proj.bias', 'deberta.encoder.layer.22.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.22.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.22.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.22.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.23.attention.self.query_proj.weight', 'deberta.encoder.layer.23.attention.self.query_proj.bias', 'deberta.encoder.layer.23.attention.self.key_proj.weight', 'deberta.encoder.layer.23.attention.self.key_proj.bias', 'deberta.encoder.layer.23.attention.self.value_proj.weight', 'deberta.encoder.layer.23.attention.self.value_proj.bias', 'deberta.encoder.layer.23.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.23.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.23.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.23.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.24.attention.self.query_proj.weight', 'deberta.encoder.layer.24.attention.self.query_proj.bias', 'deberta.encoder.layer.24.attention.self.key_proj.weight', 'deberta.encoder.layer.24.attention.self.key_proj.bias', 'deberta.encoder.layer.24.attention.self.value_proj.weight', 'deberta.encoder.layer.24.attention.self.value_proj.bias', 'deberta.encoder.layer.24.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.24.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.24.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.24.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.25.attention.self.query_proj.weight', 'deberta.encoder.layer.25.attention.self.query_proj.bias', 'deberta.encoder.layer.25.attention.self.key_proj.weight', 'deberta.encoder.layer.25.attention.self.key_proj.bias', 'deberta.encoder.layer.25.attention.self.value_proj.weight', 'deberta.encoder.layer.25.attention.self.value_proj.bias', 'deberta.encoder.layer.25.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.25.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.25.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.25.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.26.attention.self.query_proj.weight', 'deberta.encoder.layer.26.attention.self.query_proj.bias', 'deberta.encoder.layer.26.attention.self.key_proj.weight', 'deberta.encoder.layer.26.attention.self.key_proj.bias', 'deberta.encoder.layer.26.attention.self.value_proj.weight', 'deberta.encoder.layer.26.attention.self.value_proj.bias', 'deberta.encoder.layer.26.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.26.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.26.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.26.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.27.attention.self.query_proj.weight', 'deberta.encoder.layer.27.attention.self.query_proj.bias', 'deberta.encoder.layer.27.attention.self.key_proj.weight', 'deberta.encoder.layer.27.attention.self.key_proj.bias', 'deberta.encoder.layer.27.attention.self.value_proj.weight', 'deberta.encoder.layer.27.attention.self.value_proj.bias', 'deberta.encoder.layer.27.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.27.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.27.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.27.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.28.attention.self.query_proj.weight', 'deberta.encoder.layer.28.attention.self.query_proj.bias', 'deberta.encoder.layer.28.attention.self.key_proj.weight', 'deberta.encoder.layer.28.attention.self.key_proj.bias', 'deberta.encoder.layer.28.attention.self.value_proj.weight', 'deberta.encoder.layer.28.attention.self.value_proj.bias', 'deberta.encoder.layer.28.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.28.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.28.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.28.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.29.attention.self.query_proj.weight', 'deberta.encoder.layer.29.attention.self.query_proj.bias', 'deberta.encoder.layer.29.attention.self.key_proj.weight', 'deberta.encoder.layer.29.attention.self.key_proj.bias', 'deberta.encoder.layer.29.attention.self.value_proj.weight', 'deberta.encoder.layer.29.attention.self.value_proj.bias', 'deberta.encoder.layer.29.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.29.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.29.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.29.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.30.attention.self.query_proj.weight', 'deberta.encoder.layer.30.attention.self.query_proj.bias', 'deberta.encoder.layer.30.attention.self.key_proj.weight', 'deberta.encoder.layer.30.attention.self.key_proj.bias', 'deberta.encoder.layer.30.attention.self.value_proj.weight', 'deberta.encoder.layer.30.attention.self.value_proj.bias', 'deberta.encoder.layer.30.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.30.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.30.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.30.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.31.attention.self.query_proj.weight', 'deberta.encoder.layer.31.attention.self.query_proj.bias', 'deberta.encoder.layer.31.attention.self.key_proj.weight', 'deberta.encoder.layer.31.attention.self.key_proj.bias', 'deberta.encoder.layer.31.attention.self.value_proj.weight', 'deberta.encoder.layer.31.attention.self.value_proj.bias', 'deberta.encoder.layer.31.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.31.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.31.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.31.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.32.attention.self.query_proj.weight', 'deberta.encoder.layer.32.attention.self.query_proj.bias', 'deberta.encoder.layer.32.attention.self.key_proj.weight', 'deberta.encoder.layer.32.attention.self.key_proj.bias', 'deberta.encoder.layer.32.attention.self.value_proj.weight', 'deberta.encoder.layer.32.attention.self.value_proj.bias', 'deberta.encoder.layer.32.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.32.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.32.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.32.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.33.attention.self.query_proj.weight', 'deberta.encoder.layer.33.attention.self.query_proj.bias', 'deberta.encoder.layer.33.attention.self.key_proj.weight', 'deberta.encoder.layer.33.attention.self.key_proj.bias', 'deberta.encoder.layer.33.attention.self.value_proj.weight', 'deberta.encoder.layer.33.attention.self.value_proj.bias', 'deberta.encoder.layer.33.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.33.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.33.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.33.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.34.attention.self.query_proj.weight', 'deberta.encoder.layer.34.attention.self.query_proj.bias', 'deberta.encoder.layer.34.attention.self.key_proj.weight', 'deberta.encoder.layer.34.attention.self.key_proj.bias', 'deberta.encoder.layer.34.attention.self.value_proj.weight', 'deberta.encoder.layer.34.attention.self.value_proj.bias', 'deberta.encoder.layer.34.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.34.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.34.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.34.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.35.attention.self.query_proj.weight', 'deberta.encoder.layer.35.attention.self.query_proj.bias', 'deberta.encoder.layer.35.attention.self.key_proj.weight', 'deberta.encoder.layer.35.attention.self.key_proj.bias', 'deberta.encoder.layer.35.attention.self.value_proj.weight', 'deberta.encoder.layer.35.attention.self.value_proj.bias', 'deberta.encoder.layer.35.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.35.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.35.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.35.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.36.attention.self.query_proj.weight', 'deberta.encoder.layer.36.attention.self.query_proj.bias', 'deberta.encoder.layer.36.attention.self.key_proj.weight', 'deberta.encoder.layer.36.attention.self.key_proj.bias', 'deberta.encoder.layer.36.attention.self.value_proj.weight', 'deberta.encoder.layer.36.attention.self.value_proj.bias', 'deberta.encoder.layer.36.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.36.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.36.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.36.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.37.attention.self.query_proj.weight', 'deberta.encoder.layer.37.attention.self.query_proj.bias', 'deberta.encoder.layer.37.attention.self.key_proj.weight', 'deberta.encoder.layer.37.attention.self.key_proj.bias', 'deberta.encoder.layer.37.attention.self.value_proj.weight', 'deberta.encoder.layer.37.attention.self.value_proj.bias', 'deberta.encoder.layer.37.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.37.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.37.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.37.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.38.attention.self.query_proj.weight', 'deberta.encoder.layer.38.attention.self.query_proj.bias', 'deberta.encoder.layer.38.attention.self.key_proj.weight', 'deberta.encoder.layer.38.attention.self.key_proj.bias', 'deberta.encoder.layer.38.attention.self.value_proj.weight', 'deberta.encoder.layer.38.attention.self.value_proj.bias', 'deberta.encoder.layer.38.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.38.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.38.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.38.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.39.attention.self.query_proj.weight', 'deberta.encoder.layer.39.attention.self.query_proj.bias', 'deberta.encoder.layer.39.attention.self.key_proj.weight', 'deberta.encoder.layer.39.attention.self.key_proj.bias', 'deberta.encoder.layer.39.attention.self.value_proj.weight', 'deberta.encoder.layer.39.attention.self.value_proj.bias', 'deberta.encoder.layer.39.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.39.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.39.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.39.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.40.attention.self.query_proj.weight', 'deberta.encoder.layer.40.attention.self.query_proj.bias', 'deberta.encoder.layer.40.attention.self.key_proj.weight', 'deberta.encoder.layer.40.attention.self.key_proj.bias', 'deberta.encoder.layer.40.attention.self.value_proj.weight', 'deberta.encoder.layer.40.attention.self.value_proj.bias', 'deberta.encoder.layer.40.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.40.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.40.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.40.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.41.attention.self.query_proj.weight', 'deberta.encoder.layer.41.attention.self.query_proj.bias', 'deberta.encoder.layer.41.attention.self.key_proj.weight', 'deberta.encoder.layer.41.attention.self.key_proj.bias', 'deberta.encoder.layer.41.attention.self.value_proj.weight', 'deberta.encoder.layer.41.attention.self.value_proj.bias', 'deberta.encoder.layer.41.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.41.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.41.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.41.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.42.attention.self.query_proj.weight', 'deberta.encoder.layer.42.attention.self.query_proj.bias', 'deberta.encoder.layer.42.attention.self.key_proj.weight', 'deberta.encoder.layer.42.attention.self.key_proj.bias', 'deberta.encoder.layer.42.attention.self.value_proj.weight', 'deberta.encoder.layer.42.attention.self.value_proj.bias', 'deberta.encoder.layer.42.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.42.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.42.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.42.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.43.attention.self.query_proj.weight', 'deberta.encoder.layer.43.attention.self.query_proj.bias', 'deberta.encoder.layer.43.attention.self.key_proj.weight', 'deberta.encoder.layer.43.attention.self.key_proj.bias', 'deberta.encoder.layer.43.attention.self.value_proj.weight', 'deberta.encoder.layer.43.attention.self.value_proj.bias', 'deberta.encoder.layer.43.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.43.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.43.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.43.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.44.attention.self.query_proj.weight', 'deberta.encoder.layer.44.attention.self.query_proj.bias', 'deberta.encoder.layer.44.attention.self.key_proj.weight', 'deberta.encoder.layer.44.attention.self.key_proj.bias', 'deberta.encoder.layer.44.attention.self.value_proj.weight', 'deberta.encoder.layer.44.attention.self.value_proj.bias', 'deberta.encoder.layer.44.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.44.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.44.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.44.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.45.attention.self.query_proj.weight', 'deberta.encoder.layer.45.attention.self.query_proj.bias', 'deberta.encoder.layer.45.attention.self.key_proj.weight', 'deberta.encoder.layer.45.attention.self.key_proj.bias', 'deberta.encoder.layer.45.attention.self.value_proj.weight', 'deberta.encoder.layer.45.attention.self.value_proj.bias', 'deberta.encoder.layer.45.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.45.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.45.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.45.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.46.attention.self.query_proj.weight', 'deberta.encoder.layer.46.attention.self.query_proj.bias', 'deberta.encoder.layer.46.attention.self.key_proj.weight', 'deberta.encoder.layer.46.attention.self.key_proj.bias', 'deberta.encoder.layer.46.attention.self.value_proj.weight', 'deberta.encoder.layer.46.attention.self.value_proj.bias', 'deberta.encoder.layer.46.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.46.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.46.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.46.attention.self.pos_query_proj.bias', 'deberta.encoder.layer.47.attention.self.query_proj.weight', 'deberta.encoder.layer.47.attention.self.query_proj.bias', 'deberta.encoder.layer.47.attention.self.key_proj.weight', 'deberta.encoder.layer.47.attention.self.key_proj.bias', 'deberta.encoder.layer.47.attention.self.value_proj.weight', 'deberta.encoder.layer.47.attention.self.value_proj.bias', 'deberta.encoder.layer.47.attention.self.pos_key_proj.weight', 'deberta.encoder.layer.47.attention.self.pos_key_proj.bias', 'deberta.encoder.layer.47.attention.self.pos_query_proj.weight', 'deberta.encoder.layer.47.attention.self.pos_query_proj.bias']\r\n- This IS expected if you are initializing DebertaForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\r\n- This IS NOT expected if you are initializing DebertaForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\r\nSome weights of DebertaForSequenceClassification were not initialized from the model checkpoint at microsoft/deberta-xlarge-mnli and are newly initialized: ['deberta.encoder.layer.0.attention.self.q_bias', 'deberta.encoder.layer.0.attention.self.v_bias', 'deberta.encoder.layer.0.attention.self.in_proj.weight', 'deberta.encoder.layer.0.attention.self.pos_proj.weight', 'deberta.encoder.layer.0.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.0.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.1.attention.self.q_bias', 'deberta.encoder.layer.1.attention.self.v_bias', 'deberta.encoder.layer.1.attention.self.in_proj.weight', 'deberta.encoder.layer.1.attention.self.pos_proj.weight', 'deberta.encoder.layer.1.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.1.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.2.attention.self.q_bias', 'deberta.encoder.layer.2.attention.self.v_bias', 'deberta.encoder.layer.2.attention.self.in_proj.weight', 'deberta.encoder.layer.2.attention.self.pos_proj.weight', 'deberta.encoder.layer.2.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.2.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.3.attention.self.q_bias', 'deberta.encoder.layer.3.attention.self.v_bias', 'deberta.encoder.layer.3.attention.self.in_proj.weight', 'deberta.encoder.layer.3.attention.self.pos_proj.weight', 'deberta.encoder.layer.3.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.3.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.4.attention.self.q_bias', 'deberta.encoder.layer.4.attention.self.v_bias', 'deberta.encoder.layer.4.attention.self.in_proj.weight', 'deberta.encoder.layer.4.attention.self.pos_proj.weight', 'deberta.encoder.layer.4.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.4.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.5.attention.self.q_bias', 'deberta.encoder.layer.5.attention.self.v_bias', 'deberta.encoder.layer.5.attention.self.in_proj.weight', 'deberta.encoder.layer.5.attention.self.pos_proj.weight', 'deberta.encoder.layer.5.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.5.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.6.attention.self.q_bias', 'deberta.encoder.layer.6.attention.self.v_bias', 'deberta.encoder.layer.6.attention.self.in_proj.weight', 'deberta.encoder.layer.6.attention.self.pos_proj.weight', 'deberta.encoder.layer.6.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.6.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.7.attention.self.q_bias', 'deberta.encoder.layer.7.attention.self.v_bias', 'deberta.encoder.layer.7.attention.self.in_proj.weight', 'deberta.encoder.layer.7.attention.self.pos_proj.weight', 'deberta.encoder.layer.7.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.7.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.8.attention.self.q_bias', 'deberta.encoder.layer.8.attention.self.v_bias', 'deberta.encoder.layer.8.attention.self.in_proj.weight', 'deberta.encoder.layer.8.attention.self.pos_proj.weight', 'deberta.encoder.layer.8.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.8.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.9.attention.self.q_bias', 'deberta.encoder.layer.9.attention.self.v_bias', 'deberta.encoder.layer.9.attention.self.in_proj.weight', 'deberta.encoder.layer.9.attention.self.pos_proj.weight', 'deberta.encoder.layer.9.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.9.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.10.attention.self.q_bias', 'deberta.encoder.layer.10.attention.self.v_bias', 'deberta.encoder.layer.10.attention.self.in_proj.weight', 'deberta.encoder.layer.10.attention.self.pos_proj.weight', 'deberta.encoder.layer.10.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.10.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.11.attention.self.q_bias', 'deberta.encoder.layer.11.attention.self.v_bias', 'deberta.encoder.layer.11.attention.self.in_proj.weight', 'deberta.encoder.layer.11.attention.self.pos_proj.weight', 'deberta.encoder.layer.11.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.11.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.12.attention.self.q_bias', 'deberta.encoder.layer.12.attention.self.v_bias', 'deberta.encoder.layer.12.attention.self.in_proj.weight', 'deberta.encoder.layer.12.attention.self.pos_proj.weight', 'deberta.encoder.layer.12.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.12.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.13.attention.self.q_bias', 'deberta.encoder.layer.13.attention.self.v_bias', 'deberta.encoder.layer.13.attention.self.in_proj.weight', 'deberta.encoder.layer.13.attention.self.pos_proj.weight', 'deberta.encoder.layer.13.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.13.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.14.attention.self.q_bias', 'deberta.encoder.layer.14.attention.self.v_bias', 'deberta.encoder.layer.14.attention.self.in_proj.weight', 'deberta.encoder.layer.14.attention.self.pos_proj.weight', 'deberta.encoder.layer.14.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.14.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.15.attention.self.q_bias', 'deberta.encoder.layer.15.attention.self.v_bias', 'deberta.encoder.layer.15.attention.self.in_proj.weight', 'deberta.encoder.layer.15.attention.self.pos_proj.weight', 'deberta.encoder.layer.15.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.15.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.16.attention.self.q_bias', 'deberta.encoder.layer.16.attention.self.v_bias', 'deberta.encoder.layer.16.attention.self.in_proj.weight', 'deberta.encoder.layer.16.attention.self.pos_proj.weight', 'deberta.encoder.layer.16.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.16.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.17.attention.self.q_bias', 'deberta.encoder.layer.17.attention.self.v_bias', 'deberta.encoder.layer.17.attention.self.in_proj.weight', 'deberta.encoder.layer.17.attention.self.pos_proj.weight', 'deberta.encoder.layer.17.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.17.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.18.attention.self.q_bias', 'deberta.encoder.layer.18.attention.self.v_bias', 'deberta.encoder.layer.18.attention.self.in_proj.weight', 'deberta.encoder.layer.18.attention.self.pos_proj.weight', 'deberta.encoder.layer.18.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.18.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.19.attention.self.q_bias', 'deberta.encoder.layer.19.attention.self.v_bias', 'deberta.encoder.layer.19.attention.self.in_proj.weight', 'deberta.encoder.layer.19.attention.self.pos_proj.weight', 'deberta.encoder.layer.19.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.19.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.20.attention.self.q_bias', 'deberta.encoder.layer.20.attention.self.v_bias', 'deberta.encoder.layer.20.attention.self.in_proj.weight', 'deberta.encoder.layer.20.attention.self.pos_proj.weight', 'deberta.encoder.layer.20.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.20.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.21.attention.self.q_bias', 'deberta.encoder.layer.21.attention.self.v_bias', 'deberta.encoder.layer.21.attention.self.in_proj.weight', 'deberta.encoder.layer.21.attention.self.pos_proj.weight', 'deberta.encoder.layer.21.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.21.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.22.attention.self.q_bias', 'deberta.encoder.layer.22.attention.self.v_bias', 'deberta.encoder.layer.22.attention.self.in_proj.weight', 'deberta.encoder.layer.22.attention.self.pos_proj.weight', 'deberta.encoder.layer.22.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.22.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.23.attention.self.q_bias', 'deberta.encoder.layer.23.attention.self.v_bias', 'deberta.encoder.layer.23.attention.self.in_proj.weight', 'deberta.encoder.layer.23.attention.self.pos_proj.weight', 'deberta.encoder.layer.23.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.23.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.24.attention.self.q_bias', 'deberta.encoder.layer.24.attention.self.v_bias', 'deberta.encoder.layer.24.attention.self.in_proj.weight', 'deberta.encoder.layer.24.attention.self.pos_proj.weight', 'deberta.encoder.layer.24.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.24.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.25.attention.self.q_bias', 'deberta.encoder.layer.25.attention.self.v_bias', 'deberta.encoder.layer.25.attention.self.in_proj.weight', 'deberta.encoder.layer.25.attention.self.pos_proj.weight', 'deberta.encoder.layer.25.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.25.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.26.attention.self.q_bias', 'deberta.encoder.layer.26.attention.self.v_bias', 'deberta.encoder.layer.26.attention.self.in_proj.weight', 'deberta.encoder.layer.26.attention.self.pos_proj.weight', 'deberta.encoder.layer.26.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.26.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.27.attention.self.q_bias', 'deberta.encoder.layer.27.attention.self.v_bias', 'deberta.encoder.layer.27.attention.self.in_proj.weight', 'deberta.encoder.layer.27.attention.self.pos_proj.weight', 'deberta.encoder.layer.27.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.27.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.28.attention.self.q_bias', 'deberta.encoder.layer.28.attention.self.v_bias', 'deberta.encoder.layer.28.attention.self.in_proj.weight', 'deberta.encoder.layer.28.attention.self.pos_proj.weight', 'deberta.encoder.layer.28.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.28.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.29.attention.self.q_bias', 'deberta.encoder.layer.29.attention.self.v_bias', 'deberta.encoder.layer.29.attention.self.in_proj.weight', 'deberta.encoder.layer.29.attention.self.pos_proj.weight', 'deberta.encoder.layer.29.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.29.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.30.attention.self.q_bias', 'deberta.encoder.layer.30.attention.self.v_bias', 'deberta.encoder.layer.30.attention.self.in_proj.weight', 'deberta.encoder.layer.30.attention.self.pos_proj.weight', 'deberta.encoder.layer.30.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.30.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.31.attention.self.q_bias', 'deberta.encoder.layer.31.attention.self.v_bias', 'deberta.encoder.layer.31.attention.self.in_proj.weight', 'deberta.encoder.layer.31.attention.self.pos_proj.weight', 'deberta.encoder.layer.31.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.31.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.32.attention.self.q_bias', 'deberta.encoder.layer.32.attention.self.v_bias', 'deberta.encoder.layer.32.attention.self.in_proj.weight', 'deberta.encoder.layer.32.attention.self.pos_proj.weight', 'deberta.encoder.layer.32.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.32.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.33.attention.self.q_bias', 'deberta.encoder.layer.33.attention.self.v_bias', 'deberta.encoder.layer.33.attention.self.in_proj.weight', 'deberta.encoder.layer.33.attention.self.pos_proj.weight', 'deberta.encoder.layer.33.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.33.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.34.attention.self.q_bias', 'deberta.encoder.layer.34.attention.self.v_bias', 'deberta.encoder.layer.34.attention.self.in_proj.weight', 'deberta.encoder.layer.34.attention.self.pos_proj.weight', 'deberta.encoder.layer.34.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.34.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.35.attention.self.q_bias', 'deberta.encoder.layer.35.attention.self.v_bias', 'deberta.encoder.layer.35.attention.self.in_proj.weight', 'deberta.encoder.layer.35.attention.self.pos_proj.weight', 'deberta.encoder.layer.35.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.35.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.36.attention.self.q_bias', 'deberta.encoder.layer.36.attention.self.v_bias', 'deberta.encoder.layer.36.attention.self.in_proj.weight', 'deberta.encoder.layer.36.attention.self.pos_proj.weight', 'deberta.encoder.layer.36.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.36.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.37.attention.self.q_bias', 'deberta.encoder.layer.37.attention.self.v_bias', 'deberta.encoder.layer.37.attention.self.in_proj.weight', 'deberta.encoder.layer.37.attention.self.pos_proj.weight', 'deberta.encoder.layer.37.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.37.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.38.attention.self.q_bias', 'deberta.encoder.layer.38.attention.self.v_bias', 'deberta.encoder.layer.38.attention.self.in_proj.weight', 'deberta.encoder.layer.38.attention.self.pos_proj.weight', 'deberta.encoder.layer.38.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.38.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.39.attention.self.q_bias', 'deberta.encoder.layer.39.attention.self.v_bias', 'deberta.encoder.layer.39.attention.self.in_proj.weight', 'deberta.encoder.layer.39.attention.self.pos_proj.weight', 'deberta.encoder.layer.39.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.39.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.40.attention.self.q_bias', 'deberta.encoder.layer.40.attention.self.v_bias', 'deberta.encoder.layer.40.attention.self.in_proj.weight', 'deberta.encoder.layer.40.attention.self.pos_proj.weight', 'deberta.encoder.layer.40.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.40.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.41.attention.self.q_bias', 'deberta.encoder.layer.41.attention.self.v_bias', 'deberta.encoder.layer.41.attention.self.in_proj.weight', 'deberta.encoder.layer.41.attention.self.pos_proj.weight', 'deberta.encoder.layer.41.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.41.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.42.attention.self.q_bias', 'deberta.encoder.layer.42.attention.self.v_bias', 'deberta.encoder.layer.42.attention.self.in_proj.weight', 'deberta.encoder.layer.42.attention.self.pos_proj.weight', 'deberta.encoder.layer.42.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.42.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.43.attention.self.q_bias', 'deberta.encoder.layer.43.attention.self.v_bias', 'deberta.encoder.layer.43.attention.self.in_proj.weight', 'deberta.encoder.layer.43.attention.self.pos_proj.weight', 'deberta.encoder.layer.43.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.43.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.44.attention.self.q_bias', 'deberta.encoder.layer.44.attention.self.v_bias', 'deberta.encoder.layer.44.attention.self.in_proj.weight', 'deberta.encoder.layer.44.attention.self.pos_proj.weight', 'deberta.encoder.layer.44.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.44.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.45.attention.self.q_bias', 'deberta.encoder.layer.45.attention.self.v_bias', 'deberta.encoder.layer.45.attention.self.in_proj.weight', 'deberta.encoder.layer.45.attention.self.pos_proj.weight', 'deberta.encoder.layer.45.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.45.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.46.attention.self.q_bias', 'deberta.encoder.layer.46.attention.self.v_bias', 'deberta.encoder.layer.46.attention.self.in_proj.weight', 'deberta.encoder.layer.46.attention.self.pos_proj.weight', 'deberta.encoder.layer.46.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.46.attention.self.pos_q_proj.bias', 'deberta.encoder.layer.47.attention.self.q_bias', 'deberta.encoder.layer.47.attention.self.v_bias', 'deberta.encoder.layer.47.attention.self.in_proj.weight', 'deberta.encoder.layer.47.attention.self.pos_proj.weight', 'deberta.encoder.layer.47.attention.self.pos_q_proj.weight', 'deberta.encoder.layer.47.attention.self.pos_q_proj.bias']\r\nYou should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\r\n```", "Apart from the two issues mentioned above, the PR looks in a good state to me. Would you mind:\r\n\r\n- Checking what's wrong with `microsoft/deberta-xlarge-mnli`\r\n- Adding the `num_labels` field to the configuration of your MNLI models and removing the pre-load hooks\r\n- Rebasing on the current `master`\r\n\r\nI can take care of 2) and 3) if you want.", "> Apart from the two issues mentioned above, the PR looks in a good state to me. Would you mind:\r\n> \r\n> * Checking what's wrong with `microsoft/deberta-xlarge-mnli`\r\n> * Adding the `num_labels` field to the configuration of your MNLI models and removing the pre-load hooks\r\n> * Rebasing on the current `master`\r\n> \r\n> I can take care of 2) and 3) if you want.\r\n\r\nThanks @LysandreJik. I just fixed the model issue and resolved the merge conflicts. \r\nFor the hook issue, add num_labels will not fix the issue. In most of the cases we want to load a mnli fine-tuned model for another task, which has 2 or 1 labels, e.g. MRPC, STS-2, SST-B. So we still need the hook unless we get the loading issue fixed in load_pretrained_model method. One possible way is to add ignore error dictionary just like ignore_unexpected keys. But I think we should fix this in another separate PR. \r\n\r\n", "Thank you for taking care of those issues.\r\n\r\n@patrickvonplaten @sgugger, could you give this one a look?\r\n\r\nThe unresolved issue is regarding the pre-load hooks. Loading a pre-trained model that already has a classification head with a different number of labels will not work, as the weight will have the wrong numbers of parameters.\r\n\r\nUntil now, we've been doing:\r\n\r\n```py\r\nfrom transformers import DebertaV2Model, DebertaV2ForSequenceClassification\r\n\r\nseq_model = DebertaV2ForSequenceClassification.from_pretrained(\"xxx\", num_labels=4)\r\nseq_model.save_pretrained(directory)\r\n\r\nbase = DebertaV2Model.from_pretrained(directory) # Lose the head\r\nbase.save_pretrained(directory)\r\n\r\nseq_model = DebertaV2ForSequenceClassification.from_pretrained(directory, num_labels=8)\r\n```\r\n\r\nThe pre-load hook that @BigBird01 worked on drops the head instead when it finds it is ill-loaded. I'm okay to merge it like this, and I'll work on a model-agnostic approach this week. Let me know your thoughts.", "@LysandreJik Thanks for the fix. \r\nCan you merge this PR, please?", "> Awesome! Thanks so much for adding this super important model @BigBird01 ! I left a couple of comments in the `modeling_deberta_v2.py` file - it would be great if we can make the code a bit cleaner there, _e.g._:\r\n> \r\n> * remove the `use_conv` attribute\r\n> * set `output_hidden_states=False` as a default\r\n> * refactor the `MaskLayerNorm` class\r\n> \r\n> Those changes should be pretty trivial - thanks so much for all your work!\r\n\r\nThank you @patrickvonplaten! I will take a look at it soon. ", "As seen with @BigBird01, taking over the PR!", "> As seen with @BigBird01, taking over the PR!\r\n\r\nThank you @LysandreJik ! ", "My pleasure! Thank you for your work!" ]

[]

[]

[ "Yes there are some conflicts between a latest version of jax and an older version of flax (I think uninstalling both and reinstalling with pip install -e .[dev] will solve your problem). I had the same problem earlier.\r\n@patrickvonplaten It seems to have appeared with the minimum version change in jax/flax if you can have a look.\r\n", "Your workaround worked, @sgugger - thank you! \r\n\r\n> Yes there are some conflicts between a latest version of jax and an older version of flax \r\n\r\nIn which case `setup.py` needs to be updated to reflect the right combination of versions, right? I'd have sent a PR, but I don't know which min versions should be used.\r\n\r\nI also tried `pip install -e .[dev] -U` to force update, but it seems to ignore `-U` and since the requirements are met it doesn't update these libraries automatically.\r\n\r\n", "I cannot reproduce the error on my side, but the reason seems to be a mismatch of the `jax` version and `jaxlib` as shown here: https://github.com/google/jax/issues/5374 . Currently, we support `jax>=0.2.0` and in the issues it says `jax>=0.2.8` solves the issue. So I'd recommend that we also raise our minimum allowed version of jax ot `jax>=0.2.8`. What do you think?" ]

[]

[]

[ "Hello! Indeed, this is a valid request. Would you like to open a PR and take a stab at it?", "@LysandreJik Hi, thanks for the fast reply ! \r\n\r\nOk will do that :) \r\nI will comment here when the PR will be ready", "Hi @LysandreJik is there any update on this issue? If @Ierezell didn't have time, I might be able to give a shot at it in the next days", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Hi!\r\nIs this issue somewhere in consideration still?\r\nWould be awesome to be able to get tensors from the feature extraction pipeline", "I think we'd still be open to that; WDYT @Narsil?", "Sure ! \r\n\r\nWould adding an argument `return_type= \"tensors\"` be OK ? That way we can enable this feature without breaking backward compatibility ?", "I'm baffled as to why returning the features as a list is the default behavior in the first place... Isn't one common usage of feature extraction to provide an input to another model, which means it is preferred to keep it as a tensor?", "@ajsanjoaquin \r\n\r\nWell it depends, not necessarily. Another very common use case is to feed it to some feature database for querying later.\r\nThose database engines are not necessarily expecting the same kind of tensors that you are sending.\r\n\r\nBut I kind of agree that it should be at least a `numpy.array` because usually conversions between numpy and PT or TF is basically free, meaning it would be much easier to use that way.\r\n\r\nSome `pipeline` were added a long time ago where the current situation was not as clear as today, and since we are very conservative regarding breaking changes, that can explain why some defaults are the way they are.\r\n\r\nIf/When v5 is getting prepared there would be a lot of small but breaking changes in that regard." ]

[]

[ { "login": "sgugger", "id": 35901082, "node_id": "MDQ6VXNlcjM1OTAxMDgy", "avatar_url": "https://avatars.githubusercontent.com/u/35901082?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sgugger", "html_url": "https://github.com/sgugger", "followers_url": "https://api.github.com/users/sgugger/followers", "following_url": "https://api.github.com/users/sgugger/following{/other_user}", "gists_url": "https://api.github.com/users/sgugger/gists{/gist_id}", "starred_url": "https://api.github.com/users/sgugger/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sgugger/subscriptions", "organizations_url": "https://api.github.com/users/sgugger/orgs", "repos_url": "https://api.github.com/users/sgugger/repos", "events_url": "https://api.github.com/users/sgugger/events{/privacy}", "received_events_url": "https://api.github.com/users/sgugger/received_events", "type": "User", "site_admin": false } ]

[ "Also, I for one, I find it confusing that there are two different parameters that (kind of) refer to the same thing. I was lucky that I had run my code successfully with 10+ other models from the hub and it stood out that the sizes used for embeddings during training were different. " ]

[]

[]

[]

[]

[]

[ "Hi! So that we may help you, could you provide the information related to your environment as asked in the issue template?\r\n\r\nAlso, we can't understand what's happening here because we don't know what's your text, your question, and most importantly the model you used.\r\n\r\nFinally, you instantiated a model, but you didn't do so for the tokenizer. If you have mismatched models/tokenizers, then outputs are bound to be confusing.", "> Hi! So that we may help you, could you provide the information related to your environment as asked in the issue template?\r\n> \r\n> Also, we can't understand what's happening here because we don't know what's your text, your question, and most importantly the model you used.\r\n> \r\n> Finally, you instantiated a model, but you didn't do so for the tokenizer. If you have mismatched models/tokenizers, then outputs are bound to be confusing.\r\n\r\nOh,sorry. My enviroonment is as follows:\r\nOS: Macos catalina\r\npython version: python3.7.3\r\nPackage Version: transformers 3.4 \r\nThe model, which I have used, is 'ktrapeznikov/biobert_v1.1_pubmed_squad_v2'. And the code is like following:\r\n\r\n```py\r\ndata ='I live in Berkeley. I am 30 years old. And my name is Clara.'\r\nquestion = 'What's my name?'\r\npipeline = pipeline('question-answering',model = 'ktrapeznikov/biobert_v1.1_pubmed_squad_v2')\r\nanswers = pipeline(context = data, question = question)\r\n\r\nanswer is like {'score': 0.9977871179580688, 'start': 54, 'end': 59, 'answer': 'Clara.'}\r\n\r\n\r\ntokenizer = AutoTokenizer.from_pretrained('ktrapeznikov/biobert_v1.1_pubmed_squad_v2')\r\nencoding = tokenizer.encode(questions,data)\r\n\r\nlen(encoding)\r\n```\r\n\r\nWith len(encoding) I got the length of encoding 26. But from the answer I got start index as 54 and end index as 59?", "Yes, the length of encoding is the length of the list of tokens. The start and end index are the start and end index of characters, not tokens. We should clarify that in the docs.", "> Yes, the length of encoding is the length of the list of tokens. The start and end index are the start and end index of characters, not tokens. We should clarify that in the docs.\r\n\r\nokay, thats what i mean. Thx" ]

[]

[ { "login": "patrickvonplaten", "id": 23423619, "node_id": "MDQ6VXNlcjIzNDIzNjE5", "avatar_url": "https://avatars.githubusercontent.com/u/23423619?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patrickvonplaten", "html_url": "https://github.com/patrickvonplaten", "followers_url": "https://api.github.com/users/patrickvonplaten/followers", "following_url": "https://api.github.com/users/patrickvonplaten/following{/other_user}", "gists_url": "https://api.github.com/users/patrickvonplaten/gists{/gist_id}", "starred_url": "https://api.github.com/users/patrickvonplaten/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patrickvonplaten/subscriptions", "organizations_url": "https://api.github.com/users/patrickvonplaten/orgs", "repos_url": "https://api.github.com/users/patrickvonplaten/repos", "events_url": "https://api.github.com/users/patrickvonplaten/events{/privacy}", "received_events_url": "https://api.github.com/users/patrickvonplaten/received_events", "type": "User", "site_admin": false } ]

