Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

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

[]

[ "Hey @contribcode,\r\n\r\ncould you please provide a complete code snippet that would allow us to reproduce the error? Thanks!", "@patrickvonplaten thank you for your response. \r\n\r\n```python\r\nimport torch\r\ndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\r\nfrom transformers import GPT2LMHeadModel\r\nfrom transformers import GPT2TokenizerFast\r\ntokenizer = GPT2TokenizerFast.from_pretrained(\"gpt2\")\r\n\r\nspecial_tokens_dict = {'additional_special_tokens': ['[spanstarts]','[spanends]']}\r\nnum_added_toks = tokenizer.add_special_tokens(special_tokens_dict)\r\nmodel = GPT2LMHeadModel.from_pretrained(\"gpt2\", pad_token_id=tokenizer.eos_token_id)\r\nmodel.resize_token_embeddings(len(tokenizer))\r\nmodel.to(device)\r\n```\r\nI have posted the snippets about model definition, since I think the rest of the code does not affect the `generate` method. After this I just train the model. \r\n\r\nRegards.", "Hey @contribcode, \r\n\r\nI can execute the above code snippet without error -> could you include the part of the code where the error arises?", "Hey @patrickvonplaten, \r\n\r\nthe part of the code where the error occurs is the code in the original post\r\n\r\n```\r\ninput_ids = tokenizer.encode(train_texts['text'].iloc[0], return_tensors='pt')\r\nmodel.to('cuda')\r\ninput_ids.to('cuda')\r\nsample_output = model.generate(\r\n input_ids, \r\n do_sample=True, \r\n max_length=150,\r\n top_k=50,\r\n top_p=0.92\r\n)\r\n```\r\n\r\nand I get the error \r\n\r\n`RuntimeError: Input, output and indices must be on the current device`\r\n\r\nAs I mentioned in the original post, if I move the model and input_ids `to('cpu')`, it works.\r\n\r\nI also wanted to ask about parameters (`top_k`, `top_p`), are these values common for the generate function?", "Could you please post a single code snippet that I can copy/paste into a terminal and run to reproduce the error? The above snippet is not executable because `train_texts['text'].iloc[0]` is not defined. It's impossible to help you if we are not able to reproduce the error, I'm afraid.", "`train_texts['text'].iloc[0]` is just some text, for example: This man is a joke.\r\n\r\nThe whole code snippet is \r\n\r\n```python\r\nimport torch\r\ndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\r\nfrom transformers import GPT2LMHeadModel\r\nfrom transformers import GPT2TokenizerFast\r\ntokenizer = GPT2TokenizerFast.from_pretrained(\"gpt2\")\r\n\r\nspecial_tokens_dict = {'additional_special_tokens': ['[spanstarts]','[spanends]']}\r\nnum_added_toks = tokenizer.add_special_tokens(special_tokens_dict)\r\nmodel = GPT2LMHeadModel.from_pretrained(\"gpt2\", pad_token_id=tokenizer.eos_token_id)\r\nmodel.resize_token_embeddings(len(tokenizer))\r\nmodel.to(device)\r\n\r\ninput_ids = tokenizer.encode('This man is a joke.', return_tensors='pt')\r\ninput_ids.to(device)\r\nsample_output = model.generate(\r\n input_ids, \r\n do_sample=True, \r\n max_length=150,\r\n top_k=50,\r\n top_p=0.92\r\n)\r\n```\r\n\r\nand the error is what I mentioned earlier.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "I have the same issue with Transformers 4.4.2 and PyTorch 1.8.0\r\n\r\nUPD: also tried Transformers 4.5.1, the issue is still here", "Very nice @cloveranon !\r\n\r\nNow, do you know how/if GPUs can accelerate generation? Can this work in batches?\r\n\r\nMaybe, @patrickvonplaten can shed some light." ]

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

[]

