Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[ { "login": "patil-suraj", "id": 27137566, "node_id": "MDQ6VXNlcjI3MTM3NTY2", "avatar_url": "https://avatars.githubusercontent.com/u/27137566?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patil-suraj", "html_url": "https://github.com/patil-suraj", "followers_url": "https://api.github.com/users/patil-suraj/followers", "following_url": "https://api.github.com/users/patil-suraj/following{/other_user}", "gists_url": "https://api.github.com/users/patil-suraj/gists{/gist_id}", "starred_url": "https://api.github.com/users/patil-suraj/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patil-suraj/subscriptions", "organizations_url": "https://api.github.com/users/patil-suraj/orgs", "repos_url": "https://api.github.com/users/patil-suraj/repos", "events_url": "https://api.github.com/users/patil-suraj/events{/privacy}", "received_events_url": "https://api.github.com/users/patil-suraj/received_events", "type": "User", "site_admin": false } ]

[ "@patil-suraj any help is appreciated", "@patrickvonplaten", "Can anyone look into this @patil-suraj @patrickvonplaten ", "Hi @tuhinjubcse , sorry to reply only now, I've been a bit busy with the sprint and other projects so couldn't really allocate any time for this. I will get back to you by tomorrow.\r\n\r\nAlso please don't tag people who are not related to this model, it might disturb them unnecessarily.\r\n\r\nThank you for your patience.", "Thank you, it would be good to know how to finetune a many to many models with more than one lang pairs in train and validation like fairseq multilingual\r\n\r\nhttps://github.com/pytorch/fairseq/tree/master/examples/multilingual", "Okay, one issue at a time\r\n\r\nI'm taking a look at the error that you posted above.\r\n\r\nAlso, the many-to-one model was not released when we ported this model to `Transformers`, it seems to have been released recently. I will convert and push it by tomorrow.\r\n\r\nAnd regarding multi-lingual fine-tuning, I will try to write a notebook about it. What we need to do here is, say we are fine-tuning on two language pairs, in that case, we need to concatenate the two datasets or in case the two language pairs don't have the same number of examples then add some sort of sampler which will sample the example from the datasets depending on the number of examples in which one. And when processing each language pair, set the appropriate `src_lang` and `tgt_lang` tokens. The processing part is explained in the [docs](https://huggingface.co/transformers/model_doc/mbart.html#training-of-mbart-50).", "That would be really helpful if you can have a notebook which documents how to do that , or even a read me , just so that its clear", "Thanks so much for your response and looking forward to use it ", "The many to one checkpoint is now available on the hub\r\nhttps://huggingface.co/facebook/mbart-large-50-many-to-one-mmt", "Thanks for releasing this. Looking forward to the instructions to do many to one finetuning as that is what this model will be superuseful for", "Any updates on how to run many to one, can we pass --source_lang ru_RU,es_XX as a ',' separated string. Sorry I am not sure if that support is available yet. Would be really helpful if you could help here. The EMNLP arxiv deadline is super close on 17th April :) I know you are busy but this would be a huge favor\r\n", "Multilingual fine-tuning won't be included in the example script, the goal of examples is to keep them simple and let the user extend them for custom training. I'm working on the notebook, but can probably share that on Monday.\r\n\r\nAs I said in the above comment, for multilingual fine-tuning, in the simplest case you would just need to process the two datasets by setting correct `src_lang`, `tgt_lang` tokens, the rest of the training will be similar to traditional fine-tuning.\r\n\r\nFeel free to post the question on the [forum](https://discuss.huggingface.co/) as well, someone there might have better ideas for this.", "Thank you so much, if you post the notebook here by Monday that would solve my problem. I am trying on my own to do it as well", "Hi @tuhinjubcse \r\n\r\nWe just merged #11170 which now allows to fine-tune mBART-50 on **single language pair** using the `run_translation.py` script. This should resolve the issue that you posted in the first comment.", "Thanks so much", "Suraj I got multilingual to work, however, while decoding I get this error. My added token dictionary is\r\n\r\n\r\n`{\"uk_UA\": 250049, \"mk_MK\": 250036, \"mn_MN\": 250038, \"id_ID\": 250033, \"he_IL\": 250031, \"sl_SI\": 250053, \"pt_XX\": 250042, \"hr_HR\": 250032, \"th_TH\": 250047, \"tl_XX\": 250048, \"pl_PL\": 250040, \"ka_GE\": 250034, \"ta_IN\": 250045, \"km_KH\": 250035, \"te_IN\": 250046, \"xh_ZA\": 250051, \"sv_SE\": 250043, \"sw_KE\": 250044, \"ps_AF\": 250041, \"bn_IN\": 250029, \"ml_IN\": 250037, \"az_AZ\": 250027, \"af_ZA\": 250028, \"gl_ES\": 250052, \"ur_PK\": 250050, \"mr_IN\": 250039, \"fa_IR\": 250030}\r\n`\r\n\r\n File \"translate.py\", line 26, in <module>\r\n tokenizer = MBart50Tokenizer.from_pretrained(path)\r\n File \"/home/tuhin.chakr/yes/lib/python3.8/site-packages/transformers/tokenization_utils_base.py\", line 1704, in from_pretrained\r\n return cls._from_pretrained(\r\n File \"/home/tuhin.chakr/yes/lib/python3.8/site-packages/transformers/tokenization_utils_base.py\", line 1810, in _from_pretrained\r\n assert index == len(tokenizer), (\r\nAssertionError: Non-consecutive added token 'bn_IN' found. Should have index 250054 but has index 250029 in saved vocabulary.\r\n\r\n\r\nThe error comes from MBart50Tokenizer\r\nmodel = MBartForConditionalGeneration.from_pretrained(path)\r\nmodel.eval()\r\nmodel.to('cuda')\r\ntokenizer = MBart50Tokenizer.from_pretrained(path)\r\n\r\nIt works fine with MBartTokenizer\r\n\r\nI can use MBartTokenizer for common languages in mbart25 and mbart50 for my manytoone model but for languages like pt_XX i can't .", "HI @tuhinjubcse \r\n\r\nGlad you got it working.\r\n\r\nAnd this seems like a bug, I will take a look. How many new tokens did you add?", "I tried adding tokens using the `add_tokens` and `add_special_tokens` method, saved and loaded it again, I didn't observe this issue.\r\n\r\nHere's what I did\r\n```python\r\ntok = MBart50Tokenizer.from_pretrained(\"facebook/mbart-large-50\")\r\n\r\ntok.add_special_tokens({\"MY_XX\": \"MY_XX\"})\r\ntok.add_special_tokens({\"additional_special_tokens\": [\"MY2_XX\"]})\r\n\r\ntok.save_pretrained(\"./tmp\")\r\n\r\ntok = MBart50Tokenizer.from_pretrained(\"./tmp\")\r\ntok.convert_tokens_to_ids(\"MY_XX\") # 250054\r\ntok.convert_tokens_to_ids(\"MY2_XX\") # 250055\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.", "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." ]

[]

[]

[ "@patil-suraj is working on implementing GPT Neo over at https://github.com/huggingface/transformers/pull/10848! The 1.3B and 2.7B should be loadable in that architecture once finalized." ]

[]

[]

[ "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.", "One possible reason for that is bigger tokenizer cardinality and the fact that HF code predicts all tokens, not only one that are missing, so this operation might dominate overall runtime/memory while doing no useful output.\r\n\r\nIn my case it was the case and I did a small fix here https://github.com/yurymalkov/transformers/commit/0fe0725c0f7fcc13df698bba1bd01847c1494e43 that ended up in 6X larger batch at my memory usage (1060 GTX) and 3-4X times faster training of a small mbert model, but it causes some tests to fail (I think mainly due to pytorch jit failure which sees named tensor in slice, secondly to missing output of non-masked tokens - though who cares about them?)." ]

[]

[]

[ "Hi there! So the problem is a bit complex and linked to the way RoBERTa is implemented in Transformers with a small hack: its toknizer has 512 + 2 position embeddings, not 512. When you run your command, the model is randomly initialized with 512 position embeddings (the default in the config) but you still use it with that `robert-base` tokenizer which returns up to 514. This results in an index error that throws the \"device-side assert triggered\".\r\n\r\nTo fix this, you need to either use another tokenizer, or prepare your random model like this:\r\n```\r\nfrom transformers import RobertaForMaskedLM, RobertaConfig\r\n\r\nmodel = RobertaForMaskedLM(RobertaConfig(max_position_embeddings=514))\r\nmodel.save_pretrained(\"model_dir\")\r\n```\r\nthen use `model_dir` for `--model_name_or_path` when launching your script.\r\n\r\nYou can also tweak the script directly to add `max_position_embeddings=514` in [this line](https://github.com/huggingface/transformers/blob/master/examples/language-modeling/run_mlm.py#L282).", "Thank you! Now it works! :)" ]

[]

[]