[ "Hey @mshuffett, \r\n\r\nI understand your concern and I think I agree with you! Would you be interested in opening a PR to fix this for `beam_search` and `group_beam_search` ? It's essentially just moving the ` + beam_scores` line further down", "@patrickvonplaten I would be happy to but upon trying your suggested fix, it does solve the problem with the score of the first token, but the second token now has the same problem.\r\n\r\nI just moved this line below the `return_dict_in_generate` block.\r\n```python\r\nnext_token_scores = next_token_scores + beam_scores[:, None].expand_as(next_token_scores)\r\n```\r\n\r\nI believe this is still unexpected but I'm not yet sure why this is happening. If you have any thoughts that would be helpful.", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Unstale", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Ping\r\n", "Yeah I'm probably not going to be able to submit a PR for this, but I do\nthink it should be fixed.\nᐧ\n\nOn Mon, May 31, 2021 at 2:27 AM Patrick von Platen ***@***.***>\nwrote:\n\n> Ping\n>\n> —\n> You are receiving this because you were mentioned.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/10012#issuecomment-851265135>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AAIFLUJ6IGHT52SRTDKJIKDTQM25NANCNFSM4XDNCCEQ>\n> .\n>\n\n\n-- \nMichael Shuffett\nWritten with compose.ai\n", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "I had the same issue. I'm trying to replicate OpenAI's behavior when selecting \"Show probabilities: Full Spectrum\". I would like, at each step, for it to show the top N candidate tokens as well as all their probabilities. \r\n\r\nI've not quite figured out how to make it select the _top_ candidates, because sampling is selecting them randomly. It seems like beam search does indeed do what I want selecting always the top probabilities, but then I can't get the true probabilities this way because of the 1e-09 issue. ", "@monsieurpooh,\r\n\r\nCould you open a new issue for this one? I'm not quite sure whether you are interested in sampling, beam search, etc... :-) Happy to extend `generate()` to cover more important use cases", "The only thing I was trying to do was get the top 10 tokens (and their probabilities) for the next 1 token. For example: \"I went to the\" -> {\" store\": 0.25, \" park\": 0.1, ...} \r\n\r\nI do not need any beam search, sampling, etc. I used the workaround described earlier in the thread and it works perfectly.", "@patrickvonplaten can you please this question [here](https://stackoverflow.com/questions/72180737/beam-search-and-generate-are-not-consistent) . If I don't miss something, I think there is a bug in beam_search. Thanks", "Hey @rafikg,\r\n\r\nCould you please open a new issue or use the forum: https://discuss.huggingface.co/ if you have a question? Thanks!", "@monsieurpooh can you please elaborate how you implemented the workaround?\r\n\r\nLooking at master during the time of your posting in posting in March 7th, the workaround you mentioned seems already implemented?\r\n\r\nhttps://github.com/huggingface/transformers/blob/5c6f57ee75665499c8045a8bf7c73bf2415fba20/src/transformers/generation_utils.py#L2112\r\n\r\n`next_token_scores` is not invoked until after the `if return_dict_in_generate` https://github.com/huggingface/transformers/blob/5c6f57ee75665499c8045a8bf7c73bf2415fba20/src/transformers/generation_utils.py#L2115\r\n\r\n`next_token_scores_processed` is being used instead" ]

[ { "id": 2659267025, "node_id": "MDU6TGFiZWwyNjU5MjY3MDI1", "url": "https://api.github.com/repos/huggingface/transformers/labels/DeepSpeed", "name": "DeepSpeed", "color": "4D34F7", "default": false, "description": "" } ]

[ { "login": "stas00", "id": 10676103, "node_id": "MDQ6VXNlcjEwNjc2MTAz", "avatar_url": "https://avatars.githubusercontent.com/u/10676103?v=4", "gravatar_id": "", "url": "https://api.github.com/users/stas00", "html_url": "https://github.com/stas00", "followers_url": "https://api.github.com/users/stas00/followers", "following_url": "https://api.github.com/users/stas00/following{/other_user}", "gists_url": "https://api.github.com/users/stas00/gists{/gist_id}", "starred_url": "https://api.github.com/users/stas00/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/stas00/subscriptions", "organizations_url": "https://api.github.com/users/stas00/orgs", "repos_url": "https://api.github.com/users/stas00/repos", "events_url": "https://api.github.com/users/stas00/events{/privacy}", "received_events_url": "https://api.github.com/users/stas00/received_events", "type": "User", "site_admin": false } ]