[ "Hi there,\r\n\r\nI made some [integration tests](https://github.com/NielsRogge/transformers/blob/e5431da34ab2d03d6114303f18fd70192c880913/tests/test_modeling_layoutlm.py#L318) for both the base model (`LayoutLM`) as well as the model with a token classification head on top (`LayoutLMForTokenClassification`). These integration tests do not reveal any differences in terms of output on the same input data between the original implementation and the one in HuggingFace Transformers. So the implementation seems to be OK. Btw, the `segment_ids` you are referring to are called `token_type_ids` in the Transformers library.\r\n\r\nI also made a demo [notebook](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForTokenClassification_on_FUNSD.ipynb) that showcases how to fine-tune LayoutLMForTokenClassification on the FUNSD dataset, I'm getting quite good performance even though I'm not using Mask-RCNN features. Let me know if this helps you. ", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "@jethrokuan - Please try this on the transformers 4.1.0; issue looks to be fixed in this version", "Actually it looks like the problem is exasperated in newer HF. In a devbox with no CUDA:\r\n\r\n```\r\nPython 3.7.9 (default, Nov 18 2020, 11:18:33)\r\n[GCC 6.3.0 20170516] on linux\r\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\r\n>>> import transformers\r\n.../lib/python3.7/site-packages/torch/cuda/__init__.py:52: UserWarning: CUDA initialization: Found no NVIDIA driver on your system. Please check that you have an NVIDIA GPU and installed a driver from http://www.nvidia.com/Download/index.aspx (Triggered internally at /pytorch/c10/cuda/CUDAFunctions.cpp:100.)\r\n return torch._C._cuda_getDeviceCount() > 0\r\n```\r\n \r\nThis suggests that the simple action of importing anything from hf causes cuda initialization.\r\n\r\nI worked around it by deferring the import altogether:\r\n\r\n```\r\nclass Model:\r\n def __init__(self, path):\r\n self.path = path\r\n self.initialized = False\r\n\r\n def init(self):\r\n if not self.initialized:\r\n from transformers import ...\r\n self.device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\r\n self.model = AutoModelForSeq2SeqLM.from_pretrained(self.path)\r\n self.model.to(self.device)\r\n self.model.eval()\r\n self.initialized = True\r\n```\r\n " ]

[]

[]

[ "Please do not spam the issues: this is a duplicate of #9225. You can edit the title and the description, there is no need to send a new notification to everyone watching the repo.\r\nAlso, please use the [forum](https://discuss.huggingface.co/) for questions like this, we keep the issues for bugs and features requests only. In your case, this is a PyTorch command: `toch.distributed.get_world_size()` will return you the number of GPUs used.", "Hi Slvain, thank you, sorry I did not really realized the duplicate, sure, thanks " ]

[]

[]

[ "Duplicate of #9226" ]

[]

[]

[ "Hey @rabeehkarimimahabadi, \r\n\r\nOne should not call the `HfArgumentParser` with \r\n\r\n`--overwrite_output_dir True`, but just with `--overwrite_output_dir` to set the bool from `False` to `True`. If one wants to leave it as `False`, the arg should just not be passed. If the default field of this variable is already `True` and one wants to set it to `False` the argument should be passed as follows :`--no_remove_unused_columns`, see: https://github.com/huggingface/transformers/blob/master/src/transformers/hf_argparser.py#L83\r\n." ]

[]

[]

[ "this does not work since finetuner_trainer.py gets config as only one argument and local_rank is also passed with this command, is there an easy fix for this without parsing arguments from config? thanks ", "I solved it with updating finetune_trainer accepting --local_rank + config file" ]

[]

[]

[ "Maybe the author of MPNet can help here, @StillKeepTry ", "@patrickvonplaten problem is fix" ]

[]

[]