[ "Hi! You're using two `bert-base-uncased` as encoder/decoders. This is possible, but you'll need to train your resulting encoder-decoder model on a downstream task in order to obtain coherent results.\r\n\r\nThe `bert-base-uncased` checkpoint is originally from an encoder-only setup.\r\n\r\nIf I may recommend some notebooks/documentation:\r\n- [Documentation of encoder/decoder framework](https://huggingface.co/transformers/model_doc/encoderdecoder.html)\r\n- [Training a Bert2Bert model for summarization](https://github.com/patrickvonplaten/notebooks/blob/master/BERT2BERT_for_CNN_Dailymail.ipynb)\r\n- [Training a shared Roberta2Roberta for summarization](https://github.com/patrickvonplaten/notebooks/blob/master/RoBERTaShared_for_BBC_XSum.ipynb)", "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\nI also observe the similar issue with mt5 models, https://github.com/huggingface/transformers/issues/10819 , deepspeed is still not working for me due to this issue with mt5 models.\r\nI greatly appreciate having a look @patil-suraj @patrickvonplaten ", "We didn't really manage to resolve the problems with t5/mt5 + mixed precision fp16 (cc @patil-suraj). I'm not sure whether anybody has tried internally to fine-tune t5/mt5 with deepspeed (@stas00 maybe?)", "the issue arises without deepspeed, just vanilla mt5-small model. Also, I see similar nans with deepspeed with a model based on mt5-small slightly modified, please see the issue here https://github.com/huggingface/transformers/issues/10821#issuecomment-803453998, I think if the issue with fp16 option could get resolved, hopefully this will be also more stable with model changes in deepspeed as well. Thanks a lot.", "Indeed, this has nothing to do with deepspeed, other than that deepspeed trains in mixed precision and evals in full fp16 at the moment.\r\n\r\nI've started studying the bfloat16 vs. float16 numerical properties and their correlation to each other. And once I understand it well I will try to see if there some sort of magical remapping that perhaps could be done - this is my fantasy of course. I just need to finish a few other more urgent things with deepspeed stage3 integration first.\r\n\r\nBut please don't let my comment prevent you from merging the proposed fix if it already solves the problem.\r\n", "I got similar issue with mt5 model, @patrickvonplaten thanks a lot in advance for your help", "@dorost1234 + @yuvalkirstain, please kindly try this branch:\r\nhttps://github.com/huggingface/transformers/tree/t5-fp16-no-nans\r\nand let me know if it solves the problem - It seems that the problem is due to `autocast` in `T5LayerFF` so this branch tries to turn off `autocast` just for that layer. It also disables the previously added clamping.\r\n\r\nThere is also a lot of debug statements in the branch but they will be silent unless nan/inf is detected. \r\n\r\nI tested it work on a small sample with t5-small/t5-base/t5-large/google/mt5-small.\r\n\r\nThe main part of the fix is just:\r\n```\r\nclass T5LayerFF(nn.Module):\r\n def forward(self, hidden_states):\r\n with torch.cuda.amp.autocast(enabled=False):\r\n forwarded_states = self.layer_norm(hidden_states)\r\n forwarded_states = self.DenseReluDense(forwarded_states)\r\n hidden_states = hidden_states + self.dropout(forwarded_states)\r\n return hidden_states\r\n```\r\nand removing some code. So use the branch first.\r\n\r\nIf it works I guess we could just monkey patch this version for AMP or come up with some cleaner solution. Probably with `torch.is_autocast_enabled()` check", "Dear @stas00 \r\nThank you very much for taking time looking into this issue, this would be really awesome if this could fix the issue, I tried to test it, for this I got the branch, and then I install it locally with \"python setup.py develop\", then I run this command:\r\n\r\n`python run_translation.py --model_name_or_path google/mt5-small --do_train --do_eval --source_lang en --target_lang ro --dataset_name wmt16 --dataset_config_name ro-en --output_dir /temp/test --per_device_train_batch_size=4 --per_device_eval_batch_size=4 --overwrite_output_dir --predict_with_generate --logging_step 10 --fp16`\r\n\r\nI got this error:\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"run_translation.py\", line 562, in <module>\r\n main()\r\n File \"run_translation.py\", line 448, in main\r\n pad_to_multiple_of=8 if training_args.fp16 else None,\r\nTypeError: __init__() got an unexpected keyword argument 'model'\r\n```\r\nI think there is some version mismatch. I removed the model from input to the collator, as below\r\n\r\n```\r\n data_collator = DataCollatorForSeq2Seq(\r\n tokenizer,\r\n #model=model,\r\n label_pad_token_id=label_pad_token_id,\r\n pad_to_multiple_of=8 if training_args.fp16 else None,\r\n )\r\n\r\n```\r\n\r\nand then here is what I got with fp16 option:\r\n\r\n```\r\n{'loss': 23.3523, 'learning_rate': 4.999890767684712e-05, 'epoch': 0.0} \r\n{'loss': 22.5557, 'learning_rate': 4.999781535369424e-05, 'epoch': 0.0} \r\n{'loss': 25.9471, 'learning_rate': 4.999672303054136e-05, 'epoch': 0.0} \r\n{'loss': 23.0994, 'learning_rate': 4.9995630707388475e-05, 'epoch': 0.0} \r\n{'loss': 24.9974, 'learning_rate': 4.999453838423559e-05, 'epoch': 0.0} \r\n{'loss': 23.3743, 'learning_rate': 4.999344606108271e-05, 'epoch': 0.0} \r\n{'loss': 24.2147, 'learning_rate': 4.999235373792983e-05, 'epoch': 0.0} \r\n{'loss': 26.7845, 'learning_rate': 4.9991261414776954e-05, 'epoch': 0.0} \r\n{'loss': 25.2277, 'learning_rate': 4.9990169091624065e-05, 'epoch': 0.0} \r\n{'loss': 23.3156, 'learning_rate': 4.998907676847119e-05, 'epoch': 0.0} \r\n{'loss': 21.275, 'learning_rate': 4.99879844453183e-05, 'epoch': 0.0} \r\n{'loss': 23.7031, 'learning_rate': 4.9986892122165426e-05, 'epoch': 0.0} \r\n{'loss': 23.8086, 'learning_rate': 4.9985799799012544e-05, 'epoch': 0.0} \r\n{'loss': 25.8143, 'learning_rate': 4.998470747585966e-05, 'epoch': 0.0} \r\n{'loss': 24.4319, 'learning_rate': 4.998361515270678e-05, 'epoch': 0.0} \r\n{'loss': 26.8277, 'learning_rate': 4.99825228295539e-05, 'epoch': 0.0} \r\n```\r\n\r\nhere is loss without fp16:\r\n```\r\n{'loss': 27.0258, 'learning_rate': 4.999890767684712e-05, 'epoch': 0.0} \r\n{'loss': 23.141, 'learning_rate': 4.999781535369424e-05, 'epoch': 0.0} \r\n{'loss': 21.2312, 'learning_rate': 4.999672303054136e-05, 'epoch': 0.0} \r\n{'loss': 19.3567, 'learning_rate': 4.9995630707388475e-05, 'epoch': 0.0} \r\n{'loss': 18.7998, 'learning_rate': 4.999453838423559e-05, 'epoch': 0.0} \r\n{'loss': 17.9632, 'learning_rate': 4.999344606108271e-05, 'epoch': 0.0} \r\n{'loss': 17.2105, 'learning_rate': 4.999235373792983e-05, 'epoch': 0.0} \r\n{'loss': 17.5506, 'learning_rate': 4.9991261414776954e-05, 'epoch': 0.0} \r\n{'loss': 15.2566, 'learning_rate': 4.9990169091624065e-05, 'epoch': 0.0} \r\n{'loss': 14.8667, 'learning_rate': 4.998907676847119e-05, 'epoch': 0.0} \r\n{'loss': 13.7132, 'learning_rate': 4.99879844453183e-05, 'epoch': 0.0} \r\n{'loss': 13.4058, 'learning_rate': 4.9986892122165426e-05, 'epoch': 0.0\r\n```\r\n\r\nSo I think this is not optimizing the loss well. I greatly appreciate having a look. Thanks a lot.\r\n", "re errors - this is all on master - the source code and `run_translation.py`. When you install `pip install -e .` sometimes conda/pip don't clean up an old install, so it helps to do `pip uninstall transformers -y` at least 2 times!\r\n\r\nI solve such problems by running locally and not relying on the installed `transformers`, i.e.:\r\n```\r\ngit clone https://github.com/huggingface/transformers\r\ncd transformers\r\nPYTHONPATH=src python examples/seq2seq/run_translation.py ...\r\n```\r\n\r\nnow you never need to worry about what `transformers` version is installed in the environment.\r\n\r\nwrt not getting the loss going down - this is odd, I just run your code:\r\n```\r\nPYTHONPATH=src python examples/seq2seq/run_translation.py --model_name_or_path google/mt5-small --do_train --do_eval --source_lang en --target_lang ro --dataset_name wmt16 --dataset_config_name ro-en --output_dir /tmp/test --per_device_train_batch_size=4 --per_device_eval_batch_size=4 --overwrite_output_dir --predict_with_generate --logging_step 10 --fp16\r\n\r\n{'loss': 29.7519, 'learning_rate': 4.999781535369424e-05, 'epoch': 0.0} \r\n{'loss': 26.3593, 'learning_rate': 4.9995630707388475e-05, 'epoch': 0.0} \r\n{'loss': 23.4431, 'learning_rate': 4.999344606108271e-05, 'epoch': 0.0} \r\n{'loss': 21.431, 'learning_rate': 4.9991261414776954e-05, 'epoch': 0.0} \r\n{'loss': 19.2445, 'learning_rate': 4.998907676847119e-05, 'epoch': 0.0} \r\n{'loss': 17.8293, 'learning_rate': 4.9986892122165426e-05, 'epoch': 0.0} \r\n{'loss': 16.9441, 'learning_rate': 4.998470747585966e-05, 'epoch': 0.0} \r\n{'loss': 15.7572, 'learning_rate': 4.99825228295539e-05, 'epoch': 0.0}\r\n{'loss': 15.2937, 'learning_rate': 4.9980338183248135e-05, 'epoch': 0.0}\r\n{'loss': 14.4368, 'learning_rate': 4.997815353694237e-05, 'epoch': 0.0}\r\n{'loss': 14.6709, 'learning_rate': 4.997596889063661e-05, 'epoch': 0.0}\r\n{'loss': 13.2806, 'learning_rate': 4.9973784244330843e-05, 'epoch': 0.0}\r\n{'loss': 12.9245, 'learning_rate': 4.997159959802508e-05, 'epoch': 0.0}\r\n{'loss': 12.4647, 'learning_rate': 4.9969414951719316e-05, 'epoch': 0.0}\r\n{'loss': 11.4738, 'learning_rate': 4.996723030541355e-05, 'epoch': 0.0}\r\n```\r\n\r\nMust be your hardware? Try to lower the learning rate?\r\n\r\nI tried with 1 or 2 gpus and it worked in both cases.\r\n\r\n", "Hi @stas00 \r\nthank you very much for the pointers, I did it as you mentioned and now I see this is going down nicely\r\n\r\n```\r\n{'loss': 28.1802, 'learning_rate': 4.999890767684712e-05, 'epoch': 0.0} \r\n{'loss': 27.4353, 'learning_rate': 4.999781535369424e-05, 'epoch': 0.0} \r\n{'loss': 21.3904, 'learning_rate': 4.999672303054136e-05, 'epoch': 0.0} \r\n{'loss': 22.8854, 'learning_rate': 4.9995630707388475e-05, 'epoch': 0.0} \r\n{'loss': 19.6943, 'learning_rate': 4.999453838423559e-05, 'epoch': 0.0} \r\n{'loss': 21.253, 'learning_rate': 4.999344606108271e-05, 'epoch': 0.0} \r\n{'loss': 20.1937, 'learning_rate': 4.999235373792983e-05, 'epoch': 0.0} \r\n{'loss': 18.6606, 'learning_rate': 4.9991261414776954e-05, 'epoch': 0.0} \r\n{'loss': 18.0337, 'learning_rate': 4.9990169091624065e-05, 'epoch': 0.0} \r\n{'loss': 16.1259, 'learning_rate': 4.998907676847119e-05, 'epoch': 0.0} \r\n{'loss': 15.4007, 'learning_rate': 4.99879844453183e-05, 'epoch': 0.0} \r\n{'loss': 15.6753, 'learning_rate': 4.9986892122165426e-05, 'epoch': 0.0} \r\n{'loss': 15.0481, 'learning_rate': 4.9985799799012544e-05, 'epoch': 0.0} \r\n{'loss': 14.5833, 'learning_rate': 4.998470747585966e-05, 'epoch': 0.0} \r\n{'loss': 14.0758, 'learning_rate': 4.998361515270678e-05, 'epoch': 0.0} \r\n{'loss': 13.7096, 'learning_rate': 4.99825228295539e-05, 'epoch': 0.0} \r\n{'loss': 13.3216, 'learning_rate': 4.998143050640102e-05, 'epoch': 0.0} \r\n{'loss': 13.2331, 'learning_rate': 4.9980338183248135e-05, 'epoch': 0.0} \r\n{'loss': 12.1556, 'learning_rate': 4.997924586009525e-05, 'epoch': 0.0} \r\n```\r\n\r\nThis is such a great, wonderful, amazing fix. Looking forward to using it when this is pushed to the repository.\r\nFor all the hard problems, you are our only hope @stas00 \r\nThank you very much for this great fix.\r\n", "Thank you for your kind words, I'm so happy to hear that it worked, @dorost1234.\r\n\r\nI will make a proper PR after I clean this branch up.", "@yuvalkirstain, please kindly test if this PR fixes the problem: https://github.com/huggingface/transformers/pull/10956", "Thank you @stas00 !\r\nIt seems to work were my proposed fix failed with T5-Small. I will now run some additional experiments with T5-Large and update.", "Thank you for validating that, @yuvalkirstain!\r\n\r\nIndeed, I tried first local fixes but the problem would just pop-up elsewhere. \r\n\r\nI'm just thinking that perhaps we could find if it's all calls to FF that lead to the problem or only some of them, and then we could optimize the solution I proposed by only disabling `autocast` in some cases and not all. I haven't tested that yet.\r\n\r\nIf you experiment I recommend for you to try my branch, since I left the \"detector\" on and it'll immediately tell you when the first `inf` is encountered.\r\n\r\nWhat I'm most interested in is some longer runs to ensure it doesn't start overflowing at a later point. \r\n\r\nThank you for your contribution.", "Finetuned T5-Base using this branch with the standard T5 finetuning HPs on NQ (except from batch_size - used only ~26k tokens) and didn't get nans (it has been running for over 3 hours and training converged). Thanks again, I guess the issue can be closed for time being.", "Thank you for this validation, @yuvalkirstain. I still would like to see if we can find a more efficient solution before merging it, but this is great that we have one that works.\r\n\r\nThis unfortunately doesn't help with deepspeed since it doesn't use pytorch AMP and has its own version, but which doesn't use context manager so can't be turned off locally like `autocast`. So we hope to find a different solution.\r\n\r\nI linked this issue to the PR so it'll get closed automatically when it's merged.", "Well, the nans are back.\r\n\r\n`T5LayerFF: 1 has inf\r\nT5LayerNorm has inf\r\nT5LayerNorm variance has inf\r\nT5LayerNorm hidden_states has nans\r\nT5LayerNorm hidden_states before return has nans\r\nT5LayerFF: 2 has nans\r\nT5LayerFF: 3 has nans\r\nT5LayerFF: 5 has nans\r\nT5Block after T5LayerFF has nans\r\nT5Stack loop end has nans\r\nT5LayerNorm has nans\r\nT5LayerNorm variance has nans\r\nT5LayerNorm hidden_states has nans\r\nT5LayerNorm hidden_states before return has nans`\r\n\r\nThe model I used here was T5-large-ssm-nqo.\r\n@stas00 If you'd like to replicate I can send the relevant training file + command.", "Yes, please, I'm working in parallel on gpt-neo that has the same issues, so the more reproducible cases we have the higher are the chances we can find a solid fix. \r\n\r\nAlso those would be good candidates for tests (hoping that we can find a quick way to get to overflow).", "Let's continue the discussion in the PR that is trying to solve this issue: https://github.com/huggingface/transformers/pull/10956", "@dorost1234 hI, Could you please tell me how you solved this loss optimization problem. I am facing same issue", "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.", "So is this fix now in the main version of transformers?", "I found that results are different when you load like this: (first is better) \r\n\r\nmodel1a_CPU = T5ForConditionalGeneration.from_pretrained(best_model_path, low_cpu_mem_usage=True,torch_dtype=torch.float16).to(\"cuda\")\r\n\r\nthan when you load via:\r\n\r\nmodel1a_CPU = T5ForConditionalGeneration.from_pretrained(best_model_path, low_cpu_mem_usage=True)\r\nmodel1a_CPU.half()\r\nmodel1a_CPU.eval() \r\nmodel1a_CPU.to(\"cuda\")\r\n\r\nSo this could be a solution, I will compare result on /CPU versus /This versus /Half\r\n\r\n\r\n\r\n\r\n", "@seems like the solution is already implemented in this call: (model1a_CPU = T5ForConditionalGeneration.from_pretrained(best_model_path, low_cpu_mem_usage=True,torch_dtype=torch.float16).to(\"cuda\"))\r\n\r\nProbably it is trigered by torch_dtype=torch.float16. So a part of model is (likely) moved to fp32 from fp16, so it works properly, exactly the same as with FP32, and exactly the same as on CPU. \r\n\r\nOf course it does use a little bit more of memory. When you call it second way, the memory usage is around 2.5 GB for T5-large, while with first it is around 2.9GB. It is slower around 10-15 percent.\r\n\r\n\r\n\r\n\r\n\r\n\r\n" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "I'm realizing now it would be so helpful with xlsr trainings (lost models due to crash after very long training).\r\n\r\nWould you have any input or suggestions @sgugger on how I could try to implement it?", "Hi @borisdayma, I hadn't seen this issue. On our side we're more focused on how to continuously push the checkpoints to our own hub ;-).\r\n\r\nThat being said, we can definitely leverage the `on_save` event and just look for the last `checkpoint-xxx` folder, then push its content as artifact. If you have a tracking with `metric_for_best_model`, then you won't even have to look for the checkpoint, it will be in the state with `state.best_model_checkpoint`.\r\nAs for having one or separate checkpoints, I guess it really depends on what you think is best for WandB, you have more expertise than me here.", "Thanks, it makes sense!\r\nActually I imagined you may probably have some interest in a similar logic for the model hub so I wanted to work on something that would be useful for everyone.\r\n\r\nAs a side note, models could also be stored on the model hub AND tracked by W&B (some people do it with ASW S3 for example). In this way, only checksums are actually stored and the files point back to the storage space so it could be complementary.", "To do the same on the hub, my idea was to leverage the versioning system and just push the saved checkpoint every save with a commit message like \"checkpoint step xxx\". Ideally inside a Callback to avoid adding more stuff to the main training loop.\r\nI'll try to focus on this next week and see what we can easily do!", "Nice, if possible it would be cool to allow a logger method to be called right after you push your checkpoint.", "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.", "Still interested in working on it! Let me know when you have a hook for the model hub!", "We have started working on it (the Trainer gained a push_to_hub API!) the last step missing is the continuous checkpointing, will try to work on that last part soon!", "Awesome, if you can somehow have a hook after pushing to the hub (with access to the url maybe) then we could link link them to the runs.", "The call to `Trainer.push_to_hub` returns the url of the commit to the model hub (it's not part of the train for now).", "I really like where this is going!\r\n\r\nIs the goal for `TrainingArguments.push_to_hub` to be eventually directly used by the `Trainer` or will it always be handled by the scripts?\r\nAlso would it be possible to save the return of `_push_to_hub` somewhere in the `Trainer` (that way that url could be used by wandb).", "Yes, we can definitely save the URL somewhere! Would you like to make a PR with that?\r\n\r\nI'm on another project that we will release soon right now but also plan to go back to the continuous integration after (should be in two weeks!)", "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 think I don't fully agree here...`max_position_embeddings` is important even if `axial_pos_embedding` is used. If a user makes use of `axial_pos_embeddings` then `max_position_embeddings` is clearly defined by the tuple of `axial_pos_embeddings`. IMO, it's important that the user fully understands how `axial_pos_embeddigs` work and should therefore also set `max_position_embeddigns`", "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." ]