[ "The model is actually quite big so I would expect it to OOM, if you are doing multi GPU training, you could try `fairscale/deepspeed` integration for saving memory and speeding up the training, check out this blog post\r\nhttps://huggingface.co/blog/zero-deepspeed-fairscale", "hi @patil-suraj thank you for your feedback and the blog post. So would I pip install deepspeed and use it as an argument in `Seq2SeqTrainingArguments`? If so, I noticed the documentation for that kwarg says\r\n\r\n```\r\ndeepspeed (:obj:`str`, `optional`):\r\n | Use `Deepspeed <https://github.com/microsoft/deepspeed>`__. This is an experimental feature and its API may\r\n | evolve in the future. The value is the location of its json config file (usually ``ds_config.json``).\r\n```\r\n\r\nIt says to give it the location of it's json config file, but I'm not sure what that means? Would that mean 1. create a json file like [this](https://raw.githubusercontent.com/huggingface/transformers/master/examples/seq2seq/ds_config.json) and save it to disk then 2. specify the location of that json file in disk?\r\n\r\nI notice it says to also use it in command line, so would I need to run\r\n```python\r\nimport subprocess\r\nsubprocess.check_call([ \"deepspeed\"])\r\n```\r\n\r\nas far as using `Seq2SeqTrainingArguments` is there anything else that I should set for distributed training? I noticed `local_rank=-1` by default so I assumed that was all I needed. Not sure if I was supposed to set `n_gpu`, `parallel_mode` or anything else so that it knows to do distributed training", "@stas00 or surrounding community, I'd greatly appreciate any feedback on how to use deepseed. I tried pip installing it and adding deepspeed in my command line argument(in addition to `--local-rank=-1`), but I'm not sure what else I might need? I noticed `Seq2SeqTrainingArguments` also has a `deepspeed` argument, \r\n\r\n```python\r\nhelp(Seq2SeqTrainingArguments)\r\n```\r\n\r\n```\r\ndeepspeed (:obj:`str`, `optional`):\r\n | Use `Deepspeed <https://github.com/microsoft/deepspeed>`__. This is an experimental feature and its API may\r\n | evolve in the future. The value is the location of its json config file (usually ``ds_config.json``).\r\n```\r\n\r\nbut I'm not sure if I need to create my own `ds_config.json` for it, save that json file to disk and then set that file location as the string for the `deepspeed` argument in `Seq2SeqTrainingArguments`. So I tried creating a `ds_config.json` file using\r\n\r\n```python\r\nimport json\r\n\r\nds_config = {\r\n \"fp16\": {\r\n \"enabled\": \"true\",\r\n \"loss_scale\": 0,\r\n \"loss_scale_window\": 1000,\r\n \"hysteresis\": 2,\r\n \"min_loss_scale\": 1\r\n },\r\n\r\n \"zero_optimization\": {\r\n \"stage\": 2,\r\n \"allgather_partitions\": \"true\",\r\n \"allgather_bucket_size\": 2e8,\r\n \"overlap_comm\": \"true\",\r\n \"reduce_scatter\": \"true\",\r\n \"reduce_bucket_size\": 2e8,\r\n \"contiguous_gradients\": \"true\",\r\n \"cpu_offload\": \"true\"\r\n },\r\n\r\n \"zero_allow_untested_optimizer\": \"true\",\r\n\r\n \"optimizer\": {\r\n \"type\": \"AdamW\",\r\n \"params\": {\r\n \"lr\": 3e-5,\r\n \"betas\": [\r\n 0.8,\r\n 0.999\r\n ],\r\n \"eps\": 1e-8,\r\n \"weight_decay\": 3e-7\r\n }\r\n },\r\n\r\n \"scheduler\": {\r\n \"type\": \"WarmupLR\",\r\n \"params\": {\r\n \"warmup_min_lr\": 0,\r\n \"warmup_max_lr\": 3e-5,\r\n \"warmup_num_steps\": 500\r\n }\r\n },\r\n\r\n \"steps_per_print\": 2000,\r\n \"wall_clock_breakdown\": \"false\"\r\n}\r\n\r\nwith open('ds_config.json', 'w') as fp:\r\n json.dump(ds_config, fp)\r\n```\r\nthen setting\r\n\r\n```python\r\ntraining_args = Seq2SeqTrainingArguments(\r\n deepspeed=\"ds_config.json\"\r\n```\r\nbut I got an import error as far as `mpi4py`. I'm not sure if what I'm doing to use deepseed is correct. I'd greatly appreciate any help with this", "@mmoya01, let's sort it out.\r\n\r\n1. You will find the full documentation at https://huggingface.co/transformers/master/main_classes/trainer.html#deepspeed\r\n\r\nAs this is new and I haven't thought of all the use-cases please don't hesitate to flag if something is missing or unclear in the documentation and it will get sorted out.\r\n\r\n2. the `--deepspeed` cl arg (or the `deepspeed` argument of the Trainer) expects a path to a file that contains the deepspeed configuration, so your file should have just the config bit:\r\n\r\n```\r\n{\r\n \"fp16\": {\r\n \"enabled\": \"true\",\r\n \"loss_scale\": 0,\r\n \"loss_scale_window\": 1000,\r\n \"hysteresis\": 2,\r\n \"min_loss_scale\": 1\r\n },\r\n\r\n \"zero_optimization\": {\r\n \"stage\": 2,\r\n \"allgather_partitions\": \"true\",\r\n \"allgather_bucket_size\": 2e8,\r\n \"overlap_comm\": \"true\",\r\n \"reduce_scatter\": \"true\",\r\n \"reduce_bucket_size\": 2e8,\r\n \"contiguous_gradients\": \"true\",\r\n \"cpu_offload\": \"true\"\r\n },\r\n\r\n \"zero_allow_untested_optimizer\": \"true\",\r\n\r\n \"optimizer\": {\r\n \"type\": \"AdamW\",\r\n \"params\": {\r\n \"lr\": 3e-5,\r\n \"betas\": [\r\n 0.8,\r\n 0.999\r\n ],\r\n \"eps\": 1e-8,\r\n \"weight_decay\": 3e-7\r\n }\r\n },\r\n\r\n \"scheduler\": {\r\n \"type\": \"WarmupLR\",\r\n \"params\": {\r\n \"warmup_min_lr\": 0,\r\n \"warmup_max_lr\": 3e-5,\r\n \"warmup_num_steps\": 500\r\n }\r\n },\r\n\r\n \"steps_per_print\": 2000,\r\n \"wall_clock_breakdown\": \"false\"\r\n}\r\n```\r\nSo in your case if you prefer to not use the CLI arguments:\r\n```\r\ntraining_args = Seq2SeqTrainingArguments(deepspeed=\"ds_config.json\")\r\n```\r\n\r\n3. Note that the invocation of the script must change to have `deepspeed` as its launcher, please refer to one of:\r\n- https://huggingface.co/transformers/master/main_classes/trainer.html#deployment-with-multiple-gpus\r\n- https://huggingface.co/transformers/master/main_classes/trainer.html#deployment-with-one-gpu\r\n\r\nPlease give it a try and if you run into any errors please paste the exact command you used and the backtrace and we will take it from there", "Hi @stas00 , thank you for getting back to me, I greatly appreciate it. Sounds good, so I removed `deepspeed` as a cl arg and instead specified the location of the `ds_config.json` file in \r\n```python\r\n training_args = Seq2SeqTrainingArguments(\r\n predict_with_generate=True,\r\n evaluation_strategy=\"steps\",\r\n per_device_train_batch_size=batch_size,\r\n per_device_eval_batch_size=batch_size,\r\n fp16=True,\r\n fp16_backend=\"amp\",\r\n output_dir= \"/mnt/summarization_checkpoints\",\r\n logging_steps=1000,\r\n eval_steps=1000,\r\n save_steps=1000,\r\n warmup_steps=2000,\r\n save_total_limit=3,\r\n gradient_accumulation_steps=4,\r\n deepspeed=\"ds_config.json\"\r\n )\r\n```\r\n\r\nI also noticed, because of [this](https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/runtime/fp16/onebit_adam.py#L14) import in deepspeed, I ended up pip installing `mpi4py` in addition to `deepspeed` and installing [libopenmpi-dev](https://stackoverflow.com/questions/28440834/error-when-installing-mpi4py) in my cuda image. Once I did all that, I was able to get most things running up until I came across this traceback below\r\n\r\n```\r\n[1/2] c++ -MMD -MF flatten_unflatten.o.d -DTORCH_EXTENSION_NAME=utils -DTORCH_API_INCLUDE_EXTENSION_H -isystem /usr/local/lib/python3.8/dist-packages/torch/include -isystem /usr/local/lib/python3.8/dist-packages/torch/include/torch/csrc/api/include -isystem /usr/local/lib/python3.8/dist-packages/torch/include/TH -isystem /usr/local/lib/python3.8/dist-packages/torch/include/THC -isystem /usr/include/python3.8 -D_GLIBCXX_USE_CXX11_ABI=0 -fPIC -std=c++14 -c /usr/local/lib/python3.8/dist-packages/deepspeed/ops/csrc/utils/flatten_unflatten.cpp -o flatten_unflatten.o \r\n[2/2] c++ flatten_unflatten.o -shared -L/usr/local/lib/python3.8/dist-packages/torch/lib -lc10 -ltorch_cpu -ltorch -ltorch_python -o utils.so\r\nLoading extension module utils...\r\nTime to load utils op: 13.478780031204224 seconds\r\n[2021-02-09 22:26:48,901] [INFO] [stage2.py:130:__init__] Reduce bucket size 200000000.0\r\n[2021-02-09 22:26:48,901] [INFO] [stage2.py:131:__init__] Allgather bucket size 200000000.0\r\n[2021-02-09 22:26:48,901] [INFO] [stage2.py:132:__init__] CPU Offload: true\r\ngroup 0 param 0 = 459801600\r\n[2021-02-09 22:26:52,231] [INFO] [stage2.py:399:__init__] optimizer state initialized\r\n[2021-02-09 22:26:52,232] [INFO] [engine.py:586:_configure_optimizer] DeepSpeed Final Optimizer = <deepspeed.runtime.zero.stage2.FP16_DeepSpeedZeroOptimizer object at 0x7fea11ea1190>\r\n[2021-02-09 22:26:52,232] [INFO] [engine.py:405:_configure_lr_scheduler] DeepSpeed using configured LR scheduler = WarmupLR\r\n[2021-02-09 22:26:52,232] [INFO] [logging.py:60:log_dist] [Rank 0] DeepSpeed LR Scheduler = <deepspeed.runtime.lr_schedules.WarmupLR object at 0x7fe9b1759ca0>\r\n[2021-02-09 22:26:52,232] [INFO] [logging.py:60:log_dist] [Rank 0] step=0, skipped=0, lr=[3e-05], mom=[[0.8, 0.999]]\r\n\r\n[2021-02-09 22:26:52,232] [INFO] [config.py:733:print] DeepSpeedEngine configuration:\r\n[2021-02-09 22:26:52,232] [INFO] [config.py:737:print] activation_checkpointing_config <deepspeed.runtime.activation_checkpointing.config.DeepSpeedActivationCheckpointingConfig object at 0x7fe9b26b1340>\r\n[2021-02-09 22:26:52,232] [INFO] [config.py:737:print] allreduce_always_fp32 ........ False\r\n[2021-02-09 22:26:52,232] [INFO] [config.py:737:print] amp_enabled .................. False\r\n[2021-02-09 22:26:52,232] [INFO] [config.py:737:print] amp_params ................... False\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] checkpoint_tag_validation_enabled True\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] checkpoint_tag_validation_fail False\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] disable_allgather ............ False\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] dump_state ................... False\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] dynamic_loss_scale_args ...... {'init_scale': 4294967296, 'scale_window': 1000, 'delayed_shift': 2, 'min_scale': 1}\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] elasticity_enabled ........... False\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] flops_profiler_config ........ <deepspeed.profiling.config.DeepSpeedFlopsProfilerConfig object at 0x7fe9b26b1280>\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] fp16_enabled ................. true\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] global_rank .................. 0\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] gradient_accumulation_steps .. 4\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] gradient_clipping ............ 1.0\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] gradient_predivide_factor .... 1.0\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] initial_dynamic_scale ........ 4294967296\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] loss_scale ................... 0\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] memory_breakdown ............. False\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] optimizer_legacy_fusion ...... False\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] optimizer_name ............... adamw\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] optimizer_params ............. {'lr': 3e-05, 'betas': [0.8, 0.999], 'eps': 1e-08, 'weight_decay': 3e-07}\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0}\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] pld_enabled .................. False\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] pld_params ................... False\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] prescale_gradients ........... False\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] scheduler_name ............... WarmupLR\r\n[2021-02-09 22:26:52,233] [INFO] [config.py:737:print] scheduler_params ............. {'warmup_min_lr': 0, 'warmup_max_lr': 3e-05, 'warmup_num_steps': 500}\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] sparse_attention ............. None\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] sparse_gradients_enabled ..... False\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] steps_per_print .............. 2000\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] tensorboard_enabled .......... False\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] tensorboard_job_name ......... DeepSpeedJobName\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] tensorboard_output_path ...... \r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] train_batch_size ............. 8\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] train_micro_batch_size_per_gpu 2\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] wall_clock_breakdown ......... false\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] world_size ................... 1\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] zero_allow_untested_optimizer true\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] zero_config .................. {\r\n \"allgather_bucket_size\": 200000000.0,\r\n \"allgather_partitions\": \"true\",\r\n \"contiguous_gradients\": \"true\",\r\n \"cpu_offload\": \"true\",\r\n \"elastic_checkpoint\": true,\r\n \"load_from_fp32_weights\": true,\r\n \"overlap_comm\": \"true\",\r\n \"reduce_bucket_size\": 200000000.0,\r\n \"reduce_scatter\": \"true\",\r\n \"stage\": 2\r\n}\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] zero_enabled ................. True\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:737:print] zero_optimization_stage ...... 2\r\n[2021-02-09 22:26:52,234] [INFO] [config.py:739:print] json = {\r\n \"fp16\":{\r\n \"enabled\":\"true\",\r\n \"hysteresis\":2,\r\n \"loss_scale\":0,\r\n \"loss_scale_window\":1000,\r\n \"min_loss_scale\":1\r\n },\r\n \"gradient_accumulation_steps\":4,\r\n \"gradient_clipping\":1.0,\r\n \"optimizer\":{\r\n \"params\":{\r\n \"betas\":[\r\n 0.8,\r\n 0.999\r\n ],\r\n \"eps\":1e-08,\r\n \"lr\":3e-05,\r\n \"weight_decay\":3e-07\r\n },\r\n \"type\":\"AdamW\"\r\n },\r\n \"scheduler\":{\r\n \"params\":{\r\n \"warmup_max_lr\":3e-05,\r\n \"warmup_min_lr\":0,\r\n \"warmup_num_steps\":500\r\n },\r\n \"type\":\"WarmupLR\"\r\n },\r\n \"steps_per_print\":2000,\r\n \"train_micro_batch_size_per_gpu\":2,\r\n \"wall_clock_breakdown\":\"false\",\r\n \"zero_allow_untested_optimizer\":\"true\",\r\n \"zero_optimization\":{\r\n \"allgather_bucket_size\":200000000.0,\r\n \"allgather_partitions\":\"true\",\r\n \"contiguous_gradients\":\"true\",\r\n \"cpu_offload\":\"true\",\r\n \"overlap_comm\":\"true\",\r\n \"reduce_bucket_size\":200000000.0,\r\n \"reduce_scatter\":\"true\",\r\n \"stage\":2\r\n }\r\n}\r\nUsing /root/.cache/torch_extensions as PyTorch extensions root...\r\nNo modifications detected for re-loaded extension module utils, skipping build step...\r\nLoading extension module utils...\r\nTime to load utils op: 0.0004968643188476562 seconds\r\n```\r\n\r\n### Traceback\r\n```\r\n 0%| | 0/3 [00:00<?, ?it/s]/usr/local/lib/python3.8/dist-packages/nlp/utils/py_utils.py:191: UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. This means you can write to the underlying (supposedly non-writeable) NumPy array using the tensor. You may want to copy the array to protect its data or make it writeable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:141.)\r\n return function(data_struct)\r\nTraceback (most recent call last):\r\n File \"abstractive_summarization.py\", line 374, in <module>\r\n run()\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 829, in __call__\r\n return self.main(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 782, in main\r\n rv = self.invoke(ctx)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 1066, in invoke\r\n return ctx.invoke(self.callback, **ctx.params)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 610, in invoke\r\n return callback(*args, **kwargs)\r\n File \"abstractive_summarization.py\", line 349, in run\r\n trainer.train()\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/trainer.py\", line 888, in train\r\n tr_loss += self.training_step(model, inputs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/trainer.py\", line 1250, in training_step\r\n loss = self.compute_loss(model, inputs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/trainer.py\", line 1277, in compute_loss\r\n outputs = model(**inputs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/parallel/data_parallel.py\", line 155, in forward\r\n outputs = self.parallel_apply(replicas, inputs, kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/parallel/data_parallel.py\", line 165, in parallel_apply\r\n return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/parallel/parallel_apply.py\", line 85, in parallel_apply\r\n output.reraise()\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/_utils.py\", line 395, in reraise\r\n raise self.exc_type(msg)\r\nAssertionError: Caught AssertionError in replica 1 on device 1.\r\nOriginal Traceback (most recent call last):\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/parallel/parallel_apply.py\", line 60, in _worker\r\n output = module(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/deepspeed/runtime/engine.py\", line 830, in forward\r\n self.timers('forward_microstep').start()\r\n File \"/usr/local/lib/python3.8/dist-packages/deepspeed/utils/timer.py\", line 38, in start\r\n assert not self.started_, 'timer has already been started'\r\nAssertionError: timer has already been started\r\n\r\n 0%| | 0/3 [00:09<?, ?it/s]\r\n```\r\n\r\nnot sure if it's because of `checkpoint_tag_validation_fail`. I'd greatly appreciate your feedback", "Glad to hear you were able to make progress, @mmoya01 \r\n\r\nWhat was the command line you used to launch this program? You have to launch it via `deepspeed` as the docs instruct.\r\n\r\n**edit:** actually just learned that it doesn't have to be the case - will update the docs shortly, but I still need to know how you started the program. thank you.\r\n\r\n> I also noticed, because of [this](https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/runtime/fp16/onebit_adam.py#L14) import in deepspeed, I ended up pip installing `mpi4py` in addition to `deepspeed` and installing [libopenmpi-dev](https://stackoverflow.com/questions/28440834/error-when-installing-mpi4py) in my cuda image. \r\n\r\nThis is odd that you had to do it manually, DeepSpeed's pip installer should have installed all the dependencies automatically. \r\n\r\n I will see if I can reproduce that.\r\n\r\n> not sure if it's because of `checkpoint_tag_validation_fail`. I'd greatly appreciate your feedback\r\n\r\nHave you tried w/o gradient checking?\r\n\r\nThe failure is not in the transformers land so it's a bit hard to guess what has happened.\r\n\r\nI'd recommend filing an Issue with DeepSpeed: https://github.com/microsoft/DeepSpeed/issues", "This is a pure DeepSpeed domain - totally unrelated to HF Trainer integrations:\r\n\r\nI had a chance to look at the missing dependencies.\r\n\r\n> I also noticed, because of this import in deepspeed, I ended up pip installing mpi4py in addition to deepspeed and installing libopenmpi-dev in my cuda image.\r\n\r\nOK, for some reason you were trying to use `OneBitAdam` optimizer, which you haven't shown you were using above. This one requires extra dependencies that can be installed with:\r\n```\r\npip install deepspeed[1bit_adam]\r\n```\r\nI tested and it works just fine with this config file:\r\n```\r\n{\r\n \"fp16\": {\r\n \"enabled\": true,\r\n \"loss_scale\": 0,\r\n \"loss_scale_window\": 1000,\r\n \"hysteresis\": 2,\r\n \"min_loss_scale\": 1,\r\n \"initial_scale_power\": 16\r\n },\r\n\r\n \"zero_optimization\": {\r\n \"stage\": 2,\r\n \"allgather_partitions\": true,\r\n \"allgather_bucket_size\": 2e8,\r\n \"overlap_comm\": true,\r\n \"reduce_scatter\": true,\r\n \"reduce_bucket_size\": 2e8,\r\n \"contiguous_gradients\": true,\r\n \"cpu_offload\": true\r\n },\r\n\r\n \"zero_allow_untested_optimizer\": true,\r\n \"optimizer\": {\r\n \"type\": \"OneBitAdam\",\r\n \"params\": {\r\n \"lr\": 2e-4,\r\n \"weight_decay\": 0.01,\r\n \"bias_correction\": false,\r\n \"freeze_step\": 400,\r\n \"cuda_aware\": true\r\n }\r\n },\r\n\r\n \"scheduler\": {\r\n \"type\": \"WarmupLR\",\r\n \"params\": {\r\n \"warmup_min_lr\": 0,\r\n \"warmup_max_lr\": 3e-5,\r\n \"warmup_num_steps\": 500\r\n }\r\n },\r\n\r\n \"steps_per_print\": 2000,\r\n \"wall_clock_breakdown\": false\r\n}\r\n```\r\n\r\nYou shouldn't need any of these extra dependencies to run, say, `AdamW`. ", "Hello @stas00 , first, thank you again for your reply/trying to help me through this. I realized I may have set my `local_rank` incorrectly(I set it at `local_rank=-1` which I believe disables distributed training). So I tried \r\n\r\n1.) disabling gradient checkpointing\r\n\r\n```python\r\nled = AutoModelForSeq2SeqLM.from_pretrained(\r\n \"patrickvonplaten/led-large-16384-pubmed\",\r\n gradient_checkpointing=False,\r\n use_cache=False,\r\n)\r\n```\r\n2.) using this config\r\n```json\r\n{\r\n \"fp16\": {\r\n \"enabled\": \"true\",\r\n \"loss_scale\": 0,\r\n \"loss_scale_window\": 1000,\r\n \"hysteresis\": 2,\r\n \"min_loss_scale\": 1,\r\n \"initial_scale_power\": 16\r\n },\r\n\r\n \"zero_optimization\": {\r\n \"stage\": 2,\r\n \"allgather_partitions\": \"true\",\r\n \"allgather_bucket_size\": 2e8,\r\n \"overlap_comm\": \"true\",\r\n \"reduce_scatter\": \"true\",\r\n \"reduce_bucket_size\": 2e8,\r\n \"contiguous_gradients\": \"true\",\r\n \"cpu_offload\": \"true\"\r\n },\r\n\r\n\r\n \"zero_allow_untested_optimizer\": \"true\",\r\n \"optimizer\": {\r\n \"type\": \"AdamW\",\r\n \"params\": {\r\n \"lr\": 0.001,\r\n \"betas\": [0.8, 0.999],\r\n \"eps\": 1e-8,\r\n \"weight_decay\": 3e-7\r\n }\r\n },\r\n\r\n \"scheduler\": {\r\n \"type\": \"WarmupLR\",\r\n \"params\": {\r\n \"warmup_min_lr\": 0,\r\n \"warmup_max_lr\": 3e-5,\r\n \"warmup_num_steps\": 500\r\n }\r\n },\r\n\r\n \"steps_per_print\": 2000,\r\n \"wall_clock_breakdown\": \"false\"\r\n}\r\n```\r\n\r\n3.) and setting `local_rank=0` in `Seq2SeqTrainingArguments`\r\n\r\n```python\r\n training_args = Seq2SeqTrainingArguments(\r\n deepspeed=\"ds_config.json\",\r\n predict_with_generate=True,\r\n evaluation_strategy=\"steps\",\r\n per_device_train_batch_size=2,\r\n per_device_eval_batch_size=2,\r\n fp16=True,\r\n fp16_backend=\"amp\",\r\n output_dir= \"/mnt/summarization_checkpoints\",\r\n logging_steps=1000,\r\n eval_steps=1000,\r\n save_steps=1000,\r\n warmup_steps=2000,\r\n save_total_limit=3,\r\n gradient_accumulation_steps=4,\r\n local_rank = 0,\r\n # sharded_ddp = True,\r\n )\r\n```\r\n\r\nI did not specify anything else in command line. I'm not sure if I set `local_rank` correctly in `Seq2SeqTrainingArguments`. I ended up getting a memory fragmentation error\r\n\r\n```\r\n[2021-02-10 20:43:26,268] [INFO] [config.py:733:print] DeepSpeedEngine configuration:\r\n[2021-02-10 20:43:26,268] [INFO] [config.py:737:print] activation_checkpointing_config <deepspeed.runtime.activation_checkpointing.config.DeepSpeedActivationCheckpointingConfig object at 0x7f9d0b742dc0>\r\n[2021-02-10 20:43:26,268] [INFO] [config.py:737:print] allreduce_always_fp32 ........ False\r\n[2021-02-10 20:43:26,268] [INFO] [config.py:737:print] amp_enabled .................. False\r\n[2021-02-10 20:43:26,268] [INFO] [config.py:737:print] amp_params ................... False\r\n[2021-02-10 20:43:26,268] [INFO] [config.py:737:print] checkpoint_tag_validation_enabled True\r\n[2021-02-10 20:43:26,268] [INFO] [config.py:737:print] checkpoint_tag_validation_fail False\r\n[2021-02-10 20:43:26,268] [INFO] [config.py:737:print] disable_allgather ............ False\r\n[2021-02-10 20:43:26,268] [INFO] [config.py:737:print] dump_state ................... False\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] dynamic_loss_scale_args ...... {'init_scale': 65536, 'scale_window': 1000, 'delayed_shift': 2, 'min_scale': 1}\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] elasticity_enabled ........... False\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] flops_profiler_config ........ <deepspeed.profiling.config.DeepSpeedFlopsProfilerConfig object at 0x7f9d0b742e20>\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] fp16_enabled ................. true\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] global_rank .................. 0\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] gradient_accumulation_steps .. 4\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] gradient_clipping ............ 1.0\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] gradient_predivide_factor .... 1.0\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] initial_dynamic_scale ........ 65536\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] loss_scale ................... 0\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] memory_breakdown ............. False\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] optimizer_legacy_fusion ...... False\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] optimizer_name ............... adamw\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] optimizer_params ............. {'lr': 0.001, 'betas': [0.8, 0.999], 'eps': 1e-08, 'weight_decay': 3e-07}\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0}\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] pld_enabled .................. False\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] pld_params ................... False\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] prescale_gradients ........... False\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] scheduler_name ............... WarmupLR\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] scheduler_params ............. {'warmup_min_lr': 0, 'warmup_max_lr': 3e-05, 'warmup_num_steps': 500}\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] sparse_attention ............. None\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] sparse_gradients_enabled ..... False\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] steps_per_print .............. 2000\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] tensorboard_enabled .......... False\r\n[2021-02-10 20:43:26,269] [INFO] [config.py:737:print] tensorboard_job_name ......... DeepSpeedJobName\r\n[2021-02-10 20:43:26,270] [INFO] [config.py:737:print] tensorboard_output_path ...... \r\n[2021-02-10 20:43:26,270] [INFO] [config.py:737:print] train_batch_size ............. 8\r\n[2021-02-10 20:43:26,270] [INFO] [config.py:737:print] train_micro_batch_size_per_gpu 2\r\n[2021-02-10 20:43:26,270] [INFO] [config.py:737:print] wall_clock_breakdown ......... false\r\n[2021-02-10 20:43:26,270] [INFO] [config.py:737:print] world_size ................... 1\r\n[2021-02-10 20:43:26,270] [INFO] [config.py:737:print] zero_allow_untested_optimizer true\r\n[2021-02-10 20:43:26,270] [INFO] [config.py:737:print] zero_config .................. {\r\n \"allgather_bucket_size\": 200000000.0,\r\n \"allgather_partitions\": \"true\",\r\n \"contiguous_gradients\": \"true\",\r\n \"cpu_offload\": \"true\",\r\n \"elastic_checkpoint\": true,\r\n \"load_from_fp32_weights\": true,\r\n \"overlap_comm\": \"true\",\r\n \"reduce_bucket_size\": 200000000.0,\r\n \"reduce_scatter\": \"true\",\r\n \"stage\": 2\r\n}\r\n[2021-02-10 20:43:26,270] [INFO] [config.py:737:print] zero_enabled ................. True\r\n[2021-02-10 20:43:26,270] [INFO] [config.py:737:print] zero_optimization_stage ...... 2\r\n[2021-02-10 20:43:26,270] [INFO] [config.py:739:print] json = {\r\n \"fp16\":{\r\n \"enabled\":\"true\",\r\n \"hysteresis\":2,\r\n \"initial_scale_power\":16,\r\n \"loss_scale\":0,\r\n \"loss_scale_window\":1000,\r\n \"min_loss_scale\":1\r\n },\r\n \"gradient_accumulation_steps\":4,\r\n \"gradient_clipping\":1.0,\r\n \"optimizer\":{\r\n \"params\":{\r\n \"betas\":[\r\n 0.8,\r\n 0.999\r\n ],\r\n \"eps\":1e-08,\r\n \"lr\":0.001,\r\n \"weight_decay\":3e-07\r\n },\r\n \"type\":\"AdamW\"\r\n },\r\n \"scheduler\":{\r\n \"params\":{\r\n \"warmup_max_lr\":3e-05,\r\n \"warmup_min_lr\":0,\r\n \"warmup_num_steps\":500\r\n },\r\n \"type\":\"WarmupLR\"\r\n },\r\n \"steps_per_print\":2000,\r\n \"train_micro_batch_size_per_gpu\":2,\r\n \"wall_clock_breakdown\":\"false\",\r\n \"zero_allow_untested_optimizer\":\"true\",\r\n \"zero_optimization\":{\r\n \"allgather_bucket_size\":200000000.0,\r\n \"allgather_partitions\":\"true\",\r\n \"contiguous_gradients\":\"true\",\r\n \"cpu_offload\":\"true\",\r\n \"overlap_comm\":\"true\",\r\n \"reduce_bucket_size\":200000000.0,\r\n \"reduce_scatter\":\"true\",\r\n \"stage\":2\r\n }\r\n}\r\n\r\n 0%| | 0/3 [00:00<?, ?it/s]/usr/local/lib/python3.8/dist-packages/nlp/utils/py_utils.py:191: UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. This means you can write to the underlying (supposedly non-writeable) NumPy array using the tensor. You may want to copy the array to protect its data or make it writeable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:141.)\r\n return function(data_struct)\r\nUsing /root/.cache/torch_extensions as PyTorch extensions root...\r\nNo modifications detected for re-loaded extension module utils, skipping build step...\r\nLoading extension module utils...\r\nTime to load utils op: 0.0005078315734863281 seconds\r\nTraceback (most recent call last):\r\n File \"abstractive_summarization.py\", line 374, in <module>\r\n run()\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 829, in __call__\r\n return self.main(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 782, in main\r\n rv = self.invoke(ctx)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 1066, in invoke\r\n return ctx.invoke(self.callback, **ctx.params)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 610, in invoke\r\n return callback(*args, **kwargs)\r\n File \"abstractive_summarization.py\", line 349, in run\r\n trainer.train()\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/trainer.py\", line 886, in train\r\n tr_loss += self.training_step(model, inputs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/trainer.py\", line 1265, in training_step\r\n self.model_wrapped.module.backward(loss)\r\n File \"/usr/local/lib/python3.8/dist-packages/deepspeed/runtime/engine.py\", line 903, in backward\r\n self.optimizer.backward(loss)\r\n File \"/usr/local/lib/python3.8/dist-packages/deepspeed/runtime/zero/stage2.py\", line 1596, in backward\r\n buf_0 = torch.empty(int(self.reduce_bucket_size * 4.5),\r\nRuntimeError: CUDA out of memory. Tried to allocate 1.68 GiB (GPU 0; 15.78 GiB total capacity; 12.80 GiB already allocated; 1.63 GiB free; 12.97 GiB reserved in total by PyTorch)\r\n\r\n 0%| | 0/3 [00:00<?, ?it/s]\r\n```\r\n\r\nI'd greatly appreciate your advice on what I might be missing ", "I tried to run the notebook you referred to after adding the modifications to launch DeepSpeed and now I can see all the problems you were referring to.\r\n\r\nI haven't yet tried running DeepSpeed in a jupyter notebook, but only as part of a normal program, so I will sort it out and get back to you.", "It took some experimenting to figure out what it wants - basically we need to emulate the launcher, since it doesn't get run under notebooks\r\n\r\nSo I have adapted the original notebook - you will find a DeepSpeed section in it and it should be easy to see what was added\r\nhttps://colab.research.google.com/drive/1DvcbpV-g_uKKa7KWBtlwJOX5b-mQUbR-?usp=sharing\r\n\r\nI will shortly make a PR with the docs on how to do it, https://github.com/huggingface/transformers/pull/10130\r\n\r\nBut until the PR is merged you need:\r\n```\r\n\r\n# deepspeed requires a distributed environment even if one process is used\r\n# emulating distributed env with a single gpu 0\r\nimport os\r\nos.environ['CUDA_VISIBLE_DEVICES'] = \"0\"\r\nos.environ['MASTER_ADDR'] = 'localhost' #\r\nos.environ['MASTER_PORT'] = '9998'\r\nos.environ['RANK'] = \"0\"\r\nos.environ['LOCAL_RANK'] =\"0\"\r\nos.environ['WORLD_SIZE'] = \"1\"\r\n\r\ntraining_args = Seq2SeqTrainingArguments(\r\n [... normal args ...]\r\n # deepspeed-in-jupyter-notebook-special-args\r\n local_rank=0, # XXX: this won't be needed when PR is merged\r\n deepspeed=\"ds_config.json\"\r\n)\r\n\r\n# XXX: this won't be needed when PR is merged\r\ntraining_args._setup_devices\r\n\r\ntrainer = Seq2SeqTrainer(...)\r\ntrainer.train()\r\n```\r\n\r\nI don't yet know if it will help with OOM (check if perhaps you need to make the max length shorter than your dataset's entires), but this should make a smooth run otherwise.\r\n\r\nBut I think you already figured out that if you install `mpi4py` it sorts most of these things out too. I'm trying to see how to make it the simplest for the users here: https://github.com/microsoft/DeepSpeed/issues/748\r\n\r\nIf you're still getting OOM please create a notebook where I can reproduce the problem and I will have a look. Thank you.", "It's important to understand that DeepSpeed ZeRO-Offload requires an ample CPU RAM to be available, so if you're on Colab you don't get too much there and that could be the culprit - i.e. you won't benefit from the offload much - which is the main feature on a single gpu to save on gpu memory. \r\n\r\nSo I'd try one of those tricks where you make colab give you double the memory by crashing the original session with a cell:\r\n```\r\ni = []\r\nwhile(True):\r\n i.append('a')\r\n```\r\nI haven't tried it, but people report it works.\r\n\r\nAlso need to tinker and try to turn perhaps some of its features off. Also you could try to make the buffers smaller try 1e8 or even 0.5e8 in the ds config. \r\n\r\nI was able to run the notebook you started from to completion (when it didn't run out of disk space). But perhaps it was already running to completion w/o deepspeed.", "hi @stas00 , thank you so much for your help throughout this. I greatly appreciate the PR and colab notebook example. I tried following your notebook and adjusting my script based on that notebook(I'm currently running this in kubeflow with 4 v100s. Each v100 GPU has 16Gi of memory though I can increase the memory). Such as: adding `LOCAL_RANK`,`RANK` and `WORLD_SIZE` env variables, adding `training_args._setup_devices` and changing some of the kwargs in `training_args` to be more consistent with the notebook. The example below produces a fake `train` and `test` dataset and my objective is to fine tune the `patrickvonplaten/led-large-16384-pubmed` on that fake dataset. That fake `train` dataset has a sample size of 2 and the `test` dataset has a sample size of 1. The snippet below should be reproducible. However, using that snippet, I'm still running into this OOM error\r\n\r\n```\r\nRuntimeError: CUDA out of memory. Tried to allocate 1.68 GiB (GPU 0; 15.78 GiB total capacity; 12.80 GiB already allocated; 1.63 GiB free; 12.97 GiB reserved in total by PyTorch)\r\n\r\n 0%| | 0/1 [00:00<?, ?it/s]\r\n```\r\nI'd greatly appreciate your two cents on what I might be missing in the snippet below\r\n```python\r\nimport datasets\r\nfrom datasets import load_dataset, load_metric\r\n\r\nimport click\r\nimport torch\r\nimport logging\r\nimport boto3\r\nimport json\r\n\r\nfrom io import BytesIO\r\nimport pandas as pd\r\n\r\nimport pyarrow as pa\r\nimport pyarrow.parquet as pq\r\nfrom nlp import arrow_dataset\r\n\r\nimport glob\r\nimport os\r\nimport tarfile\r\nimport os.path\r\nfrom transformers import (\r\n AutoTokenizer,\r\n AutoModelForSeq2SeqLM,\r\n Seq2SeqTrainer,\r\n Seq2SeqTrainingArguments,\r\n AutoTokenizer,\r\n AutoModelForSeq2SeqLM,\r\n)\r\n\r\n\r\nimport torch.utils.checkpoint\r\n\r\n\r\n\r\n\r\nlogger = logging.getLogger(__name__)\r\nlogger.setLevel(logging.INFO)\r\nlogging.basicConfig(\r\n level=logging.INFO, format=\"[%(levelname)s] %(asctime)s %(module)s: %(message)s\"\r\n)\r\n\r\n\r\nrouge = load_metric(\"rouge\")\r\n\r\n\r\nMODEL_NAME = \"patrickvonplaten/led-large-16384-pubmed\"\r\n\r\n\r\nds_config = {\r\n \"fp16\": {\r\n \"enabled\": \"true\",\r\n \"loss_scale\": 0,\r\n \"loss_scale_window\": 1000,\r\n \"hysteresis\": 2,\r\n \"min_loss_scale\": 1\r\n },\r\n\r\n \"zero_optimization\": {\r\n \"stage\": 2,\r\n \"allgather_partitions\": \"true\",\r\n \"allgather_bucket_size\": 2e8,\r\n \"overlap_comm\": \"true\",\r\n \"reduce_scatter\": \"true\",\r\n \"reduce_bucket_size\": 2e8,\r\n \"contiguous_gradients\": \"true\",\r\n \"cpu_offload\": \"true\"\r\n },\r\n\r\n \"zero_allow_untested_optimizer\": \"true\",\r\n\r\n \"optimizer\": {\r\n \"type\": \"AdamW\",\r\n \"params\": {\r\n \"lr\": 3e-5,\r\n \"betas\": [0.8, 0.999],\r\n \"eps\": 1e-8,\r\n \"weight_decay\": 3e-7\r\n }\r\n },\r\n\r\n \"scheduler\": {\r\n \"type\": \"WarmupLR\",\r\n \"params\": {\r\n \"warmup_min_lr\": 0,\r\n \"warmup_max_lr\": 3e-5,\r\n \"warmup_num_steps\": 500\r\n }\r\n },\r\n\r\n \"steps_per_print\": 2000,\r\n \"wall_clock_breakdown\": \"false\"\r\n}\r\n\r\nwith open('ds_config.json', 'w') as fp:\r\n json.dump(ds_config, fp)\r\n\r\n\r\n\r\n\r\nlogger.info(f\"load tokenizer using {MODEL_NAME}\")\r\ntokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)\r\n\r\nlogger.info(f\"Load {MODEL_NAME}. IMPORTANT NOTE:I'm enabling gradient checkpointing to save memory.\")\r\n# load model + enable gradient checkpointing & disable cache for checkpointing\r\nled = AutoModelForSeq2SeqLM.from_pretrained(\r\n MODEL_NAME,\r\n gradient_checkpointing=True,\r\n use_cache=False,\r\n)\r\n\r\n\r\n# max encoder length is 2048 for PubMed\r\nencoder_max_length = 2048\r\ndecoder_max_length = 256\r\nbatch_size = 2\r\n\r\n\r\n# set decoding params\r\nled.config.num_beams = 2\r\nled.config.max_length = 256\r\nled.config.min_length = 100\r\nled.config.length_penalty = 2.0\r\nled.config.early_stopping = True\r\nled.config.no_repeat_ngram_size = 3\r\n\r\ndef process_data_to_model_inputs(batch):\r\n # tokenize the inputs and labels\r\n inputs = tokenizer(\r\n batch[\"extractive_summary\"],\r\n padding=\"max_length\",\r\n truncation=True,\r\n max_length=encoder_max_length,\r\n )\r\n outputs = tokenizer(\r\n batch[\"reference_summary\"],\r\n padding=\"max_length\",\r\n truncation=True,\r\n max_length=decoder_max_length,\r\n )\r\n\r\n batch[\"input_ids\"] = inputs.input_ids\r\n batch[\"attention_mask\"] = inputs.attention_mask\r\n\r\n # create 0 global_attention_mask lists\r\n batch[\"global_attention_mask\"] = len(batch[\"input_ids\"]) * [\r\n [0 for _ in range(len(batch[\"input_ids\"][0]))]\r\n ]\r\n\r\n # since above lists are references, the following line changes the 0 index for all samples\r\n batch[\"global_attention_mask\"][0][0] = 1\r\n batch[\"labels\"] = outputs.input_ids\r\n\r\n # We have to make sure that the PAD token is ignored\r\n batch[\"labels\"] = [\r\n [-100 if token == tokenizer.pad_token_id else token for token in labels]\r\n for labels in batch[\"labels\"]\r\n ]\r\n\r\n return batch\r\n\r\ndef compute_metrics(pred):\r\n labels_ids = pred.label_ids\r\n pred_ids = pred.predictions\r\n\r\n pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)\r\n labels_ids[labels_ids == -100] = tokenizer.pad_token_id\r\n label_str = tokenizer.batch_decode(labels_ids, skip_special_tokens=True)\r\n\r\n rouge_output = rouge.compute(\r\n predictions=pred_str, references=label_str, rouge_types=[\"rouge2\"]\r\n )[\"rouge2\"].mid\r\n\r\n return {\r\n \"rouge2_precision\": round(rouge_output.precision, 4),\r\n \"rouge2_recall\": round(rouge_output.recall, 4),\r\n \"rouge2_fmeasure\": round(rouge_output.fmeasure, 4),\r\n }\r\n\r\ndef run():\r\n\r\n logger.info(\"create fictious train and test data\")\r\n train = pd.DataFrame({\"reference_summary\": [' '.join([\"I am a reference summary\"] * 200),\r\n ' '.join([\"I am another reference summary\"] * 200)],\r\n \"extractive_summary\": [' '.join([\"hello\"] * 200), ' '.join([\"goodbye\"] * 200)]})\r\n test = pd.DataFrame({\"reference_summary\": [' '.join([\"I am another reference summary\"] * 200)],\r\n \"extractive_summary\": [' '.join([\"goodbye\"] * 200)]})\r\n\r\n train = pa.Table.from_pandas(train)\r\n train = arrow_dataset.Dataset(train)\r\n\r\n test = pa.Table.from_pandas(test)\r\n test = arrow_dataset.Dataset(test)\r\n logger.info(\"map train data\")\r\n train = train.map(\r\n process_data_to_model_inputs,\r\n batched=True,\r\n batch_size=batch_size,\r\n remove_columns=[\"reference_summary\", \"extractive_summary\"],\r\n )\r\n\r\n logger.info(\"map test data\")\r\n test = test.map(\r\n process_data_to_model_inputs,\r\n batched=True,\r\n batch_size=batch_size,\r\n remove_columns=[\"reference_summary\", \"extractive_summary\"],\r\n\r\n )\r\n\r\n logger.info(\"set Python list in train to PyTorch tensor\")\r\n train.set_format(\r\n type=\"torch\",\r\n columns=[\"input_ids\", \"attention_mask\", \"global_attention_mask\", \"labels\"],\r\n )\r\n\r\n logger.info(\"set Python list in test to PyTorch tensor\")\r\n test.set_format(\r\n type=\"torch\",\r\n columns=[\"input_ids\", \"attention_mask\", \"global_attention_mask\", \"labels\"],\r\n )\r\n\r\n logger.info(\"enable fp16 amp training\")\r\n logger.info(f\"checkpoint files will be written to a pvc mount\")\r\n \r\n\r\n #define env variables required for training\r\n os.environ['RANK'] = \"0\"\r\n os.environ['LOCAL_RANK'] = \"0\"\r\n os.environ['WORLD_SIZE'] = \"1\"\r\n\r\n checkpoint_dir_path = \"/mnt/summarization_checkpoints\"\r\n training_args = Seq2SeqTrainingArguments(\r\n predict_with_generate=True,\r\n evaluation_strategy=\"steps\",\r\n per_device_train_batch_size=batch_size,\r\n per_device_eval_batch_size=batch_size,\r\n fp16=True,\r\n output_dir=checkpoint_dir_path,\r\n logging_steps=5,\r\n eval_steps=10,\r\n save_steps=10,\r\n save_total_limit=1,\r\n gradient_accumulation_steps=4,\r\n num_train_epochs=1,\r\n local_rank=0,\r\n deepspeed=\"ds_config.json\"\r\n )\r\n\r\n training_args._setup_devices\r\n\r\n os.makedirs(checkpoint_dir_path, exist_ok=True)\r\n logger.info(\"instantiate trainer\")\r\n trainer = Seq2SeqTrainer(\r\n model=led,\r\n tokenizer=tokenizer,\r\n args=training_args,\r\n compute_metrics=compute_metrics,\r\n train_dataset=train,\r\n eval_dataset=test,\r\n )\r\n\r\n\r\n\r\n logger.info(\"start training\")\r\n trainer.train()\r\n\r\n\r\n\r\nif __name__ == \"__main__\":\r\n run()\r\n```\r\nthank you for your help with this nonetheless", "Thank you for supplying the reproducible script, @mmoya01 - it worked with some small tweaks.\r\n\r\nLet's take a step back and go back to your original problem. That is let's remove the DeepSpeed for now.\r\n\r\nI modified your script to have 1000 smaller train records instead of 1 and if I run it it doesn't use more than 9GB of GPU RAM including cuda kernels - the actual Peak memory used: 7116MB - with your original one it was around 9GB peak and under 11GB total gpu RAM.\r\n\r\nSo may be it's worthwhile to sort it out first and then see if you actually need DeepSpeed in this case. We need to find what eats up the rest of your GPU memory.