[ "cc @NielsRogge ", "It doesn't work because `type_vocab_sizes=[3, 256, 256, 2, 256, 256, 10]` which means the default max size of a table is 256*256. However, the tokenizer doesn't check it. \r\nStrictly speaking, it's not a bug but I think in that kind of tasks, many tables' size can exceed the default value.\r\nMaybe it's good to add codes to handle this problem.", "Ok I've investigated this a bit more. The problem here is that there's a column called \"Area (acres, 1872)\" in the table whose values cause the `column_ranks` token types to exceed the vocab size of 256. Deleting this column resolves the issue. Also, the table that you provide is actually (way) too large for TAPAS. Note that it only has a max sequence length of 512 tokens, and that in the original paper, only tables having at most 500 cells were considered. \r\n\r\nTo correctly truncate the table, you have to initialize TapasTokenizer as follows:\r\n\r\n```\r\nfrom transformers import TapasTokenizer\r\n\r\ntokenizer = TapasTokenizer.from_pretrained(\"google/tapas-base-finetuned-tabfact\", drop_rows_to_fit=True)\r\n```\r\n\r\nand then encode the table + query as follows:\r\n\r\n```\r\ninputs = tokenizer(table=table, queries=queries,\r\n padding=\"max_length\", \r\n truncation=True, return_tensors=\"pt\")\r\n```\r\n\r\ncc @LysandreJik \r\n\r\nIt's maybe a bit weird that people have to initialize `TapasTokenizer` with `drop_rows_to_fit` to True and set `truncation` to True when calling it.\r\n", "Thank you for the reply. It's really helpful!", "You're right @NielsRogge; should we switch the `drop_rows_to_fit` to `True` by default given that no models can handle it set to `False` when overflowing?", "I guess it should only be set to `True` when calling the tokenizer with `truncation` set to `True` or to `drop_rows_to_fit`. If the user does not specify truncation, and the table is too large, then an error will be thrown as shown [here](https://github.com/huggingface/transformers/blob/8eb7f26d5d9ce42eb88be6f0150b22a41d76a93d/src/transformers/models/tapas/tokenization_tapas.py#L1426).", "Fixed by #9507 ", "https://stackoverflow.com/questions/76849532/hugging-face-google-tapas-base-finetuned-wtq-indexerror-index-out-of-range" ]

[]

[]

[ "@ratthachat - thanks a lot for your PR! I agree that option 2) is the better option and think the general approach is exactly right. I left a couple of statements making that improve readability a bit IMO. Also could we add one test here: \r\nhttps://github.com/huggingface/transformers/blob/f38c4ad302dd34faef1137b13e9636f4408b0462/tests/test_modeling_rag.py#L502\r\ncalled `test_model_generate_from_context_input_ids` that verifies that using `input_ids` and `context_input_ids` yields the same result?", "@ratthachat Don't worry about `make style` I can fix that later!", "Thanks for the review, Patrick! I pretty much agree on every point, esp. the test case. I will do it. :)", "Hi Patrick, I think I addressed every suggestion :)\r\n[ Still fails on code-quality even already applied `make style` again :D ] @patrickvonplaten ", "Great job @ratthachat - I'm currently running all RAG slow tests to make sure nothing was unintentionally broken. Will merge once they all pass ", "Slow tests all passing", "Nice @patrickvonplaten !" ]

[]

[]

[ "@sgugger can you maybe review as well and merge if ok?" ]

[]

[]