[]

[]

[ "This appears to be the proper way to run it.\r\n\r\n`python3 -m transformers.convert_graph_to_onnx --framework pt --model bert-base-cased bert-base-cased.onnx`" ]

[]

[]

[]

[]

[]

[]

[ { "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 } ]

[ "I'm able to reproduce your issue, @dorost1234 \r\n\r\nSo this particular error happens because you change something in the model structure after the checkpoint is saved and before it's resumed. For example at the resume point your model is different than what it was just before you saved the checkpoint.\r\n\r\nThe reason you encountered this problem with deepspeed and did not without it, is because deepspeed by default saves the optimizer state and resumes from it. So here is a quick hack to overcome this in short-term:\r\n\r\n```\r\n--- a/seq2seq/third_party/trainers/trainer.py\r\n+++ b/seq2seq/third_party/trainers/trainer.py\r\n@@ -1166,7 +1166,7 @@ class Trainer:\r\n\r\n if self.deepspeed:\r\n # Not sure how to check if there is a saved deepspeed checkpoint, but since it just return None if it fails to find a deepspeed checkpoint this is sort of a check-n-load function\r\n- self.deepspeed.load_checkpoint(checkpoint, load_optimizer_states=True, load_lr_scheduler_states=True)\r\n+ self.deepspeed.load_checkpoint(checkpoint, load_optimizer_states=False, load_lr_scheduler_states=False)\r\n\r\n def hyperparameter_search(\r\n self,\r\n\r\ndiff --git a/seq2seq/ds_config.json b/seq2seq/ds_config.json\r\nindex 18ce5a3..44b5cc0 100644\r\n--- a/seq2seq/ds_config.json\r\n+++ b/seq2seq/ds_config.json\r\n@@ -15,7 +15,8 @@\r\n \"reduce_scatter\": true,\r\n \"reduce_bucket_size\": 2e8,\r\n \"contiguous_gradients\": true,\r\n- \"cpu_offload\": false\r\n+ \"cpu_offload\": false,\r\n+ \"load_from_fp32_weights\": false\r\n },\r\n\r\n \"zero_allow_untested_optimizer\": true,\r\n\r\n```\r\n\r\nBasically, we are telling deepspeed not to resume the optimizer/scheduler states and ignore the fp32 weights as well. The last one is not great and may not do what you want, as it's likely to impact the precision.\r\n\r\nI added some debug code to deepspeed and the key it fails on in your traceback is \"exp_avg\" . If it helps here is what I did:\r\n```\r\ndiff --git a/deepspeed/runtime/zero/stage2.py b/deepspeed/runtime/zero/stage2.py\r\nindex e0ca4f0..02d904b 100755\r\n--- a/deepspeed/runtime/zero/stage2.py\r\n+++ b/deepspeed/runtime/zero/stage2.py\r\n@@ -1802,7 +1802,11 @@ class FP16_DeepSpeedZeroOptimizer(object):\r\n p = group['params'][0]\r\n for key, saved in base_optimizer_group_states[i].items():\r\n if torch.is_tensor(self.optimizer.state[p][key]):\r\n- self.optimizer.state[p][key].data.copy_(saved.data)\r\n+ try:\r\n+ self.optimizer.state[p][key].data.copy_(saved.data)\r\n+ except:\r\n+ print(f\"failed with key={key}\")\r\n+ raise\r\n else:\r\n self.optimizer.state[p][key] = saved\r\n```\r\n\r\n------------\r\n\r\nAlso: unrelated to this particular issue, the code base your forked from is quite outdated and many deepspeed-related bug fixes were applied since then, so I highly recommend that you sync your fork with the current trainer. If possible try to subclass the Trainer rather than hacking it directly, so you always get the most up-to-date code base you inherit from. Deepspeed is very fresh and expect many more changes happening in the next few weeks/months.\r\n\r\n-----------\r\n\r\nNow to actually solving the problem.\r\n\r\nStudy your code and see where your model's structure gets modified between its creation and the checkpoint saving. Trainer tries to resume from the checkpoint as soon as `train()` starts and it appears that at that point the model is different structurally (dimensions are different) then it is later when it gets saved. So you need your model to be in an identical shape during resume and saving points.\r\n\r\nPlease let me know if this helps.\r\n\r\nI'd attack it as simple as dumping the model's param dimensions just before the checkpoint is loaded and just before it's saved, comparing the two - most likely finding the mismatch and then going forward from loading or backward from saving in the code and finding a spot where you modify the model. Then move the modification to before you resume from the checkpoint. i.e. before you call `train()`.\r\n\r\n----------\r\n\r\na minor comment - your repro instructions are 95% working, I had to do a few fixes to make it work (e.g. your `setup.py` is slightly broken), so it's always a good idea to re-validate that it's actually reproducible ;) But it's all good, I figured it out. It was very helpful to have what you provided to reproduce the problem.", "Dear @stas00 \r\nThank you very much for taking your precious time to look into this issue to assist me, I am indebted to you, and for all the incredible job you do. About reproducibility, I very honestly created a new condo environment and tested it before sending it out, and was working on my side, please accept my sincere apologies for any shortcomings and if I missed something without realizing it. \r\n\r\nI will investigate the issue with the great pointer you shared with me and I will keep this issue updated.\r\n\r\nThank you very much again for the great help. ", "Dear @stas00 \r\nI will be indebted to also ask also about the nans I get with deepspeed just with the same code you run the loss is nan. I very much appreciate any suggestion you might have for me to try to resolve the issue I face with deepspeed, I am using mt5-small. It would be a great help to me if I could use the great work you have done in deepspeed in huggingface repo and overcome the nan issue. Thank you.", "can we close this one now? we are dealing with nans at https://github.com/huggingface/transformers/pull/10956/files", "Dear Stas\nI unfortunately could not still figure this out. I am a bit confused by\nwhere fp16 casting is applied in MT5, specially with the new PR, where this\nis disabled in forward path. To me, in the middle there is some fp16\ncasting, which causes this, but I could not figure this out so far, I was\nwondering if you could give me more time. I appreciate any hints. thanks\n\nOn Tue, Mar 30, 2021 at 12:52 AM Stas Bekman ***@***.***>\nwrote:\n\n> can we close this one now? we are dealing with nans at\n> https://github.com/huggingface/transformers/pull/10956/files\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/10821#issuecomment-809771691>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AS37NMTDDH2FJGNOVF6ELVDTGEAELANCNFSM4ZQQFMIA>\n> .\n>\n", "I had a closer look and - you have a very dated trainer code with multiple bugs that have been fixed since then - you will need to update your code base to `transformers` master. I'm pretty sure that once you had it synced this problem will no longer be there.\r\n\r\nPlease do update me if this is still not the case and then we will fix it then. Thank you!\r\n\r\n", "Dear @stas00 \r\nThank you very much for the help. Much appreciated.\r\nI upgraded the codes to the last version of the codes in huggingface repository and I am still having the same issue. \r\nI will make an updated the repository asap and keep you updated on this.\r\nThank you very much.", "Yes, let me know when you have a repo I can reproduce the issue with. Thank you.", "Hi @stas00 \r\nI finally found this bug, this is the issue reported also here https://github.com/huggingface/transformers/issues/11294 \r\nI was freezing some parameters, and during checkpoiting, since huggingface codes does not handle the freezed parameters properly and has a bug currently regarding this, those parameters were not freezed when its loads from the checkpoints, and caused the difference in the number of parameters for the deepspeed.\r\nthanks a lot for all the hints and helps on this." ]