\r\n\r\nI added this at the end of the script:\r\n\r\n```\r\n import torch\r\n print(f\"Peak memory used: {torch.cuda.max_memory_reserved()>>20}MB\")\r\n import time\r\n time.sleep(10) # check nvidia-smi\r\n```\r\nmay be put some pauses through the script and observe if you get your gpu memory partially used up before the training starts?\r\n\r\nand to make 1000 entries:\r\n```\r\n n_recs = 1000\r\n frames = {\"reference_summary\": [' '.join([f\"{i} I am a reference summary\"] * 200) for i in range(n_recs)],\r\n \"extractive_summary\": [' '.join([f\"{i} hello\"] * 200) for i in range(n_recs)],\r\n }\r\n train = pd.DataFrame(frames)\r\n test = pd.DataFrame({\"reference_summary\": [' '.join([\"I am another reference summary\"] * 200)],\r\n \"extractive_summary\": [' '.join([\"goodbye\"] * 200)]})\r\n```\r\n\r\nSo if you have 16GB of gpu RAM, this should be more than enough. What are we missing here setup difference-wise? Do you have something else that consumes GPU RAM? Try to print the peak mem usage stats as I suggested above. But of course this might not work if you OOM.\r\n\r\nI'm using: pt-nightly and transformers master for this test.\r\n\r\n```\r\nPyTorch version: 1.8.0.dev20210202+cu110\r\nCUDA used to build PyTorch: 11.0\r\nPython version: 3.8 (64-bit runtime)\r\n```\r\n\r\n**edit:** \r\nI changed the mods to create the larger dataset to a cleaner way.\r\n\r\nI have a feeling this has to do with your dataset.\r\n\r\nI will get back to it shortly - will post an update.", "hi @stas00 , thank you again for the update. The image I'm using uses `nvidia/cuda:10.2-devel-ubuntu18.04` and `torch==1.6.0`. I used your tweak of 1000 examples and I also tried looking at\r\n\r\n```python\r\n if device.type == \"cuda\":\r\n logger.info(torch.cuda.get_device_name(0))\r\n logger.info(\"Memory Usage:\")\r\n logger.info(\r\n f\"Allocated: \"\r\n + str(round(torch.cuda.memory_allocated(0) / 1024 ** 3, 1))\r\n + \" GB\"\r\n )\r\n logger.info(\r\n \"Cached: \" + str(round(torch.cuda.memory_reserved(0) / 1024 ** 3, 1)) + \" GB\"\r\n )\r\n logger.info(\"number of GPUs available: \"+str(torch.cuda.device_count()))\r\n \r\n \r\n logger.info(f\"Peak memory used: {torch.cuda.max_memory_reserved()>>20}MB\")\r\n```\r\n\r\nwhich gave me\r\n\r\n```\r\n[INFO] 2021-02-11 22:21:51,155 abstractive_summarization: Using device: cuda\r\n[INFO] 2021-02-11 22:21:51,164 abstractive_summarization: Tesla V100-SXM2-16GB\r\n[INFO] 2021-02-11 22:21:51,164 abstractive_summarization: Memory Usage:\r\n[INFO] 2021-02-11 22:21:51,165 abstractive_summarization: Allocated: 0.0 GB\r\n[INFO] 2021-02-11 22:21:51,165 abstractive_summarization: Cached: 0.0 GB\r\n[INFO] 2021-02-11 22:21:51,165 abstractive_summarization: number of GPUs available: 4\r\n[INFO] 2021-02-11 22:21:51,165 abstractive_summarization: Peak memory used: 0MB\r\n```\r\nIf I omit deepspeed, I run into memory fragment error using those 1000 examples. I'm not sure why I might be getting 0MB peak memory, 0 GB cached memory and no memory usage. My full logs gave me the following:\r\n\r\n\r\n```\r\n[INFO] 2021-02-11 22:21:51,155 abstractive_summarization: Using device: cuda\r\n[INFO] 2021-02-11 22:21:51,164 abstractive_summarization: Tesla V100-SXM2-16GB\r\n[INFO] 2021-02-11 22:21:51,164 abstractive_summarization: Memory Usage:\r\n[INFO] 2021-02-11 22:21:51,165 abstractive_summarization: Allocated: 0.0 GB\r\n[INFO] 2021-02-11 22:21:51,165 abstractive_summarization: Cached: 0.0 GB\r\n[INFO] 2021-02-11 22:21:51,165 abstractive_summarization: number of GPUs available: 4\r\n[INFO] 2021-02-11 22:21:51,165 abstractive_summarization: Peak memory used: 0MB\r\n[INFO] 2021-02-11 22:21:51,216 abstractive_summarization: map train data\r\n\r\n 0%| | 0/500 [00:00<?, ?it/s]\r\n 1%| | 4/500 [00:00<00:15, 32.53it/s]\r\n 2%|▏ | 8/500 [00:00<00:14, 32.86it/s]\r\n 2%|▏ | 12/500 [00:00<00:14, 32.72it/s]\r\n 3%|▎ | 16/500 [00:00<00:14, 32.76it/s]\r\n 4%|▍ | 20/500 [00:00<00:14, 32.54it/s]\r\n 5%|▍ | 24/500 [00:00<00:15, 31.73it/s]\r\n 6%|▌ | 28/500 [00:00<00:14, 32.05it/s]\r\n 6%|▋ | 32/500 [00:01<00:15, 30.78it/s]\r\n 7%|▋ | 36/500 [00:01<00:14, 31.31it/s]\r\n 8%|▊ | 40/500 [00:01<00:14, 31.41it/s]\r\n 9%|▉ | 44/500 [00:01<00:14, 31.86it/s]\r\n 10%|▉ | 48/500 [00:01<00:14, 31.81it/s]\r\n 10%|█ | 52/500 [00:01<00:13, 32.03it/s]\r\n 11%|█ | 56/500 [00:01<00:13, 32.17it/s]\r\n 12%|█▏ | 60/500 [00:01<00:13, 32.33it/s]\r\n 13%|█▎ | 64/500 [00:01<00:13, 32.35it/s]\r\n 14%|█▎ | 68/500 [00:02<00:13, 32.44it/s]\r\n 14%|█▍ | 72/500 [00:02<00:13, 32.37it/s]\r\n 15%|█▌ | 76/500 [00:02<00:13, 32.48it/s]\r\n 16%|█▌ | 80/500 [00:02<00:12, 32.35it/s]\r\n 17%|█▋ | 84/500 [00:02<00:12, 32.06it/s]\r\n 18%|█▊ | 88/500 [00:02<00:12, 31.89it/s]\r\n 18%|█▊ | 92/500 [00:02<00:13, 31.01it/s]\r\n 19%|█▉ | 96/500 [00:03<00:12, 31.47it/s]\r\n 20%|██ | 100/500 [00:03<00:12, 31.91it/s]\r\n 21%|██ | 104/500 [00:03<00:12, 32.16it/s]\r\n 22%|██▏ | 108/500 [00:03<00:12, 31.08it/s]\r\n 22%|██▏ | 112/500 [00:03<00:12, 30.71it/s]\r\n 23%|██▎ | 116/500 [00:03<00:12, 30.61it/s]\r\n 24%|██▍ | 120/500 [00:03<00:12, 31.19it/s]\r\n 25%|██▍ | 124/500 [00:03<00:11, 31.47it/s]\r\n 26%|██▌ | 128/500 [00:04<00:11, 31.78it/s]\r\n 26%|██▋ | 132/500 [00:04<00:11, 32.01it/s]\r\n 27%|██▋ | 136/500 [00:04<00:11, 32.11it/s]\r\n 28%|██▊ | 140/500 [00:04<00:11, 32.19it/s]\r\n 29%|██▉ | 144/500 [00:04<00:11, 31.53it/s]\r\n 30%|██▉ | 148/500 [00:04<00:11, 31.84it/s]\r\n 30%|███ | 152/500 [00:04<00:11, 31.18it/s]\r\n 31%|███ | 156/500 [00:04<00:10, 31.40it/s]\r\n 32%|███▏ | 160/500 [00:05<00:10, 31.59it/s]\r\n 33%|███▎ | 164/500 [00:05<00:11, 29.86it/s]\r\n 34%|███▎ | 168/500 [00:05<00:10, 30.59it/s]\r\n 34%|███▍ | 172/500 [00:05<00:10, 31.01it/s]\r\n 35%|███▌ | 176/500 [00:05<00:10, 30.73it/s]\r\n 36%|███▌ | 180/500 [00:05<00:10, 31.21it/s]\r\n 37%|███▋ | 184/500 [00:05<00:10, 31.02it/s]\r\n 38%|███▊ | 188/500 [00:05<00:09, 31.41it/s]\r\n 38%|███▊ | 192/500 [00:06<00:09, 31.29it/s]\r\n 39%|███▉ | 196/500 [00:06<00:09, 31.29it/s]\r\n 40%|████ | 200/500 [00:06<00:09, 31.12it/s]\r\n 41%|████ | 204/500 [00:06<00:09, 31.56it/s]\r\n 42%|████▏ | 208/500 [00:06<00:09, 31.78it/s]\r\n 42%|████▏ | 212/500 [00:06<00:09, 31.95it/s]\r\n 43%|████▎ | 216/500 [00:06<00:08, 32.01it/s]\r\n 44%|████▍ | 220/500 [00:06<00:08, 31.80it/s]\r\n 45%|████▍ | 224/500 [00:07<00:08, 31.63it/s]\r\n 46%|████▌ | 228/500 [00:07<00:08, 31.41it/s]\r\n 46%|████▋ | 232/500 [00:07<00:08, 31.10it/s]\r\n 47%|████▋ | 236/500 [00:07<00:08, 30.91it/s]\r\n 48%|████▊ | 240/500 [00:07<00:08, 30.88it/s]\r\n 49%|████▉ | 244/500 [00:07<00:08, 30.87it/s]\r\n 50%|████▉ | 248/500 [00:07<00:08, 30.78it/s]\r\n 50%|█████ | 252/500 [00:07<00:07, 31.05it/s]\r\n 51%|█████ | 256/500 [00:08<00:07, 30.93it/s]\r\n 52%|█████▏ | 260/500 [00:08<00:07, 30.62it/s]\r\n 53%|█████▎ | 264/500 [00:08<00:07, 30.72it/s]\r\n 54%|█████▎ | 268/500 [00:08<00:07, 30.68it/s]\r\n 54%|█████▍ | 272/500 [00:08<00:07, 30.62it/s]\r\n 55%|█████▌ | 276/500 [00:08<00:07, 28.52it/s]\r\n 56%|█████▌ | 280/500 [00:08<00:07, 29.09it/s]\r\n 57%|█████▋ | 284/500 [00:09<00:07, 29.45it/s]\r\n 58%|█████▊ | 288/500 [00:09<00:07, 29.80it/s]\r\n 58%|█████▊ | 292/500 [00:09<00:06, 30.08it/s]\r\n 59%|█████▉ | 296/500 [00:09<00:06, 30.19it/s]\r\n 60%|██████ | 300/500 [00:09<00:06, 30.23it/s]\r\n 61%|██████ | 304/500 [00:09<00:06, 29.57it/s]\r\n 61%|██████▏ | 307/500 [00:09<00:06, 29.58it/s]\r\n 62%|██████▏ | 311/500 [00:09<00:06, 29.21it/s]\r\n 63%|██████▎ | 315/500 [00:10<00:06, 29.40it/s]\r\n 64%|██████▎ | 318/500 [00:10<00:06, 29.50it/s]\r\n 64%|██████▍ | 322/500 [00:10<00:05, 29.75it/s]\r\n 65%|██████▌ | 326/500 [00:10<00:06, 28.45it/s]\r\n 66%|██████▌ | 329/500 [00:10<00:06, 27.29it/s]\r\n 66%|██████▋ | 332/500 [00:10<00:06, 27.94it/s]\r\n 67%|██████▋ | 336/500 [00:10<00:05, 28.73it/s]\r\n 68%|██████▊ | 340/500 [00:10<00:05, 29.01it/s]\r\n 69%|██████▊ | 343/500 [00:11<00:05, 29.18it/s]\r\n 69%|██████▉ | 347/500 [00:11<00:05, 29.44it/s]\r\n 70%|███████ | 351/500 [00:11<00:04, 29.95it/s]\r\n 71%|███████ | 354/500 [00:11<00:04, 29.88it/s]\r\n 71%|███████▏ | 357/500 [00:11<00:04, 29.84it/s]\r\n 72%|███████▏ | 360/500 [00:11<00:04, 29.28it/s]\r\n 73%|███████▎ | 364/500 [00:11<00:04, 29.68it/s]\r\n 74%|███████▎ | 368/500 [00:11<00:04, 29.95it/s]\r\n 74%|███████▍ | 372/500 [00:12<00:04, 30.12it/s]\r\n 75%|███████▌ | 376/500 [00:12<00:04, 29.80it/s]\r\n 76%|███████▌ | 379/500 [00:12<00:04, 29.83it/s]\r\n 77%|███████▋ | 383/500 [00:12<00:03, 30.09it/s]\r\n 77%|███████▋ | 387/500 [00:12<00:03, 30.03it/s]\r\n 78%|███████▊ | 391/500 [00:12<00:03, 29.54it/s]\r\n 79%|███████▉ | 394/500 [00:12<00:03, 29.49it/s]\r\n 80%|███████▉ | 398/500 [00:12<00:03, 29.42it/s]\r\n 80%|████████ | 402/500 [00:13<00:03, 29.05it/s]\r\n 81%|████████ | 406/500 [00:13<00:03, 29.39it/s]\r\n 82%|████████▏ | 410/500 [00:13<00:03, 29.72it/s]\r\n 83%|████████▎ | 413/500 [00:13<00:02, 29.78it/s]\r\n 83%|████████▎ | 416/500 [00:13<00:02, 29.82it/s]\r\n 84%|████████▍ | 419/500 [00:13<00:02, 29.21it/s]\r\n 85%|████████▍ | 423/500 [00:13<00:02, 29.58it/s]\r\n 85%|████████▌ | 427/500 [00:13<00:02, 29.75it/s]\r\n 86%|████████▌ | 431/500 [00:14<00:02, 29.95it/s]\r\n 87%|████████▋ | 434/500 [00:14<00:02, 29.72it/s]\r\n 87%|████████▋ | 437/500 [00:14<00:02, 29.68it/s]\r\n 88%|████████▊ | 440/500 [00:14<00:02, 29.66it/s]\r\n 89%|████████▉ | 444/500 [00:14<00:01, 29.78it/s]\r\n 90%|████████▉ | 448/500 [00:14<00:01, 29.78it/s]\r\n 90%|█████████ | 451/500 [00:14<00:01, 29.51it/s]\r\n 91%|█████████ | 455/500 [00:14<00:01, 29.71it/s]\r\n 92%|█████████▏| 458/500 [00:14<00:01, 29.76it/s]\r\n 92%|█████████▏| 461/500 [00:15<00:01, 28.39it/s]\r\n 93%|█████████▎| 465/500 [00:15<00:01, 29.07it/s]\r\n 94%|█████████▎| 468/500 [00:15<00:01, 28.43it/s]\r\n 94%|█████████▍| 471/500 [00:15<00:01, 28.80it/s]\r\n 95%|█████████▌| 475/500 [00:15<00:00, 29.40it/s]\r\n 96%|█████████▌| 479/500 [00:15<00:00, 29.63it/s]\r\n 96%|█████████▋| 482/500 [00:15<00:00, 29.16it/s]\r\n 97%|█████████▋| 486/500 [00:15<00:00, 29.70it/s]\r\n 98%|█████████▊| 490/500 [00:16<00:00, 29.87it/s]\r\n 99%|█████████▉| 494/500 [00:16<00:00, 30.03it/s]\r\n100%|█████████▉| 498/500 [00:16<00:00, 30.17it/s]\r\n100%|██████████| 500/500 [00:16<00:00, 30.52it/s]\r\n[INFO] 2021-02-11 22:22:07,639 arrow_writer: Done writing 1000 examples in 51224000 bytes .\r\n[INFO] 2021-02-11 22:22:07,647 abstractive_summarization: map test data\r\n\r\n 0%| | 0/1 [00:00<?, ?it/s]\r\n100%|██████████| 1/1 [00:00<00:00, 91.30it/s]\r\n[INFO] 2021-02-11 22:22:07,664 arrow_writer: Done writing 1 examples in 51232 bytes .\r\n[INFO] 2021-02-11 22:22:07,665 abstractive_summarization: set Python list in train to PyTorch tensor\r\n[INFO] 2021-02-11 22:22:07,665 arrow_dataset: Set __getitem__(key) output type to torch for ['input_ids', 'attention_mask', 'global_attention_mask', 'labels'] columns (when key is int or slice) and don't output other (un-formated) columns.\r\n[INFO] 2021-02-11 22:22:07,665 abstractive_summarization: set Python list in test to PyTorch tensor\r\n[INFO] 2021-02-11 22:22:07,665 arrow_dataset: Set __getitem__(key) output type to torch for ['input_ids', 'attention_mask', 'global_attention_mask', 'labels'] columns (when key is int or slice) and don't output other (un-formated) columns.\r\n[INFO] 2021-02-11 22:22:07,665 abstractive_summarization: enable fp16 amp training\r\n[INFO] 2021-02-11 22:22:07,665 abstractive_summarization: file will be written to /workspace\r\n[2021-02-11 22:22:08,008] [INFO] [distributed.py:36:init_distributed] Not using the DeepSpeed or torch.distributed launchers, attempting to detect MPI environment...\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\n[2021-02-11 22:22:08,356] [INFO] [distributed.py:83:mpi_discovery] Discovered MPI settings of world_rank=0, local_rank=0, world_size=1, master_addr=10.23.29.192, master_port=29500\r\n[2021-02-11 22:22:08,356] [INFO] [distributed.py:46:init_distributed] Initializing torch distributed with backend: nccl\r\n[INFO] 2021-02-11 22:22:08,359 abstractive_summarization: instantiate trainer\r\n[INFO] 2021-02-11 22:22:11,706 abstractive_summarization: start training\r\n[2021-02-11 22:22:11,706] [INFO] [logging.py:60:log_dist] [Rank 0] DeepSpeed info: version=0.3.11, git-hash=unknown, git-branch=unknown\r\n[2021-02-11 22:22:11,732] [INFO] [engine.py:73:_initialize_parameter_parallel_groups] data_parallel_size: 1, parameter_parallel_size: 1\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\nUsing /root/.cache/torch_extensions as PyTorch extensions root...\r\nCreating extension directory /root/.cache/torch_extensions/cpu_adam...\r\nDetected CUDA files, patching ldflags\r\nEmitting ninja build file /root/.cache/torch_extensions/cpu_adam/build.ninja...\r\nBuilding extension module cpu_adam...\r\nAllowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\n[1/3] /usr/local/cuda/bin/nvcc -DTORCH_EXTENSION_NAME=cpu_adam -DTORCH_API_INCLUDE_EXTENSION_H -I/usr/local/lib/python3.8/dist-packages/deepspeed/ops/csrc/includes -I/usr/local/cuda/include -isystem /usr/local/lib/python3.8/dist-packages/torch/include -isystem /usr/local/lib/python3.8/dist-packages/torch/include/torch/csrc/api/include -isystem /usr/local/lib/python3.8/dist-packages/torch/include/TH -isystem /usr/local/lib/python3.8/dist-packages/torch/include/THC -isystem /usr/local/cuda/include -isystem /usr/include/python3.8 -D_GLIBCXX_USE_CXX11_ABI=0 -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_70,code=sm_70 --compiler-options '-fPIC' -O3 --use_fast_math -std=c++14 -U__CUDA_NO_HALF_OPERATORS__ -U__CUDA_NO_HALF_CONVERSIONS__ -U__CUDA_NO_HALF2_OPERATORS__ -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_70,code=compute_70 -c /usr/local/lib/python3.8/dist-packages/deepspeed/ops/csrc/adam/custom_cuda_kernel.cu -o custom_cuda_kernel.cuda.o \r\n[2/3] c++ -MMD -MF cpu_adam.o.d -DTORCH_EXTENSION_NAME=cpu_adam -DTORCH_API_INCLUDE_EXTENSION_H -I/usr/local/lib/python3.8/dist-packages/deepspeed/ops/csrc/includes -I/usr/local/cuda/include -isystem /usr/local/lib/python3.8/dist-packages/torch/include -isystem /usr/local/lib/python3.8/dist-packages/torch/include/torch/csrc/api/include -isystem /usr/local/lib/python3.8/dist-packages/torch/include/TH -isystem /usr/local/lib/python3.8/dist-packages/torch/include/THC -isystem /usr/local/cuda/include -isystem /usr/include/python3.8 -D_GLIBCXX_USE_CXX11_ABI=0 -fPIC -std=c++14 -O3 -std=c++14 -L/usr/local/cuda/lib64 -lcudart -lcublas -g -Wno-reorder -march=native -fopenmp -D__AVX256__ -c /usr/local/lib/python3.8/dist-packages/deepspeed/ops/csrc/adam/cpu_adam.cpp -o cpu_adam.o \r\n[3/3] c++ cpu_adam.o custom_cuda_kernel.cuda.o -shared -L/usr/local/lib/python3.8/dist-packages/torch/lib -lc10 -lc10_cuda -ltorch_cpu -ltorch_cuda -ltorch -ltorch_python -L/usr/local/cuda/lib64 -lcudart -o cpu_adam.so\r\nAdam Optimizer #0 is created with AVX2 arithmetic capability.\r\nLoading extension module cpu_adam...\r\nTime to load cpu_adam op: 23.714597702026367 seconds\r\n[2021-02-11 22:22:39,771] [INFO] [engine.py:551:_configure_optimizer] Using DeepSpeed Optimizer param name adamw as basic optimizer\r\n[2021-02-11 22:22:39,771] [INFO] [engine.py:556:_configure_optimizer] DeepSpeed Basic Optimizer = DeepSpeedCPUAdam (\r\nParameter Group 0\r\n amsgrad: False\r\n betas: [0.8, 0.999]\r\n bias_correction: True\r\n eps: 1e-08\r\n lr: 3e-05\r\n weight_decay: 3e-07\r\n)\r\nChecking ZeRO support for optimizer=DeepSpeedCPUAdam type=<class 'deepspeed.ops.adam.cpu_adam.DeepSpeedCPUAdam'>\r\n[2021-02-11 22:22:39,771] [INFO] [engine.py:672:_configure_zero_optimizer] Creating fp16 ZeRO stage 2 optimizer\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\nConfig: alpha=0.000030, betas=(0.800000, 0.999000), weight_decay=0.000000, adam_w=1\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\nUsing /root/.cache/torch_extensions as PyTorch extensions root...\r\nCreating extension directory /root/.cache/torch_extensions/utils...\r\nEmitting ninja build file /root/.cache/torch_extensions/utils/build.ninja...\r\nBuilding extension module utils...\r\nAllowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\n[1/2] c++ -MMD -MF flatten_unflatten.o.d -DTORCH_EXTENSION_NAME=utils -DTORCH_API_INCLUDE_EXTENSION_H -isystem /usr/local/lib/python3.8/dist-packages/torch/include -isystem /usr/local/lib/python3.8/dist-packages/torch/include/torch/csrc/api/include -isystem /usr/local/lib/python3.8/dist-packages/torch/include/TH -isystem /usr/local/lib/python3.8/dist-packages/torch/include/THC -isystem /usr/include/python3.8 -D_GLIBCXX_USE_CXX11_ABI=0 -fPIC -std=c++14 -c /usr/local/lib/python3.8/dist-packages/deepspeed/ops/csrc/utils/flatten_unflatten.cpp -o flatten_unflatten.o \r\n[2/2] c++ flatten_unflatten.o -shared -L/usr/local/lib/python3.8/dist-packages/torch/lib -lc10 -ltorch_cpu -ltorch -ltorch_python -o utils.so\r\nLoading extension module utils...\r\nTime to load utils op: 13.4954514503479 seconds\r\n[2021-02-11 22:22:53,267] [INFO] [stage2.py:130:__init__] Reduce bucket size 200000000.0\r\n[2021-02-11 22:22:53,267] [INFO] [stage2.py:131:__init__] Allgather bucket size 200000000.0\r\n[2021-02-11 22:22:53,267] [INFO] [stage2.py:132:__init__] CPU Offload: true\r\ngroup 0 param 0 = 459801600\r\n[2021-02-11 22:22:56,596] [INFO] [stage2.py:399:__init__] optimizer state initialized\r\n[2021-02-11 22:22:56,597] [INFO] [engine.py:586:_configure_optimizer] DeepSpeed Final Optimizer = <deepspeed.runtime.zero.stage2.FP16_DeepSpeedZeroOptimizer object at 0x7f9302607190>\r\n[2021-02-11 22:22:56,597] [INFO] [engine.py:405:_configure_lr_scheduler] DeepSpeed using configured LR scheduler = WarmupLR\r\n[2021-02-11 22:22:56,597] [INFO] [logging.py:60:log_dist] [Rank 0] DeepSpeed LR Scheduler = <deepspeed.runtime.lr_schedules.WarmupLR object at 0x7f9354837850>\r\n[2021-02-11 22:22:56,597] [INFO] [logging.py:60:log_dist] [Rank 0] step=0, skipped=0, lr=[3e-05], mom=[[0.8, 0.999]]\r\n[2021-02-11 22:22:56,597] [INFO] [config.py:733:print] DeepSpeedEngine configuration:\r\n[2021-02-11 22:22:56,597] [INFO] [config.py:737:print] activation_checkpointing_config <deepspeed.runtime.activation_checkpointing.config.DeepSpeedActivationCheckpointingConfig object at 0x7f93016d3310>\r\n[2021-02-11 22:22:56,597] [INFO] [config.py:737:print] allreduce_always_fp32 ........ False\r\n[2021-02-11 22:22:56,597] [INFO] [config.py:737:print] amp_enabled .................. False\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] amp_params ................... False\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] checkpoint_tag_validation_enabled True\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] checkpoint_tag_validation_fail False\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] disable_allgather ............ False\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] dump_state ................... False\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] dynamic_loss_scale_args ...... {'init_scale': 4294967296, 'scale_window': 1000, 'delayed_shift': 2, 'min_scale': 1}\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] elasticity_enabled ........... False\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] flops_profiler_config ........ <deepspeed.profiling.config.DeepSpeedFlopsProfilerConfig object at 0x7f93016d3370>\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] fp16_enabled ................. true\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] global_rank .................. 0\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] gradient_accumulation_steps .. 4\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] gradient_clipping ............ 1.0\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] gradient_predivide_factor .... 1.0\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] initial_dynamic_scale ........ 4294967296\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] loss_scale ................... 0\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] memory_breakdown ............. False\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] optimizer_legacy_fusion ...... False\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] optimizer_name ............... adamw\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] optimizer_params ............. {'lr': 3e-05, 'betas': [0.8, 0.999], 'eps': 1e-08, 'weight_decay': 3e-07}\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0}\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] pld_enabled .................. False\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] pld_params ................... False\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] prescale_gradients ........... False\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] scheduler_name ............... WarmupLR\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] scheduler_params ............. {'warmup_min_lr': 0, 'warmup_max_lr': 3e-05, 'warmup_num_steps': 500}\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] sparse_attention ............. None\r\n[2021-02-11 22:22:56,598] [INFO] [config.py:737:print] sparse_gradients_enabled ..... False\r\n[2021-02-11 22:22:56,599] [INFO] [config.py:737:print] steps_per_print .............. 2000\r\n[2021-02-11 22:22:56,599] [INFO] [config.py:737:print] tensorboard_enabled .......... False\r\n[2021-02-11 22:22:56,599] [INFO] [config.py:737:print] tensorboard_job_name ......... DeepSpeedJobName\r\n[2021-02-11 22:22:56,599] [INFO] [config.py:737:print] tensorboard_output_path ...... \r\n[2021-02-11 22:22:56,599] [INFO] [config.py:737:print] train_batch_size ............. 8\r\n[2021-02-11 22:22:56,599] [INFO] [config.py:737:print] train_micro_batch_size_per_gpu 2\r\n[2021-02-11 22:22:56,599] [INFO] [config.py:737:print] wall_clock_breakdown ......... false\r\n[2021-02-11 22:22:56,599] [INFO] [config.py:737:print] world_size ................... 1\r\n[2021-02-11 22:22:56,599] [INFO] [config.py:737:print] zero_allow_untested_optimizer true\r\n[2021-02-11 22:22:56,599] [INFO] [config.py:737:print] zero_config .................. {\r\n \"allgather_bucket_size\": 200000000.0,\r\n \"allgather_partitions\": \"true\",\r\n \"contiguous_gradients\": \"true\",\r\n \"cpu_offload\": \"true\",\r\n \"elastic_checkpoint\": true,\r\n \"load_from_fp32_weights\": true,\r\n \"overlap_comm\": \"true\",\r\n \"reduce_bucket_size\": 200000000.0,\r\n \"reduce_scatter\": \"true\",\r\n \"stage\": 2\r\n}\r\n[2021-02-11 22:22:56,599] [INFO] [config.py:737:print] zero_enabled ................. True\r\n[2021-02-11 22:22:56,599] [INFO] [config.py:737:print] zero_optimization_stage ...... 2\r\n[2021-02-11 22:22:56,599] [INFO] [config.py:739:print] json = {\r\n \"fp16\":{\r\n \"enabled\":\"true\",\r\n \"hysteresis\":2,\r\n \"loss_scale\":0,\r\n \"loss_scale_window\":1000,\r\n \"min_loss_scale\":1\r\n },\r\n \"gradient_accumulation_steps\":4,\r\n \"gradient_clipping\":1.0,\r\n \"optimizer\":{\r\n \"params\":{\r\n \"betas\":[\r\n 0.8,\r\n 0.999\r\n ],\r\n \"eps\":1e-08,\r\n \"lr\":3e-05,\r\n \"weight_decay\":3e-07\r\n },\r\n \"type\":\"AdamW\"\r\n },\r\n \"scheduler\":{\r\n \"params\":{\r\n \"warmup_max_lr\":3e-05,\r\n \"warmup_min_lr\":0,\r\n \"warmup_num_steps\":500\r\n },\r\n \"type\":\"WarmupLR\"\r\n },\r\n \"steps_per_print\":2000,\r\n \"train_micro_batch_size_per_gpu\":2,\r\n \"wall_clock_breakdown\":\"false\",\r\n \"zero_allow_untested_optimizer\":\"true\",\r\n \"zero_optimization\":{\r\n \"allgather_bucket_size\":200000000.0,\r\n \"allgather_partitions\":\"true\",\r\n \"contiguous_gradients\":\"true\",\r\n \"cpu_offload\":\"true\",\r\n \"overlap_comm\":\"true\",\r\n \"reduce_bucket_size\":200000000.0,\r\n \"reduce_scatter\":\"true\",\r\n \"stage\":2\r\n }\r\n}\r\nUsing /root/.cache/torch_extensions as PyTorch extensions root...\r\nNo modifications detected for re-loaded extension module utils, skipping build step...\r\nLoading extension module utils...\r\nTime to load utils op: 0.0005064010620117188 seconds\r\n\r\n 0%| | 0/125 [00:00<?, ?it/s]/usr/local/lib/python3.8/dist-packages/nlp/utils/py_utils.py:191: UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. This means you can write to the underlying (supposedly non-writeable) NumPy array using the tensor. You may want to copy the array to protect its data or make it writeable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:141.)\r\n return function(data_struct)\r\nTraceback (most recent call last):\r\n File \"abstractive_summarization.py\", line 396, in <module>\r\n run()\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 829, in __call__\r\n return self.main(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 782, in main\r\n rv = self.invoke(ctx)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 1066, in invoke\r\n return ctx.invoke(self.callback, **ctx.params)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 610, in invoke\r\n return callback(*args, **kwargs)\r\n File \"abstractive_summarization.py\", line 371, in run\r\n trainer.train()\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/trainer.py\", line 886, in train\r\n tr_loss += self.training_step(model, inputs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/trainer.py\", line 1265, in training_step\r\n self.model_wrapped.module.backward(loss)\r\n File \"/usr/local/lib/python3.8/dist-packages/deepspeed/runtime/engine.py\", line 903, in backward\r\n self.optimizer.backward(loss)\r\n File \"/usr/local/lib/python3.8/dist-packages/deepspeed/runtime/zero/stage2.py\", line 1596, in backward\r\n buf_0 = torch.empty(int(self.reduce_bucket_size * 4.5),\r\nRuntimeError: CUDA out of memory. Tried to allocate 1.68 GiB (GPU 0; 15.78 GiB total capacity; 12.80 GiB already allocated; 1.63 GiB free; 12.97 GiB reserved in total by PyTorch)\r\n\r\n 0%| | 0/125 [00:00<?, ?it/s]\r\n```", "> I'm not sure why I might be getting 0MB peak memory, 0 GB cached memory and no memory usage\r\n\r\nAh, yes, the older pytorch is buggy and you need to use the device context manager to get the correct numbers, e.g:\r\n\r\n```\r\n def get_current_gpu_memory_use():\r\n \"\"\" returns a list of cuda memory allocations per GPU in MBs\"\"\"\r\n\r\n per_device_memory = []\r\n for id in range(torch.cuda.device_count()):\r\n with torch.cuda.device(id):\r\n per_device_memory.append(torch.cuda.memory_allocated() >> 20)\r\n\r\n return per_device_memory\r\n```\r\n\r\n`pynvml` is another way, and it's more useful in this context since it shows the full memory usage and not just the pytorch's allocation - there are other things happening on the gpu that pytorch doesn't account for - primarily 0.5-1.5GB of cuda kernels preloading.\r\n\r\nIf you're working with notebooks you may want to consider using https://github.com/stas00/ipyexperiments/ and it'll tell you cell by cell all the memory usage stats automatically. It takes its measurements via `pynvml`.\r\n\r\nBut you can also use its util functions in a standalone script, e.g.: after `pip install ipyexperiments`\r\n```\r\npython -c \"from ipyexperiments.utils.mem import gpu_mem_get_mbs; print(gpu_mem_get_mbs())\"\r\nGPUMemory(total=8119, free=8115, used=4)\r\n```\r\n\r\nThis will give you identical numbers to `nvidia-smi` and not `torch.cuda` memory API. The latter is always smaller since it doesn't account for the cuda kernels.", "> If I omit deepspeed, I run into memory fragment error using those 1000 examples. \r\n\r\nBased on the log - you're not omitting deepspeed, you're running the same thing.\r\n\r\nSince you keep getting the exact same error - something is telling me that you're editing one thing but running another thing - find a way to make sure that the script that you run is actually up-to-date with your edits.\r\n", "I tried playing with your script w/o DeepSpeed and I'm not sure how you're getting a much higher GPU memory usage, it shouldn't be very different regardless of gpu, as I suggested - is it possible that you modify one script but run another? \r\n\r\ne.g. what happens if you set `decoder_max_length = 64` - it should cut off a few GBs for bs=2 that you're trying to get in.\r\n\r\nThe other thing I'd check is using a more recent pytorch version.\r\n\r\nalso, https://github.com/huggingface/transformers/pull/10130 is merged now, so you don't need to pass `local_rank=0` to trainer args class if you update to transformers master.\r\n", "hello @stas00 thank you for the update! I tried testing it without deepspeed. I also tried checking out the following:\r\n```python\r\n nvmlInit()\r\n h = nvmlDeviceGetHandleByIndex(0)\r\n info = nvmlDeviceGetMemoryInfo(h)\r\n logger.info(f'GPU total Memory : {info.total}')\r\n logger.info(f'GPU free Memory : {info.free}')\r\n logger.info(f'GPU Memory used : {info.used}')\r\n```\r\nand I got\r\n\r\n```\r\n[INFO] 2021-02-12 02:02:42,596 abstractive_summarization: GPU total Memory : 16945512448\r\n[INFO] 2021-02-12 02:02:42,596 abstractive_summarization: GPU free Memory : 16941842432\r\n[INFO] 2021-02-12 02:02:42,596 abstractive_summarization: GPU Memory used : 3670016\r\n```\r\n\r\nbut after running the snippet below, I still run into\r\n```\r\nRuntimeError: CUDA out of memory. Tried to allocate 194.00 MiB (GPU 0; 15.78 GiB total capacity; 14.12 GiB already allocated; 146.00 MiB free; 14.47 GiB reserved in total by PyTorch)\r\n\r\n 0%| | 0/125 [00:00<?, ?it/s]\r\n```\r\nit looks like I'm able to fine tune`MODEL_NAME='allenai/led-base-16384'` as the base model(currently testing it out) , but I run into issues when trying to fine tune `patrickvonplaten/led-large-16384-pubmed` using the snippet below. I'd greatly appreciate any other suggestions you might have\r\n\r\n```python\r\nimport datasets\r\nfrom datasets import load_dataset, load_metric\r\n\r\nimport click\r\nimport torch\r\nimport logging\r\nimport boto3\r\nimport json\r\n\r\nfrom io import BytesIO\r\nimport pandas as pd\r\n\r\nimport pyarrow as pa\r\nimport pyarrow.parquet as pq\r\nfrom nlp import arrow_dataset\r\n\r\nimport glob\r\nimport os\r\nimport tarfile\r\nimport os.path\r\nfrom transformers import (\r\n AutoTokenizer,\r\n AutoModelForSeq2SeqLM,\r\n Seq2SeqTrainer,\r\n Seq2SeqTrainingArguments,\r\n AutoTokenizer,\r\n AutoModelForSeq2SeqLM,\r\n)\r\n\r\n\r\nimport torch.utils.checkpoint\r\nfrom pynvml import *\r\n\r\n\r\n\r\n\r\nlogger = logging.getLogger(__name__)\r\nlogger.setLevel(logging.INFO)\r\nlogging.basicConfig(\r\n level=logging.INFO, format=\"[%(levelname)s] %(asctime)s %(module)s: %(message)s\"\r\n)\r\n\r\n\r\n\r\n\r\nrouge = load_metric(\"rouge\")\r\n\r\n\r\nMODEL_NAME = \"patrickvonplaten/led-large-16384-pubmed\"\r\n\r\n\r\n# ds_config = {\r\n# \"fp16\": {\r\n# \"enabled\": \"true\",\r\n# \"loss_scale\": 0,\r\n# \"loss_scale_window\": 1000,\r\n# \"hysteresis\": 2,\r\n# \"min_loss_scale\": 1\r\n# },\r\n\r\n# \"zero_optimization\": {\r\n# \"stage\": 2,\r\n# \"allgather_partitions\": \"true\",\r\n# \"allgather_bucket_size\": 2e8,\r\n# \"overlap_comm\": \"true\",\r\n# \"reduce_scatter\": \"true\",\r\n# \"reduce_bucket_size\": 2e8,\r\n# \"contiguous_gradients\": \"true\",\r\n# \"cpu_offload\": \"true\"\r\n# },\r\n\r\n# \"zero_allow_untested_optimizer\": \"true\",\r\n\r\n# \"optimizer\": {\r\n# \"type\": \"AdamW\",\r\n# \"params\": {\r\n# \"lr\": 3e-5,\r\n# \"betas\": [0.8, 0.999],\r\n# \"eps\": 1e-8,\r\n# \"weight_decay\": 3e-7\r\n# }\r\n# },\r\n\r\n# \"scheduler\": {\r\n# \"type\": \"WarmupLR\",\r\n# \"params\": {\r\n# \"warmup_min_lr\": 0,\r\n# \"warmup_max_lr\": 3e-5,\r\n# \"warmup_num_steps\": 500\r\n# }\r\n# },\r\n\r\n# \"steps_per_print\": 2000,\r\n# \"wall_clock_breakdown\": \"false\"\r\n# }\r\n\r\n# with open('ds_config.json', 'w') as fp:\r\n# json.dump(ds_config, fp)\r\n\r\n\r\n\r\n\r\nlogger.info(f\"load tokenizer using {MODEL_NAME}\")\r\ntokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)\r\n\r\nlogger.info(f\"Load {MODEL_NAME}. IMPORTANT NOTE:I'm enabling gradient checkpointing to save memory.\")\r\n# load model + enable gradient checkpointing & disable cache for checkpointing\r\nled = AutoModelForSeq2SeqLM.from_pretrained(\r\n MODEL_NAME,\r\n gradient_checkpointing=False,\r\n use_cache=False,\r\n)\r\n\r\n\r\n# max encoder length is 2048 for PubMed\r\nencoder_max_length = 2048\r\ndecoder_max_length = 256\r\nbatch_size = 2\r\n\r\n\r\n# set decoding params\r\nled.config.num_beams = 2\r\nled.config.max_length = 256\r\nled.config.min_length = 100\r\nled.config.length_penalty = 2.0\r\nled.config.early_stopping = True\r\nled.config.no_repeat_ngram_size = 3\r\n\r\n\r\n\r\ndef process_data_to_model_inputs(batch):\r\n # tokenize the inputs and labels\r\n inputs = tokenizer(\r\n batch[\"extractive_summary\"],\r\n padding=\"max_length\",\r\n truncation=True,\r\n max_length=encoder_max_length,\r\n )\r\n outputs = tokenizer(\r\n batch[\"reference_summary\"],\r\n padding=\"max_length\",\r\n truncation=True,\r\n max_length=decoder_max_length,\r\n )\r\n\r\n batch[\"input_ids\"] = inputs.input_ids\r\n batch[\"attention_mask\"] = inputs.attention_mask\r\n\r\n # create 0 global_attention_mask lists\r\n batch[\"global_attention_mask\"] = len(batch[\"input_ids\"]) * [\r\n [0 for _ in range(len(batch[\"input_ids\"][0]))]\r\n ]\r\n\r\n # since above lists are references, the following line changes the 0 index for all samples\r\n batch[\"global_attention_mask\"][0][0] = 1\r\n batch[\"labels\"] = outputs.input_ids\r\n\r\n # We have to make sure that the PAD token is ignored\r\n batch[\"labels\"] = [\r\n [-100 if token == tokenizer.pad_token_id else token for token in labels]\r\n for labels in batch[\"labels\"]\r\n ]\r\n\r\n return batch\r\n\r\ndef compute_metrics(pred):\r\n labels_ids = pred.label_ids\r\n pred_ids = pred.predictions\r\n\r\n pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)\r\n labels_ids[labels_ids == -100] = tokenizer.pad_token_id\r\n label_str = tokenizer.batch_decode(labels_ids, skip_special_tokens=True)\r\n\r\n rouge_output = rouge.compute(\r\n predictions=pred_str, references=label_str, rouge_types=[\"rouge2\"]\r\n )[\"rouge2\"].mid\r\n\r\n return {\r\n \"rouge2_precision\": round(rouge_output.precision, 4),\r\n \"rouge2_recall\": round(rouge_output.recall, 4),\r\n \"rouge2_fmeasure\": round(rouge_output.fmeasure, 4),\r\n }\r\n\r\ndef run():\r\n nvmlInit()\r\n h = nvmlDeviceGetHandleByIndex(0)\r\n info = nvmlDeviceGetMemoryInfo(h)\r\n logger.info(f'GPU total Memory : {info.total}')\r\n logger.info(f'GPU free Memory : {info.free}')\r\n logger.info(f'GPU Memory used : {info.used}')\r\n\r\n logger.info(\"create fictious train and test data\")\r\n n_recs = 1000\r\n frames = [\r\n {\"reference_summary\": [' '.join([f\"{i} I am a reference summary\"] * 200),\r\n ' '.join([\"I am another reference summary\"] * 200)],\r\n \"extractive_summary\": [' '.join([f\"{i} hello\"] * 200), ' '.join([\"goodbye\"] * 200)]} for i in range(n_recs)]\r\n train = pd.DataFrame(frames)\r\n test = pd.DataFrame({\"reference_summary\": [' '.join([\"I am another reference summary\"] * 200)],\r\n \"extractive_summary\": [' '.join([\"goodbye\"] * 200)]})\r\n\r\n train = pa.Table.from_pandas(train)\r\n train = arrow_dataset.Dataset(train)\r\n\r\n test = pa.Table.from_pandas(test)\r\n test = arrow_dataset.Dataset(test)\r\n logger.info(\"map train data\")\r\n train = train.map(\r\n process_data_to_model_inputs,\r\n batched=True,\r\n batch_size=batch_size,\r\n remove_columns=[\"reference_summary\", \"extractive_summary\"],\r\n )\r\n\r\n logger.info(\"map test data\")\r\n test = test.map(\r\n process_data_to_model_inputs,\r\n batched=True,\r\n batch_size=batch_size,\r\n remove_columns=[\"reference_summary\", \"extractive_summary\"],\r\n\r\n )\r\n\r\n logger.info(\"set Python list in train to PyTorch tensor\")\r\n train.set_format(\r\n type=\"torch\",\r\n columns=[\"input_ids\", \"attention_mask\", \"global_attention_mask\", \"labels\"],\r\n )\r\n\r\n logger.info(\"set Python list in test to PyTorch tensor\")\r\n test.set_format(\r\n type=\"torch\",\r\n columns=[\"input_ids\", \"attention_mask\", \"global_attention_mask\", \"labels\"],\r\n )\r\n\r\n logger.info(\"enable fp16 amp training\") \r\n\r\n #define env variables required for training\r\n os.environ['MASTER_ADDR'] = \"10.23.29.192\"\r\n os.environ['MASTER_PORT'] = \"29500\"\r\n os.environ['RANK'] = \"0\"\r\n os.environ['LOCAL_RANK'] = \"0\"\r\n os.environ['WORLD_SIZE'] = \"1\"\r\n\r\n checkpoint_dir_path = \"/mnt/summarization_checkpoints\"\r\n training_args = Seq2SeqTrainingArguments(\r\n predict_with_generate=True,\r\n evaluation_strategy=\"steps\",\r\n per_device_train_batch_size=batch_size,\r\n per_device_eval_batch_size=batch_size,\r\n fp16=True,\r\n output_dir=checkpoint_dir_path,\r\n logging_steps=5,\r\n eval_steps=10,\r\n save_steps=10,\r\n save_total_limit=1,\r\n gradient_accumulation_steps=4,\r\n num_train_epochs=1,\r\n local_rank=0,\r\n# deepspeed=\"ds_config.json\"\r\n )\r\n\r\n training_args._setup_devices\r\n\r\n os.makedirs(checkpoint_dir_path, exist_ok=True)\r\n logger.info(\"instantiate trainer\")\r\n trainer = Seq2SeqTrainer(\r\n model=led,\r\n tokenizer=tokenizer,\r\n args=training_args,\r\n compute_metrics=compute_metrics,\r\n train_dataset=train,\r\n eval_dataset=test,\r\n )\r\n\r\n\r\n\r\n logger.info(\"start training\")\r\n trainer.train()\r\n\r\n\r\nif __name__ == \"__main__\":\r\n run()\r\n```\r\n\r\n```\r\n[INFO] 2021-02-12 02:02:16,547 filelock: Lock 139661825384256 released on /root/.cache/huggingface/transformers/85a878681daf8945866e644056c360d1fefe287fc88b31b48c20478be4d12b24.d2560ecf8e14415e1113077ca8941c38e7512a1e8b82e19e4150c7ab9e45350a.lock\r\n[INFO] 2021-02-12 02:02:42,587 abstractive_summarization: Using device: cuda\r\n[INFO] 2021-02-12 02:02:42,596 abstractive_summarization: GPU total Memory : 16945512448\r\n[INFO] 2021-02-12 02:02:42,596 abstractive_summarization: GPU free Memory : 16941842432\r\n[INFO] 2021-02-12 02:02:42,596 abstractive_summarization: GPU Memory used : 3670016\r\n[INFO] 2021-02-12 02:02:42,673 abstractive_summarization: map train data\r\n\r\n 0%| | 0/500 [00:00<?, ?it/s]\r\n 1%| | 4/500 [00:00<00:15, 31.16it/s]\r\n 2%|▏ | 8/500 [00:00<00:15, 32.18it/s]\r\n 2%|▏ | 12/500 [00:00<00:15, 32.37it/s]\r\n 3%|▎ | 16/500 [00:00<00:15, 32.17it/s]\r\n 4%|▍ | 20/500 [00:00<00:14, 32.20it/s]\r\n 5%|▍ | 24/500 [00:00<00:14, 32.11it/s]\r\n 6%|▌ | 28/500 [00:00<00:15, 30.96it/s]\r\n 6%|▋ | 32/500 [00:01<00:15, 31.08it/s]\r\n 7%|▋ | 36/500 [00:01<00:14, 31.49it/s]\r\n 8%|▊ | 40/500 [00:01<00:14, 31.94it/s]\r\n 9%|▉ | 44/500 [00:01<00:14, 31.91it/s]\r\n 10%|▉ | 48/500 [00:01<00:14, 32.20it/s]\r\n 10%|█ | 52/500 [00:01<00:13, 32.33it/s]\r\n 11%|█ | 56/500 [00:01<00:13, 32.40it/s]\r\n 12%|█▏ | 60/500 [00:01<00:13, 32.55it/s]\r\n 13%|█▎ | 64/500 [00:01<00:13, 32.58it/s]\r\n 14%|█▎ | 68/500 [00:02<00:13, 32.64it/s]\r\n 14%|█▍ | 72/500 [00:02<00:13, 32.75it/s]\r\n 15%|█▌ | 76/500 [00:02<00:12, 32.69it/s]\r\n 16%|█▌ | 80/500 [00:02<00:12, 32.68it/s]\r\n 17%|█▋ | 84/500 [00:02<00:12, 32.17it/s]\r\n 18%|█▊ | 88/500 [00:02<00:12, 32.16it/s]\r\n 18%|█▊ | 92/500 [00:02<00:12, 32.27it/s]\r\n 19%|█▉ | 96/500 [00:02<00:12, 32.32it/s]\r\n 20%|██ | 100/500 [00:03<00:12, 32.41it/s]\r\n 21%|██ | 104/500 [00:03<00:12, 32.52it/s]\r\n 22%|██▏ | 108/500 [00:03<00:12, 32.44it/s]\r\n 22%|██▏ | 112/500 [00:03<00:11, 32.57it/s]\r\n 23%|██▎ | 116/500 [00:03<00:11, 32.33it/s]\r\n 24%|██▍ | 120/500 [00:03<00:11, 31.91it/s]\r\n 25%|██▍ | 124/500 [00:03<00:12, 30.94it/s]\r\n 26%|██▌ | 128/500 [00:03<00:11, 31.47it/s]\r\n 26%|██▋ | 132/500 [00:04<00:11, 31.89it/s]\r\n 27%|██▋ | 136/500 [00:04<00:11, 32.22it/s]\r\n 28%|██▊ | 140/500 [00:04<00:11, 32.55it/s]\r\n 29%|██▉ | 144/500 [00:04<00:10, 32.57it/s]\r\n 30%|██▉ | 148/500 [00:04<00:10, 32.65it/s]\r\n 30%|███ | 152/500 [00:04<00:10, 32.65it/s]\r\n 31%|███ | 156/500 [00:04<00:11, 31.24it/s]\r\n 32%|███▏ | 160/500 [00:04<00:10, 31.56it/s]\r\n 33%|███▎ | 164/500 [00:05<00:10, 31.00it/s]\r\n 34%|███▎ | 168/500 [00:05<00:10, 31.50it/s]\r\n 34%|███▍ | 172/500 [00:05<00:10, 31.58it/s]\r\n 35%|███▌ | 176/500 [00:05<00:10, 31.86it/s]\r\n 36%|███▌ | 180/500 [00:05<00:09, 32.15it/s]\r\n 37%|███▋ | 184/500 [00:05<00:09, 32.31it/s]\r\n 38%|███▊ | 188/500 [00:05<00:09, 32.32it/s]\r\n 38%|███▊ | 192/500 [00:05<00:09, 32.16it/s]\r\n 39%|███▉ | 196/500 [00:06<00:09, 32.09it/s]\r\n 40%|████ | 200/500 [00:06<00:09, 31.76it/s]\r\n 41%|████ | 204/500 [00:06<00:09, 31.90it/s]\r\n 42%|████▏ | 208/500 [00:06<00:09, 31.94it/s]\r\n 42%|████▏ | 212/500 [00:06<00:09, 31.84it/s]\r\n 43%|████▎ | 216/500 [00:06<00:08, 31.90it/s]\r\n 44%|████▍ | 220/500 [00:06<00:08, 31.43it/s]\r\n 45%|████▍ | 224/500 [00:06<00:08, 31.20it/s]\r\n 46%|████▌ | 228/500 [00:07<00:08, 31.09it/s]\r\n 46%|████▋ | 232/500 [00:07<00:08, 30.88it/s]\r\n 47%|████▋ | 236/500 [00:07<00:08, 30.69it/s]\r\n 48%|████▊ | 240/500 [00:07<00:08, 30.71it/s]\r\n 49%|████▉ | 244/500 [00:07<00:08, 30.81it/s]\r\n 50%|████▉ | 248/500 [00:07<00:08, 30.49it/s]\r\n 50%|█████ | 252/500 [00:07<00:08, 30.63it/s]\r\n 51%|█████ | 256/500 [00:08<00:08, 30.16it/s]\r\n 52%|█████▏ | 260/500 [00:08<00:07, 30.22it/s]\r\n 53%|█████▎ | 264/500 [00:08<00:07, 30.17it/s]\r\n 54%|█████▎ | 268/500 [00:08<00:07, 30.11it/s]\r\n 54%|█████▍ | 272/500 [00:08<00:07, 30.21it/s]\r\n 55%|█████▌ | 276/500 [00:08<00:07, 29.75it/s]\r\n 56%|█████▌ | 280/500 [00:08<00:07, 29.45it/s]\r\n 57%|█████▋ | 284/500 [00:08<00:07, 29.73it/s]\r\n 57%|█████▋ | 287/500 [00:09<00:07, 29.79it/s]\r\n 58%|█████▊ | 291/500 [00:09<00:06, 30.13it/s]\r\n 59%|█████▉ | 295/500 [00:09<00:06, 30.11it/s]\r\n 60%|█████▉ | 299/500 [00:09<00:06, 30.29it/s]\r\n 61%|██████ | 303/500 [00:09<00:06, 30.54it/s]\r\n 61%|██████▏ | 307/500 [00:09<00:06, 30.60it/s]\r\n 62%|██████▏ | 311/500 [00:09<00:06, 30.46it/s]\r\n 63%|██████▎ | 315/500 [00:10<00:06, 29.67it/s]\r\n 64%|██████▎ | 318/500 [00:10<00:06, 29.63it/s]\r\n 64%|██████▍ | 321/500 [00:10<00:06, 29.68it/s]\r\n 65%|██████▌ | 325/500 [00:10<00:05, 29.86it/s]\r\n 66%|██████▌ | 328/500 [00:10<00:06, 28.25it/s]\r\n 66%|██████▋ | 332/500 [00:10<00:05, 29.00it/s]\r\n 67%|██████▋ | 336/500 [00:10<00:05, 29.48it/s]\r\n 68%|██████▊ | 339/500 [00:10<00:05, 29.49it/s]\r\n 68%|██████▊ | 342/500 [00:10<00:05, 29.58it/s]\r\n 69%|██████▉ | 346/500 [00:11<00:05, 29.82it/s]\r\n 70%|██████▉ | 349/500 [00:11<00:05, 29.74it/s]\r\n 71%|███████ | 353/500 [00:11<00:04, 30.13it/s]\r\n 71%|███████▏ | 357/500 [00:11<00:04, 29.24it/s]\r\n 72%|███████▏ | 360/500 [00:11<00:04, 29.36it/s]\r\n 73%|███████▎ | 364/500 [00:11<00:04, 29.53it/s]\r\n 73%|███████▎ | 367/500 [00:11<00:04, 29.56it/s]\r\n 74%|███████▍ | 371/500 [00:11<00:04, 29.89it/s]\r\n 75%|███████▍ | 374/500 [00:12<00:04, 29.64it/s]\r\n 76%|███████▌ | 378/500 [00:12<00:04, 29.90it/s]\r\n 76%|███████▋ | 382/500 [00:12<00:03, 30.15it/s]\r\n 77%|███████▋ | 386/500 [00:12<00:03, 30.31it/s]\r\n 78%|███████▊ | 390/500 [00:12<00:03, 30.44it/s]\r\n 79%|███████▉ | 394/500 [00:12<00:03, 30.53it/s]\r\n 80%|███████▉ | 398/500 [00:12<00:03, 30.31it/s]\r\n 80%|████████ | 402/500 [00:12<00:03, 30.13it/s]\r\n 81%|████████ | 406/500 [00:13<00:03, 30.27it/s]\r\n 82%|████████▏ | 410/500 [00:13<00:03, 29.79it/s]\r\n 83%|████████▎ | 413/500 [00:13<00:02, 29.24it/s]\r\n 83%|████████▎ | 416/500 [00:13<00:02, 29.16it/s]\r\n 84%|████████▍ | 419/500 [00:13<00:02, 29.09it/s]\r\n 85%|████████▍ | 423/500 [00:13<00:02, 29.44it/s]\r\n 85%|████████▌ | 427/500 [00:13<00:02, 29.74it/s]\r\n 86%|████████▌ | 431/500 [00:13<00:02, 29.89it/s]\r\n 87%|████████▋ | 435/500 [00:14<00:02, 30.06it/s]\r\n 88%|████████▊ | 439/500 [00:14<00:02, 30.15it/s]\r\n 89%|████████▊ | 443/500 [00:14<00:01, 30.08it/s]\r\n 89%|████████▉ | 447/500 [00:14<00:01, 29.99it/s]\r\n 90%|█████████ | 451/500 [00:14<00:01, 30.03it/s]\r\n 91%|█████████ | 455/500 [00:14<00:01, 30.05it/s]\r\n 92%|█████████▏| 459/500 [00:14<00:01, 30.04it/s]\r\n 93%|█████████▎| 463/500 [00:14<00:01, 30.14it/s]\r\n 93%|█████████▎| 467/500 [00:15<00:01, 30.10it/s]\r\n 94%|█████████▍| 471/500 [00:15<00:00, 29.80it/s]\r\n 95%|█████████▍| 474/500 [00:15<00:00, 29.67it/s]\r\n 96%|█████████▌| 478/500 [00:15<00:00, 29.75it/s]\r\n 96%|█████████▋| 482/500 [00:15<00:00, 29.95it/s]\r\n 97%|█████████▋| 486/500 [00:15<00:00, 30.07it/s]\r\n 98%|█████████▊| 490/500 [00:15<00:00, 29.73it/s]\r\n 99%|█████████▉| 494/500 [00:16<00:00, 29.84it/s]\r\n100%|█████████▉| 498/500 [00:16<00:00, 30.03it/s]\r\n100%|██████████| 500/500 [00:16<00:00, 30.82it/s]\r\n[INFO] 2021-02-12 02:02:58,936 arrow_writer: Done writing 1000 examples in 51224000 bytes .