[ "Hey @mukhtar-algezoli,\r\n\r\nIt seems like you are using tensorflow, but the script `run_clm.py` only works for PyTorch -> I'm afraid you have to use PyTorch in order to use `run_clm.py`.", "@patrickvonplaten thank you.", "hi @patrickvonplaten ,\r\nif i want use tensorflow to pretrain bert model with transformers , are there some examples like run_mlm.py ?", "https://github.com/huggingface/transformers/tree/master/examples/tensorflow/language-modeling\r\n\r\nuse tensorflow examples get the same error, How can i fix it?", "@Rocketknight1 for TF examples :-)", "@hrdxwandg This is an odd issue! Can you please install the most recent version of transformers (e.g. with `pip install --upgrade transformers`) and then try running the following in a console or notebook and tell me what you see?\r\n\r\n```\r\nfrom transformers import MODEL_FOR_CAUSAL_LM_MAPPING\r\n\r\nprint(MODEL_FOR_CAUSAL_LM_MAPPING)\r\n```", "> print(MODEL_FOR_CAUSAL_LM_MAPPING)\r\n\r\nOrderedDict([(<class 'transformers.models.roformer.configuration_roformer.RoFormerConfig'>, <class 'transformers.models.roformer.modeling_roformer.RoFormerForCausalLM'>), (<class 'transformers.models.bigbird_pegasus.configuration_bigbird_pegasus.BigBirdPegasusConfig'>, <class 'transformers.models.bigbird_pegasus.modeling_bigbird_pegasus.BigBirdPegasusForCausalLM'>), (<class 'transformers.models.gpt_neo.configuration_gpt_neo.GPTNeoConfig'>, <class 'transformers.models.gpt_neo.modeling_gpt_neo.GPTNeoForCausalLM'>), (<class 'transformers.models.big_bird.configuration_big_bird.BigBirdConfig'>, <class 'transformers.models.big_bird.modeling_big_bird.BigBirdForCausalLM'>), (<class 'transformers.models.camembert.configuration_camembert.CamembertConfig'>, <class 'transformers.models.camembert.modeling_camembert.CamembertForCausalLM'>), (<class 'transformers.models.xlm_roberta.configuration_xlm_roberta.XLMRobertaConfig'>, <class 'transformers.models.xlm_roberta.modeling_xlm_roberta.XLMRobertaForCausalLM'>), (<class 'transformers.models.roberta.configuration_roberta.RobertaConfig'>, <class 'transformers.models.roberta.modeling_roberta.RobertaForCausalLM'>), (<class 'transformers.models.bert.configuration_bert.BertConfig'>, <class 'transformers.models.bert.modeling_bert.BertLMHeadModel'>), (<class 'transformers.models.openai.configuration_openai.OpenAIGPTConfig'>, <class 'transformers.models.openai.modeling_openai.OpenAIGPTLMHeadModel'>), (<class 'transformers.models.gpt2.configuration_gpt2.GPT2Config'>, <class 'transformers.models.gpt2.modeling_gpt2.GPT2LMHeadModel'>), (<class 'transformers.models.transfo_xl.configuration_transfo_xl.TransfoXLConfig'>, <class 'transformers.models.transfo_xl.modeling_transfo_xl.TransfoXLLMHeadModel'>), (<class 'transformers.models.xlnet.configuration_xlnet.XLNetConfig'>, <class 'transformers.models.xlnet.modeling_xlnet.XLNetLMHeadModel'>), (<class 'transformers.models.xlm.configuration_xlm.XLMConfig'>, <class 'transformers.models.xlm.modeling_xlm.XLMWithLMHeadModel'>), (<class 'transformers.models.ctrl.configuration_ctrl.CTRLConfig'>, <class 'transformers.models.ctrl.modeling_ctrl.CTRLLMHeadModel'>), (<class 'transformers.models.reformer.configuration_reformer.ReformerConfig'>, <class 'transformers.models.reformer.modeling_reformer.ReformerModelWithLMHead'>), (<class 'transformers.models.bert_generation.configuration_bert_generation.BertGenerationConfig'>, <class 'transformers.models.bert_generation.modeling_bert_generation.BertGenerationDecoder'>), (<class 'transformers.models.xlm_prophetnet.configuration_xlm_prophetnet.XLMProphetNetConfig'>, <class 'transformers.models.xlm_prophetnet.modeling_xlm_prophetnet.XLMProphetNetForCausalLM'>), (<class 'transformers.models.prophetnet.configuration_prophetnet.ProphetNetConfig'>, <class 'transformers.models.prophetnet.modeling_prophetnet.ProphetNetForCausalLM'>), (<class 'transformers.models.bart.configuration_bart.BartConfig'>, <class 'transformers.models.bart.modeling_bart.BartForCausalLM'>), (<class 'transformers.models.mbart.configuration_mbart.MBartConfig'>, <class 'transformers.models.mbart.modeling_mbart.MBartForCausalLM'>), (<class 'transformers.models.pegasus.configuration_pegasus.PegasusConfig'>, <class 'transformers.models.pegasus.modeling_pegasus.PegasusForCausalLM'>), (<class 'transformers.models.marian.configuration_marian.MarianConfig'>, <class 'transformers.models.marian.modeling_marian.MarianForCausalLM'>), (<class 'transformers.models.blenderbot.configuration_blenderbot.BlenderbotConfig'>, <class 'transformers.models.blenderbot.modeling_blenderbot.BlenderbotForCausalLM'>), (<class 'transformers.models.blenderbot_small.configuration_blenderbot_small.BlenderbotSmallConfig'>, <class 'transformers.models.blenderbot_small.modeling_blenderbot_small.BlenderbotSmallForCausalLM'>), (<class 'transformers.models.megatron_bert.configuration_megatron_bert.MegatronBertConfig'>, <class 'transformers.models.megatron_bert.modeling_megatron_bert.MegatronBertForCausalLM'>)])", "That's extremely strange - the error above said that `MODEL_FOR_CAUSAL_LM_MAPPING` was `None`, but your code there clearly shows that it's importing the dict fine.\r\n\r\nAre you sure you're getting the same error as the commenter above?", "> That's extremely strange - the error above said that `MODEL_FOR_CAUSAL_LM_MAPPING` was `None`, but your code there clearly shows that it's importing the dict fine.\r\n> \r\n> Are you sure you're getting the same error as the commenter above?\r\n\r\nsorry, I don't check clearly. My error msg is below:\r\nTraceback (most recent call last):\r\n File \"run_mlm_tf.py\", line 63, in <module>\r\n MODEL_CONFIG_CLASSES = list(MODEL_FOR_MASKED_LM_MAPPING.keys())\r\nAttributeError: 'NoneType' object has no attribute 'keys'\r\n\r\nI use your method and recheck:\r\nfrom transformers import MODEL_FOR_MASKED_LM_MAPPING\r\nprint(MODEL_FOR_MASKED_LM_MAPPING)\r\n\r\nthe output is None", "Hi, sorry! I lost track of this issue because we marked it as closed, but I see we didn't resolve @hrdxwandg's problem. Are you still encountering the same error?", "@Rocketknight1 ; sorry to bump this, but I am getting the same error with the latest transformers version" ]