[]

[]

[ "I think you just have to name your file with a \".json\" extension for the script to work.\r\n\r\nThere is no support for other formats that will be added to this script as it's not easy to add the csv format while continue to support all the translation datasets in Datasets. You should just tweak the data processing of your example (for instance by doing the same as in `run_summarization`) to your needs if you need to use a csv file.", "Thank you very much! It works. \r\nBecause the extension for JSONLINES format is `jsonl`, it's better to explain it in the readme.", "This actually might be a good idea to change the extension to `.jsonl` to make it less ambiguous and it would require no special documentation then." ]

[]

[]

[ "Duplicate of https://github.com/huggingface/transformers/issues/10830", "Hi @patrickvonplaten this is not exact duplicate, I am using mt5-small and the other user in #10830 is using t5-large, I appreciate considering both thank you ", "@dorost1234, please kindly test if this PR fixes the problem: https://github.com/huggingface/transformers/pull/10956", "@stas00 thank you very much for the contributions, it now works for me for the mt5-small, I am running some more experiments with it and update.", "Dear @stas00 \r\nI tested more codes, without deepspeed, it works fine with setting the feedforward layer to float32, as suggested in the PR, but the moment I switch to deepspeed I still get nan issue in my codes. I greatly appreciate if you can spare some moments from your precious time and provide me with a suggestion for the case of deepspeed for the same problem. Thank you very much\r\n\r\nI also used your debug codes:\r\n```\r\n^M 0%| | 0/38600 [00:00<?, ?it/s]WARNING:seq2seq.third_party.models.t5.debug_utils:gelu 5 has inf\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5Block after T5LayerFF has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5Block after T5LayerFF has inf\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5Stack loop end has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5Stack loop start has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5Block has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5LayerNorm has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5LayerNorm variance has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5LayerNorm hidden_states has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5LayerNorm hidden_states before return has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5Block after T5LayerSelfAttention has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5Block before T5LayerFF has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5LayerNorm has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5LayerNorm variance has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5LayerNorm hidden_states has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:T5LayerNorm hidden_states before return has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:gelu 1 has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:gelu 2 has nans\r\nWARNING:seq2seq.third_party.models.t5.debug_utils:gelu 3 has nans\r\n\r\n```\r\n", "I was just thinking about it, so thank you for confirming that. \r\n\r\nDeepspeed is not using `autocast` so in essence the proposed fixed makes no difference under Deepspeed as we aren't running under `autocast` in the first place. Let's ask the DeepSpeed developers https://github.com/microsoft/DeepSpeed/issues/908\r\n\r\nThough let's continue the discussion on the deepspeed in the other issue you opened, since these are related but different problems. That's we may fix one but not the other, or the fixes may come at different times, so it's easier to track separate issues. \r\n\r\nOr if there is not one specific issue to t5/mt5+deepspeed please open one. Thank you.\r\n", "Dear @stas00 \r\nSure, thank you very much for coming back to me. Having your permission I will open up an issue on this. \r\nThank you very much.", "I already did - please see the link in my last comment. Please do not worry, we will surely find one way or another to resolve this.", "oh, great, thank you very much ", "Dear @stas00 \r\nI tested the code more (without deepspeed) on larger scale and when I train on opus100 (I train on 20 languages of it), after 2000 iterations with mt5-small, after applying the fix, this gets nan still. I will share with you a reproducible code soon. thanks a lot for all the great work. ", "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": 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" } ]

[]

[]

[]

[]

[ "Ah actually the dependabot PR was earlier in my notifications so I merged it without seeing you had opened a PR here. Sorry about that, closing as already taken care of in https://github.com/huggingface/transformers/commit/dbfe3795147e1360b3afac53a9ee0e14374d2ea6", "Actually, the proposed `>=` is probably better, so fixing conflicts and merging this. Thanks @stas00!" ]

[ { "id": 1990918270, "node_id": "MDU6TGFiZWwxOTkwOTE4Mjcw", "url": "https://api.github.com/repos/huggingface/transformers/labels/Good%20First%20Issue", "name": "Good First Issue", "color": "bbf794", "default": false, "description": "" }, { "id": 2392046359, "node_id": "MDU6TGFiZWwyMzkyMDQ2MzU5", "url": "https://api.github.com/repos/huggingface/transformers/labels/Good%20Second%20Issue", "name": "Good Second Issue", "color": "dd935a", "default": false, "description": "Issues that are more difficult to do than \"Good First\" issues - give it a try if you want!" } ]

[]