\r\n[INFO] 2021-02-12 02:02:58,945 abstractive_summarization: map test data\r\n\r\n 0%| | 0/1 [00:00<?, ?it/s]\r\n100%|██████████| 1/1 [00:00<00:00, 91.93it/s]\r\n[INFO] 2021-02-12 02:02:58,961 arrow_writer: Done writing 1 examples in 51232 bytes .\r\n[INFO] 2021-02-12 02:02:58,962 abstractive_summarization: set Python list in train to PyTorch tensor\r\n[INFO] 2021-02-12 02:02:58,962 arrow_dataset: Set __getitem__(key) output type to torch for ['input_ids', 'attention_mask', 'global_attention_mask', 'labels'] columns (when key is int or slice) and don't output other (un-formated) columns.\r\n[INFO] 2021-02-12 02:02:58,962 abstractive_summarization: set Python list in test to PyTorch tensor\r\n[INFO] 2021-02-12 02:02:58,962 arrow_dataset: Set __getitem__(key) output type to torch for ['input_ids', 'attention_mask', 'global_attention_mask', 'labels'] columns (when key is int or slice) and don't output other (un-formated) columns.\r\n[INFO] 2021-02-12 02:02:58,962 abstractive_summarization: enable fp16 amp training\r\n[INFO] 2021-02-12 02:02:58,962 abstractive_summarization: file will be written to /workspace\r\n[INFO] 2021-02-12 02:02:59,261 abstractive_summarization: instantiate trainer\r\n[INFO] 2021-02-12 02:03:02,626 abstractive_summarization: start training\r\n\r\n 0%| | 0/125 [00:00<?, ?it/s]/usr/local/lib/python3.8/dist-packages/nlp/utils/py_utils.py:191: UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. This means you can write to the underlying (supposedly non-writeable) NumPy array using the tensor. You may want to copy the array to protect its data or make it writeable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:141.)\r\n return function(data_struct)\r\nTraceback (most recent call last):\r\n File \"abstractive_summarization.py\", line 408, in <module>\r\n run()\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 829, in __call__\r\n return self.main(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 782, in main\r\n rv = self.invoke(ctx)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 1066, in invoke\r\n return ctx.invoke(self.callback, **ctx.params)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 610, in invoke\r\n return callback(*args, **kwargs)\r\n File \"abstractive_summarization.py\", line 383, in run\r\n trainer.train()\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/trainer.py\", line 938, in train\r\n tr_loss += self.training_step(model, inputs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/trainer.py\", line 1302, in training_step\r\n loss = self.compute_loss(model, inputs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/trainer.py\", line 1334, in compute_loss\r\n outputs = model(**inputs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/parallel/distributed.py\", line 511, in forward\r\n output = self.module(*inputs[0], **kwargs[0])\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 2344, in forward\r\n outputs = self.led(\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 2193, in forward\r\n encoder_outputs = self.encoder(\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 1831, in forward\r\n layer_outputs = encoder_layer(\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 907, in forward\r\n attn_outputs = self.self_attn(\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 718, in forward\r\n self_outputs = self.longformer_self_attn(\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 276, in forward\r\n attn_output = self._compute_attn_output_with_global_indices(\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 597, in _compute_attn_output_with_global_indices\r\n attn_output_without_global = self._sliding_chunks_matmul_attn_probs_value(\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 505, in _sliding_chunks_matmul_attn_probs_value\r\n chunked_attn_probs = self._pad_and_diagonalize(chunked_attn_probs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 356, in _pad_and_diagonalize\r\n chunked_hidden_states = F.pad(\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/functional.py\", line 3552, in _pad\r\n return _VF.constant_pad_nd(input, pad, value)\r\nRuntimeError: CUDA out of memory. Tried to allocate 194.00 MiB (GPU 0; 15.78 GiB total capacity; 14.12 GiB already allocated; 146.00 MiB free; 14.47 GiB reserved in total by PyTorch)\r\n\r\n 0%| | 0/125 [00:00<?, ?it/s]\r\n```", "Have you read the suggestions at https://github.com/huggingface/transformers/issues/10011#issuecomment-777918847?\r\n\r\n", "Hi @stas00 thank you for the update and merge! If possible, I'm trying to avoid reducing the decoder output. We would love summaries that are around 200 tokens in length. \r\n\r\nI'm noticing, if I try using deepspeed, it's now hanging on here:\r\n```\r\n[2021-02-12 16:55:53,106] [INFO] [distributed.py:46:init_distributed] Initializing torch distributed with backend: nccl\r\n```\r\n\r\nand then times out\r\n```\r\nTraceback (most recent call last):\r\n File \"abstractive_summarization.py\", line 407, in <module>\r\n run()\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 829, in __call__\r\n return self.main(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 782, in main\r\n rv = self.invoke(ctx)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 1066, in invoke\r\n return ctx.invoke(self.callback, **ctx.params)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 610, in invoke\r\n return callback(*args, **kwargs)\r\n File \"abstractive_summarization.py\", line 349, in run\r\n training_args = Seq2SeqTrainingArguments(\r\n File \"<string>\", line 61, in __init__\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/training_args.py\", line 478, in __post_init__\r\n if is_torch_available() and self.device.type != \"cuda\" and self.fp16:\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/file_utils.py\", line 1346, in wrapper\r\n return func(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/training_args.py\", line 583, in device\r\n return self._setup_devices\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/file_utils.py\", line 1336, in __get__\r\n cached = self.fget(obj)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/file_utils.py\", line 1346, in wrapper\r\n return func(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/training_args.py\", line 551, in _setup_devices\r\n deepspeed.init_distributed()\r\n File \"/usr/local/lib/python3.8/dist-packages/deepspeed/utils/distributed.py\", line 49, in init_distributed\r\n torch.distributed.init_process_group(backend=dist_backend,\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/distributed/distributed_c10d.py\", line 422, in init_process_group\r\n store, rank, world_size = next(rendezvous_iterator)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/distributed/rendezvous.py\", line 172, in _env_rendezvous_handler\r\n store = TCPStore(master_addr, master_port, world_size, start_daemon, timeout)\r\nRuntimeError: connect() timed out.\r\n```\r\n\r\nif I don't use deepspeed, I get\r\n```\r\n[INFO] 2021-02-12 17:44:39,161 filelock: Lock 140104053693120 released on /root/.cache/huggingface/transformers/85a878681daf8945866e644056c360d1fefe287fc88b31b48c20478be4d12b24.d2560ecf8e14415e1113077ca8941c38e7512a1e8b82e19e4150c7ab9e45350a.lock\r\n[INFO] 2021-02-12 17:45:05,102 abstractive_summarization: Using device: cuda\r\n[INFO] 2021-02-12 17:45:05,111 abstractive_summarization: GPU total Memory : 16945512448\r\n[INFO] 2021-02-12 17:45:05,111 abstractive_summarization: GPU free Memory : 16941842432\r\n[INFO] 2021-02-12 17:45:05,111 abstractive_summarization: GPU Memory used : 3670016\r\n[INFO] 2021-02-12 17:45:05,166 abstractive_summarization: map train data\r\n\r\n 0%| | 0/500 [00:00<?, ?it/s]\r\n 1%| | 3/500 [00:00<00:17, 28.33it/s]\r\n 1%|▏ | 7/500 [00:00<00:16, 29.83it/s]\r\n 2%|▏ | 11/500 [00:00<00:15, 31.06it/s]\r\n 3%|▎ | 15/500 [00:00<00:15, 31.43it/s]\r\n 4%|▍ | 19/500 [00:00<00:15, 31.97it/s]\r\n 5%|▍ | 23/500 [00:00<00:14, 32.19it/s]\r\n 5%|▌ | 27/500 [00:00<00:14, 32.25it/s]\r\n 6%|▌ | 31/500 [00:00<00:14, 32.31it/s]\r\n 7%|▋ | 35/500 [00:01<00:14, 31.67it/s]\r\n 8%|▊ | 39/500 [00:01<00:14, 31.92it/s]\r\n 9%|▊ | 43/500 [00:01<00:14, 31.44it/s]\r\n 9%|▉ | 47/500 [00:01<00:14, 31.64it/s]\r\n 10%|█ | 51/500 [00:01<00:14, 30.68it/s]\r\n 11%|█ | 55/500 [00:01<00:14, 31.12it/s]\r\n 12%|█▏ | 59/500 [00:01<00:14, 31.44it/s]\r\n 13%|█▎ | 63/500 [00:01<00:13, 31.84it/s]\r\n 13%|█▎ | 67/500 [00:02<00:13, 32.09it/s]\r\n 14%|█▍ | 71/500 [00:02<00:13, 32.37it/s]\r\n 15%|█▌ | 75/500 [00:02<00:13, 31.68it/s]\r\n 16%|█▌ | 79/500 [00:02<00:13, 31.91it/s]\r\n 17%|█▋ | 83/500 [00:02<00:13, 31.98it/s]\r\n 17%|█▋ | 87/500 [00:02<00:12, 32.10it/s]\r\n 18%|█▊ | 91/500 [00:02<00:12, 32.28it/s]\r\n 19%|█▉ | 95/500 [00:02<00:12, 32.27it/s]\r\n 20%|█▉ | 99/500 [00:03<00:12, 31.89it/s]\r\n 21%|██ | 103/500 [00:03<00:12, 31.60it/s]\r\n 21%|██▏ | 107/500 [00:03<00:12, 31.75it/s]\r\n 22%|██▏ | 111/500 [00:03<00:12, 31.95it/s]\r\n 23%|██▎ | 115/500 [00:03<00:11, 32.12it/s]\r\n 24%|██▍ | 119/500 [00:03<00:11, 32.21it/s]\r\n 25%|██▍ | 123/500 [00:03<00:11, 32.23it/s]\r\n 25%|██▌ | 127/500 [00:03<00:11, 32.28it/s]\r\n 26%|██▌ | 131/500 [00:04<00:11, 31.77it/s]\r\n 27%|██▋ | 135/500 [00:04<00:11, 32.01it/s]\r\n 28%|██▊ | 139/500 [00:04<00:11, 32.07it/s]\r\n 29%|██▊ | 143/500 [00:04<00:11, 32.29it/s]\r\n 29%|██▉ | 147/500 [00:04<00:10, 32.43it/s]\r\n 30%|███ | 151/500 [00:04<00:10, 32.43it/s]\r\n 31%|███ | 155/500 [00:04<00:10, 32.27it/s]\r\n 32%|███▏ | 159/500 [00:04<00:10, 32.26it/s]\r\n 33%|███▎ | 163/500 [00:05<00:10, 30.81it/s]\r\n 33%|███▎ | 167/500 [00:05<00:10, 31.26it/s]\r\n 34%|███▍ | 171/500 [00:05<00:10, 31.56it/s]\r\n 35%|███▌ | 175/500 [00:05<00:10, 31.68it/s]\r\n 36%|███▌ | 179/500 [00:05<00:10, 31.88it/s]\r\n 37%|███▋ | 183/500 [00:05<00:09, 31.87it/s]\r\n 37%|███▋ | 187/500 [00:05<00:09, 32.08it/s]\r\n 38%|███▊ | 191/500 [00:06<00:09, 31.48it/s]\r\n 39%|███▉ | 195/500 [00:06<00:09, 31.16it/s]\r\n 40%|███▉ | 199/500 [00:06<00:09, 30.59it/s]\r\n 41%|████ | 203/500 [00:06<00:09, 30.72it/s]\r\n 41%|████▏ | 207/500 [00:06<00:09, 31.31it/s]\r\n 42%|████▏ | 211/500 [00:06<00:09, 31.58it/s]\r\n 43%|████▎ | 215/500 [00:06<00:08, 31.79it/s]\r\n 44%|████▍ | 219/500 [00:06<00:08, 31.72it/s]\r\n 45%|████▍ | 223/500 [00:07<00:08, 31.47it/s]\r\n 45%|████▌ | 227/500 [00:07<00:08, 31.32it/s]\r\n 46%|████▌ | 231/500 [00:07<00:08, 31.12it/s]\r\n 47%|████▋ | 235/500 [00:07<00:08, 30.91it/s]\r\n 48%|████▊ | 239/500 [00:07<00:08, 30.54it/s]\r\n 49%|████▊ | 243/500 [00:07<00:08, 30.43it/s]\r\n 49%|████▉ | 247/500 [00:07<00:08, 30.45it/s]\r\n 50%|█████ | 251/500 [00:07<00:08, 30.46it/s]\r\n 51%|█████ | 255/500 [00:08<00:07, 30.80it/s]\r\n 52%|█████▏ | 259/500 [00:08<00:07, 30.63it/s]\r\n 53%|█████▎ | 263/500 [00:08<00:07, 30.51it/s]\r\n 53%|█████▎ | 267/500 [00:08<00:07, 30.46it/s]\r\n 54%|█████▍ | 271/500 [00:08<00:07, 30.45it/s]\r\n 55%|█████▌ | 275/500 [00:08<00:07, 30.01it/s]\r\n 56%|█████▌ | 279/500 [00:08<00:07, 30.10it/s]\r\n 57%|█████▋ | 283/500 [00:09<00:07, 30.22it/s]\r\n 57%|█████▋ | 287/500 [00:09<00:07, 30.12it/s]\r\n 58%|█████▊ | 291/500 [00:09<00:06, 30.30it/s]\r\n 59%|█████▉ | 295/500 [00:09<00:06, 29.63it/s]\r\n 60%|█████▉ | 298/500 [00:09<00:06, 29.62it/s]\r\n 60%|██████ | 302/500 [00:09<00:06, 29.92it/s]\r\n 61%|██████ | 305/500 [00:09<00:06, 29.47it/s]\r\n 62%|██████▏ | 309/500 [00:09<00:06, 29.59it/s]\r\n 62%|██████▏ | 312/500 [00:09<00:06, 29.58it/s]\r\n 63%|██████▎ | 315/500 [00:10<00:06, 29.65it/s]\r\n 64%|██████▍ | 319/500 [00:10<00:06, 29.88it/s]\r\n 65%|██████▍ | 323/500 [00:10<00:05, 30.03it/s]\r\n 65%|██████▌ | 326/500 [00:10<00:06, 28.54it/s]\r\n 66%|██████▌ | 329/500 [00:10<00:05, 28.77it/s]\r\n 67%|██████▋ | 333/500 [00:10<00:05, 29.18it/s]\r\n 67%|██████▋ | 336/500 [00:10<00:05, 29.37it/s]\r\n 68%|██████▊ | 339/500 [00:10<00:05, 29.50it/s]\r\n 68%|██████▊ | 342/500 [00:11<00:05, 29.59it/s]\r\n 69%|██████▉ | 345/500 [00:11<00:05, 27.98it/s]\r\n 70%|██████▉ | 348/500 [00:11<00:05, 28.37it/s]\r\n 70%|███████ | 352/500 [00:11<00:05, 29.10it/s]\r\n 71%|███████ | 355/500 [00:11<00:04, 29.15it/s]\r\n 72%|███████▏ | 359/500 [00:11<00:04, 29.51it/s]\r\n 73%|███████▎ | 363/500 [00:11<00:04, 29.80it/s]\r\n 73%|███████▎ | 367/500 [00:11<00:04, 30.16it/s]\r\n 74%|███████▍ | 371/500 [00:11<00:04, 30.30it/s]\r\n 75%|███████▌ | 375/500 [00:12<00:04, 30.22it/s]\r\n 76%|███████▌ | 379/500 [00:12<00:03, 30.29it/s]\r\n 77%|███████▋ | 383/500 [00:12<00:03, 30.30it/s]\r\n 77%|███████▋ | 387/500 [00:12<00:03, 30.32it/s]\r\n 78%|███████▊ | 391/500 [00:12<00:03, 30.33it/s]\r\n 79%|███████▉ | 395/500 [00:12<00:03, 30.35it/s]\r\n 80%|███████▉ | 399/500 [00:12<00:03, 29.86it/s]\r\n 81%|████████ | 403/500 [00:13<00:03, 29.92it/s]\r\n 81%|████████ | 406/500 [00:13<00:03, 29.08it/s]\r\n 82%|████████▏ | 409/500 [00:13<00:03, 29.31it/s]\r\n 82%|████████▏ | 412/500 [00:13<00:03, 28.97it/s]\r\n 83%|████████▎ | 415/500 [00:13<00:03, 27.09it/s]\r\n 84%|████████▍ | 419/500 [00:13<00:02, 28.02it/s]\r\n 85%|████████▍ | 423/500 [00:13<00:02, 28.80it/s]\r\n 85%|████████▌ | 427/500 [00:13<00:02, 29.20it/s]\r\n 86%|████████▌ | 430/500 [00:13<00:02, 29.32it/s]\r\n 87%|████████▋ | 433/500 [00:14<00:02, 29.41it/s]\r\n 87%|████████▋ | 436/500 [00:14<00:02, 29.29it/s]\r\n 88%|████████▊ | 439/500 [00:14<00:02, 28.79it/s]\r\n 88%|████████▊ | 442/500 [00:14<00:01, 29.08it/s]\r\n 89%|████████▉ | 446/500 [00:14<00:01, 29.51it/s]\r\n 90%|█████████ | 450/500 [00:14<00:01, 29.84it/s]\r\n 91%|█████████ | 454/500 [00:14<00:01, 30.09it/s]\r\n 92%|█████████▏| 458/500 [00:14<00:01, 30.08it/s]\r\n 92%|█████████▏| 462/500 [00:15<00:01, 30.16it/s]\r\n 93%|█████████▎| 466/500 [00:15<00:01, 30.25it/s]\r\n 94%|█████████▍| 470/500 [00:15<00:00, 30.31it/s]\r\n 95%|█████████▍| 474/500 [00:15<00:00, 30.32it/s]\r\n 96%|█████████▌| 478/500 [00:15<00:00, 30.30it/s]\r\n 96%|█████████▋| 482/500 [00:15<00:00, 30.25it/s]\r\n 97%|█████████▋| 486/500 [00:15<00:00, 30.28it/s]\r\n 98%|█████████▊| 490/500 [00:15<00:00, 30.23it/s]\r\n 99%|█████████▉| 494/500 [00:16<00:00, 30.14it/s]\r\n100%|█████████▉| 498/500 [00:16<00:00, 30.12it/s]\r\n100%|██████████| 500/500 [00:16<00:00, 30.63it/s]\r\n[INFO] 2021-02-12 17:45:21,532 arrow_writer: Done writing 1000 examples in 51224000 bytes .\r\n[INFO] 2021-02-12 17:45:21,539 abstractive_summarization: map test data\r\n\r\n 0%| | 0/1 [00:00<?, ?it/s]\r\n100%|██████████| 1/1 [00:00<00:00, 91.35it/s]\r\n[INFO] 2021-02-12 17:45:21,556 arrow_writer: Done writing 1 examples in 51232 bytes .\r\n[INFO] 2021-02-12 17:45:21,557 abstractive_summarization: set Python list in train to PyTorch tensor\r\n[INFO] 2021-02-12 17:45:21,557 arrow_dataset: Set __getitem__(key) output type to torch for ['input_ids', 'attention_mask', 'global_attention_mask', 'labels'] columns (when key is int or slice) and don't output other (un-formated) columns.\r\n[INFO] 2021-02-12 17:45:21,557 abstractive_summarization: set Python list in test to PyTorch tensor\r\n[INFO] 2021-02-12 17:45:21,557 arrow_dataset: Set __getitem__(key) output type to torch for ['input_ids', 'attention_mask', 'global_attention_mask', 'labels'] columns (when key is int or slice) and don't output other (un-formated) columns.\r\n[INFO] 2021-02-12 17:45:21,557 abstractive_summarization: enable fp16 amp training\r\n[INFO] 2021-02-12 17:45:21,557 abstractive_summarization: file will be written to /workspace\r\n[INFO] 2021-02-12 17:45:21,882 abstractive_summarization: instantiate trainer\r\n[INFO] 2021-02-12 17:45:25,224 abstractive_summarization: start training\r\n\r\n 0%| | 0/31 [00:00<?, ?it/s]/usr/local/lib/python3.8/dist-packages/nlp/utils/py_utils.py:191: UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. This means you can write to the underlying (supposedly non-writeable) NumPy array using the tensor. You may want to copy the array to protect its data or make it writeable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:141.)\r\n return function(data_struct)\r\nTraceback (most recent call last):\r\n File \"abstractive_summarization.py\", line 407, in <module>\r\n run()\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 829, in __call__\r\n return self.main(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 782, in main\r\n rv = self.invoke(ctx)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 1066, in invoke\r\n return ctx.invoke(self.callback, **ctx.params)\r\n File \"/usr/local/lib/python3.8/dist-packages/click/core.py\", line 610, in invoke\r\n return callback(*args, **kwargs)\r\n File \"abstractive_summarization.py\", line 382, in run\r\n trainer.train()\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/trainer.py\", line 940, in train\r\n tr_loss += self.training_step(model, inputs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/trainer.py\", line 1302, in training_step\r\n loss = self.compute_loss(model, inputs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/trainer.py\", line 1334, in compute_loss\r\n outputs = model(**inputs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/parallel/data_parallel.py\", line 155, in forward\r\n outputs = self.parallel_apply(replicas, inputs, kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/parallel/data_parallel.py\", line 165, in parallel_apply\r\n return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/parallel/parallel_apply.py\", line 85, in parallel_apply\r\n output.reraise()\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/_utils.py\", line 395, in reraise\r\n raise self.exc_type(msg)\r\nRuntimeError: Caught RuntimeError in replica 0 on device 0.\r\nOriginal Traceback (most recent call last):\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/parallel/parallel_apply.py\", line 60, in _worker\r\n output = module(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 2344, in forward\r\n outputs = self.led(\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 2193, in forward\r\n encoder_outputs = self.encoder(\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 1831, in forward\r\n layer_outputs = encoder_layer(\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 907, in forward\r\n attn_outputs = self.self_attn(\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 718, in forward\r\n self_outputs = self.longformer_self_attn(\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 722, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 201, in forward\r\n attn_scores = self._sliding_chunks_query_key_matmul(\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 431, in _sliding_chunks_query_key_matmul\r\n diagonal_chunked_attention_scores = self._pad_and_transpose_last_two_dims(\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/led/modeling_led.py\", line 329, in _pad_and_transpose_last_two_dims\r\n hidden_states_padded = F.pad(\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/functional.py\", line 3552, in _pad\r\n return _VF.constant_pad_nd(input, pad, value)\r\nRuntimeError: CUDA out of memory. Tried to allocate 386.00 MiB (GPU 0; 15.78 GiB total capacity; 14.09 GiB already allocated; 162.00 MiB free; 14.42 GiB reserved in total by PyTorch)\r\n\r\n 0%| | 0/31 [00:09<?, ?it/s]\r\n```", "> Hi @stas00 , I'm trying to avoid reducing the decoder output if possible. We would love summaries that are around 200 tokens in length. Thank you for the update and merge!\r\n\r\nFor sure, we are trying to get things running first - removing OOM, then comes the optimization. \r\n\r\n> I'm noticing, if I try using deepspeed, it's now hanging on here:\r\n> \r\n> ```\r\n> [2021-02-12 16:55:53,106] [INFO] [distributed.py:46:init_distributed] Initializing torch distributed with backend: nccl\r\n\r\nlooks like distributed gets stuck there - you might be having another instance using the same port, try using a different \r\n`os.environ['MASTER_PORT']` or kill any run-away processes. \r\n\r\nwhen pre-1.8.0 pytorch crashes it often leave zombies, you have to kill those manually. this has been fixed in pt-1.8.0.\r\n\r\nThe zombies also consume gpu ram - this could be your problem too. might also help to watch nvidia-smi \r\n```\r\nwatch -n 1 nvidia-smi\r\n```\r\nto ensure you have no memory used by other programs when you start a new one.\r\n\r\nAs I mentioned earlier, you don't need DeepSpeed here, you need to figure out why your setup takes much more gpu ram than if I run the same script. Can you try a more recent pytorch version?\r\n\r\n> if I don't use deepspeed, I get\r\n\r\n> RuntimeError: CUDA out of memory. Tried to allocate 386.00 MiB (GPU 0; 15.78 GiB total capacity; 14.09 GiB already allocated; 162.00 MiB free; 14.42 GiB reserved in total by PyTorch)\r\n\r\nHere we are going in circles - if you didn't change anything in the program how would this change?\r\n\r\nTo repeat using the latest pytorch release the memory consumption appears to be much smaller than what you get - so if possible try to to upgrade it?\r\n\r\ne.g. have you tried running the same on colab? It also gives you a 16GB gpu if you use the freebie version.", "oh okay, so I tried testing this in colab\r\n```python\r\nimport datasets\r\nfrom datasets import load_dataset, load_metric\r\n\r\nimport click\r\nimport torch\r\nimport logging\r\nimport json\r\n\r\nfrom io import BytesIO\r\nimport pandas as pd\r\n\r\nimport pyarrow as pa\r\nimport pyarrow.parquet as pq\r\nfrom nlp import arrow_dataset\r\n\r\nimport os\r\nfrom transformers import (\r\n AutoTokenizer,\r\n AutoModelForSeq2SeqLM,\r\n Seq2SeqTrainer,\r\n Seq2SeqTrainingArguments,\r\n AutoTokenizer,\r\n AutoModelForSeq2SeqLM,\r\n)\r\n\r\n\r\nimport torch.utils.checkpoint\r\nfrom pynvml import *\r\n\r\n\r\n\r\n\r\nlogger = logging.getLogger(__name__)\r\nlogger.setLevel(logging.INFO)\r\nlogging.basicConfig(\r\n level=logging.INFO, format=\"[%(levelname)s] %(asctime)s %(module)s: %(message)s\"\r\n)\r\n\r\n\r\n\r\n\r\nrouge = load_metric(\"rouge\")\r\n\r\n\r\nMODEL_NAME = \"patrickvonplaten/led-large-16384-pubmed\"\r\n\r\n\r\n# ds_config = {\r\n# \"fp16\": {\r\n# \"enabled\": \"true\",\r\n# \"loss_scale\": 0,\r\n# \"loss_scale_window\": 1000,\r\n# \"hysteresis\": 2,\r\n# \"min_loss_scale\": 1\r\n# },\r\n\r\n# \"zero_optimization\": {\r\n# \"stage\": 2,\r\n# \"allgather_partitions\": \"true\",\r\n# \"allgather_bucket_size\": 1e8,\r\n# \"overlap_comm\": \"true\",\r\n# \"reduce_scatter\": \"true\",\r\n# \"reduce_bucket_size\": 1e8,\r\n# \"contiguous_gradients\": \"true\",\r\n# \"cpu_offload\": \"true\"\r\n# },\r\n\r\n# \"zero_allow_untested_optimizer\": \"true\",\r\n\r\n# \"optimizer\": {\r\n# \"type\": \"AdamW\",\r\n# \"params\": {\r\n# \"lr\": 3e-5,\r\n# \"betas\": [0.8, 0.999],\r\n# \"eps\": 1e-8,\r\n# \"weight_decay\": 3e-7\r\n# }\r\n# },\r\n\r\n# \"scheduler\": {\r\n# \"type\": \"WarmupLR\",\r\n# \"params\": {\r\n# \"warmup_min_lr\": 0,\r\n# \"warmup_max_lr\": 3e-5,\r\n# \"warmup_num_steps\": 500\r\n# }\r\n# },\r\n\r\n# \"steps_per_print\": 2000,\r\n# \"wall_clock_breakdown\": \"false\"\r\n# }\r\n\r\n# with open('ds_config.json', 'w') as fp:\r\n# json.dump(ds_config, fp)\r\n\r\n\r\n\r\n\r\nlogger.info(f\"load tokenizer using {MODEL_NAME}\")\r\ntokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)\r\n\r\nlogger.info(f\"Load {MODEL_NAME}. IMPORTANT NOTE:I'm enabling gradient checkpointing to save memory.\")\r\n# load model + enable gradient checkpointing & disable cache for checkpointing\r\nled = AutoModelForSeq2SeqLM.from_pretrained(\r\n MODEL_NAME,\r\n gradient_checkpointing=False,\r\n use_cache=False,\r\n)\r\n\r\n\r\n# max encoder length is 2048 for PubMed\r\nencoder_max_length = 2048\r\ndecoder_max_length = 64\r\nbatch_size = 2\r\n\r\n\r\n# set decoding params\r\nled.config.num_beams = 2\r\nled.config.max_length = 256\r\nled.config.min_length = 100\r\nled.config.length_penalty = 2.0\r\nled.config.early_stopping = True\r\nled.config.no_repeat_ngram_size = 3\r\n\r\n\r\n\r\ndef make_tarfile(output_filename, source_dir):\r\n with tarfile.open(output_filename, \"w:gz\") as tar:\r\n tar.add(source_dir, arcname=os.path.basename(source_dir))\r\n\r\n\r\n\r\ndef process_data_to_model_inputs(batch):\r\n # tokenize the inputs and labels\r\n inputs = tokenizer(\r\n batch[\"extractive_summary\"],\r\n padding=\"max_length\",\r\n truncation=True,\r\n max_length=encoder_max_length,\r\n )\r\n outputs = tokenizer(\r\n batch[\"reference_summary\"],\r\n padding=\"max_length\",\r\n truncation=True,\r\n max_length=decoder_max_length,\r\n )\r\n\r\n batch[\"input_ids\"] = inputs.input_ids\r\n batch[\"attention_mask\"] = inputs.attention_mask\r\n\r\n # create 0 global_attention_mask lists\r\n batch[\"global_attention_mask\"] = len(batch[\"input_ids\"]) * [\r\n [0 for _ in range(len(batch[\"input_ids\"][0]))]\r\n ]\r\n\r\n # since above lists are references, the following line changes the 0 index for all samples\r\n batch[\"global_attention_mask\"][0][0] = 1\r\n batch[\"labels\"] = outputs.input_ids\r\n\r\n # We have to make sure that the PAD token is ignored\r\n batch[\"labels\"] = [\r\n [-100 if token == tokenizer.pad_token_id else token for token in labels]\r\n for labels in batch[\"labels\"]\r\n ]\r\n\r\n return batch\r\n\r\ndef compute_metrics(pred):\r\n labels_ids = pred.label_ids\r\n pred_ids = pred.predictions\r\n\r\n pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)\r\n labels_ids[labels_ids == -100] = tokenizer.pad_token_id\r\n label_str = tokenizer.batch_decode(labels_ids, skip_special_tokens=True)\r\n\r\n rouge_output = rouge.compute(\r\n predictions=pred_str, references=label_str, rouge_types=[\"rouge2\"]\r\n )[\"rouge2\"].mid\r\n\r\n return {\r\n \"rouge2_precision\": round(rouge_output.precision, 4),\r\n \"rouge2_recall\": round(rouge_output.recall, 4),\r\n \"rouge2_fmeasure\": round(rouge_output.fmeasure, 4),\r\n }\r\n\r\n# def run():\r\nnvmlInit()\r\nh = nvmlDeviceGetHandleByIndex(0)\r\ninfo = nvmlDeviceGetMemoryInfo(h)\r\nlogger.info(f'GPU total Memory : {info.total}')\r\nlogger.info(f'GPU free Memory : {info.free}')\r\nlogger.info(f'GPU Memory used : {info.used}')\r\n\r\nlogger.info(\"create fictious train and test data\")\r\nn_recs = 1000\r\nframes = [\r\n {\"reference_summary\": [' '.join([f\"{i} I am a reference summary\"] * 200),\r\n ' '.join([\"I am another reference summary\"] * 200)],\r\n \"extractive_summary\": [' '.join([f\"{i} hello\"] * 200), ' '.join([\"goodbye\"] * 200)]} for i in range(n_recs)]\r\ntrain = pd.DataFrame(frames)\r\ntest = pd.DataFrame({\"reference_summary\": [' '.join([\"I am another reference summary\"] * 200)],\r\n \"extractive_summary\": [' '.join([\"goodbye\"] * 200)]})\r\n\r\ntrain = pa.Table.from_pandas(train)\r\ntrain = arrow_dataset.Dataset(train)\r\n\r\ntest = pa.Table.from_pandas(test)\r\ntest = arrow_dataset.Dataset(test)\r\nlogger.info(\"map train data\")\r\ntrain = train.map(\r\n process_data_to_model_inputs,\r\n batched=True,\r\n batch_size=batch_size,\r\n remove_columns=[\"reference_summary\", \"extractive_summary\"],\r\n)\r\n\r\nlogger.info(\"map test data\")\r\ntest = test.map(\r\n process_data_to_model_inputs,\r\n batched=True,\r\n batch_size=batch_size,\r\n remove_columns=[\"reference_summary\", \"extractive_summary\"],\r\n\r\n)\r\n\r\nlogger.info(\"set Python list in train to PyTorch tensor\")\r\ntrain.set_format(\r\n type=\"torch\",\r\n columns=[\"input_ids\", \"attention_mask\", \"global_attention_mask\", \"labels\"],\r\n)\r\n\r\nlogger.info(\"set Python list in test to PyTorch tensor\")\r\ntest.set_format(\r\n type=\"torch\",\r\n columns=[\"input_ids\", \"attention_mask\", \"global_attention_mask\", \"labels\"],\r\n)\r\n\r\nlogger.info(\"enable fp16 amp training\")\r\nlogger.info(f\"file will be written to {os.getcwd()}\")\r\n\r\n\r\n#define env variables required for training\r\nos.environ['MASTER_ADDR'] = 'localhost'\r\nos.environ['MASTER_PORT'] = '9994'\r\nos.environ['RANK'] = \"0\"\r\nos.environ['LOCAL_RANK'] = \"0\"\r\nos.environ['WORLD_SIZE'] = \"1\"\r\n\r\ncheckpoint_dir_path = \"/mnt/summarization_checkpoints\"\r\ntraining_args = Seq2SeqTrainingArguments(\r\n predict_with_generate=True,\r\n evaluation_strategy=\"steps\",\r\n per_device_train_batch_size=batch_size,\r\n per_device_eval_batch_size=batch_size,\r\n fp16=True,\r\n output_dir=checkpoint_dir_path,\r\n logging_steps=5,\r\n eval_steps=10,\r\n save_steps=10,\r\n save_total_limit=1,\r\n gradient_accumulation_steps=4,\r\n num_train_epochs=1,\r\n # deepspeed=\"ds_config.json\"\r\n)\r\n\r\n# training_args._setup_devices\r\n\r\nos.makedirs(checkpoint_dir_path, exist_ok=True)\r\nlogger.info(\"instantiate trainer\")\r\ntrainer = Seq2SeqTrainer(\r\n model=led,\r\n tokenizer=tokenizer,\r\n args=training_args,\r\n compute_metrics=compute_metrics,\r\n train_dataset=train,\r\n eval_dataset=test,\r\n)\r\n\r\n\r\n\r\nlogger.info(\"start training\")\r\ntrainer.train()\r\n```\r\n\r\n\r\nand setting the decoder max length to 64 but it's still giving me memory issues:\r\n\r\nhttps://colab.research.google.com/drive/1IN1tHkey0It_LWZHvOuCbbcgtglGizw4?usp=sharing", "This is great, so that we can work on the same environment. I will work on it later today and hopefully find the culprit. I will keep you posted, @mmoya01 ", "I started working on it but haven't figured it out yet - colab is not very friendly to debug OOM - not better than running a script - have to restart it all the time - will continue tomorrow - hopefully will have a resolution soon.\r\n\r\n", "Hi @stas00 thank you for the update and for looking into this", "OK, so I experimented a bit and sat with various profilers to make sense out of it all, since there are many different nuances to understand.\r\n\r\nHere is what I have to share with you.\r\n\r\n1. DeepSpeed's primary use is for distributed training (multi-gpu), and while it can shine on a single gpu - it needs general RAM - which collab doesn't have much of - you can't do anything serious with 12GB of RAM for the whole vm. It just kept on crashing. If on your original setup you have much more RAM then it's definitely worth trying to deploy DeepSpeed.\r\n\r\n I have several extra things to experiment with in the DeepSpeed-land hopefully in the next few days which may help a bit, but since I haven't tried it yet, I can't tell.\r\n\r\n2. Now let's look at reality - you took a notebook that was tuned to fit into the available 15GB gpu and swapped in a model that is ~3x bigger. So there is not much you can do given the RAM limitation.\r\n\r\nI did multiple experiments and found this to fit very snugly - i.e. a few bytes away from OOM:\r\n```\r\nencoder_max_length = 2048\r\ndecoder_max_length = 64\r\n\r\nbatch_size = 1\r\ngradient_accumulation_steps=8\r\nGPU Memory used : 15802040320\r\n```\r\n\r\nSo your effective batch is 8, but `decoder_max_length` is unsatisfactory. I am aware of that.\r\n\r\nAlso I added to the notebook `ipyexperiments` which memory profiles each cell automatically for you. So that you can easily see what's happening w/o needing to manually add printouts.\r\n\r\nhttps://colab.research.google.com/drive/1rEspdkR839xZzh561OwSYLtFnnKhQdEl?usp=sharing\r\n\r\nNote that it reports the memory at current and also the delta that was consumed and peaked. So if after training it shows a lot more memory still left, it's after clearing the cache - so if you take the used memory + peaked delta you will get the total peak memory the program reached during that cell.\r\n\r\nRunning the same experiments on a larger gpu, they all surpass 15GB peak memory with bs=2. In one of my very first reports I suggested that I get much less memory used on my larger card, but I was wrong, I didn't account for the peak memory in my first measurements.\r\n\r\nJust in case you are not familiar with the term - Peak memory - is when a program consumes some memory temporarily and then releases it, so the reported total is less.\r\n\r\n3. Research if perhaps someone has made a distilled model of the same, in which case it'll be less of everything and probably fit better. I see other models finetuned on pubmed on the datasets hub - I don't know if they fit your needs.\r\n\r\n4. In your experiments be aware that colab is terrible at gpu memory management, and doesn't quite free memory, so it's full restart on each experiment :( I'm mentioning that so that you won't be getting false negatives if you decided to re-run the same cell that trains.\r\n\r\nAs I mentioned earlier there is at least one more thing I hope to try in the next few days. If I succeed I will send you an update.\r\n", "One other thing you may want to try is fp16 training. I have no idea how LED takes to that.\r\n\r\n```\r\npip install apex\r\n```\r\n \r\n```\r\ntraining_args = Seq2SeqTrainingArguments(\r\n [...]\r\n fp16=True,\r\n fp16_backend=\"apex\",\r\n fp16_opt_level=\"O3\",\r\n```\r\n\r\nThis will use significantly less memory, but your training may or may not converge.\r\n\r\nIt's very likely that you will want to keep batch norm at fp32 though - but the current trainer doesn't have a way to enable that from the user side. So either you need to change the trainer source code\r\n```\r\n# trainer.py\r\n def _wrap_model(self, model, training=True):\r\n # Mixed precision training with apex (torch < 1.6)\r\n if self.use_apex and training:\r\n model, self.optimizer = amp.initialize(model, self.optimizer, opt_level=self.args.fp16_opt_level, keep_batchnorm_fp32=True)\r\n```\r\nI added a new argument `keep_batchnorm_fp32=True` there.\r\n\r\nor perhaps it's easier to monkey patch `amp` in your script/notebook:\r\n\r\n```\r\nfrom apex import amp\r\norig_amp_init = amp.initialize\r\ndef new_amp_init(model, optimiser, **kwargs):\r\n return orig_amp_init(model, optimiser, keep_batchnorm_fp32=True, **kwargs)\r\namp.initialize = new_amp_init\r\n\r\ntrainer = ...\r\n```\r\n\r\nor the same can be done in a simpler way with `partial`:\r\n```\r\nfrom functools import partial\r\nfrom apex import amp\r\namp.initialize = partial(amp.initialize, keep_batchnorm_fp32=True)\r\n\r\ntrainer = ...\r\n```\r\njust don't re-run this cell more than once per session\r\n\r\n**edit:** transformers doesn't actually use batchnorm so that 2nd part was irrelevant.\r\n\r\nTo understand exactly what I proposed see: https://nvidia.github.io/apex/amp.html#o3-fp16-training\r\n", "ok, figured it out - I suggested for you try to disable the gradient checkpointing in the context of being unable to use Deepspeed, but I didn't think of asking you to restore this config...\r\n\r\nSo enable `from_pretrained(MODEL_NAME, gradient_checkpointing=True,...`\r\n\r\nAnd voila, this config works just fine:\r\n```\r\nencoder_max_length = 2048\r\ndecoder_max_length = 256\r\nbatch_size = 4\r\n```\r\n\r\nYou can go for even larger length, it should have a very small impact. And I think your batch size can now be even larger, so that you can remove `gradient_accumulation_steps` if wanted - or reduce it.\r\n\r\n\r\nI updated the notebook, so you can see it working:\r\nhttps://colab.research.google.com/drive/1rEspdkR839xZzh561OwSYLtFnnKhQdEl?usp=sharing", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Hi @ioana-blue ,\r\n\r\n`max_seq_length` is the naming \"convention\" when talking about the tokenization side, e.g. when you tokenize your tweets they will be converted to ids and normally padded to a `max_seq_length`.\r\n\r\n`max_position_embeddings` is the naming convention when talking about pre-training the model, so e.g. BERT has seen 512 subtokens during pre-training phase.\r\n\r\nSo `max_seq_length` should be less or equal `max_position_embeddings`.\r\n\r\nIn your case it seems that the model has seen 130 subtokens during pre-training phase (which is ok, because tweets usually are much shorter than 512 subtokens). \r\n\r\nCould you check your tokenization part and the number of subtokens that you're later passing to the model :thinking: ", "It would also help if you could paste the tokenization part (e.g. converting plain text/tweets into model ids) here, so we can have a look into it!", "Ok, I thought that was one potential issue. So I ran with `max_seq_length` of 100 and I still get the same problem. I'll probably ask for a feature request to print an error when trying to run with a `max_seq_length` that is higher than `max_possition_embeddings` (I used to run it with 512 for `max_seq_length` and there is no complain). \r\n\r\nI'm using slightly modified version of the glue example in the code, so I didn't modify any of the tokenization part. The only thing that I added is the data processors/loaders. \r\n\r\nLet me know if I could provide any more info to help debug this issue. Greatly appreciate your help!\r\n", "I just wanted to reproduce the error message with an example after my :pizza: , but it is working with the GLUE example:\r\n\r\n```bash\r\npython3 run_glue.py \\\r\n --model_name_or_path vinai/bertweet-base \\\r\n --task_name $TASK_NAME \\\r\n --do_train \\\r\n --do_eval \\\r\n --max_seq_length 128 \\\r\n --per_device_train_batch_size 32 \\\r\n --learning_rate 2e-5 \\\r\n --num_train_epochs 3 \\\r\n --use_fast False \\\r\n --output_dir /tmp/$TASK_NAME/\r\n```\r\n\r\nImportant argument is to pass `--use_fast False` to avoid an error message. I set `TASK_NAME` to the `wnli` task.\r\n\r\nCould you specify what version of Transformers you're using :thinking: I'm using a 4.3 version (d5888ef0ab949ec82ed4768556c2b2743e3ca1df).", "Is it also possible that you paste the trainer output that shows e.g.:\r\n\r\n```bash\r\n02/04/2021 20:59:33 - INFO - __main__ - Sample 281 of the training set: {'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 'idx': 281, 'input_ids': [0, 1717, 966, 9, 329, 2\r\n125, 24, 6, 52562, 7, 42, 58, 8215, 29, 41118, 7939, 4, 2, 2, 2125, 8215, 29, 41118, 7939, 4, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1], 'label': 0, 'sentence1': \"Paul tried to call George on the phone, but he wasn't successful.\", 'sentence2': \"George wasn't su\r\nccessful.\", 'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0]}.\r\n```\r\n\r\n", "Oh, and could you also pass `--overwrite_cache` to the training script, this is really helpful espec. when you experiment with different sequence lengths 😀", "Thanks for your help! \r\n\r\nI tried the code with 3.4 and 4.2, similar behavior. \r\n\r\nThe encoding look fine to me:\r\n\r\n```\r\n02/04/2021 16:15:57 - INFO - util_processors - *** Example ***\r\n02/04/2021 16:15:57 - INFO - util_processors - guid: test-1178818409812746240_twitter\r\n02/04/2021 16:15:57 - INFO - util_processors - features: InputFeatures(input_ids=[0, 10584, 56843, 241, 66, 103, 6, 289, 1389, 32, 38, 97, 11, 23465, 72, 618, 27658, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], attention_mask=[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], token_type_ids=[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], label=0)\r\n02/04/2021 16:15:57 - INFO - util_processors - *** Example ***\r\n02/04/2021 16:15:57 - INFO - util_processors - guid: test-19346774_gab\r\n02/04/2021 16:15:57 - INFO - util_processors - features: InputFeatures(input_ids=[0, 52, 112, 52, 37, 1621, 11, 8812, 15, 634, 5230, 37, 116, 45, 96, 11, 3559, 25, 37, 56, 140, 28748, 701, 24, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], attention_mask=[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], token_type_ids=[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], label=0)\r\n02/04/2021 16:15:57 - INFO - util_processors - *** Example ***\r\n02/04/2021 16:15:57 - INFO - util_processors - guid: test-1165819983701643266_twitter\r\n02/04/2021 16:15:57 - INFO - util_processors - features: InputFeatures(input_ids=[0, 557, 31, 39, 94, 11, 397, 31, 1844, 46154, 13, 1190, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], attention_mask=[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], token_type_ids=[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], label=1)\r\n```\r\n\r\n`--use_fast` doesn't work for me (no such param). \r\n\r\nI always use overwrite cache 👍 \r\n\r\nThis is my command line:\r\n\r\n```\r\npython ../models/jigsaw/tr-3.4//run_puppets.py --model_name_or_path vinai/bertweet-base --task_name binary_hatex --do_train --do_eval --do_logits --do_predict --data_dir /dccstor/redrug_ier/ioana/fairnlp/toxi-data/hatex/processed/ --max_seq_length 128 --per_device_train_batch_size 32 --learning_rate 2e-5 --num_train_epochs 1 --output_dir /dccstor/redrug_ier/ioana/fairnlp/toxi-data/hatex/run-results/runpred_e1_binary_hatex_bertweet_20210204_12_03_58 --cache_dir /dccstor/redrug_ier/ioana/fairnlp/toxi-data/hatex/run-results/runpred_e1_binary_hatex_bertweet_20210204_12_03_58/cache/ --overwrite_cache --logging_steps 10000 --save_steps 200000\r\n```\r\nI implemented some command line args for predicting and printing logits, etc., but it doesn't get there, the problem is in the training. \r\n\r\nIf I feel adventurous, I will probably try to step through the training and see if I notice any issues. It looks like an out of bounds indexing somewhere. ", "I'm going to try the run_glue on my side to see if I can reproduce your successful run. \r\n", "Ayayay. I got a successful run with 3.4 and the command line above (my own code, I mean). Strange. ", "Yep, I can confirm running with different seq size and it works. I think what happened was the following:\r\n- Initially I was running with a seq size that was too large. \r\n- I upgraded the transformers to 4.2 and also realized the seq size problem. I started using smaller seq sizes, but there was a problem. I'm guessing the problem comes from some backward-compatibility (my code was inspired by the sample code from version 3.4; I'm guessing something changed that breaks the code with 4.2)\r\n- Once I went back to 3.4 AND small seq size, it worked. I'll open a feature request. I don't think runs should be allowed with `max_seq_size > max_position_embeddings`\r\n\r\nThanks for your help, much appreciated. I'm closing this one. \r\n" ]