[ "Hi @stas00,\r\nPlease let me know if this is still open and I can contribute.", "Yes, please.", "I reproduced this in colab and got 28% slowness but still figuring out the cause, \r\nEarlier my assumption was this bit reduction/quantization was a device-specific thing.", "Usually in such situations I try to either go from the bottom up or in reverse. That is just take the `model(**inputs)` and measure the speed w/ `model` vs `model.half()` - if it's the same go one level up into `generate`, etc. Or starting from the top (`generate`) and then removing big chunks of code until you find the part that contributes to the slow down.\r\n\r\nYou can use this tracker to bracket the operation you measure.\r\n\r\nhttps://github.com/huggingface/transformers/blob/335c0ca35c159f88d73198bdac928e61a4d480c7/src/transformers/trainer_utils.py#L258\r\n\r\nBut a totally different approach which might get to the core of the issue much faster is to use a python profiler, .e.g. `cProfile` - that way you get the full analytics on each function call and if you compare these side by side w/ and w/o `half()` you might get an instant answer. Actually now that I wrote this I'd say start with this approach.\r\n", "I have done a few measures on 2 different cards (a 3090 and a 2080 Ti) using various evaluation batch sizes, and I haven't observed a single culprit for this problem. Instead, I'm seeing that all the operations in the `forward` pass are somewhat slower with `fp16`, and consistently so.\r\n\r\nSetup\r\n* Evaluation batch size in {4, 8, 16, 32, 64, 128}\r\n* 128 evaluation samples. Since I'm using powers of 2 for the batch sizes, this allows us to test from 1 batch to many batches of the same size.\r\n* `max_length` = `min_length` = 128. Setting `min_length` to 128 increases processing time.\r\n\r\nThese are the results for the main operations inside the `forward` method of `T5Block` (total seconds spent in the corresponding areas; figures from the 3090 and the 3 first batch sizes for brevity):\r\n\r\n<img width=\"558\" alt=\"image\" src=\"https://user-images.githubusercontent.com/1177582/115751446-7bf66080-a399-11eb-828a-097ea8cb1308.png\">\r\n\r\nThe time difference depends on the batch size, but `fp16` is always between 15% (for bs=64) and 26% (bs=16) slower.\r\n\r\n---\r\n\r\nToday I discovered [this thread](https://github.com/pytorch/pytorch/issues/50153) in the PyTorch forums, and repeated the test using a version of **PyTorch compiled from source**. Amazingly, processing is now **almost twice as fast**, but the difference is still there:\r\n\r\n<img width=\"558\" alt=\"image\" src=\"https://user-images.githubusercontent.com/1177582/115752012-fde68980-a399-11eb-868d-fa37b3effd54.png\">\r\n\r\nIn this case, using a batch size of 128 (1 batch) is about 13% slower, while a batch size of 16 is 27% slower.\r\n\r\nI'm not sure how to proceed. Does this ring a bell for anyone?", "Thank you for researching and profiling, @pcuenca!\r\n\r\nI think the next step is the new pytorch profiler:\r\nhttps://pytorch.org/blog/introducing-pytorch-profiler-the-new-and-improved-performance-tool/\r\n\r\nUnfortunately, at the moment I have no time to dig into it, so I hope someone will beat me to it.\r\n\r\n-------------\r\n\r\nre: building from source:\r\n\r\nIndeed, I recently built pytorch from source and I don't know if it's that or something else since 1 month passed since OP was made, but I'm getting 2x speed improvement (rtx-3090) on training this task. eval is only slightly faster, but is still 25% slower @ fp16.\r\n\r\nAlso adapted the cmd line to the recently changed examples:\r\n\r\n```\r\n\r\nexport BS=16; rm -r output_dir; PYTHONPATH=src USE_TF=0 CUDA_VISIBLE_DEVICES=0 \\\r\n./examples/pytorch/translation/run_translation.py --model_name_or_path t5-small --output_dir /tmp/zero3 \\\r\n--overwrite_output_dir --max_train_samples 10 --max_eval_samples 100 --max_source_length 12 \\\r\n--max_target_length 128 --do_train --num_train_epochs 1 \\\r\n--per_device_train_batch_size 2 --learning_rate 3e-3 --warmup_steps 8 --predict_with_generate \\\r\n--logging_steps 0 --save_steps 2 --eval_steps 1 --group_by_length --adafactor --dataset_name wmt16 \\\r\n--dataset_config ro-en --source_lang en --target_lang ro \\\r\n--source_prefix \"translate English to Romanian: \" --do_eval \r\n\r\n***** train metrics *****\r\n epoch = 1.0\r\n init_mem_cpu_alloc_delta = 1254MB\r\n init_mem_cpu_peaked_delta = 155MB\r\n init_mem_gpu_alloc_delta = 230MB\r\n init_mem_gpu_peaked_delta = 0MB\r\n train_mem_cpu_alloc_delta = 1382MB\r\n train_mem_cpu_peaked_delta = 125MB\r\n train_mem_gpu_alloc_delta = 231MB\r\n train_mem_gpu_peaked_delta = 194MB\r\n train_runtime = 0:00:04.19\r\n train_samples = 10\r\n train_samples_per_second = 1.191\r\n\r\n***** eval metrics *****\r\n epoch = 1.0\r\n eval_bleu = 2.2434\r\n eval_gen_len = 15.69\r\n eval_loss = 3.7374\r\n eval_mem_cpu_alloc_delta = 1MB\r\n eval_mem_cpu_peaked_delta = 0MB\r\n eval_mem_gpu_alloc_delta = 0MB\r\n eval_mem_gpu_peaked_delta = 171MB\r\n eval_runtime = 0:00:04.33\r\n eval_samples = 100\r\n eval_samples_per_second = 23.051\r\n```\r\n\r\nadd `--fp16_full_eval`\r\n\r\n```\r\nexport BS=16; rm -r output_dir; PYTHONPATH=src USE_TF=0 CUDA_VISIBLE_DEVICES=0 \\\r\n./examples/pytorch/translation/run_translation.py --model_name_or_path t5-small --output_dir /tmp/zero3 \\\r\n--overwrite_output_dir --max_train_samples 10 --max_eval_samples 100 --max_source_length 12 \\\r\n--max_target_length 128 --do_train --num_train_epochs 1 \\\r\n--per_device_train_batch_size 2 --learning_rate 3e-3 --warmup_steps 8 --predict_with_generate \\\r\n--logging_steps 0 --save_steps 2 --eval_steps 1 --group_by_length --adafactor --dataset_name wmt16 \\\r\n--dataset_config ro-en --source_lang en --target_lang ro \\\r\n--source_prefix \"translate English to Romanian: \" --do_eval --fp16_full_eval\r\n\r\n***** train metrics *****\r\n epoch = 1.0\r\n init_mem_cpu_alloc_delta = 1259MB\r\n init_mem_cpu_peaked_delta = 155MB\r\n init_mem_gpu_alloc_delta = 230MB\r\n init_mem_gpu_peaked_delta = 0MB\r\n train_mem_cpu_alloc_delta = 1380MB\r\n train_mem_cpu_peaked_delta = 125MB\r\n train_mem_gpu_alloc_delta = 231MB\r\n train_mem_gpu_peaked_delta = 194MB\r\n train_runtime = 0:00:03.76\r\n train_samples = 10\r\n train_samples_per_second = 1.326\r\n\r\n***** eval metrics *****\r\n epoch = 1.0\r\n eval_bleu = 2.2434\r\n eval_gen_len = 15.69\r\n eval_loss = 3.7383\r\n eval_mem_cpu_alloc_delta = 4MB\r\n eval_mem_cpu_peaked_delta = 0MB\r\n eval_mem_gpu_alloc_delta = -231MB\r\n eval_mem_gpu_peaked_delta = 262MB\r\n eval_runtime = 0:00:05.32\r\n eval_samples = 100\r\n eval_samples_per_second = 18.778\r\n\r\n```", "By running everything with `CUDA_LAUNCH_BLOCKING=1` under the line profiler, I found that [this](https://github.com/huggingface/transformers/blob/ff5cdc086be1e0c3e2bbad8e3469b34cffb55a85/src/transformers/models/t5/modeling_t5.py#L677) and [this](https://github.com/huggingface/transformers/blob/ff5cdc086be1e0c3e2bbad8e3469b34cffb55a85/src/transformers/models/t5/modeling_t5.py#L692) check for infinite values take up more time than I expected.\r\n\r\nAfter removing those checks, this is what I end up with: \r\n```\r\n$ export BS=16; rm -r output_dir; PYTHONPATH=src USE_TF=0 CUDA_VISIBLE_DEVICES=0 \\\r\npython -m cProfile -o profile.prof ./examples/pytorch/translation/run_translation.py --model_name_or_path t5-small --output_dir /tmp/zero3 \\\r\n--overwrite_output_dir --max_train_samples 10 --max_eval_samples 1600 --max_source_length 12 \\\r\n--max_target_length 128 --do_train --num_train_epochs 1 \\\r\n--per_device_train_batch_size 4 --per_device_eval_batch_size $BS --learning_rate 3e-3 --warmup_steps 8 --predict_with_generate \\\r\n--logging_steps 0 --save_steps 2 --eval_steps 1 --group_by_length --adafactor --dataset_name wmt16 \\\r\n--dataset_config ro-en --source_lang en --target_lang ro \\\r\n--source_prefix \"translate English to Romanian: \" --do_eval\r\n...\r\n***** eval metrics *****\r\n epoch = 1.0\r\n eval_bleu = 0.3251\r\n eval_gen_len = 10.2375\r\n eval_loss = 3.6796\r\n eval_runtime = 0:01:03.89\r\n eval_samples = 1600\r\n eval_samples_per_second = 25.04\r\n eval_steps_per_second = 1.565\r\n```\r\n\r\nThe same with `--fp16_full_eval`:\r\n```\r\n***** eval metrics *****\r\n epoch = 1.0\r\n eval_bleu = 0.3258\r\n eval_gen_len = 10.2406\r\n eval_loss = 3.6797\r\n eval_runtime = 0:01:01.43\r\n eval_samples = 1600\r\n eval_samples_per_second = 26.043\r\n eval_steps_per_second = 1.628\r\n```\r\n\r\nNote that I had to dial up the number of eval examples since this measurement was quite noisy on the shared system I used. However, the FP16 was faster most of the time. If someone could double check these observations under more reliable circumstances, that'll be great. ", "Thank you for looking into it, @dsuess! \r\n\r\nI'm trying to figure out torch.profiler to get a better understanding using native tools.\r\n\r\nGreat to hear you found those checks to be slowdowns. Need to investigate these closer with torch.profiler.\r\n\r\nAnd I also found https://github.com/huggingface/transformers/blob/ff5cdc086be1e0c3e2bbad8e3469b34cffb55a85/src/transformers/models/t5/modeling_t5.py#L504 to be another point of slowdown. It's possible that the upcast can be removed completely, which should speed things up. But definitely a slightly faster version is to:\r\n```\r\n attn_weights = nn.functional.softmax(scores.float(), dim=-1).type_as(scores)\r\n attn_weights = nn.functional.softmax(scores.float(), dim=-1, dtype=scores.dtype)\r\n```\r\nfor fp16 (it makes no difference for fp32)\r\n\r\nI will look closer into the 2 points you suggested.\r\n\r\nbut also we should run under a more realistic configuration of at least seqlen 512 and not 12 like I had it originally, with large seqlen things change quite a lot. That is `--max_source_length 512 --max_target_length 512` (or even better 1024). ", "Thanks for your feedback @stas00. I finally got the time to have a closer look with the pytorch profiler. I'd summarize what I found with:\r\n- the speedup we're getting for matmuls in fp16 aren't that great. This might be due to fewer kernels being executed on Tensor cores when using FP16 (31% of kernels) compared to FP32 (74% of kernels).\r\n- this is made worse by additional copy/conversion operations as can be seen in the device self time for FP16 (left) vs FP32 (right): \r\n<img width=\"937\" alt=\"image\" src=\"https://user-images.githubusercontent.com/5870291/150110417-2e8baf04-6904-45b9-960c-7cc12a16ee03.png\">\r\n\r\nThese conversions happen in the [layer norm](https://github.com/huggingface/transformers/blob/master/src/transformers/models/t5/modeling_t5.py#L246) and before the [softmax](https://github.com/huggingface/transformers/blob/master/src/transformers/models/t5/modeling_t5.py#L513), which matches with your observation. I also double checked the layer norm with this [micro benchmark](https://github.com/dsuess/transformers/blob/10816-fp16_eval_performance/tests/benchmark_modeling_t5.py), which runs ~30% slower in FP16.\r\nThere's a [tiny improvement](https://github.com/dsuess/transformers/commit/63f039329434e5b57051111be9b8466c87689159), which makes the eval-example run ~1% faster, but it doesn't even register in the micro benchmark. \r\n\r\nJudging from [the issue](https://github.com/pytorch/pytorch/issues/66707) you raised, we can't run layer norm in FP16. I'd expect the same to be true for softmax, so I am unsure if we can get rid of those conversions. We may have a chance to get more out of the matmuls, so I'll try to figure out why those kernels don't run on Tensor cores despite being eligible.\r\n\r\n---\r\nI've done all these experiments on a 3080Ti with `--max_source_length 512 --max_target_length 512`", "This is fantastic work, @dsuess!\r\n\r\nHere is an additional profiling report of the same issue but under tf32: https://github.com/huggingface/transformers/issues/14608#issuecomment-1001257392\r\n\r\nThis appears to be specific to t5 and derived models. And yes the problem is that it uses RMSNorm which pytorch doesn't provide and that's why it's slow.\r\n\r\nI made a request to make an RMSNorm fused kernel here: https://github.com/NVIDIA/apex/issues/1271 and once this is done to ask to upstream it into pytorch. I hope this should solve this issue.\r\n\r\nI also tried to avoid re-casting using some tricks here by trying to deploy the existing fused functions: https://github.com/huggingface/transformers/pull/14656 but I couldn't find a faster way using the existing pytorch python API.\r\n\r\nHave you by chance tried any other architectures using the same benchmarks? e.g. gpt2 and bert as they are very distinct from t5.\r\n", "> Here is an additional profiling report of the same issue but under tf32: [#14608 (comment)](https://github.com/huggingface/transformers/issues/14608#issuecomment-1001257392)\r\n\r\nGreat benchmark of the different data types, thanks for sharing.\r\n\r\n> Have you by chance tried any other architectures using the same benchmarks? e.g. gpt2 and bert as they are very distinct from t5.\r\n\r\nI've just tested the same script with some of the mbart variants and as expected, fp16 is faster for those.\r\n" ]

[]

[]

[ "Hi @stas00 \r\nI tested this PR and for me this becomes really slower with mt5-small model after adding the modifications in this PR, here is the command I run, I am using transformer=4.4.2. I will be grateful to your expert knowledge to have the speed issue also fixed. Thank you very much for the incredible jobs you do. \r\n\r\n`deepspeed run_translation.py --model_name_or_path google/mt5-small --do_train --do_eval --source_lang en --target_lang ro --dataset_name wmt16 --dataset_config_name ro-en --output_dir test/tst-t1ranslatieeeon --per_device_train_batch_size=4 --per_device_eval_batch_size=4 --predict_with_generate --max_train_samples 100 --fp16 --deepspeed ds_config.json --max_val_samples 100 --logging_step 10\r\n`", "I can't possibly see how this PR could impact the speed since it changes the label_smoother and your command line doesn't have `--label_smoothing 0.1` so the modified code in this PR doesn't get to run.\r\n\r\nThat said when you use this PR you're in a way using `master`, so perhaps you were testing with some other `transformers` version before and noticed a regression in `master`. Try to test with whatever version you were using before and then retest with the `master` branch and see whether you can reproduce your issue. \r\n\r\nIf it is, do you know how to use `git bisect`? You can then in a matter of a few runs find out the commit that impacted the performance. If you can't figure it out just give me the last good transformers version and I will help you from there.\r\n\r\n-----\r\n\r\nAlso you're not telling me what's inside `ds_config.json` - I assume it's zero2 configuration. zero3 isn't quite ready yet." ]

[]

[]

[]

[]

[]

[ "That is not a bug:\r\nhttps://stackoverflow.com/questions/66625945/huggingfaces-reformerformaskedlm-configuration-issue/66636363#66636363", "Unfortunately that did not help. Adding:\r\n```from transformers import ReformerTokenizer, ReformerForMaskedLM, ReformerConfig\r\nimport torch\r\n\r\ntokenizer = ReformerTokenizer.from_pretrained('google/reformer-crime-and-punishment')\r\nconfig = ReformerConfig.from_pretrained('google/reformer-crime-and-punishment')\r\nconfig.is_decoder=False\r\nmodel = ReformerForMaskedLM.from_pretrained('google/reformer-crime-and-punishment', config=config)\r\n\r\ninputs = tokenizer(\"The capital of France is [MASK].\", return_tensors=\"pt\")\r\nlabels = tokenizer(\"The capital of France is Paris.\", return_tensors=\"pt\")[\"input_ids\"]\r\n\r\noutputs = model(**inputs, labels=labels)\r\nloss = outputs.loss\r\nlogits = outputs.logits```\r\n\r\ncaused this exception:\r\n```Traceback (most recent call last):\r\n File \"testReformers.py\", line 13, in <module>\r\n outputs = model(**inputs, labels=labels)\r\n File \"/mnt/nfs/d4nvme0/userhomes/ksrinivs/anaconda3/envs/reformers/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 727, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/mnt/nfs/d4nvme0/userhomes/ksrinivs/anaconda3/envs/reformers/lib/python3.8/site-packages/transformers/models/reformer/modeling_reformer.py\", line 2367, in forward\r\n masked_lm_loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))\r\n File \"/mnt/nfs/d4nvme0/userhomes/ksrinivs/anaconda3/envs/reformers/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 727, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/mnt/nfs/d4nvme0/userhomes/ksrinivs/anaconda3/envs/reformers/lib/python3.8/site-packages/torch/nn/modules/loss.py\", line 961, in forward\r\n return F.cross_entropy(input, target, weight=self.weight,\r\n File \"/mnt/nfs/d4nvme0/userhomes/ksrinivs/anaconda3/envs/reformers/lib/python3.8/site-packages/torch/nn/functional.py\", line 2468, in cross_entropy\r\n return nll_loss(log_softmax(input, 1), target, weight, None, ignore_index, None, reduction)\r\n File \"/mnt/nfs/d4nvme0/userhomes/ksrinivs/anaconda3/envs/reformers/lib/python3.8/site-packages/torch/nn/functional.py\", line 2261, in nll_loss\r\n raise ValueError('Expected input batch_size ({}) to match target batch_size ({}).'\r\nValueError: Expected input batch_size (21) to match target batch_size (17).```.\r\n\r\nSomething seems amiss here given that there is a single sentence being passed in and it seems to think we have batch sizes of 21 and 17?", "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, thanks for opening an issue! We try to keep the github issues for bugs/feature requests.\r\nCould you ask your question on the [forum](https://discusss.huggingface.co) instead?\r\n\r\nThanks!", "Thank you very much.\r\nI will contact the forum.\r\nRegards,\r\nlematmat\r\n" ]