[]

[]

[ "Good question! That's how the original `roberta-base` model was pretrained. I recommend you stick to the way the model was pretrained in order to obtain best performance.\r\n\r\nYou can find the [roberta paper here](https://arxiv.org/pdf/1907.11692.pdf). I believe the section 4.2 contains information regarding model inputs.", "Alright, thanks!" ]

[ { "id": 2796628563, "node_id": "MDU6TGFiZWwyNzk2NjI4NTYz", "url": "https://api.github.com/repos/huggingface/transformers/labels/WIP", "name": "WIP", "color": "234C99", "default": false, "description": "Label your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progress" } ]

[]

[ "I'm in favor of this! It's also a bit annoying to me that downloading takes that much time and I think for people that try out a bunch of different checkpoitns to decide on which model they want to use, a 60% speed-up in downloading would be very nice.", "A related issue is that t5-11b is actually too large to be served by Cloudfront (they have a limit of 20GB for a single file) so we have to fallback to serving using S3, which is way, way slower. (up to a few MB/s depending on where you are, vs. basically saturating your downlink when from Cloudfront)\r\n\r\nIf large models are here to stay, then we probably need to support the **splitting of models** in `save_pretrained`/`from_pretrained`\r\n\r\nalso cc @mfuntowicz @n1t0 ", "> If large models are here to stay, then we probably need to support the splitting of models in save_pretrained/from_pretrained\r\n\r\nIf we go with compression why not do the normal volumes of whatever common compressor tool we choose, as in:\r\n\r\n```\r\npytorch.bin.rar00\r\npytorch.bin.rar01\r\n```\r\n\r\nSo you kill 2 birds at the same time, get the compression and the splitting.\r\n\r\nAgain, the user doesn't need to do anything. the compression and splitting can be triggered upon the upload.\r\n\r\ne.g. with tar.gz:\r\n\r\non upload:\r\n```\r\ntar cvzf - pytorch_model.bin | split -b 10G - pytorch_model.bin.tar.gz.\r\n```\r\nwhich should give:\r\n```\r\npytorch_model.bin.tar.gz.aa\r\npytorch_model.bin.tar.gz.ab\r\npytorch_model.bin.tar.gz.ac\r\n```\r\non download:\r\n```\r\ncat pytorch_model.bin.tar.gz.a* | tar xzvf -\r\n```\r\n\r\nwell, we don't need tar here - it's just one file. so just gzip would be enough.\r\n\r\nJust need to choose which compression is good enough, and doesn't take too long to decompress - e.d. don't use the highest compression possible and 100% available on all clients - so gzip and uncompress for sure, 7zip/rar can't be trusted to have, but if there is a python client that can handle it, it may work anywhere?", "Operationally I'm wondering if instead of doing it at rest in an async way (which might prove difficult to easily scale to a much larger number of models), we should probably handle this in the `save_pretrained` (which means users will upload their models already in the supported format)", "Is the intention to upload both compressed and uncompressed or just the former?\r\n\r\nI propose to manage compression/decompression transparently on the server side and leave everything as is on the client side (other than download of the compressed version). \r\n\r\nHere are some quick pros/cons for 3 different scenarios I see.\r\n\r\n### 1. Having only the compressed version on the client side:\r\n\r\nCons:\r\n1. Will create a constant overheard of compression on `save_pretrained` and check pointing\r\n2. Will create a constant wasteful overhead of decompression during `from_pretrained`\r\n3. Should the max split size change - how do you tell the users that they all need to update their repo?\r\n\r\nPros:\r\n1. Will make the upload faster\r\n\r\n\r\n### 2. Having only the decompressed version on the client side:\r\n\r\nCons:\r\n1. More expensive upload\r\n\r\nPros:\r\n1. Everything else is simple\r\n\r\n\r\n\r\n### 3. Having both versions on the client side:\r\n\r\nThis one is like case 2, but with an one additional change in each up/down direction:\r\n\r\nPros: same as in case 2\r\n\r\nCons:\r\n1. Could be confusing to the user during upload if they need to upload only the compressed files\r\n2. More to upload \r\n\r\nExtra notes:\r\n1. Need to make sure that the decompression will happen once upon download and not on every `from_pretrained()` call. \r\n\r\n\r\nPlease feel free to edit this post directly as I'm sure I've missed some aspects.", "I agree that 2) is simpler, and is in line with the goal of keeping things simple for the user. As a PyTorch user I would prefer seeing my files in the native PyTorch format rather than a compressed format I don't know about, on which I'll need to apply additional pre-processing before using it in a custom model. Especially since we've seen users use `torch.load` instead of `from_pretrained` in the past.", "Should we keep this one alive? Is this on someone's TODO list?", "i think checkpoint-splitting (#13548) is going to be a better/more future-proof solution than compression (on top of already compressed binary files) where the size delta is going to be rather minimal\r\n\r\nSo I'd vote to close this issue and focus on #13548", "Sounds good, Julien. Let's close this one." ]

[]

[]

[ "> a) tf.tile should be used instead of tf.broadcast_to &\r\n\r\nThere are two reasons for this, the first one is because `broadcast_to` does `reshape` + `tile`, here we don't need to reshape, just `tile` is enough. The second reason is that `broadcast_to` is not compliant with ONNXRuntime.\r\n\r\n> b) why we cannot simply use the shape of attn_probs since we apply the mask on attn_probs itself? So we know that shape_list(masked_index) == shape_list(attn_probs)\r\n\r\nThis part is a bit tricky to explain. The issue here was that `attn_probs` was not always the same shape, if `is_global_attn` is True, then the shape of `attn_probs` is `[batch_size, seq_len, self.num_heads, self.one_sided_attn_window_size * 2 + max_num_global_attn_indices + 1]`, while if it equals False its shape is `[batch_size, seq_len, self.num_heads, self.one_sided_attn_window_size * 2 + 1]`. Now, because the shape is never potentially the same during the execution when run in graph mode, the pre-computed shape for `attn_probs` by the TF tracing was `[batch_size, seq_len, self.num_heads, variable]`, where `variable` cannot be computed. The consequence of this was that `attn_probs` had never the proper shape at the end and creates a conflict in the `tf.where`. To solve this we had to also create a mask of a fixed shape that depends on `is_global_attn`.\r\n\r\nI don't know if it is clear enough or not. Don't hesitate to tell me if there is something you don't get.", "> > a) tf.tile should be used instead of tf.broadcast_to &\r\n> \r\n> There are two reasons for this, the first one is because `broadcast_to` does `reshape` + `tile`, here we don't need to reshape, just `tile` is enough. The second reason is that `broadcast_to` is not compliant with ONNXRuntime.\r\n> \r\n> > b) why we cannot simply use the shape of attn_probs since we apply the mask on attn_probs itself? So we know that shape_list(masked_index) == shape_list(attn_probs)\r\n> \r\n> This part is a bit tricky to explain. The issue here was that `attn_probs` was not always the same shape, if `is_global_attn` is True, then the shape of `attn_probs` is `[batch_size, seq_len, self.num_heads, self.one_sided_attn_window_size * 2 + max_num_global_attn_indices + 1]`, while if it equals False its shape is `[batch_size, seq_len, self.num_heads, self.one_sided_attn_window_size * 2 + 1]`. Now, because the shape is never potentially the same during the execution when run in graph mode, the pre-computed shape for `attn_probs` by the TF tracing was `[batch_size, seq_len, self.num_heads, variable]`, where `variable` cannot be computed. The consequence of this was that `attn_probs` had never the proper shape at the end and creates a conflict in the `tf.where`. To solve this we had to also create a mask of a fixed shape that depends on `is_global_attn`.\r\n> \r\n> I don't know if it is clear enough or not. Don't hesitate to tell me if there is something you don't get.\r\n\r\nThanks for the explanation - just tried it out and cool to see that your change fixes the test!", "The entire list of slow tests are ok!", "@sgugger Feel free to merge if it looks ok for you!" ]

[]

[]

[ "I just ran `run.sh`, but did not see this error. Could you maybe post the full stack trace?", "Hi @patil-suraj, thanks for replying. The full stack trace is posted below.\r\n```\r\nTraceback (most recent call last):\r\n File \"run_origin.py\", line 437, in <module>\r\n main()\r\n File \"run_origin.py\", line 310, in main\r\n load_from_cache_file=not data_args.overwrite_cache,\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/datasets/dataset_dict.py\", line 303, in map\r\n for k, dataset in self.items()\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/datasets/dataset_dict.py\", line 303, in <dictcomp>\r\n for k, dataset in self.items()\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/datasets/arrow_dataset.py\", line 1259, in map\r\n update_data=update_data,\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/datasets/arrow_dataset.py\", line 157, in wrapper\r\n out: Union[\"Dataset\", \"DatasetDict\"] = func(self, *args, **kwargs)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/datasets/fingerprint.py\", line 158, in wrapper\r\n self._fingerprint, transform, kwargs_for_fingerprint\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/datasets/fingerprint.py\", line 105, in update_fingerprint\r\n hasher.update(transform_args[key])\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/datasets/fingerprint.py\", line 57, in update\r\n self.m.update(self.hash(value).encode(\"utf-8\"))\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/datasets/fingerprint.py\", line 53, in hash\r\n return cls.hash_default(value)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/datasets/fingerprint.py\", line 46, in hash_default\r\n return cls.hash_bytes(dumps(value))\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/datasets/utils/py_utils.py\", line 389, in dumps\r\n dump(obj, file)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/datasets/utils/py_utils.py\", line 361, in dump\r\n Pickler(file, recurse=True).dump(obj)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 409, in dump\r\n self.save(obj)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 476, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/datasets/utils/py_utils.py\", line 556, in save_function\r\n obj=obj,\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 610, in save_reduce\r\n save(args)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 476, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 751, in save_tuple\r\n save(element)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 476, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 751, in save_tuple\r\n save(element)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 476, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/dill/_dill.py\", line 1129, in save_cell\r\n pickler.save_reduce(_create_cell, (f,), obj=obj)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 610, in save_reduce\r\n save(args)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 476, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 736, in save_tuple\r\n save(element)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 521, in save\r\n self.save_reduce(obj=obj, *rv)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 605, in save_reduce\r\n save(cls)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 476, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/dill/_dill.py\", line 1315, in save_type\r\n obj.__bases__, _dict), obj=obj)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 610, in save_reduce\r\n save(args)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 476, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 751, in save_tuple\r\n save(element)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 476, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/dill/_dill.py\", line 902, in save_module_dict\r\n StockPickler.save_dict(pickler, obj)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 821, in save_dict\r\n self._batch_setitems(obj.items())\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 847, in _batch_setitems\r\n save(v)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 476, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/dill/_dill.py\", line 902, in save_module_dict\r\n StockPickler.save_dict(pickler, obj)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 821, in save_dict\r\n self._batch_setitems(obj.items())\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 847, in _batch_setitems\r\n save(v)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 521, in save\r\n self.save_reduce(obj=obj, *rv)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 634, in save_reduce\r\n save(state)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 476, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 736, in save_tuple\r\n save(element)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 476, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/dill/_dill.py\", line 902, in save_module_dict\r\n StockPickler.save_dict(pickler, obj)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 821, in save_dict\r\n self._batch_setitems(obj.items())\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 847, in _batch_setitems\r\n save(v)\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/pickle.py\", line 476, in save\r\n f(self, obj) # Call unbound method with explicit self\r\n File \"/Users/bytedance/opt/anaconda3/lib/python3.6/site-packages/dill/_dill.py\", line 1148, in save_dictproxy\r\n raise ReferenceError(\"%s does not reference a class __dict__\" % obj)\r\nReferenceError: {'help': 'The name of the task (ner, pos...).'} does not reference a class __dict__\r\n\r\n```", "looks like you are running your own script `run_origin.py`, so the issue is not with `run_ner.py`", "Hey @gongel ,\r\n\r\ncould you confirm that you are really using latest 4.3 version :thinking: \r\n\r\nFor me the example is working with`run_ner.sh`.", "@patil-suraj, I just renamed run_ner.py to run_origin.py.", "Hi @stefan-it , Yes\r\n```\r\n(base) C02D925LMD6R:transformers gong$ pip show transformers\r\nName: transformers\r\nVersion: 4.3.0.dev0\r\nSummary: State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch\r\nHome-page: https://github.com/huggingface/transformers\r\nAuthor: Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Sam Shleifer, Patrick von Platen, Sylvain Gugger, Google AI Language Team Authors, Open AI team Authors, Facebook AI Authors, Carnegie Mellon University Authors\r\nAuthor-email: [email protected]\r\nLicense: Apache\r\nLocation: /Users/bytedance/transformers/src\r\nRequires: filelock, numpy, packaging, regex, requests, sacremoses, tokenizers, tqdm, dataclasses, importlib-metadata\r\nRequired-by: sentence-transformers\r\n```", "Could you run the `run_ner.py` script using master? as Stefan said your version seems old.", "I tried 4.3.0.dev0, 4.4.0.dev0 and 4.2.2 .\r\nThey all didn't work. 😭", "You might have an issue if your version of `datasets` is old. In any case, the whole serialization error is linked to the datasets library, so pinging @lhoestq in case he has a better idea :-)", "Hi ! \r\nCan you try updating `dill` ?\r\nIt looks like [one of their issues](https://github.com/uqfoundation/dill/issues/312) from 2019 that has been fixed now.", "Thank you, @lhoestq @sgugger @patil-suraj @stefan-it \r\nIt works by updating ```dill``` from ```0.2.9``` to ```0.3.3```." ]

[ { "id": 1843244711, "node_id": "MDU6TGFiZWwxODQzMjQ0NzEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/New%20model", "name": "New model", "color": "fbca04", "default": false, "description": "" } ]

[]

[ "Hey @guillaume-be, glad to know that you are offering Rust implementation of this model :)\r\n\r\nThere's no license currently, but I'll add MIT license to this model. ", "Hello @patil-suraj ,\r\nCould you please share an update on this issue? \r\nThank you!", "Hi @guillaume-be \r\nI just added an MIT license https://huggingface.co/valhalla/longformer-base-4096-finetuned-squadv1/blob/main/LICENSE", "@patil-suraj also referenced it from your model card's YAML: https://huggingface.co/valhalla/longformer-base-4096-finetuned-squadv1/commit/1ad74ed17896eb4d3a314b1acedefbfc184cc582 so that it's reflected in the model tags etc.", "Thanks Julien ! ", "This is great thank you!" ]

[]

[]

[ "Also interested in this question! ", "Hmm, Wav2Vec2 is still a very recent addition and I don't have a good idea on an ETA for full ONNX support. However, I think your error above is due to the input that's passed to `Wav2Vec2Tokenizer` being a string instead of a speech input. So in order to make the conversion work, you will have to tweak the script `convert_graph_to_onnx` yourself a bit for Wav2Vec2 - I think the only different should be that instead of passing it `\"This is a sample output\"` you should pass it a 1D float array.", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Hi!\r\nvery interested in this question! did anyone managed to make it work ? \r\n", "Hey @OthmaneJ,\r\n\r\nThink @ccoreilly managed to get it to work here: https://github.com/ccoreilly/wav2vec2-service/blob/master/convert_torch_to_onnx.py", "@patrickvonplaten thanks! 👌", "hi @patrickvonplaten \r\n is there any way to transform mms asr model to onnx?\r\nif yes, how?\r\nthank you very much!" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Hello, running the code does not raise an error:\r\n\r\n```py\r\n>>> from transformers import DistilBertTokenizer, DistilBertModel\r\n... import torch\r\n... \r\n... tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-german-cased')\r\n... model = DistilBertModel.from_pretrained('distilbert-base-german-cased')\r\nDownloading: 100%|██████████| 240k/240k [00:00<00:00, 690kB/s]\r\nDownloading: 100%|██████████| 464/464 [00:00<00:00, 199kB/s]\r\nDownloading: 100%|██████████| 270M/270M [00:07<00:00, 36.6MB/s]\r\n```\r\n\r\nPlease put the error in your issue, otherwise it's impossible to help you.", "Sorry, @LysandreJik , it was a copy paste error. Added it to the issue now:\r\n```\r\nTraceback (most recent call last):\r\n File \"/home/tarask/Desktop/Work/Code/Git/probabilistic-gesticulator/my_code/data_processing/annotations/encode_text.py\", line 5, in <module>\r\n model = DistilBertModel.from_pretrained('distilbert-base-german-cased')\r\n File \"/home/tarask/anaconda3/envs/gesture_flow/lib/python3.7/site-packages/transformers/modeling_utils.py\", line 1034, in from_pretrained\r\n model = cls(config, *model_args, **model_kwargs)\r\n File \"/home/tarask/anaconda3/envs/gesture_flow/lib/python3.7/site-packages/transformers/models/distilbert/modeling_distilbert.py\", line 419, in __init__\r\n self.embeddings = Embeddings(config) # Embeddings\r\n File \"/home/tarask/anaconda3/envs/gesture_flow/lib/python3.7/site-packages/transformers/models/distilbert/modeling_distilbert.py\", line 89, in __init__\r\n n_pos=config.max_position_embeddings, dim=config.dim, out=self.position_embeddings.weight\r\n File \"/home/tarask/anaconda3/envs/gesture_flow/lib/python3.7/site-packages/transformers/models/distilbert/modeling_distilbert.py\", line 76, in create_sinusoidal_embeddings\r\n out[:, 0::2] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))\r\nRuntimeError: a view of a leaf Variable that requires grad is being used in an in-place operation.\r\n```", "Ah, I see, thanks! This is because you're using the latest PyTorch 1.8+. We patched this issue yesterday in https://github.com/huggingface/transformers/pull/9917, if you install from source you shouldn't see this error anymore.", "Ah, I see, thanks! So I should install `transformers` from the source, right?", "Reverting Pytorch to 1.7 also fixed this error. \r\nThank you so much for such prompt help @LysandreJik !", "Glad you could solve it!" ]

[]

[]

[ "The failing model templates test is untrue. The second model template test (that succeeds) is true. I'll fix the github-actions YAML in a second PR.", "No worries!" ]

[]

[]

[ "I'll have a look at the git issue in the evening", "Thanks for the PR, a few quick comments:\r\n\r\n> This is also why I defined an additional ModelOutput called BaseModelOutputWithCrossAttentionsAndIntermediateHiddenStates, which adds intermediate activations of the decoder layers as output.\r\n\r\nI will strongly object to a name that long as a matter of principle :sweat_smile: But jsut so I understand what it adds, are those intermediate activations of the decoder layers not in the `hidden_states` attribute already?\r\n\r\n> I wonder whether DETR needs a tokenizer.\r\n\r\nI think the \"tokenization\" file (we can rename it if we want) should exist and contain the `NestedTensor` class and the utilities for padding. Like Wav2Vec2 Patrick added recently, the tokenizer call would only take care of the padding, resizing to a max size (if given) and normalizing. The tokenizer could also have a method that loads the images from a filename and accept in its call one or a list of decoded images (as np.array or tensor) or one or a list of filenames (and decode them with PIL for instance). \r\nIt could also have a `decode` method which would in this case do the rescale of bounding boxes and map label IDs to label names, so it's easier to then plot the results.\r\n\r\nThe inputs of the models should completely be renamed to reflect the types of objects expected (so probably `pixel_values` and `pixel_mask` would be better names than `input_ids` etc) and the tokenizer call should output a dictionary with those names as keys (so we can use the usual API of feeding directly to the model the output of the tokenizer).\r\n\r\nI imagine something like as a final easy API:\r\n```\r\ninputs = tokenizer([filename1, filename2])\r\noutputs = model(**inputs)\r\npreocessed_outputs = tokenizer.decode(outputs)\r\n```", "> will strongly object to a name that long as a matter of principle 😅 But jsut so I understand what it adds, are those intermediate activations of the decoder layers not in the `hidden_states` attribute already?\r\n\r\nYes, the intermediate activations are the hidden states of the decoder layers, each of them followed by a `LayerNorm`. I agree that the name is too long 😅 \r\n\r\n> I think the \"tokenization\" file (we can rename it if we want) should exist and contain the `NestedTensor` class and the utilities for padding. Like Wav2Vec2 Patrick added recently, the tokenizer call would only take care of the padding, resizing to a max size (if given) and normalizing. The tokenizer could also have a method that loads the images from a filename and accept in its call one or a list of decoded images (as np.array or tensor) or one or a list of filenames (and decode them with PIL for instance).\r\n\r\nI've created a first draft of `DetrTokenizer` as you requested. The API looks as follows:\r\n\r\n```\r\nfrom PIL import Image\r\nimport requests\r\nfrom transformers import DetrTokenizer\r\n\r\nurl = 'http://images.cocodataset.org/val2017/000000039769.jpg'\r\nimage = Image.open(requests.get(url, stream=True).raw)\r\n\r\ntokenizer = DetrTokenizer() # later, this is gonna be .from_pretrained(\"facebook/detr-resnet-50\")\r\nencoding = tokenizer(image)\r\n```\r\nCurrently it accepts PIL images, Numpy arrays and PyTorch tensors. The `encoding` (which is a `BatchEncoding`) has 2 keys, namely `pixel_values` and `pixel_mask`. You can call the tokenizer with the following parameters:\r\n* `resize`: whether to resize images to a given size.\r\n* `size`: arbitrary integer to which you want to resize the images \r\n* `max_size`: the largest size an image dimension can have (otherwise it's capped). \r\n* `normalize`: whether to apply mean-std normalization.\r\n\r\nAn additional complexity with object detection is that if you resize images, the annotated bounding boxes must be resized accordingly. So if you want to prepare data for training, you can also pass in annotations in the `__call__` method of `DetrTokenizer`. In that case, the `encoding` will also include a key named `labels`.", "Resolution of the git issue: https://github.com/huggingface/transformers/pull/10119", "> Currently it accepts PIL images, Numpy arrays and PyTorch tensors.\r\n\r\nPretty cool! Can we strings or pathlib.Paths too?\r\n\r\nAbout the general API, not sure if we should inherit from `PreTrainedTokenizer` since the `from_pretrained`/`save_pretrained` methods are not going to work. Wdyt @LysandreJik ? This is also not a tokenizer, more like an `AnnotatedImagePreProcessor` or something like that.\r\n\r\n> An additional complexity with object detection is that if you resize images, the annotated bounding boxes must be resized accordingly. So if you want to prepare data for training, you can also pass in annotations in the __call__ method of DetrTokenizer\r\n\r\nYes, this is expected. Maybe we could create a new type a bit like `BatchEncoding` that groups together the image (on all possible formats, string, PIL, array, tensor) with its annotation, so we can then just pass that object (or a list of those objects) to the tokenizer. What do you think?", "> Pretty cool! Can we strings or pathlib.Paths too?\r\n> \r\n> About the general API, not sure if we should inherit from `PreTrainedTokenizer` since the `from_pretrained`/`save_pretrained` methods are not going to work. Wdyt @LysandreJik ? This is also not a tokenizer, more like an `AnnotatedImagePreProcessor` or something like that.\r\n\r\nSure, it's best to make a similar API for ViT, right? (And more Transformer-based image models that will come after that). I've heard some people are working on ViT? To be fair, I could write a conversion script for ViT if you want, I see it's available in timm.\r\n\r\n\r\n\r\n> Yes, this is expected. Maybe we could create a new type a bit like `BatchEncoding` that groups together the image (on all possible formats, string, PIL, array, tensor) with its annotation, so we can then just pass that object (or a list of those objects) to the tokenizer. What do you think?\r\n\r\nYou mean pass that object to the model, rather than the tokenizer? For me, `BatchEncoding` seems like a good name.", "> Sure, it's best to make a similar API for ViT, right? (And more Transformer-based image models that will come after that). I\r\n\r\nSince ViT is not ported yet, this is where we decide the API that will be used for other vision/multi-model models :-)\r\n\r\n> You mean pass that object to the model, rather than the tokenizer? For me, `BatchEncoding` seems like a good name.\r\n\r\nNo, I meant to the tokenizer (though I'm not too sure about this part, it may end up over-complicating things). `BatchEncoding` comes with its text-related methods (`word_ids`, `sequence_ids` etc) so I don't think it should be used here since they won't be available.\r\n", "Regarding the tokenizer I think we can have a bit more freedom here than we would with NLP models as it's the first vision model, but as you've said @sgugger I think that it should still be somewhat aligned with NLP tokenizers: \r\n\r\n- It should take care of all the pre-processing steps\r\n - Creation of batches of images, with padding & truncation\r\n - All the functionalities you mentionned @NielsRogge `resize`/`size`/`normalize`, etc\r\n- Ideally it should have a very similar API to existing NLP tokenizers. Applying processing with the `__call__` method, loading/saving with `from_pretrained`/`save_pretrained`. I didn't dive in the implementation, but if parameters like `resize`/`size`/`normalize` etc are checkpoint-specific, then it's a good opportunity to save these configuration values in the `tokenizer_config.json`, leveraging the loading/saving methods mentioned above.\r\n- If there needs to be some decoding done after the model has processed the image, then that object should be able to handle it as well.\r\n\r\n@sgugger regarding what the tokenizer accepts, I'm not sure I see the advantage of handling paths directly. We don't handle paths to text files or paths to CSVs in our other tokenizers. We don't handle paths to sound files either for `Wav2Vec2`, for all of that we rely on external tools and I think that's fine.\r\n\r\nFurthermore, handling images directly in the tokenizer sounds especially memory-heavy, and relying on the `datasets` library, which can handle memory mapping, seems like a better approach than leveraging the tokenizer to load files into memory.", "Yes at least the normalize statistics (mean and std) are checkpoint-specific so should be loaded/saved with the usual API.\r\n\r\n> @sgugger regarding what the tokenizer accepts, I'm not sure I see the advantage of handling paths directly. We don't handle paths to text files or paths to CSVs in our other tokenizers. We don't handle paths to sound files either for Wav2Vec2, for all of that we rely on external tools and I think that's fine.\r\n\r\nThe difference is that a tokenizer accepts strings which is a universal type, whereas this image processor accepts PIL images, which is the format given by one specific library (so you can't load your image with openCV and feed it to the tokenizer). Since we already have a privileged image preprocessing library I really think it makes sense to let it also accept filenames. An alternative is to accept only numpy arrays and tensors, but there is the conversion back to PIL images inside the function (we could avoid it and do everything on tensors if we wanted to btw) so I don't think it makes sense.\r\n\r\nIn any case the user can still use their own preprocessing and pass the final numpy array/torch tensor with the API so I don't see the downside in accepting filenames. Usual tokenizers would have a hard time making the difference between a string that is a text and a string that is a path but this is not the case for images (or sounds, we could have that API there too and I think we should). It's just free functionality.\r\n\r\nIn NLP we have datasets as lists of texts since text is light in memory, but in CV all the datasets will come as lists of filenames that you have to load lazily (except maybe CIFAR10 and MNIST since they are tiny). Just trying to make it as easy as possible to the user.\r\n\r\n> Furthermore, handling images directly in the tokenizer sounds especially memory-heavy\r\n\r\nThe memory will be used in any case as the images passed to the tokenizer are already loaded if you don't pass filenames. The use shouldn't change between passing n filenames and n images.", "I think this goes against the API we've defined up to now for all existing modalities (text, speech, tabular), and it adds additional work on the tokenizer whereas I think data loading should be handled by PyTorch dataloaders/Datasets, or with `datasets`.\r\n\r\nHowever, your points echo with me and I have less experience than you both in vision, so if you feel that such an API is what would be best for vision, then happy to drop it and feel free to implement it this way.", "Let's not add the file supports for now and discuss it at our next internal meeting then. I agree it is a new functionality that would be different from our other APIs.", "Any update on this?\r\n\r\nThe tokenizer (I know we should rename it to something else) that I currently implemented accepts images as PIL images, Numpy arrays or PyTorch tensors, and creates 2 things: `pixel_values` and `pixel_mask`. It could be used for both DETR and ViT.\r\n\r\nWe should probably define some base utils similar to what Patrick did for the speech models.\r\n\r\ncc @LysandreJik @sgugger @patrickvonplaten ", "Thanks for reaching out!\r\n\r\nSo the \"tokenizer\" as you wrote it is good, but it should be renamed to a `DetrFeatureExtractor` and subclass `PreTrainedFeatureExtractor` (following the example of Wav2Vec2). All the necessary info to create one should be in one json file in the model repo (basically the same API as Wav2Vec2, but just the feature extractor part since there is no tokenizer in DETR). For ViT we can copy the same (we will refactor down the road if there are many models sharing the same functionality but for now we'll just use copies with # Copied from xxx markers).\r\n\r\nThere is no need for new base utils, the base utils Patrick defined are the ones to use for this case. As for the inputs, we agreed to stay with PIL Images, NumPy arrays and torch Tensors, so all good on this side.", "The [PreTrainedFeatureExtractor](https://github.com/huggingface/transformers/blob/11655fafdd42eb56ad94e09ecd84d4dc2d1041ae/src/transformers/feature_extraction_utils.py#L195) seems to be quite specifically defined for speech recognition (it requires a `sampling_rate` for instance at initialization). ", "cc @patrickvonplaten but I thought this one was supposed to be generic.", "Talked offline with Patrick and I misunderstood the plan. `PreTrainedFeatureExtractor` is for all kinds of inputs that are representable as 1d arrays of floats (like speech). For images, we should create a new base class that will implement the same methods. If you can take inspiration on `PreTrainedFeatureExtractor` to create an `ImageProcessor`, it would be great! The only thing that should be exactly the same is the name of the saved config: `preprocessing_config.json`.\r\n\r\nDoes that make sense?", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[]