[]

[]

[]

[]

[]

[ "Hi! I believe this is was overlook on our part. Your change looks reasonable to me, do you want to open a PR with your proposed fix?\r\n\r\nAnd thank you for opening such a detailed and well-structured issue!" ]

[ { "id": 2934977194, "node_id": "MDU6TGFiZWwyOTM0OTc3MTk0", "url": "https://api.github.com/repos/huggingface/transformers/labels/Flax", "name": "Flax", "color": "4862AD", "default": false, "description": "" } ]

[]

[ "> Great work! My only concern is to make sure we don't lose any performance by not using `nn.linen.SelfAttention`. If we are just using the same code as its implementation, there is no reason for that but it's good to double-check.\r\n> Otherwise, I agree it's better to re-implement it than to have custom weight loading logic..\r\n\r\nGreat! Yeah, I'll talk with @avital about this next week (hopefully) :-) " ]

[]

[]

[ "Ok actually this is ready to merge" ]

[]

[]

[ "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." ]

[]

[]

[]

[]

[]

[ "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." ]

[]

[]

[ "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 don't know the answer, but hoping that it works after this PR got merged https://github.com/huggingface/transformers/pull/10753\r\n\r\nDo you already use deepspeed ?", "No, i did not" ]

[]

[]

[]

[]

[]

[ "To address your comments on the `MakeFile`, I have removed some checks from `extra_quality_checks` because they are checks that modify content and `make quality` is only supposed to check, not change.\r\n\r\nTo have `make fixup` still work as intended, I put the checks that change content in `extra_style_checks` that is called both by `make fixup` and `make style`. Could you double-check it looks okay @LysandreJik and @stas00 ? Thanks!" ]

[]

[]

[ "Hello, thanks for opening an issue! We try to keep the github issues for bugs/feature requests.\r\nCould you ask your question on the [forum](https://discusss.huggingface.co) instead?\r\n\r\nThanks!" ]

[]

[]

[ "I'd be super-handy to link directly to suitable datasets and models for each example as in \r\n- https://huggingface.co/datasets?search=squad\r\n- https://huggingface.co/models?filter=squad\r\n\r\nmay be this could be an easy first good issue.\r\n\r\nSome of the keywords and whether to use `?filter=` or `?search=` will require some investigation since the former is hidden and packs some power missing from the latter.", "The first may be helpful, but the second is not necessarily: it shows the models that have been fine-tuned on a squad dataset, not the models that can be fine-tuned on it. There is no way to filter all the models that have an architecture containing a question-answering head as far as I know, which is what we would want to show.", "Would this be at least in the right direction? https://huggingface.co/models?pipeline_tag=question-answering \r\n", "Mmm, those seem to be models fine-tuned on a question-answering task, not all models with a QuestionAnswering arch available (for instance, you should see all BERT checkpoints, all distilBERT checkpoints etc).", "OK, then it won't work.\r\n\r\nIt'd be really awesome if in the future we had a filter to filter models by architecture - and sub-architecture in this case - that is without the model-specific part of the class name." ]

[]

[]

[ "Hello, thanks for opening an issue! We try to keep the github issues for bugs/feature requests.\r\nCould you ask your question on the [forum](https://discusss.huggingface.co) instead?\r\n\r\nThanks!" ]

[]

[]

[ "Hello! This is weird, you should be gotten an error before even being able to instantiate the tokenizer with `from_pretrained`. Such an error:\r\n```\r\nImportError: \r\nXLNetTokenizer requires the SentencePiece library but it was not found in your environment. Checkout the instructions on the\r\ninstallation page of its repo: https://github.com/google/sentencepiece#installation and follow the ones\r\nthat match your environment.\r\n```\r\n\r\nCould you install SentencePiece `pip install sentencepiece` and let me know if it fixes your issue?", "Hi! I had actually did the `pip install sentencepiece`. I was getting `None` after it.\r\n\r\nI saw the source code and the embedding size used was `None` there. You can check it <a href=\"https://github.com/gauravsharma-97/transformers/blob/master/src/transformers/models/xlnet/tokenization_xlnet.py#L41-L44\">here</a>. I think that is the issue and it should be some integer like BertTokenizer uses 512 as embedding size.", "I may be wrong, but I think this can happen if you're in a colab environment and you install SentencePiece, but don't reload the kernel before re-running your cell. \r\n\r\nYou say you're on Ubuntu, I managed to obtain a similar result by re-running the code you mentioned twice in the same Python runtime, by installing `sentencepiece` between the two code statements. Since sentencepiece is loaded on the fly, this can be the result.\r\n\r\nI stand by what I say that this is due to `sentencepiece` not being installed. If it's correctly installed in your environment, running your statement results in:\r\n```\r\nPreTrainedTokenizer(name_or_path='xlnet-base-cased', vocab_size=32000, model_max_len=1000000000000000019884624838656, is_fast=False, padding_side='left', special_tokens={'bos_token': '<s>', 'eos_token': '</s>', 'unk_token': '<unk>', 'sep_token': '<sep>', 'pad_token': '<pad>', 'cls_token': '<cls>', 'mask_token': AddedToken(\"<mask>\", rstrip=False, lstrip=True, single_word=False, normalized=True), 'additional_special_tokens': ['<eop>', '<eod>']})\r\n``` \r\n\r\nYou're mentioning the embedding size which is `None`, this is on purpose. The XLNet model uses relative positional embeddings, and has therefore no limitations on the size of the input (note the `model_max_len` in the above code statement); which isn't the case for BERT, that uses absolute positional embeddings which are limited to 512.", "Yes you are correct. I was running this on colab and it might have required reloading the kernel. But funnily enough, its working today without reloading it.\r\n\r\nYesterday might have been an isolated incident, although I did try to get it to run for very long before posting the issue.\r\n\r\nAnyway, thanks for the help @LysandreJik and for the explanation on embeddings. " ]

[]

[]

[ "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.", "cc @patrickvonplaten ", "Hey @merrymercy - super sorry, I saw the PR too late and it was actually already fixed.", "Thanks for your effort on Jax integration! @patrickvonplaten \r\nCould you also add some doc for these examples https://github.com/huggingface/transformers/tree/master/examples/flax/language-modeling?" ]

[]

[]

[]

[]

[]

[ "Do you want to take a look @patrickvonplaten?", "thanks a lot！！！" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "I think the `DataCollatorForSeq2Seq` should be used in all cases as it does more than just padding. If you want to suggest a PR with the fix, that would be more than welcome!", "Assuming the goal is:\r\n- using DataCollatorForSeq2Seq in Seq2SeqTrainer as default when no data_collator is provided, while keeping the remaining functionality unchanged, \r\n\r\nthe first approach could be:\r\n- providing Seq2SeqTrainer with an `__init__` method: \r\n - instantiating a DataCollatorForSeq2Seq if no collator provided, and\r\n - calling Trainer's `__init__` and passing the instance along with other parameters. \r\n \r\nSomething like:\r\n\r\n```\r\nclass Seq2SeqTrainer(Trainer):\r\n \r\n def __init__(\r\n self,\r\n model: Union[PreTrainedModel, torch.nn.Module] = None,\r\n args: TrainingArguments = None,\r\n data_collator: Optional[DataCollator] = None,\r\n train_dataset: Optional[Dataset] = None,\r\n eval_dataset: Optional[Dataset] = None,\r\n tokenizer: Optional[\"PreTrainedTokenizerBase\"] = None,\r\n model_init: Callable[[], PreTrainedModel] = None,\r\n compute_metrics: Optional[Callable[[EvalPrediction], Dict]] = None,\r\n callbacks: Optional[List[TrainerCallback]] = None,\r\n optimizers: Tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR] = (None, None),\r\n ): \r\n \"\"\"\r\n Setting DataCollatorForSeq2Seq as default if no data_collator is provided.\r\n \"\"\"\r\n\r\n if data_collator is None:\r\n # Perform validation and overwrite model with model_init before passing to collator,\r\n # as done in Trainer\r\n if tokenizer is None:\r\n raise RuntimeError(\r\n \"`tokenizer` parameter is required by the default `DataCollatorForSeq2Seq`\"\r\n )\r\n if model is None and model_init is None:\r\n raise RuntimeError(\r\n \"`Trainer` requires either a `model` or `model_init` argument\"\r\n )\r\n model_collator = model\r\n if model_init is not None:\r\n # No parameter handling for hyper-parameter search (trial)\r\n # Only passing the prepare_decoder_input_ids_from_labels function\r\n model_collator = model_init()\r\n\r\n data_collator = DataCollatorForSeq2Seq(tokenizer, model=model_collator)\r\n\r\n super().__init__(\r\n model,\r\n args,\r\n data_collator,\r\n train_dataset,\r\n eval_dataset,\r\n tokenizer,\r\n model_init,\r\n compute_metrics,\r\n callbacks,\r\n optimizers,\r\n ) \r\n \r\n```\r\n\r\nOf course, I would need to look further into the code and the handling of other DataCollatorForSeq2Seq \r\nparameters like: `pad_to_multiple_of=8 if training_args.fp16 else None`\r\n\r\n@sgugger, Thanks for the suggestion, it is very interesting;)", "Mmm, I was thinking of an easier fix to just use that in the example script without necessary changing the default in `Seq2SeqTrainer`." ]

[]

[]

[ "I guess this is related to URL issue #10744 ? and one should change model URL", "I resolved this by updating URL's to models, this is my current code:\r\n```PRETRAINED_VOCAB_FILES_MAP = {\r\n \"vocab_file\": {\"allegro/herbert-base-cased\": \"https://huggingface.co/allegro/herbert-base-cased/resolve/main/vocab.json\"},\r\n \"merges_file\": {\"allegro/herbert-base-cased\": \"https://huggingface.co/allegro/herbert-base-cased/resolve/main/merges.txt\"},\r\n}``` \r\nIs there a way to fix this to maintain backward compatibility? @LysandreJik", "Cross-posting the Forum thread: https://discuss.huggingface.co/t/delete-organizations-models-from-the-hub/954/40" ]

[ { "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 } ]