[ { "id": 2659267025, "node_id": "MDU6TGFiZWwyNjU5MjY3MDI1", "url": "https://api.github.com/repos/huggingface/transformers/labels/DeepSpeed", "name": "DeepSpeed", "color": "4D34F7", "default": false, "description": "" } ]

[ { "login": "stas00", "id": 10676103, "node_id": "MDQ6VXNlcjEwNjc2MTAz", "avatar_url": "https://avatars.githubusercontent.com/u/10676103?v=4", "gravatar_id": "", "url": "https://api.github.com/users/stas00", "html_url": "https://github.com/stas00", "followers_url": "https://api.github.com/users/stas00/followers", "following_url": "https://api.github.com/users/stas00/following{/other_user}", "gists_url": "https://api.github.com/users/stas00/gists{/gist_id}", "starred_url": "https://api.github.com/users/stas00/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/stas00/subscriptions", "organizations_url": "https://api.github.com/users/stas00/orgs", "repos_url": "https://api.github.com/users/stas00/repos", "events_url": "https://api.github.com/users/stas00/events{/privacy}", "received_events_url": "https://api.github.com/users/stas00/received_events", "type": "User", "site_admin": false } ]

[ "Well, I'm closing this right away, since it's not a bug, but feel free to comment or ask questions in the comments.", "(I'm adding to this issue, even though it's closed, because it's directly related)\r\n\r\nI am seeing OOM trying to get this to work: 1 GPU, SeqLength 128 (originally tried 256), buffers {2e8, 3e8, 5e8} (just changes the epoch of the OOM), BS=1. \r\n\r\n@stas00 , I kept track of the GPU memory (as reported in nvidia-smi) to see if it's a progressive memory leak, but I don't think it is:\r\n- 23.2gb after loading model weights\r\n- 33.8gb @ epoch ~1\r\n- 33.8gb @ epoch 25 \r\n- long pause at epoch 26, then dies with OOM\r\n\r\n\r\nRunscript:\r\n(Note I am using unifiedqa-t5-11b, which is just a fine-tuned t5-11b -- I don't think that should change anything)\r\n```\r\nexport DATADIR=/home/pajansen/11b-data/ \\\r\nexport SEQLEN=128 \\\r\nexport OUTPUTDIR=output_dir \\\r\n\r\nexport BS=1; rm -rf $OUTPUTDIR; PYTHONPATH=../../src USE_TF=0 /usr/bin/time -v deepspeed --num_gpus=1 ./finetune_trainer.py --model_name_or_path allenai/unifiedqa-t5-11b --output_dir $OUTPUTDIR --adam_eps 1e-06 --data_dir $DATADIR \\\r\n--do_eval --do_predict --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 \\\r\n--logging_first_step --logging_steps 1000 --max_source_length $SEQLEN --max_target_length $SEQLEN --num_train_epochs 2 \\\r\n--overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS \\\r\n--predict_with_generate --sortish_sampler \\\r\n--test_max_target_length $SEQLEN --val_max_target_length $SEQLEN \\\r\n--warmup_steps 5 \\\r\n--deepspeed ds_config.json --fp16 \\\r\n```\r\n\r\nConda environment:\r\n```\r\n# Make new environment\r\nconda create --name transformers-feb4-2020 python=3.8\r\nconda activate transformers-feb4-2020\r\n\r\n# Clone transformers\r\ngit clone https://github.com/huggingface/transformers.git\r\ncd transformers\r\n\r\n# Install nightly build of Pytorch\r\npip install --pre torch torchvision -f https://download.pytorch.org/whl/nightly/cu110/torch_nightly.html -U\r\n\r\n# Install seq2seq transformers requirements\r\npip install -r examples/seq2seq/requirements.txt\r\n\r\n# Install transformers\r\npip install -e .\r\n\r\n# Install DeepSpeed from source for the A100 support\r\ncd ..\r\ngit clone https://github.com/microsoft/DeepSpeed.git\r\ncd DeepSpeed/\r\n./install.sh\r\npip install .\r\n\r\n```\r\n\r\nThe monster output:\r\n[oom-feb4-t5-11b.txt](https://github.com/huggingface/transformers/files/5928851/oom-feb4-t5-11b.txt)\r\n\r\nJust the last bit of the output:\r\n(the overflow errors are probably noteworthy?)\r\n```\r\nUsing /home/pajansen/.cache/torch_extensions as PyTorch extensions root...\r\nNo modifications detected for re-loaded extension module utils, skipping build step...\r\nLoading extension module utils...\r\nTime to load utils op: 0.0005221366882324219 seconds\r\n[INFO|trainer.py:837] 2021-02-04 15:05:54,964 >> ***** Running training *****\r\n[INFO|trainer.py:838] 2021-02-04 15:05:54,964 >> Num examples = 592\r\n[INFO|trainer.py:839] 2021-02-04 15:05:54,964 >> Num Epochs = 2\r\n[INFO|trainer.py:840] 2021-02-04 15:05:54,964 >> Instantaneous batch size per device = 1\r\n[INFO|trainer.py:841] 2021-02-04 15:05:54,964 >> Total train batch size (w. parallel, distributed & accumulation) = 1\r\n[INFO|trainer.py:842] 2021-02-04 15:05:54,964 >> Gradient Accumulation steps = 1\r\n[INFO|trainer.py:843] 2021-02-04 15:05:54,964 >> Total optimization steps = 1184\r\n 0%| | 0/1184 [00:00<?, ?it/s][2021-02-04 15:05:58,447] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 4294967296, reducing to 4294967296\r\n{'loss': inf, 'learning_rate': 0.0, 'epoch': 0.0} \r\n 0%|▏ | 1/1184 [00:03<1:08:20, 3.47s/it][2021-02-04 15:06:02,124] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 4294967296, reducing to 2147483648.0\r\n 0%|▎ | 2/1184 [00:07<1:09:31, 3.53s/it][2021-02-04 15:06:05,853] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 2147483648.0, reducing to 1073741824.0\r\n 0%|▍ | 3/1184 [00:10<1:10:38, 3.59s/it][2021-02-04 15:06:09,757] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 1073741824.0, reducing to 536870912.0\r\n 0%|▋ | 4/1184 [00:14<1:12:26, 3.68s/it][2021-02-04 15:06:13,120] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 536870912.0, reducing to 268435456.0\r\n 0%|▊ | 5/1184 [00:18<1:10:29, 3.59s/it][2021-02-04 15:06:16,495] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 268435456.0, reducing to 134217728.0\r\n 1%|▉ | 6/1184 [00:21<1:09:10, 3.52s/it][2021-02-04 15:06:19,825] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 134217728.0, reducing to 67108864.0\r\n 1%|█ | 7/1184 [00:24<1:07:59, 3.47s/it][2021-02-04 15:06:23,182] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 67108864.0, reducing to 33554432.0\r\n 1%|█▎ | 8/1184 [00:28<1:07:17, 3.43s/it][2021-02-04 15:06:26,854] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 33554432.0, reducing to 16777216.0\r\n 1%|█▍ | 9/1184 [00:31<1:08:37, 3.50s/it][2021-02-04 15:06:30,436] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 16777216.0, reducing to 8388608.0\r\n 1%|█▌ | 10/1184 [00:35<1:09:01, 3.53s/it][2021-02-04 15:06:33,801] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 8388608.0, reducing to 4194304.0\r\n 1%|█▋ | 11/1184 [00:38<1:08:00, 3.48s/it][2021-02-04 15:06:37,147] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 4194304.0, reducing to 2097152.0\r\n 1%|█▉ | 12/1184 [00:42<1:07:10, 3.44s/it][2021-02-04 15:06:40,510] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 2097152.0, reducing to 1048576.0\r\n 1%|██ | 13/1184 [00:45<1:06:40, 3.42s/it][2021-02-04 15:06:43,887] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 1048576.0, reducing to 524288.0\r\n 1%|██▏ | 14/1184 [00:48<1:06:23, 3.40s/it][2021-02-04 15:06:47,250] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 524288.0, reducing to 262144.0\r\n 1%|██▎ | 15/1184 [00:52<1:06:05, 3.39s/it][2021-02-04 15:06:50,615] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 262144.0, reducing to 131072.0\r\n 1%|██▌ | 16/1184 [00:55<1:05:52, 3.38s/it][2021-02-04 15:06:53,976] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 131072.0, reducing to 65536.0\r\n 1%|██▋ | 17/1184 [00:58<1:05:41, 3.38s/it][2021-02-04 15:06:57,313] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 65536.0, reducing to 32768.0\r\n 2%|██▊ | 18/1184 [01:02<1:05:23, 3.36s/it][2021-02-04 15:07:00,672] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 32768.0, reducing to 16384.0\r\n 2%|███ | 19/1184 [01:05<1:05:18, 3.36s/it][2021-02-04 15:07:04,003] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 16384.0, reducing to 8192.0\r\n 2%|███▏ | 20/1184 [01:09<1:05:03, 3.35s/it][2021-02-04 15:07:07,382] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 8192.0, reducing to 4096.0\r\n 2%|███▎ | 21/1184 [01:12<1:05:08, 3.36s/it][2021-02-04 15:07:10,753] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 4096.0, reducing to 2048.0\r\n 2%|███▍ | 22/1184 [01:15<1:05:09, 3.36s/it][2021-02-04 15:07:14,118] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 2048.0, reducing to 1024.0\r\n 2%|███▋ | 23/1184 [01:19<1:05:06, 3.36s/it][2021-02-04 15:07:17,475] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 1024.0, reducing to 512.0\r\n 2%|███▊ | 24/1184 [01:22<1:05:00, 3.36s/it][2021-02-04 15:07:20,816] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 512.0, reducing to 256.0\r\n 2%|███▉ | 25/1184 [01:25<1:04:49, 3.36s/it][2021-02-04 15:07:24,174] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 256.0, reducing to 128.0\r\n 2%|████ | 26/1184 [01:29<1:04:46, 3.36s/it]Killing subprocess 3319579\r\nTraceback (most recent call last):\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/runpy.py\", line 194, in _run_module_as_main\r\n return _run_code(code, main_globals, None,\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/runpy.py\", line 87, in _run_code\r\n exec(code, run_globals)\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/deepspeed/launcher/launch.py\", line 171, in <module>\r\n main()\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/deepspeed/launcher/launch.py\", line 161, in main\r\n sigkill_handler(signal.SIGTERM, None) # not coming back\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/deepspeed/launcher/launch.py\", line 139, in sigkill_handler\r\n raise subprocess.CalledProcessError(returncode=last_return_code, cmd=cmd)\r\nsubprocess.CalledProcessError: Command '['/home/pajansen/anaconda3/envs/transformers-feb4-2020/bin/python', '-u', './finetune_trainer.py', '--local_rank=0', '--model_name_or_path', 'allenai/unifiedqa-t5-11b', '--output_dir', 'output_dir_compexpl-feb4-epoch2-uqa-11b-wholetree-rev', '--adam_eps', '1e-06', '--data_dir', '/home/pajansen/github/compositional-expl/data/feb4-initialtest-q693/wholetree-rev/', '--do_eval', '--do_predict', '--do_train', '--evaluation_strategy=steps', '--freeze_embeds', '--label_smoothing', '0.1', '--learning_rate', '3e-5', '--logging_first_step', '--logging_steps', '1000', '--max_source_length', '128', '--max_target_length', '128', '--num_train_epochs', '2', '--overwrite_output_dir', '--per_device_eval_batch_size', '1', '--per_device_train_batch_size', '1', '--predict_with_generate', '--sortish_sampler', '--test_max_target_length', '128', '--val_max_target_length', '128', '--warmup_steps', '5', '--deepspeed', 'ds_config.json', '--fp16']' died with <Signals.SIGSEGV: 11>.\r\n Command being timed: \"deepspeed --num_gpus=1 ./finetune_trainer.py --model_name_or_path allenai/unifiedqa-t5-11b --output_dir output_dir_compexpl-feb4-epoch2-uqa-11b-wholetree-rev --adam_eps 1e-06 --data_dir /home/pajansen/github/compositional-expl/data/feb4-initialtest-q693/wholetree-rev/ --do_eval --do_predict --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 2 --overwrite_output_dir --per_device_eval_batch_size 1 --per_device_train_batch_size 1 --predict_with_generate --sortish_sampler --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 5 --deepspeed ds_config.json --fp16\"\r\n User time (seconds): 1152.16\r\n System time (seconds): 746.75\r\n Percent of CPU this job got: 396%\r\n Elapsed (wall clock) time (h:mm:ss or m:ss): 7:58.47\r\n Average shared text size (kbytes): 0\r\n Average unshared data size (kbytes): 0\r\n Average stack size (kbytes): 0\r\n Average total size (kbytes): 0\r\n Maximum resident set size (kbytes): 233292336\r\n Average resident set size (kbytes): 0\r\n Major (requiring I/O) page faults: 0\r\n Minor (reclaiming a frame) page faults: 108071918\r\n Voluntary context switches: 38621\r\n Involuntary context switches: 588867\r\n Swaps: 0\r\n File system inputs: 0\r\n File system outputs: 48\r\n Socket messages sent: 0\r\n Socket messages received: 0\r\n Signals delivered: 0\r\n Page size (bytes): 4096\r\n Exit status: 0\r\n```\r\n", "Thank you for the report and the details, @PeterAJansen\r\n\r\nIn the future, let's try to have a dedicated issue for each unique problem, but since the OP wasn't really an issue, it is now ;) so all is good.\r\n\r\nLet me see if I can reproduce the problem with your changes, perhaps my data sample was too short.\r\n\r\nThe other difference I see is that you're not using `--task` which then defaults to `summarization` - so we surely don't test the exact same thing.\r\n\r\nThe `allenai/unifiedqa-t5-11b` model looks of identical size to `t5-11b`, but let me download the former to make sure that I'm doing an exact reproduction. \r\n\r\nLet me see \r\n1. if I can get it to OOM with the translation task that I have been testing with first \r\n2. and if that fails, I will try one of the local summarization datasets, \r\n3. and if all runs fine still will need to see what's different about your dataset.\r\n\r\n> (the overflow errors are probably noteworthy?)\r\n\r\nthese are normal. not a problem.", "OK, I'm able to reproduce it. The GPU memory usage grows slowly at some times and jumps at quick bump ups of several GBs at other times. \r\n\r\nI used buffers of 1e8 and cmd:\r\n```\r\nexport BS=2; rm -rf output_dir; PYTHONPATH=../../src USE_TF=0 /usr/bin/time -v deepspeed --num_gpus=1 ./finetune_trainer.py --model_name_or_path allenai/unifiedqa-t5-11b --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_predict --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 5 --deepspeed ds_config.json --fp16\r\n```\r\n\r\nWhich means that either transformers (trainer or model) or DeepSpeed or both leak memory. I'm going to switch to a much smaller model size as with this model it takes ages for it to just start - can't develop like this and try to detect where the leak is coming from.\r\n\r\nBTW, here is a tip. Currently transformers performs a silly thing - it inits the model, inits the weights, and overwrites all this work with pretrained weights. Which with this model takes like 10 minutes. You can shortcut it with:\r\n\r\n```\r\n--- a/src/transformers/modeling_utils.py\r\n+++ b/src/transformers/modeling_utils.py\r\n@@ -747,7 +747,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin):\r\n Initializes and prunes weights if needed.\r\n \"\"\"\r\n # Initialize weights\r\n- self.apply(self._init_weights)\r\n+ #self.apply(self._init_weights)\r\n\r\n # Prune heads if needed\r\n if self.config.pruned_heads:\r\n```\r\nwhich skips 90% of the pointless of weight inits.\r\n\r\nI'm trying to advocate for this to be a feature here: https://github.com/huggingface/transformers/issues/9205", "Heh, we were assuming it was OOM, but it got SIGSEGV - I didn't bother to look closer - so pytorch w/Deepspeed segfaults pretty much at step 22. Investigating...\r\n\r\nNo useful info in the core bt. Stripped binaries.\r\n\r\nI eliminated the possibility that the issue could be with pytorch.\r\n\r\nMost likely a regression in DS. \r\n\r\nDowngrading `pip install deepspeed==0.3.10` solves the segfault\r\n\r\nI must have been using an old DS yesterday and that's why it was working for me.\r\n\r\nTrying to locate the faulty commit in DS\r\n\r\nAnd the reason it was happening always at step 22 was because AdamW wasn't running until this step, this is all those skipping step overflow reports:\r\n\r\n```\r\n[2021-02-04 22:40:47,424] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 2048.0, reducing to 1024.0\r\n 0%| | 23/60000 [01:18<55:05:44, 3.31s/it][2021-02-04 22:40:50,837] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 1024.0, reducing to 512.0\r\n 0%| | 24/60000 [01:21<55:37:22, 3.34s/it][2021-02-04 22:40:54,255] [INFO] [stage2.py:1357:step] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 512.0, reducing to 256.0\r\n```\r\n\r\nAs soon as it run it segfaulted.\r\n\r\nHopefully we will have a fix soon, but until then please use `deepspeed==0.3.10` ", "Thanks @stas00 ! \r\n\r\nI have downgraded to deepspeed 0.3.10 and I'm going to leave Transformers running overnight on a proper training job to see if it crashes (it's currently about 20% completed, so that's promising). Though it does appear that the GPU memory usage periodically moves from ~34GB up to nearly the entire 40GB minus a few hundred MB, so it's a real nail biter watching it: \r\n\r\n\r\n\r\nTransformers+DeepSpeed really doesn't believe in wasting RAM... :) \r\n", "update: DeepSpeed yanked 0.3.11 from pypi, so a normal pip install should now result in a good working 0.3.10 installed until this issue is fixed.", "Update on my end: with DeepSpeed 0.3.10 it did run successfully through the night on a full job, successfully training and generating the predictions. Amazing work @stas00 et al. \r\n", "@stas00 I'm not sure if this is a bug or if I'm just not doing it correctly given how fast most of this is moving, but I'm trying to evaluate/generate predictions post-training and getting not-on-device errors. I should not that it worked fine when I did the whole thing in one command (train/eval/predict) overnight, but now I'm trying to use the fine-tuned model to generate predictions on other data. \r\n\r\nI have (a) just removed the --do_train flag from the call to finetune_trainer (and, set the model path to the output path of the fine-tuned model), and this gives an error (below). I've also (b) tried CPU-based eval (--device cpu) with the official instructions in examples/seq2seq/, which gave a different error (but I've not done non-cuda eval before, so that might be my issue). \r\n\r\nHere's the error from (A):\r\n```\r\n[2021-02-05 12:00:30,238] [WARNING] [runner.py:117:fetch_hostfile] Unable to find hostfile, will proceed with training with local resources only.\r\n[2021-02-05 12:00:30,586] [INFO] [runner.py:355:main] cmd = /home/pajansen/anaconda3/envs/transformers-feb4-2020/bin/python -u -m deepspeed.launcher.launch --world_info=eyJsb2NhbGhvc3QiOiBbMCwgMSwgMiwgM119 --master_addr=127.0.0.1 --master_port=29500 ./finetune_trainer.py --model_name_or_path output_dir_compexpl-feb4-epoch3-uqa-11b-wholetree-rev --output_dir output_dir_compexpl-feb4-epoch3-uqa-11b-wholetree-rev-unannotated --adam_eps 1e-06 --data_dir /home/pajansen/github/compexpl/data/feb4-initialtest-q693/unannotated/ --do_eval --do_predict --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 256 --max_target_length 256 --num_train_epochs 3 --overwrite_output_dir --per_device_eval_batch_size 1 --per_device_train_batch_size 1 --predict_with_generate --sortish_sampler --test_max_target_length 256 --val_max_target_length 256 --warmup_steps 5 --deepspeed ds_config.json --fp16\r\n[2021-02-05 12:00:31,464] [INFO] [launch.py:78:main] WORLD INFO DICT: {'localhost': [0, 1, 2, 3]}\r\n[2021-02-05 12:00:31,464] [INFO] [launch.py:84:main] nnodes=1, num_local_procs=4, node_rank=0\r\n[2021-02-05 12:00:31,464] [INFO] [launch.py:99:main] global_rank_mapping=defaultdict(<class 'list'>, {'localhost': [0, 1, 2, 3]})\r\n[2021-02-05 12:00:31,464] [INFO] [launch.py:100:main] dist_world_size=4\r\n[2021-02-05 12:00:31,464] [INFO] [launch.py:102:main] Setting CUDA_VISIBLE_DEVICES=0,1,2,3\r\n[2021-02-05 12:00:33,681] [INFO] [distributed.py:39:init_distributed] Initializing torch distributed with backend: nccl\r\n[2021-02-05 12:00:33,788] [INFO] [distributed.py:39:init_distributed] Initializing torch distributed with backend: nccl\r\n[2021-02-05 12:00:33,908] [INFO] [distributed.py:39:init_distributed] Initializing torch distributed with backend: nccl\r\n[2021-02-05 12:00:34,042] [INFO] [distributed.py:39:init_distributed] Initializing torch distributed with backend: nccl\r\nWARNING:__main__:Process rank: 0, device: cuda:0, n_gpu: 1, distributed training: True, 16-bits training: True\r\n[INFO|configuration_utils.py:447] 2021-02-05 12:00:34,625 >> loading configuration file output_dir_compexpl-feb4-epoch3-uqa-11b-wholetree-rev/config.json\r\n[INFO|configuration_utils.py:485] 2021-02-05 12:00:34,626 >> Model config T5Config {\r\n \"_name_or_path\": \"allenai/unifiedqa-t5-11b\",\r\n \"architectures\": [\r\n \"T5ForConditionalGeneration\"\r\n ],\r\n \"d_ff\": 65536,\r\n \"d_kv\": 128,\r\n \"d_model\": 1024,\r\n \"decoder_start_token_id\": 0,\r\n \"dropout_rate\": 0.1,\r\n \"early_stopping\": true,\r\n \"eos_token_id\": 1,\r\n \"feed_forward_proj\": \"relu\",\r\n \"initializer_factor\": 1.0,\r\n \"is_encoder_decoder\": true,\r\n \"layer_norm_epsilon\": 1e-06,\r\n \"length_penalty\": 2.0,\r\n \"max_length\": 200,\r\n \"min_length\": 30,\r\n \"model_type\": \"t5\",\r\n \"n_positions\": 512,\r\n \"no_repeat_ngram_size\": 3,\r\n \"num_beams\": 4,\r\n \"num_decoder_layers\": 24,\r\n \"num_heads\": 128,\r\n \"num_layers\": 24,\r\n \"output_past\": true,\r\n \"pad_token_id\": 0,\r\n \"prefix\": \"summarize: \",\r\n \"relative_attention_num_buckets\": 32,\r\n \"task_specific_params\": {\r\n \"summarization\": {\r\n \"early_stopping\": true,\r\n \"length_penalty\": 2.0,\r\n \"max_length\": 200,\r\n \"min_length\": 30,\r\n \"no_repeat_ngram_size\": 3,\r\n \"num_beams\": 4,\r\n \"prefix\": \"summarize: \"\r\n },\r\n \"translation_en_to_de\": {\r\n \"early_stopping\": true,\r\n \"max_length\": 300,\r\n \"num_beams\": 4,\r\n \"prefix\": \"translate English to German: \"\r\n },\r\n \"translation_en_to_fr\": {\r\n \"early_stopping\": true,\r\n \"max_length\": 300,\r\n \"num_beams\": 4,\r\n \"prefix\": \"translate English to French: \"\r\n },\r\n \"translation_en_to_ro\": {\r\n \"early_stopping\": true,\r\n \"max_length\": 300,\r\n \"num_beams\": 4,\r\n \"prefix\": \"translate English to Romanian: \"\r\n }\r\n },\r\n \"transformers_version\": \"4.3.0.dev0\",\r\n \"use_cache\": true,\r\n \"vocab_size\": 32128\r\n}\r\n\r\n[INFO|configuration_utils.py:447] 2021-02-05 12:00:34,626 >> loading configuration file output_dir_compexpl-feb4-epoch3-uqa-11b-wholetree-rev/config.json\r\n[INFO|configuration_utils.py:485] 2021-02-05 12:00:34,627 >> Model config T5Config {\r\n \"_name_or_path\": \"allenai/unifiedqa-t5-11b\",\r\n \"architectures\": [\r\n \"T5ForConditionalGeneration\"\r\n ],\r\n \"d_ff\": 65536,\r\n \"d_kv\": 128,\r\n \"d_model\": 1024,\r\n \"decoder_start_token_id\": 0,\r\n \"dropout_rate\": 0.1,\r\n \"early_stopping\": true,\r\n \"eos_token_id\": 1,\r\n \"feed_forward_proj\": \"relu\",\r\n \"initializer_factor\": 1.0,\r\n \"is_encoder_decoder\": true,\r\n \"layer_norm_epsilon\": 1e-06,\r\n \"length_penalty\": 2.0,\r\n \"max_length\": 200,\r\n \"min_length\": 30,\r\n \"model_type\": \"t5\",\r\n \"n_positions\": 512,\r\n \"no_repeat_ngram_size\": 3,\r\n \"num_beams\": 4,\r\n \"num_decoder_layers\": 24,\r\n \"num_heads\": 128,\r\n \"num_layers\": 24,\r\n \"output_past\": true,\r\n \"pad_token_id\": 0,\r\n \"prefix\": \"summarize: \",\r\n \"relative_attention_num_buckets\": 32,\r\n \"task_specific_params\": {\r\n \"summarization\": {\r\n \"early_stopping\": true,\r\n \"length_penalty\": 2.0,\r\n \"max_length\": 200,\r\n \"min_length\": 30,\r\n \"no_repeat_ngram_size\": 3,\r\n \"num_beams\": 4,\r\n \"prefix\": \"summarize: \"\r\n },\r\n \"translation_en_to_de\": {\r\n \"early_stopping\": true,\r\n \"max_length\": 300,\r\n \"num_beams\": 4,\r\n \"prefix\": \"translate English to German: \"\r\n },\r\n \"translation_en_to_fr\": {\r\n \"early_stopping\": true,\r\n \"max_length\": 300,\r\n \"num_beams\": 4,\r\n \"prefix\": \"translate English to French: \"\r\n },\r\n \"translation_en_to_ro\": {\r\n \"early_stopping\": true,\r\n \"max_length\": 300,\r\n \"num_beams\": 4,\r\n \"prefix\": \"translate English to Romanian: \"\r\n }\r\n },\r\n \"transformers_version\": \"4.3.0.dev0\",\r\n \"use_cache\": true,\r\n \"vocab_size\": 32128\r\n}\r\n\r\n[INFO|tokenization_utils_base.py:1685] 2021-02-05 12:00:34,627 >> Model name 'output_dir_compexpl-feb4-epoch3-uqa-11b-wholetree-rev' not found in model shortcut name list (t5-small, t5-base, t5-large, t5-3b, t5-11b). Assuming 'output_dir_compexpl-feb4-epoch3-uqa-11b-wholetree-rev' is a path, a model identifier, or url to a directory containing tokenizer files.\r\n[INFO|tokenization_utils_base.py:1721] 2021-02-05 12:00:34,627 >> Didn't find file output_dir_compexpl-feb4-epoch3-uqa-11b-wholetree-rev/tokenizer.json. We won't load it.\r\n[INFO|tokenization_utils_base.py:1721] 2021-02-05 12:00:34,627 >> Didn't find file output_dir_compexpl-feb4-epoch3-uqa-11b-wholetree-rev/added_tokens.json. We won't load it.\r\n[INFO|tokenization_utils_base.py:1784] 2021-02-05 12:00:34,627 >> loading file output_dir_compexpl-feb4-epoch3-uqa-11b-wholetree-rev/spiece.model\r\n[INFO|tokenization_utils_base.py:1784] 2021-02-05 12:00:34,627 >> loading file None\r\n[INFO|tokenization_utils_base.py:1784] 2021-02-05 12:00:34,627 >> loading file None\r\n[INFO|tokenization_utils_base.py:1784] 2021-02-05 12:00:34,627 >> loading file output_dir_compexpl-feb4-epoch3-uqa-11b-wholetree-rev/special_tokens_map.json\r\n[INFO|tokenization_utils_base.py:1784] 2021-02-05 12:00:34,627 >> loading file output_dir_compexpl-feb4-epoch3-uqa-11b-wholetree-rev/tokenizer_config.json\r\nWARNING:__main__:Process rank: 1, device: cuda:1, n_gpu: 1, distributed training: True, 16-bits training: True\r\nWARNING:__main__:Process rank: 3, device: cuda:3, n_gpu: 1, distributed training: True, 16-bits training: True\r\nWARNING:__main__:Process rank: 2, device: cuda:2, n_gpu: 1, distributed training: True, 16-bits training: True\r\n[INFO|modeling_utils.py:1025] 2021-02-05 12:00:34,753 >> loading weights file output_dir_compexpl-feb4-epoch3-uqa-11b-wholetree-rev/pytorch_model.bin\r\n[INFO|modeling_utils.py:1143] 2021-02-05 12:04:48,021 >> All model checkpoint weights were used when initializing T5ForConditionalGeneration.\r\n\r\n[INFO|modeling_utils.py:1151] 2021-02-05 12:04:48,034 >> All the weights of T5ForConditionalGeneration were initialized from the model checkpoint at output_dir_compexpl-feb4-epoch3-uqa-11b-wholetree-rev.\r\nIf your task is similar to the task the model of the checkpoint was trained on, you can already use T5ForConditionalGeneration for predictions without further training.\r\n[INFO|trainer.py:348] 2021-02-05 12:04:48,080 >> Using amp fp16 backend\r\n[INFO|trainer.py:1600] 2021-02-05 12:04:48,080 >> ***** Running Evaluation *****\r\n[INFO|trainer.py:1601] 2021-02-05 12:04:48,080 >> Num examples = 1950\r\n[INFO|trainer.py:1602] 2021-02-05 12:04:48,080 >> Batch size = 1\r\nTraceback (most recent call last):\r\n File \"./finetune_trainer.py\", line 367, in <module>\r\n main()\r\n File \"./finetune_trainer.py\", line 327, in main\r\n metrics = trainer.evaluate(metric_key_prefix=\"val\")\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/trainer.py\", line 1506, in evaluate\r\n output = self.prediction_loop(\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/trainer.py\", line 1630, in prediction_loop\r\n loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/examples/seq2seq/seq2seq_trainer.py\", line 220, in prediction_step\r\n generated_tokens = self.model.generate(\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/autograd/grad_mode.py\", line 27, in decorate_context\r\n return func(*args, **kwargs)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/generation_utils.py\", line 847, in generate\r\n model_kwargs = self._prepare_encoder_decoder_kwargs_for_generation(input_ids, model_kwargs)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/generation_utils.py\", line 379, in _prepare_encoder_decoder_kwargs_for_generation\r\n model_kwargs[\"encoder_outputs\"]: ModelOutput = encoder(input_ids, return_dict=True, **encoder_kwargs)\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/models/t5/modeling_t5.py\", line 878, in forward\r\n inputs_embeds = self.embed_tokens(input_ids)\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/modules/sparse.py\", line 145, in forward\r\n return F.embedding(\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/functional.py\", line 1921, in embedding\r\n return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)\r\nRuntimeError: Input, output and indices must be on the current device\r\nTraceback (most recent call last):\r\n File \"./finetune_trainer.py\", line 367, in <module>\r\n main()\r\n File \"./finetune_trainer.py\", line 327, in main\r\n metrics = trainer.evaluate(metric_key_prefix=\"val\")\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/trainer.py\", line 1506, in evaluate\r\n output = self.prediction_loop(\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/trainer.py\", line 1630, in prediction_loop\r\n loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/examples/seq2seq/seq2seq_trainer.py\", line 220, in prediction_step\r\n generated_tokens = self.model.generate(\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/autograd/grad_mode.py\", line 27, in decorate_context\r\n return func(*args, **kwargs)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/generation_utils.py\", line 847, in generate\r\n model_kwargs = self._prepare_encoder_decoder_kwargs_for_generation(input_ids, model_kwargs)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/generation_utils.py\", line 379, in _prepare_encoder_decoder_kwargs_for_generation\r\n model_kwargs[\"encoder_outputs\"]: ModelOutput = encoder(input_ids, return_dict=True, **encoder_kwargs)\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/models/t5/modeling_t5.py\", line 878, in forward\r\n inputs_embeds = self.embed_tokens(input_ids)\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/modules/sparse.py\", line 145, in forward\r\n return F.embedding(\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/functional.py\", line 1921, in embedding\r\n return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)\r\nRuntimeError: Input, output and indices must be on the current device\r\nTraceback (most recent call last):\r\n File \"./finetune_trainer.py\", line 367, in <module>\r\n main()\r\n File \"./finetune_trainer.py\", line 327, in main\r\n metrics = trainer.evaluate(metric_key_prefix=\"val\")\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/trainer.py\", line 1506, in evaluate\r\n output = self.prediction_loop(\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/trainer.py\", line 1630, in prediction_loop\r\n loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/examples/seq2seq/seq2seq_trainer.py\", line 220, in prediction_step\r\n generated_tokens = self.model.generate(\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/autograd/grad_mode.py\", line 27, in decorate_context\r\n return func(*args, **kwargs)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/generation_utils.py\", line 847, in generate\r\n model_kwargs = self._prepare_encoder_decoder_kwargs_for_generation(input_ids, model_kwargs)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/generation_utils.py\", line 379, in _prepare_encoder_decoder_kwargs_for_generation\r\n model_kwargs[\"encoder_outputs\"]: ModelOutput = encoder(input_ids, return_dict=True, **encoder_kwargs)\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/models/t5/modeling_t5.py\", line 878, in forward\r\n inputs_embeds = self.embed_tokens(input_ids)\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/modules/sparse.py\", line 145, in forward\r\n return F.embedding(\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/functional.py\", line 1921, in embedding\r\n return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)\r\nRuntimeError: Input, output and indices must be on the current device\r\nTraceback (most recent call last):\r\n File \"./finetune_trainer.py\", line 367, in <module>\r\n main()\r\n File \"./finetune_trainer.py\", line 327, in main\r\n metrics = trainer.evaluate(metric_key_prefix=\"val\")\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/trainer.py\", line 1506, in evaluate\r\n output = self.prediction_loop(\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/trainer.py\", line 1630, in prediction_loop\r\n loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/examples/seq2seq/seq2seq_trainer.py\", line 220, in prediction_step\r\n generated_tokens = self.model.generate(\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/autograd/grad_mode.py\", line 27, in decorate_context\r\n return func(*args, **kwargs)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/generation_utils.py\", line 847, in generate\r\n model_kwargs = self._prepare_encoder_decoder_kwargs_for_generation(input_ids, model_kwargs)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/generation_utils.py\", line 379, in _prepare_encoder_decoder_kwargs_for_generation\r\n model_kwargs[\"encoder_outputs\"]: ModelOutput = encoder(input_ids, return_dict=True, **encoder_kwargs)\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/pajansen/github/transformers-feb4-2021/transformers/src/transformers/models/t5/modeling_t5.py\", line 878, in forward\r\n inputs_embeds = self.embed_tokens(input_ids)\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/modules/sparse.py\", line 145, in forward\r\n return F.embedding(\r\n File \"/home/pajansen/anaconda3/envs/transformers-feb4-2020/lib/python3.8/site-packages/torch/nn/functional.py\", line 1921, in embedding\r\n return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)\r\nRuntimeError: Input, output and indices must be on the current device\r\n```", "Are you on master and not by chance on my experimental t5-pipeline branch? If it's the latter then it's very likely that you'd hit that \"not on the current device\" error. Please make sure you're using the master `transformers`.", "Definitely on the master :)", "Update: I did figure out the CPU eval error -- I had --fp16 set (as in the example script), which currently throws an esoteric pytorch error on CPU (\"threshold_cpu\" not implemented for 'Half'). Removing this lets it run on CPU, but with 64 cores T5-11B is evaluating at 150 seconds *per generation*, instead of less than 1 sec with the GPU, so I think I'll kill that. ", "> @PeterAJansen want to confirm with you one detail, is your setup with Intel or AMD cpu?\r\n\r\nIt's AMD.\r\n\r\nI'm using Peter's machine for debugging this, so you can ask me anything. \r\n\r\n------------\r\n\r\n@PeterAJansen, glad you sorted it out - let me see if I can reproduce that and we could ensure that we prevent the erroneous fp16/cpu combination in first place.\r\n\r\n------------\r\n\r\nUpdate on DeepSpeed: it looks like the segfault over CPU ADAM problem is specific to AMD, which is the case on your computer, so the DeepSpeed team are working on figuring that out and hopefully will have a new release some time soon that will do the right thing on AMD and be fast too.", "@PeterAJansen, \r\n\r\n- I have fixed the first bug where you went for inference without training - please use this PR branch if it's not merged https://github.com/huggingface/transformers/pull/10039\r\nWell basically we aren't using deepspeed at the moment at all if `--do_train` wasn't run - need to think how to benefit from Deepspeed for pure inference. I will experiment with that.\r\n\r\n- wrt `--device cpu` could you please explain how you managed to use it? Since it's not a valid flag for `finetune_trainer.py`, so if you could share the full cmd that would help to reproduce the problem. \r\n\r\nThank you!\r\n", "@PeterAJansen, for the future let's do this:\r\n\r\n- Try new things - if they fail assume it's 99% a bug in our code - things should either work or give a user-friendly message so that you know it's your error - if it's anything else we should be fixing it.\r\n- Please do file a new issue every time - while all these bugs are totally related it is very difficult to track when it's one pile\r\n- Always paste the full cmd that you used\r\n- Ideally try to use generic datasets/models to make it easy to reproduce the problem\r\n\r\nThen:\r\n1. I reproduce\r\n2. I write a new test\r\n3. I fix the bug\r\n4. You try new things \r\n5. Rinse and repeat\r\n\r\n\r\n\r\n;)\r\n ", "> @PeterAJansen,\r\n> \r\n> * I have fixed the first bug where you went for inference without training - please use this PR branch if it's not merged #10039\r\n> Well basically we aren't using deepspeed at the moment at all if `--do_train` wasn't run - need to think how to benefit from Deepspeed for pure inference. I will experiment with that.\r\n\r\nThanks!\r\n\r\n> * wrt `--device cpu` could you please explain how you managed to use it? Since it's not a valid flag for `finetune_trainer.py`, so if you could share the full cmd that would help to reproduce the problem.\r\n> \r\n> Thank you!\r\n\r\nApologies, I think in my exhilaration that it's running T5-11B on 40G cards that I forgot proper issue submission procedures. The --fp16 error is submitted as isssue #10040 :)", "both issues have been fixed https://github.com/huggingface/transformers/pull/10039 and https://github.com/huggingface/transformers/pull/10041", "@stas00 have you tried profiling Hugging Face models with DeepSpeed's `FlopsProfiler`? I'm curious to see what kind of stats you get, especially for decoder-only models such as `GPT2LMHeadModel` as you increase the model size.", "I haven't tried yet - as I'm busy at the moment at figuring out the pipeline, but I logged that idea here https://github.com/huggingface/transformers/issues/9606 for a later time or if someone else is moved to do it before I get a chance to do so. \r\n\r\nI appreciate the suggestion, @g-karthik. I'm like a kid in a candy store, so many things to try, so little time.", "@stas00 not sure if this issue is closed and/or I should start a new thread. But my question is very much related. Here goes:\r\n\r\nI followed the instructions mentioned here (same deepspeed version, t5-11b. everything same). However on 1x 40GB gpu w/ Deepspeed (A100-SXM4-40GB) it goes OOM. **Does not train even with BS=1 using deepspeed.**\r\n\r\nStill wondering how you were able to train this on 1x A100-SXM4-40GB since the t5-11b downloaded (automatically by huggingface), pytorch.bin model file itself has a size of ≈ 45GB (raw file size). Just loading the model itself will cause OOM on a 40GB 1x A100-SXM4-40GB.\r\n\r\nAm I missing something? or did the t5-11b model size change since this post?\r\n\r\nSrikar ", "Hi @srikar2097,\r\n\r\ndeepspeed does `model.half()` by default so you are only loading 22.5GB in weights. though it did add support for fp32 since that post.\r\n\r\nMost likely your seq_len is much larger than the test that I did. Does it work if you reduce it?\r\n\r\nAlso this is really old now, and you have the offload available so if you have lots of RAM you shouldn't have a problem loading t5-11b on A100-50GB.\r\n\r\nIf you are still struggling, then yes, by all means please open a new issue and full details on how to reproduce the problem. and tag me please.\r\n\r\n", "FWIW, I remember having a specific commit that seemed to work for T5-11B in the 40gb A100s, and it not working after -- and me mostly using the T5-3B model for speed, so I haven't tried it recently to see if it still works (without the offloading). ", "@stas00 thanks for the tips. I did try with seq_len=512 with BS=1. Then with seq_len=128 with BS=1 (both times OOM).\r\n\r\nFor T5-11b on a A100-40B, I guess sticking to fp16 is the way to go since fp32 will load entire model into GPU mem? (which will surely cause OOM since raw model file itself is 45GB). \r\n\r\nmy host has 1TB RAM, so you suggest to use offload? Do you have some comments on if using offload would slow down training? (since optimizer-states/gradients has to flow back-and-forth between GPU <-> CPU)... \r\n\r\n@PeterAJansen I am using T5-3b for now since I haven't yet cracked the code with T5-11b.. appreciate re-affirming my comments that T5-11b is not working for you too... \r\n", "> @stas00 thanks for the tips. I did try with seq_len=512 with BS=1. Then with seq_len=128 with BS=1 (both times OOM).\r\n\r\nPlease file a new Issue with a full report with config file and command line and then I'd be happy to try to diagnose this with you.\r\n\r\nThank you for experimenting with shorter seq_len.\r\n\r\n@PeterAJansen do you remember which commit or perhaps it's logged somewhere in the Issue comments? Could probably `git bisect` to find it.\r\n\r\n> For T5-11b on a A100-40B, I guess sticking to fp16 is the way to go since fp32 will load entire model into GPU mem? (which will surely cause OOM since raw model file itself is 45GB).\r\n\r\ncorrect!\r\n\r\n> my host has 1TB RAM, so you suggest to use offload? Do you have some comments on if using offload would slow down training? (since optimizer-states/gradients has to flow back-and-forth between GPU <-> CPU)...\r\n\r\nI don't have numbers to share yet, but the offload protocol is written to pre-fetch data, so the overhead in theory should be minimal. so absolutely yes to offload.\r\n", "@stas00 I have a feeling it might be `c130e67d` , or failing that something on or around February 12th 2021. ", "OK, I'm able to train t5-11b on a single A100-SXM4-40GB with seq len 1024 with BS=4 at about 40GB gpu mem usage with deepspeed zero2:\r\n```\r\nexport BS=4; rm -r output_dir; PYTHONPATH=src USE_TF=0 CUDA_VISIBLE_DEVICES=0 deepspeed --num_gpus=1 \\\r\nexamples/pytorch/translation/run_translation.py --model_name_or_path t5-11b --output_dir output_dir \\\r\n--adam_eps 1e-06 --evaluation_strategy=steps --do_train --label_smoothing 0.1 --learning_rate 3e-5 \\\r\n--logging_first_step --logging_steps 500 --max_source_length 1024 --max_target_length 1024 \\\r\n--num_train_epochs 1 --overwrite_output_dir --per_device_train_batch_size $BS \\\r\n--predict_with_generate --sortish_sampler --source_lang en --target_lang ro --dataset_name wmt16 \\\r\n--dataset_config \"ro-en\" --source_prefix \"translate English to Romanian: \" --val_max_target_length \\\r\n128 --warmup_steps 50 --max_train_samples 2000 --max_eval_samples 50 --deepspeed \\\r\ntests/deepspeed/ds_config_zero2.json --fp16\r\n```\r\n\r\nlet's log for posterity (both master HEAD as of this writing)\r\n- PyTorch version: 1.8.1\r\n- cuda: 11.1\r\n\r\n```\r\n$ cd transformers\r\n$ git rev-parse --short HEAD\r\n61c506349\r\n\r\n$ cd ../deepspeed\r\nccc522c\r\n```\r\n\r\nsurprisingly zero3 with full offload OOMs! Need to figure that one out.\r\n\r\nThanks to @PeterAJansen for letting me use his rig.\r\n", "OK, @samyam helped me to figure out ZeRO-3 - getting a 3.5x larger BS than with zero2. The key was to lower:\r\n\r\n```\r\n\"sub_group_size\": 1e9,\r\n```\r\nfrom `1e14`.\r\n\r\nSo, I'm able to train t5-11b on a single A100-SXM4-40GB with seq len 1024 with **BS=14** with deepspeed ZeRO-3:\r\n\r\n```\r\nexport BS=14; rm -r output_dir; PYTHONPATH=src USE_TF=0 CUDA_VISIBLE_DEVICES=0 deepspeed --num_gpus=1 \\\r\nexamples/pytorch/translation/run_translation.py --model_name_or_path t5-11b --output_dir output_dir \\\r\n--adam_eps 1e-06 --evaluation_strategy=steps --do_train --label_smoothing 0.1 --learning_rate 3e-5 \\\r\n--logging_first_step --logging_steps 500 --max_source_length 1024 --max_target_length 1024 \\\r\n--num_train_epochs 1 --overwrite_output_dir --per_device_train_batch_size $BS \\\r\n--predict_with_generate --sortish_sampler --source_lang en --target_lang ro --dataset_name wmt16 \\\r\n--dataset_config \"ro-en\" --source_prefix \"translate English to Romanian: \" --val_max_target_length \\\r\n128 --warmup_steps 50 --max_train_samples 2000 --max_eval_samples 50 --deepspeed \\\r\ntests/deepspeed/ds_config_zero3.json --fp16\r\n```\r\n\r\neverything else is the same as in the zero-2 post above, and config file is too from transformers @ 61c506349 , but `ds_config_zero3.json` needs to be changed as shown above.\r\n\r\n\r\n", "I'd like to mention that the code above uses dynamic padding, which doesn't pad to length 1024, so the input and output are not 1024. Turning on \"--pad_to_max_length True\" results in OOM, unfortunately, with even low batch size of 1. I tried length 512 as well with batch size 1 but also got out of memory.\r\n\r\nIs there a way to use zero stage 3 for applications where long sequences are needed (512+)?", "Thank you for this report, @benathi \r\n\r\nFirst I just want to validate that you're referring to the setup from my most [recent comment](https://github.com/huggingface/transformers/issues/9996#issuecomment-856384448) and not the OP.\r\n\r\nSo what you're suggesting is that being able to use a largish BS was nothing but a fluke since the dataset entries happened to be quite short, correct?\r\n\r\nHave you tried using a smaller BS?\r\n\r\nAlso do you have access to a single card only?", "Yes I refer to your most recent comment. I tried 1 GPU (using A100 same as\nyou) and 2 and 8.\n\nI tried using batch size as small as 1 for length 512 (input 512 output\n512) but ran into memory issues for 1,2,8 GPUs\n\nI suspect that for it is due to memory surge during attention computation,\nwhich can be quite a lot for long sequence. Im not sure what is needed to\novercome this. I tried changing the bucket size in the config to no avail.\n\nIf I don’t use “—pad_to_max_length True”, I can run your exact script\n(input 1024 output 1024) just fine with 1,2,8 GPUs.\n\nBest,\nBen\n\nOn Thu, Sep 16, 2021 at 11:02 PM Stas Bekman ***@***.***>\nwrote:\n\n> Thank you for this report, @benathi <https://github.com/benathi>\n>\n> First I just want to validate that you're referring to the setup from my\n> most recent comment\n> <https://github.com/huggingface/transformers/issues/9996#issuecomment-856384448>\n> and not the OP.\n>\n> So what you're suggesting is that being able to use a largish BS was\n> nothing but a fluke since the dataset entries happened to be quite short,\n> correct?\n>\n> Have you tried using a smaller BS?\n>\n> Also do you have access to a single card only?\n>\n> —\n> You are receiving this because you were mentioned.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/9996#issuecomment-921417244>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AA5DMSZM2YQIB5E3BXWB2O3UCKVSXANCNFSM4XCHBJ4A>\n> .\n> Triage notifications on the go with GitHub Mobile for iOS\n> <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675>\n> or Android\n> <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.\n>\n>\n" ]

[]

[]

[ "@LysandreJik Hey, Thank you for merging my first pull request. Happy to help! That's exactly what happened, took me a second to realize, but it became pretty clear when I read through the make file." ]

[]

[]

[ "Hey @itssimon \r\n\r\nFrom a quick look at your code, it seems that the fairseq model is on GPU, but the transformers model is on CPU, which could explain the huge speed difference. Could you try running it on GPU ?\r\n", "Oh dear, how embarassing. That's it! Thanks!" ]

[]

[]

[ "This is breaking sadly: with this change someone using `trainer.model` after instantiating a `Trainer` won't have it on the GPU anymore, which will make code fail. It's also best IMO if an OOM error happens sooner rather than later.\r\n\r\nNow for deepspeed I understand why this would be necessary, so we can move the `model.to` in that case. I don't see other cases when this is useful (mixed precision with APEX and AMP keep a copy of the model in full precision)", "oh, that's no problem for now. Let's do it just for deepspeed then. Fairscale might join down the road.\r\n\r\nActually Deepspeed doesn't even need the `.to()` call at all. So it's even simpler.\r\n\r\nSo basically this skipping `.to()` is needed for all extensions that partition or tweak the model size, so MP/DeepSpeed and this will be so for PP as well.\r\n" ]

[ { "id": 2648621985, "node_id": "MDU6TGFiZWwyNjQ4NjIxOTg1", "url": "https://api.github.com/repos/huggingface/transformers/labels/Feature%20request", "name": "Feature request", "color": "FBCA04", "default": false, "description": "Request for a new feature" } ]

[]

[ "Interesting thread, thank you for posting it! You could also post it on [the forums](https://discuss.huggingface.co) to reach more users!" ]

[]

[]

[ "That's tricky. We already have way too many channels between the forums, the blog, the documentation and soon the course so I don't want to add a new one. You can create wiki posts on the forum, so maybe use that for the iterative process where you want some collaboration? We can then link those from the doc if relevant.\r\n\r\nDown the road, once such a document is stable it should be converted in a doc page though.", "\r\n> That's tricky. We already have way too many channels between the forums, the blog, the documentation and soon the course so I don't want to add a new one. You can create wiki posts on the forum, so maybe use that for the iterative process where you want some collaboration? We can then link those from the doc if relevant.\r\n\r\nOh, I was thinking not to add a new channel but re-use the available ones - I was just thinking how to link it to the main docs while it's a work in progress. \r\n\r\nI'm thinking of a much simpler approach - one of:\r\n1. transformers github wiki - would be limited to hf members - less direct input, but easier to manage\r\n2. forums wiki - would be open to all - but potentially require much more effort to manage\r\nand then linking one of these to the docs website menu - is that possible? and once the doc is strong it can migrate to a real .md doc.\r\n\r\n> Down the road, once such a document is stable it should be converted in a doc page though.\r\n\r\nThat!\r\n", "Hi @stas00 ,\r\n\r\ncould maybe have a look at: https://hackmd.io/\r\n\r\nSo you can just edit your markdown/README file, invite other collaborators and when everything is ready you could open a PR for the final submission into Transformers :)", "Thank you, @stefan-it.\r\n\r\nIt's not so much about where to collaborate on it, but how to potentially do it long term while keeping the doc easily found with all the other transformers docs, while it's a work in progress.\r\n\r\nI think the question is simple - @sgugger - would you support linking from the https://huggingface.co/transformers/ to some docs in progress until they are mature enough to import them as a normal doc? Then we can look at what would be the easiest way to collaborate.\r\n\r\nOr to keep things on the website, perhaps an iframe that remains on https://huggingface.co/transformers/ but includes the off-site doc? Not asking for anything complicated at all, whatever the easy/quick solution works. This is just an idea.", "stale" ]

[]

[]

[ "Hi @LysandreJik I was wondering does the test will be for both encoder and decoder?\r\n", "Yes, that would be for the best!" ]

[]

[ { "login": "LysandreJik", "id": 30755778, "node_id": "MDQ6VXNlcjMwNzU1Nzc4", "avatar_url": "https://avatars.githubusercontent.com/u/30755778?v=4", "gravatar_id": "", "url": "https://api.github.com/users/LysandreJik", "html_url": "https://github.com/LysandreJik", "followers_url": "https://api.github.com/users/LysandreJik/followers", "following_url": "https://api.github.com/users/LysandreJik/following{/other_user}", "gists_url": "https://api.github.com/users/LysandreJik/gists{/gist_id}", "starred_url": "https://api.github.com/users/LysandreJik/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/LysandreJik/subscriptions", "organizations_url": "https://api.github.com/users/LysandreJik/orgs", "repos_url": "https://api.github.com/users/LysandreJik/repos", "events_url": "https://api.github.com/users/LysandreJik/events{/privacy}", "received_events_url": "https://api.github.com/users/LysandreJik/received_events", "type": "User", "site_admin": false } ]

[ "Hello! It seems this is simply passing the test?", "> Hello! It seems this is simply passing the test?\r\n\r\nI made it to claim the issue and work on it. ", "@LysandreJik Lxmert requires `visual_feats` and `visual_pos` could I change the `model.config.visual_feats_dim` to a smaller value like 5 or 10. \r\nEdit: or we could use a `seed` and generate a random tensor with original `visual_feats_dim`", "@LysandreJik in this I used `np.random.seed` to fix the `visual_feats` and `visual_pos`, otherwise we could load lxmertmodel by lxmertconfig while changing `visual_feat_dim` to something manageable. \r\n\r\nWhat do you suggest?", "@LysandreJik is there something I can help with here?", "> Could you try to replace the `np.random.rand` by the `ids_tensor`\r\nwell the `visual_feats` is torch.float will the `ids_tensor` returns int32.\r\n\r\n\r\n@LysandreJik I think another alternative, is to make the `model.config.visual_feat_dim` smaller then we can have a fixed `visual_feats` \r\n\r\n\r\nwhat do you think?\r\n", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "@LysandreJik I did raise an issue to have a context manager that will fix a seed #10143, do you think it will be usefull here?", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Hello @sadakmed, sorry for taking a while to merge this. Do you mind rebasing on `master`, and running `make fixup` at the root of your clone? There's an issue with the code quality.\r\n\r\nWill merge this right after.", "Thanks a lot for your contribution @sadakmed!" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Training with constant_warmup would be an option since it does not do learning rate decay with respect to dataset size. But I am a bit afraid of end having a poor trained model after 72H of training.", "Closed since the new `dataset.set_transform()` lazy loading. Thanks!" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "Hello! Have you taken a look at how we compute the loss in the [DistilbertForTokenClassificationModel](https://github.com/huggingface/transformers/blob/master/src/transformers/models/distilbert/modeling_distilbert.py#L807-L819)? If you pass the `labels` to the model, your loss will get computed automatically. \r\n\r\nIf you want to compute your loss yourself, I would advise to copy/paste the loss computation as shown here and adapt it to your own loss!", "Hey @LysandreJik thank you for helping me out. I've implemented it and it is working perfectly fine.\r\n\r\n```\r\n Epoch | Batch | Train Loss | Val Loss | Val Acc | Elapsed \r\n----------------------------------------------------------------------\r\n 1 | 20 | 0.000000 | - | - | 273.43 \r\n 1 | 40 | 0.000000 | - | - | 262.14 \r\n 1 | 60 | 0.000000 | - | - | 258.93 \r\n 1 | 80 | 0.000000 | - | - | 266.20 \r\n 1 | 84 | 0.000000 | - | - | 50.22 \r\n----------------------------------------------------------------------\r\n 1 | - | 0.000000 | 0.299704 | 19.37 | 1201.87 \r\n----------------------------------------------------------------------\r\n\r\n\r\n Epoch | Batch | Train Loss | Val Loss | Val Acc | Elapsed \r\n----------------------------------------------------------------------\r\n 2 | 20 | 0.000000 | - | - | 273.85 \r\n 2 | 40 | 0.000000 | - | - | 264.77 \r\n 2 | 60 | 0.000000 | - | - | 263.98 \r\n 2 | 80 | 0.000000 | - | - | 263.12 \r\n 2 | 84 | 0.000000 | - | - | 50.64 \r\n----------------------------------------------------------------------\r\n 2 | - | 0.000000 | 0.230533 | 19.39 | 1207.72 \r\n----------------------------------------------------------------------\r\n```\r\n\r\nNotebook: https://colab.research.google.com/drive/1FWPEV_5eOhveiT2AQyuSYm1Ka1pgeY2f?usp=sharing\r\n\r\n\r\nWhat I noticed is\r\n\r\n1. It is taking too long even in colab. Is it usual? ( `1207.72/60 = 20mins`)\r\n2. Accuracy (sum/total even if there will be more `O` tags) is not improving that much on [wnut17train.conll](http://noisy-text.github.io/2017/files/wnut17train.conll). Is there something I might be doing wrong?", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "> If you think this still needs to be addressed please comment on this thread.\r\n\r\nYeah, need to be.", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Before merging, please take a look at the failing tests.", "> LGTM, this is indeed a clean fix. Do we know why our BlenderBot still behaves incorrectly compared to ParlAI?\r\n> \r\nI need to look deeper, by default they use FP16 and final scores are still different in order of magnitude (I'm expecting they correspond to different things), but when looking at the full beam searches they still look similar.\r\n\r\nI've done step by step debugging and scores withing the beam search are super close for a lot of steps.\r\nThis fix is the major drift that would occur pretty fast.\r\n\r\n> Regarding personas, this could probably be handled directly in the `ConversationalPipeline`?\r\n\r\nYes exactly my opinion.\r\n", "@sgugger Can you take a look please? ", "@LysandreJik figured it out. Its' because of some logic within ConversationPipeline which is invalid for `blenderbot`.\r\n\r\nComing up with a follow-up PR." ]

[]

[]

[ "@LysandreJik Need your help here. Not sure, why test cases are failing.", "opening new PR." ]

[]

[]

[ "@LysandreJik Need your help here. Not sure, why test cases are failing.", "closing this PR, due to git conflict." ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ { "login": "sgugger", "id": 35901082, "node_id": "MDQ6VXNlcjM1OTAxMDgy", "avatar_url": "https://avatars.githubusercontent.com/u/35901082?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sgugger", "html_url": "https://github.com/sgugger", "followers_url": "https://api.github.com/users/sgugger/followers", "following_url": "https://api.github.com/users/sgugger/following{/other_user}", "gists_url": "https://api.github.com/users/sgugger/gists{/gist_id}", "starred_url": "https://api.github.com/users/sgugger/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sgugger/subscriptions", "organizations_url": "https://api.github.com/users/sgugger/orgs", "repos_url": "https://api.github.com/users/sgugger/repos", "events_url": "https://api.github.com/users/sgugger/events{/privacy}", "received_events_url": "https://api.github.com/users/sgugger/received_events", "type": "User", "site_admin": false } ]

[ "For this use case, it's best to subclass the `Trainer` and override the `evaluate` method. An example of this is given for question-answering [here](https://github.com/huggingface/transformers/blob/master/examples/question-answering/trainer_qa.py) where we need to post-process the predictions using the original dataset (a bit like your use case since the attention masks will be in the dataset). The predictions returned by the Trainer are in the same order as the elements of your dataset, so you're safe with that.", "Thanks a lot Sylvain.\r\nActually I realised that the `predictions.label_ids` coming from the training loop were padded with the value `-100`. By using the same padding value in my preprocessing, I can recover `the attention_mask` by putting a threshold.\r\nCheers ", "Oh even easier then! Can we close the issue?", "> For this use case, it's best to subclass the `Trainer` and override the `evaluate` method. An example of this is given for question-answering [here](https://github.com/huggingface/transformers/blob/master/examples/question-answering/trainer_qa.py) where we need to post-process the predictions using the original dataset (a bit like your use case since the attention masks will be in the dataset). The predictions returned by the Trainer are in the same order as the elements of your dataset, so you're safe with that.\r\n\r\nHello, can I ask where I can find the proper example again? Because the example link you mentioned has an error with 404 now. @sgugger ", "Hi @saekomdalkom, you can find it here: https://github.com/huggingface/transformers/blob/main/examples/pytorch/question-answering/trainer_qa.py", "> Hi @saekomdalkom, you can find it here: https://github.com/huggingface/transformers/blob/main/examples/pytorch/question-answering/trainer_qa.py\r\n\r\nThanks a lot!!" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "Reopened as this might still be in the works.", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Unstale", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "@abhi1thakur - should we still try to merge this PR? ", "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.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[]

[]

[]

[ "You shouldn't unsqueeze your labels, because the `labels` should just be a tensor of shape `(batch_size,)`. " ]

[]

[]

[]

[]

[]

[ "And remove a repeated sentence in README" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "Hello! Why did you close your branch? The integration test looks good, you only need to run `make fixup` at the root of your clone to apply the quality requirements.", "@LysandreJik Hi, sorry I was getting an error when i ran 'make fixup' and I was trying to figure it out. Ill finish it up tonight, unless you know whats wrong? Thank you for responding.\r\n\r\n```\r\n File \"utils/get_modified_files.py\", line 28\r\n modified_files = subprocess.check_output(f\"git diff --name-only {fork_point_sha}\".split()).decode(\"utf-8\").split()\r\n ^\r\nSyntaxError: invalid syntax\r\nNo library .py files were modified\r\n File \"setup.py\", line 192\r\n entries = \"\\n\".join([f' \"{k}\": \"{v}\",' for k, v in deps.items()])\r\n ^\r\nSyntaxError: invalid syntax\r\n```" ]

[ { "id": 1905493434, "node_id": "MDU6TGFiZWwxOTA1NDkzNDM0", "url": "https://api.github.com/repos/huggingface/transformers/labels/dependencies", "name": "dependencies", "color": "0366d6", "default": false, "description": "Pull requests that update a dependency file" } ]

[]

[ "Looks like bleach is no longer a dependency, so this is no longer needed." ]

[ { "id": 2690307185, "node_id": "MDU6TGFiZWwyNjkwMzA3MTg1", "url": "https://api.github.com/repos/huggingface/transformers/labels/Performance", "name": "Performance", "color": "207F32", "default": false, "description": "" }, { "id": 2796628563, "node_id": "MDU6TGFiZWwyNzk2NjI4NTYz", "url": "https://api.github.com/repos/huggingface/transformers/labels/WIP", "name": "WIP", "color": "234C99", "default": false, "description": "Label your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progress" } ]

[ { "login": "stas00", "id": 10676103, "node_id": "MDQ6VXNlcjEwNjc2MTAz", "avatar_url": "https://avatars.githubusercontent.com/u/10676103?v=4", "gravatar_id": "", "url": "https://api.github.com/users/stas00", "html_url": "https://github.com/stas00", "followers_url": "https://api.github.com/users/stas00/followers", "following_url": "https://api.github.com/users/stas00/following{/other_user}", "gists_url": "https://api.github.com/users/stas00/gists{/gist_id}", "starred_url": "https://api.github.com/users/stas00/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/stas00/subscriptions", "organizations_url": "https://api.github.com/users/stas00/orgs", "repos_url": "https://api.github.com/users/stas00/repos", "events_url": "https://api.github.com/users/stas00/events{/privacy}", "received_events_url": "https://api.github.com/users/stas00/received_events", "type": "User", "site_admin": false } ]

[ "I did a quick benchmark, with `--sharded_ddp --fp16` and just `--fp16` and there is no visible difference . Perhaps it is more visible in a different kind of training/model combination.\r\n\r\nTesting HF `AdamW` vs. `torch.optim._multi_tensor.AdamW`\r\n\r\n```\r\n# benchmark with just --fp16\r\n\r\n# baseline HF `AdamW`\r\n\r\nexport BS=16; rm -r output_dir; PYTHONPATH=../../src USE_TF=0 python -m torch.distributed.launch --nproc_per_node=2 ./finetune_trainer.py --model_name_or_path t5-large --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_train --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_train_batch_size $BS --sortish_sampler --task translation_en_to_ro --warmup_steps 500 --n_train 20000 --fp16\r\n\r\n{'train_runtime': 226.5618, 'train_samples_per_second': 2.759, 'epoch': 1.0}\r\n\r\n# w/ torch.optim._multi_tensor.AdamW\r\n\r\nexport BS=16; rm -r output_dir; PYTHONPATH=../../src USE_TF=0 python -m torch.distributed.launch --nproc_per_node=2 ./finetune_trainer.py --model_name_or_path t5-large --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_train --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_train_batch_size $BS --sortish_sampler --task translation_en_to_ro --warmup_steps 500 --n_train 20000 --fp16\r\n\r\n{'train_runtime': 226.1715, 'train_samples_per_second': 2.763, 'epoch': 1.0}\r\n```\r\n\r\nThe change I did was:\r\n```\r\n--- a/examples/seq2seq/seq2seq_trainer.py\r\n+++ b/examples/seq2seq/seq2seq_trainer.py\r\n@@ -24,7 +24,6 @@ from transformers.integrations import is_fairscale_available\r\n from transformers.models.fsmt.configuration_fsmt import FSMTConfig\r\n from transformers.optimization import (\r\n Adafactor,\r\n- AdamW,\r\n get_constant_schedule,\r\n get_constant_schedule_with_warmup,\r\n get_cosine_schedule_with_warmup,\r\n@@ -32,6 +31,7 @@ from transformers.optimization import (\r\n get_linear_schedule_with_warmup,\r\n get_polynomial_decay_schedule_with_warmup,\r\n )\r\n+from torch.optim._multi_tensor import AdamW\r\n from transformers.trainer_pt_utils import get_tpu_sampler\r\n from transformers.training_args import ParallelMode\r\n```\r\n\r\nand this is from pytorch-nightly from today.\r\n", "you must have a really strange bottleneck in that test, neither the latest fairscale nor these are changing anything ? These optimizers are measurably faster in isolation, and sure enough we see a difference in fairscale CI, even on a dummy job / small model ([see for instance, two last jobs](https://app.circleci.com/pipelines/github/facebookresearch/fairscale/1522/workflows/e95cd0af-9582-4021-8176-beafa306f147/jobs/7130))", "testing with the same command, I see a vastly varying throughput depending on `num_train_epochs`, which seems a bit strange to me", "To share with others, @blefaudeux and his team made speed improvements in fairscale (master) recently, which should have been quite visible, but a few days ago we tested this same script with `--sharded_ddp` and saw no improvement whatsoever. So something odd is going on.", "I will leave this issue open for now as an incentive to profile this script and identify the bottleneck.", "@stas00 Do you think this should be revisited given the [discussion](https://github.com/pytorch/pytorch/issues/71274) in upstream PyTorch?", "Yes, I was just about to revisit it. \r\n\r\nedit: I thought you might have wanted to work on that, but the pytorch team asks to run a profiler on it and all, so I probably will look into testing it out again.\r\n\r\n--- original comment --- \r\n\r\nDo you want to take a lead on this experiment, @jaketae?\r\n\r\nThe new `--optim` HF Trainer just got merged, so you can quickly implement `--optim adamw_torch_multi_tensor` in the same way `--optim adamw`\r\n\r\nYou can use this tool for benchmarking https://github.com/huggingface/transformers/pull/14934 if it helps. I think it's pretty stable now, I will propose to PR it.\r\n\r\n" ]