[ "I'm getting a similar problem after training BERT with MLM using DeepSpeed where all the saved weights are of size 1. The same `run_mlm` script worked as expected if I didn't use DeepSpeed.\r\n\r\n`RuntimeError: Error(s) in loading state_dict for BertForSequenceClassification:\r\n size mismatch for bert.embeddings.word_embeddings.weight: copying a param with shape torch.Size([1]) from checkpoint, the shape in current model is torch.Size([119547, 768]).\r\n size mismatch for bert.embeddings.position_embeddings.weight: copying a param with shape torch.Size([1]) from checkpoint, the shape in current model is torch.Size([512, 768]).\r\n size mismatch for bert.encoder.layer.0.attention.self.query.weight: copying a param with shape torch.Size([1]) from checkpoint, the shape in current model is torch.Size([768, 768]).`", "Since this is using DeepSpeed, maybe @stas00 has an idea?", "Just tried loading a model trained with `sharded_ddp` and got a different error:\r\n\r\n```[INFO|modeling_utils.py:1044] 2021-03-18 12:56:04,792 >> loading weights file fs-test-mlm-mbert/checkpoint-1000/pytorch_model.bin\r\nTraceback (most recent call last):\r\n File \"/export/proj/code/transformers/src/transformers/modeling_utils.py\", line 1057, i\r\nn from_pretrained\r\n state_dict = torch.load(resolved_archive_file, map_location=\"cpu\")\r\n File \"/export/proj/env_cuda11_1/lib/python3.7/site-packages/torch/serialization.py\", line 593, in load\r\n return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)\r\n File \"/export/proj/env_cuda11_1/lib/python3.7/site-packages/torch/serialization.py\", line 762, in _legacy_load\r\n magic_number = pickle_module.load(f, **pickle_load_args)\r\nEOFError: Ran out of input\r\n```\r\nIt seems the model saving might not be happening properly for these two integrations? I also noticed that only the config and weights were being saved when using `--sharded_ddp`.\r\n\r\n\r\nUPDATE: It's actually the checkpoint saving getting stuck that's causing this issue. Started another run to confirm and it got stuck while saving as well. \r\n\r\nUPDATE 2: This only happens with `zero_dp_2` and `zero_dp_3`. `simple` appears to work fine. For DeepSpeed, using stage 2 appears to fix the problem (I was previously using stage 3).", "@samsontmr I have changed DeepSpeed stage to 2 and it seems works well - checkpoints are saved properly. I also used DeepSpeed stage 3 before.\r\n\r\nIt seems problems are in Stage 3 integration. Maybe @stas00 could help, he did previous integration of DeepSpeed into trainer.", "DeepSpeed Stage 3 integration is not finished yet, a wip PR is here if you'd like to try it - though it has a ton of debug statements still and a few more features are still missing.\r\nhttps://github.com/huggingface/transformers/pull/10753\r\n\r\nMake sure you are using the latest deepspeed since zero3 had problems with saving checkpoint but the 0.3.13 release should be good.\r\n\r\nBut I am pretty sure the issue is different, as I literally merged the code that generated the error you quoted 2 days ago:\r\nIf it worked before please roll back to any sha before https://github.com/huggingface/transformers/pull/10760 and let me know if it works.\r\n\r\nThe problem with DeepSpeed is that it doesn't currently have a way to save a fp32 checkpoint that can be loaded normally and not via DeepSpeed, https://github.com/microsoft/DeepSpeed/issues/800 so when you save a model you only get an fp16 version. However its special checkpoint (see e.g. `global-step10` folder in the checkpoint folder) contains all the right data and thus if you want to load deepspeed model you need to `train(resume_from_checkpoint)` instead. \r\n\r\nSo if you want to resume training you can't use `from_pretrained()` at the moment, unless fp16 weights are sufficient for your work. And it sounds that it's broken at the moment.\r\n\r\nLet me know if any of this makes sense and let's see how we can make your code work with what we have.\r\n\r\nI'd be happy to adapt my recent changes to meet your needs.\r\n\r\n\r\n", "Thanks for the detailed reply @stas00! Is the issue with the fp32 checkpoint saving only happening with zero3 or also with stage 2? My fine-tuning step started with no issues when I used the checkpoint from a stage 2 training run (hasn't completed yet so I'm not sure how it'll end up).", "> Is the issue with the fp32 checkpoint saving only happening with zero3 or also with stage 2? \r\n\r\nIt's an issue with any zero stage under deepspeed. \r\n\r\nAre you saying that the problem emerged once switching to zero3 config? I'm not at all sure it can resume from zero2 checkpoint to zero3 config - those are quite different setups. So we really need to get the fp32 saving sorted out\r\n\r\nLet's see if we can ask to make this a higher priority at https://github.com/huggingface/transformers/issues/10789\r\n\r\n", "> Are you saying that the problem emerged once switching to zero3 config? I'm not at all sure it can resume from zero2 checkpoint to zero3 config - those are quite different setups. So we really need to get the fp32 saving sorted out\r\n\r\nYup, I didn't try going from zero2 to zero3; I just restarted my training using zero2, then fine-tuned the model without deepspeed... which somehow managed to load just by using `.from_pretrained`", "As I tried to explain you were getting only fp16 weights when using from `from_pretrained` which may or may not be good enough for your needs. It mostly should be OK. Except some metrics or feature may break under fp16 if they weren't coded for it.\r\ne.g. https://github.com/huggingface/transformers/issues/10674\r\n\r\nSo let's lay out a test that I need to work on to reproduce your issues. Could you please lay out a sequence of events - ideally in code but pseudo-code will work too and then I will try to see where the breakage is.\r\n\r\nThe PR I referred to includes several save/resume tests, so the saving is normal, and resume uses `train(resume_from_checkpoint)` and it works too. Though I need to add zero3 test as well. Only tested zero2 so far. The resume test is here:\r\nhttps://github.com/huggingface/transformers/blob/008672e6e5fb0f2d2fc6fbd367ab6e135eea3f2d/examples/tests/deepspeed/test_deepspeed.py#L279\r\n\r\nYou shouldn't get:\r\n```\r\nValueError: [deepspeed] failed to resume from checkpoint ./templates/siamese-t5-small-v1_1-template\r\n```\r\n\r\nif you're not trying to do `train(resume_from_checkpoint)`, you can see where it gets triggered:\r\nhttps://github.com/huggingface/transformers/blob/008672e6e5fb0f2d2fc6fbd367ab6e135eea3f2d/src/transformers/integrations.py#L452\r\n", "As for me: I fixed my problem with unnessesary checkpoint load, where I get load error, but it still has an save error on DeepSpeed stage 3 mode. If you @stas00 could help me, I would appreciate.\r\n\r\n\r\nHere is steps to reproduce my error with model save:\r\n\r\n- Clone this repo:\r\nhttps://github.com/exelents/try_t5_siamese\r\n- Extract folder \"qasc\" from this archive:\r\nhttps://drive.google.com/file/d/1gwvFiPzWW0JLr0XLS25PuG2Br5S4fPbR/view?usp=sharing\r\n- Go to clonned repo folder and run ./create-siamese-template.sh - it will create siamese NN from two t5-small encoders in folder ./templates/siamese-t5-small-template\r\n- then you can run ./run-siamese-small.sh - you will see normal behaviour, in folder ./siamese_train_deepspeed/output_dir/ you will find there will be stored checkpoints every 3 steps? and you will can see a sight that weights are stored:\r\n weights files like ./siamese_train_deepspeed/output_dir/checkpoint-6/left/pytorch_model.bin will have size around hundred megabytes.\r\n\r\n- Then to see a problem open ./run-siamese-small.sh and change \"ds_config.json\" to \"ds_config_stage3.json\" and rerun training. You will see that weights files, like ./siamese_train_deepspeed/output_dir/checkpoint-6/left/pytorch_model.bin will have size for a few kilobytes, and you couldn't load model from that checkpoint. There is a probleb, and it appears only if I turn on \"stage 3\" mode in config.", "Thank you for the detailed instructions, @exelents.\r\n\r\nLet me adapt the existing test first to zero3 so I am sure it's working and then will try your sequence. I will keep you posted.", "I can reproduce the saved model size problem. `pytorch_model.bin` with:\r\n- zero2 135M \r\n- zero3 38K \r\n\r\nbut as I mentioned currently Deepspeed doesn't provide a proper way to save a model on its own.\r\n\r\nIt saves the model state in its own sub-folder, e.g., in your case:\r\n```\r\nls -l output_dir/checkpoint-6/global_step6/\r\ntotal 809M\r\n-rw-rw-r-- 1 stas stas 53K Mar 18 14:03 zero_pp_rank_0_mp_rank_00_model_states.pt\r\n-rw-rw-r-- 1 stas stas 809M Mar 18 14:03 zero_pp_rank_0_mp_rank_00_optim_states.pt\r\n```\r\nas you can see the optimizer states dict has everything in it. So you should be able to resume from it.\r\n\r\nYour script is a bit old and based on an old example - so it doesn't support the current mechanism of doing resume from command line using https://github.com/huggingface/transformers/blob/master/examples/README.md#resuming-training\r\n\r\nSo for resume to currently work, you either need to bring your script up-to-date, by probably checking the latest version of the example you used as a base for your work.\r\n\r\nThe key is `train(resume_from_checkpoint)` if you passed this as `output_dir/checkpoint-6` deepspeed reloads where it left on and continues on its merry way.\r\n\r\nTo help you think the new script in your case is this and I pointed to where the critical part is:\r\n\r\nhttps://github.com/huggingface/transformers/blob/dcebe254fadfe142b6f0d6301cc8a875dca7d603/examples/seq2seq/run_translation.py#L500\r\n\r\n(this is on master)\r\n\r\nSo if you could bring your script up-to-date with the current way it'd automatically work, or you can adapt it manually as I suggested above.\r\n\r\nIf any of my comments are unclear please don't hesitate to ask for clarifications.\r\n\r\nMeanwhile I will investigate why the model state_dict is almost empty under zero3 - this looks like a bug - making it work might help you move on w/o needing you to change your code.\r\n\r\nI will get back to you.\r\n\r\n", "I investigated and `model.state_dict()` returns some sort of placeholder with `tensor([1.],` for each weights and no real data, that's why `pytorch_model.bin` is tiny. Filed a request: https://github.com/microsoft/DeepSpeed/issues/872\r\n\r\nSo until we find a way to reconstruct it, I suggest to stick to zero2 otherwise you will remain locked in into DeepSpeed data files, that is you should be able to continue training but not being able to use it w/o deepspeed.\r\n", "While the Deepspeed team is sorting the addition of a method to extract model weights from its checkpoint, here is an update for you.\r\n\r\nDeepspeed stores the model weights in its checkpoint file (a file per gpu) which at the moment can only be loaded via its `deepspeed.load_checkpoint`. Therefore please adapt your code to rely on that to save and resume your custom models. Do not rely on `save_pretrained` and then expect `from_pretrained` to work, since the model weights won't be there.\r\n\r\nThe new method we are discussing will be able to convert the deepspeed checkpoint into consolidated from multiple gpus model weights. This is quite expensive so it shouldn't happen on each checkpoint saving and definitely shouldn't be the default because there might not be enough memory to do the consolidation (e.g. a model spread out over dozens of gpus).\r\n\r\nBottom line, should you choose to use deepspeed zero-3 things aren't as straightforward. And we will work out a solution in this case.\r\n\r\nI suppose it's a similar story with fairscale Sharded DDP, but I am working on DeepSpeed only at the moment and can't comment on the former. Unless @sgugger who did the initial integration of fairscale beats me to it I will be able to look at it once I complete the integration of DeepSpeed ZeRO-3, which is coming along nicely but requires changes on the DeepSpeed side - so it'll take some time.\r\n", "@exelents, here is how to solve your specific problem of:\r\n```\r\nclass T5Siamese(T5PreTrainedModel):\r\n[....]\r\n def init_from_base_t5_model(model_name_or_path='t5-base', output_root='./'):\r\n [...]\r\n model_left = T5EncoderModel.from_pretrained(MODEL)\r\n model_right = T5EncoderModel.from_pretrained(MODEL)\r\n```\r\nwith DeepSpeed zero-3.\r\n\r\nIf you don't mind continuing training and not being to retrieve the final weights until https://github.com/microsoft/DeepSpeed/issues/872 is addressed, here is what you can do immediately to be able to move forward:\r\n\r\nDo the above only when starting \"cold\", but when resuming from a checkpoint don't do that and let instead `T5Siamese` be restored from the deepspeed checkpoint at once. \r\n\r\nOnce we get the method to extract the model weights out of the DeepSpeed checkpoint, you can then recover both sub-model weights if you want to upload them to the hub or to take them elsewhere.\r\n\r\nPlease let me know if this solution resonates with you. Or if you run into any hiccups I haven't considered.\r\n\r\nNote that currently under zero-2 you're only recovering fp16 weights, so it is also not ideal either. So you want to use this solution for both cases.\r\n", "@samsontmr, would you kindly open a separate issue since while this is related the use-case is quite different. Please tag me and we will work on solving your use case there. Thank you!\r\n\r\np.s. also when you test please make sure you are using the `transformers` and `deepspeeed` master since there are constant fixes merged into it. ", "@stas00 Thank you for the explanation. So, to load stage-3 checkpoint I should make \"cold load\" from original T5 weights, and then load actual weights via `deepspeed.load_checkpoint` . The question is: is it possible to use this model in usual jupyter notebook, or usual python script, if I load model weights using deepspeed function? Or if I trained model via deepspeed once, I will be bound to it's runner forever?", "> So, to load stage-3 checkpoint I should make \"cold load\" from original T5 weights, and then load actual weights via deepspeed.load_checkpoint . \r\n\r\nI haven't tested it, but I can't think of any reason why it won't work. If you run into problems that I haven't considered please let me know.\r\n\r\n> The question is: is it possible to use this model in usual jupyter notebook, or usual python script, if I load model weights using deepspeed function? \r\n\r\nYes, of course.\r\n\r\nJust note that if you use the notebook directly and don't launch an external process which launches the distributed environment, you will be limited to 1 gpu and you will have to emulate the distributed environment like so:\r\n```\r\nimport os\r\ndist_env_1_gpu = dict(MASTER_ADDR=\"localhost\", MASTER_PORT=\"10999\", RANK=\"0\", LOCAL_RANK=\"0\", WORLD_SIZE=\"1\")\r\nfor k,v in dist_env_1_gpu.items():\r\n os.environ[k] = v\r\n```\r\nand please make sure you're on the master or very recent `transformers` version for this to work.\r\n\r\nBut if you just use the notebook to open a shell with the `deepspeed` launcher then you have no limitation of one gpu, e.g. see: https://github.com/stas00/porting/blob/master/transformers/deepspeed/DeepSpeed_on_colab_CLI.ipynb\r\n\r\n> Or if I trained model via deepspeed once, I will be bound to it's runner forever?\r\n\r\nI'm not sure what you ask here, as I don't know whether you refer to the `deepspeed` launcher, or something else. \r\n\r\n1. The `deepspeed` launcher is a more elaborate equivalent of `python -m torch.distributed.launch`. In simple cases of a single node you can use the latter. Here all DeepSpeed needs is to have a dedicated process per gpu and the distributed env set up (even in the case of one gpu).\r\n\r\n2. If you're asking whether your data will be locked into the deepspeed checkpoints, then at the moment the answer is yes.\r\nOnce https://github.com/microsoft/DeepSpeed/issues/872 is resolved you will be able to recover the consolidated weights and use them in any way you want.", "Ok, thank you for the explanation. I'm not sure if I could test these changes on my code soon, but I'll do it sooner or later.", "I just proposed yet another API in https://github.com/microsoft/DeepSpeed/issues/872:\r\n\r\n> being able to call `deepspeed.consolidate_weights()` in the rank0 process which would give users full weights back (perhaps with a bool arg of whether they want the fp16 or fp32 version). So now they can just save the model as they do with any other pytorch tools. This would only be practical for small-ish models. The key here is that while this would be somewhat costly they will be able to use their code almost w/o any change if they train in various ways and not just with deepspeed.\r\n\r\nSo if that was added then your current code would also work with just adding this newly proposed API. Let's see.", "@stas00 thanks! My problem is solved for now since I'm also using fp16 during fine-tuning so the current stage2 saves are good enough for me.", "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, @stas00. I have created an issue due to problems with converting model to fp32. Can you say something about it?\r\nhttps://github.com/microsoft/DeepSpeed/issues/1009" ]

[]

[]

[ "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! You might find this reply https://github.com/huggingface/transformers/issues/10789#issuecomment-802100991 by @stas00 of interest.", "@avionkmh, very soon it'll be supported, you may want to track: https://github.com/huggingface/transformers/pull/10753", "@stas00, @LysandreJik, thank you for your kind information. I'm going to track the issues you recommended. ", "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." ]

[]

[]

[]

[]

[]

[ "Indeed! Pinging @patil-suraj ", "urgh, my bad. Thanks for pointing it out. fixed!" ]

[]

[]

[ "Hello, thanks for opening an issue! We try to keep the github issues for bugs/feature requests.\r\nCould you ask your question on the [forum](https://discusss.huggingface.co) instead?\r\n\r\n@patrickvonplaten @stas00 \r\n\r\nThanks!", "@LysandreJik Thanks for pointing me in the right direction, I've moved the post over to the forum." ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Hi! Could you provide a way to benchmark the change so that we can see in which situations is the speedup visible? Thank you!", "> Hi! Could you provide a way to benchmark the change so that we can see in which situations is the speedup visible? Thank you!\r\n\r\nHi @LysandreJik , I just updated the benchmark code snippets and sample output in the description. Hopefully it can validate the change.", "Hi @LysandreJik , just wanted to follow up on this, is there anything else you would like to see on this PR? I also tried to run the slow test. It looked ok, but with a few connection errors: `ValueError: Connection error, and we cannot find the requested files in the cached path. Please try again or make sure your Internet connection is on`, and re-run didn't fix it. I was wondering how the slow tests look like on your end. Thanks!", "Awesome, thanks to both of you!" ]

[ { "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" } ]

[]

[ "Hey @LarsHill, \r\n\r\nYes, we should fix this indeed :-) I'll try to open a PR for this this week!", "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." ]

[]

[]

[ "Thanks @mansimane !\r\nI added the test and realized the implementation was not working as expected, so fixed it (I had forgotten this does not behave like the `DistributedSampler` that takes one sample every other num_replicas but sliced the indices at the beginning). If you want to have a last look to check I didn't do anything bad that would be great.\r\n\r\nThe rest of the changes are just the result of our styling scripts", "Changes look good to me too. I tested with microbatch size 2." ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "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." ]

[]

[]

[]

[]

[ { "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 } ]

[ "@patrickvonplaten \r\nWho should I tag for S2T ?", "@patil-suraj for s2t", "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 may be mistaken, but this test has appeared since the ProphetNet refactor, right? Is it due to that refactor, or is it a newly added test? " ]

[]

[]

[ "Yes you can. TAPAS is an encoder-model, and can be used in an encoder-decoder set-up, like so:\r\n\r\n```\r\nfrom transformers import EncoderDecoderModel\r\n\r\nmodel = EncoderDecoderModel.from_encoder_decoder_pretrained(\"google/tapas-base\", \"bert-base-cased\")\r\n```\r\nYou can specify any decoder you want, here I'm using BERT as a decoder, but you can also use GPT-2, etc (any model that supports the `is_decoder` logic).\r\n\r\nFor more information, see the [docs](https://huggingface.co/transformers/model_doc/encoderdecoder.html) of `EncoderDecoderModel`. ", "Thanks @NielsRogge ,\r\n\r\nI went through the concept of `EncoderDecoderModel` and I have a doubt in implementing it for TAPAS - \r\n\r\nUnlike normal BERT models, TAPAS tokenizer takes `table`, `queries` and `answers` for fine-tuning. So if I want to generate questions, should I skip questions for TAPAS (currently using `google/tapas-large`) encoder and give them to decoder (currently using `GPT-2-medium`) instead?", "Yes, if you want to generate questions given a table, then you should only encode the table (you can set `queries=None` when providing a table to `TapasTokenizer`). ", "Thanks @NielsRogge ,\r\n\r\nI'll implement and let you know", "Great, I'm curious to see the results. Another use case could be to generate answers given a question + table with an EncoderDecoder set-up. ", "Hi @NielsRogge ,\r\n\r\nI tried with the above approach by passing `table` to encoder and `queries` to decoder. But while encoding, it's giving warning as - **TAPAS is a question answering model but you have not passed a query. Please be aware that the model will probably not behave correctly** which is because of [this](https://github.com/huggingface/transformers/blob/master/src/transformers/models/tapas/tokenization_tapas.py#L1014). I thought passing 'queries' to TAPAS is mandatory.\r\n\r\nAnyhow I trained the model but it's not performing as expected. While inferencing, it is giving same question (not fully formed) for any input that I pass. Below is the sample snippet\r\n\r\n\r\n", "Yeah that warning is shown because TAPAS has been pre-trained on text-table pairs. You can ignore that warning, because we can still encode just the tables.\r\n\r\nWhat kind of generation method are you using? Greedy decoding, beam search? (See [this post](https://huggingface.co/blog/how-to-generate) for the different arguments you can pass to `.generate()`). ", "@NielsRogge ,\r\n\r\nI'm not specifying any generation method so it should be Greedy itself.", "Can you provide a notebook?", "Hi @NielsRogge ,\r\n\r\n[Here](https://colab.research.google.com/drive/1d8m_hmipL-1ZzU15LfA2XHmypKipvnJR?usp=sharing) is the colab link for the replica of my work. As I cannot share or upload any files from my Office' VPN, I created this notebook which is same as the one I'm working with in our VM's. Only change is I used `google/tapas-base` and `gpt2` in colab whereas I'm using `google/tapas-large` and `gpt2-medium` for my official work.", "Hi @NielsRogge \r\n\r\nI'm able to generate decent questions but only one generic question per table. How can I extend this to generate a question based on a particular cell value? \r\n\r\nBecause currently it's giving very basic ones like - \r\n`what are all of the countries?`\r\n`what are the names of all the drivers?`", "Ok great :) sorry I didn't have the time yet to look at your notebook. Looking at it now, it looks really clean!\r\n\r\nI think the questions that it will generate highly depend on the training data you provide. I see you're currently training on SQA questions, and only those for which `position==0`. These questions are almost always very generic, because SQA is a dataset involving conversational questions, which means that the first question (with position 0) is most of the time a very generic question regarding a table, and the ones that come after (with position 1, 2, 3) are then more specific follow-up questions (regarding particular cell values).\r\n\r\nSo either you can also add those follow-up questions to your training dataset, or consider train on questions of the [WTQ](https://nlp.stanford.edu/blog/wikitablequestions-a-complex-real-world-question-understanding-dataset/) dataset? (Note that there is an overlap between WTQ and SQA questions - SQA was created based on WTQ). Or maybe questions from the WikiSQL dataset (which is available in HuggingFace datasets)?\r\n\r\nVery nice use case!", "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 @NielsRogge, \r\nThanks for the TAPAS implementation!\r\n\r\nI'm trying to follow this use-case in order train the model to perform conditional generation from tables.\r\nSince TAPAS can encode the semi-structured meaning in tables, I guessed it was a good choice to use it as an encoder and say GPT2 as decoder.\r\n\r\nI however encountered a problem when trying to generate from that EncoderDecoder model:\r\nHere is the relevant pieces of code, this: \r\n\r\nresults in this error:\r\n\r\n\r\nI guess this is since model.generate() for EncoderDecoder does not expect to have the extra `token_type_ids` that TAPAS has. Can you think of a way I can make this work?\r\n\r\nThanks!" ]