[]

[]

[ "Thanks for merging @LysandreJik but I want to know why the errors occurred when I ran `make style ` as I see in https://github.com/huggingface/transformers/pull/9217#discussion_r548898170" ]

[]

[]

[ "In particular I need to define a callback with on_save function, for this I need to be able to acces \"is_world_process_zero\" and I am not sure how to do it. If I want to pass trainer to callback, how can I make sure this is passed? thanks", "solved with understanding is_world_process_zero is accessible via state.is_world_process_zero " ]

[]

[]

[ "As I tried to explain earlier it's not ready yet for general consumption.\r\n\r\nIf you want it sooner see: https://github.com/huggingface/transformers/issues/9156#issuecomment-748501582" ]

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

[]

[ "If your bert model is an instance of `BertModel` then you should be able to save it using `self.your_model.bert.save_pretrained(path)` and load using `.from_pretrained`", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

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

[]

[ "hello, i convert bert model from tensorflow to pytorch to fine tune it and in train function specially in loss , pred_masks = model(inputs, labels) i got this error not enough values to unpack (expected 2, got 1) why model return one only it should return two one for loss and the second for pred_masks", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

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

[]

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

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

[]

[ "Thanks @stas00 for putting this together! I think there might be a few things we can do on the deepspeed side to smooth a few pain points out. Most of us our out of office until the new year but will definitely be taking a close look at this soon and help where we can.", "Thank you, @jeffra! \r\n\r\nNew year sounds perfect as a time for you to make suggestions if any, but meanwhile I think it's coming along nicely. ", "OK, so this PR also introduces a concept of `self.wrapped_model` so that we have less confusion about which is which in the trainer and user code. \r\n\r\n* `self.model` is always transformers model - this is the internal model\r\n* `self.wrapped_model` is a wrapped model - always the most external model which could be DDP(Transformers Model), DDP(Deepspeed(Transformers Model)), etc.\r\n\r\nIt's not documented yet, but @sgugger when you get a chance could you please check that it looks correct what I did here https://github.com/huggingface/transformers/pull/9211/commits/1510444399a4a088e6b14076bd857a47c3f2b1eb\r\n\r\nQuestions: \r\n\r\n1. is it correct that I set it to `None` if there is no wrapped model?\r\n2. would it be better to call it `model_wrapped` - so the two better align side by side during debug or IDE prediction completion engines?\r\n3. I'm not sure where to document this? And we should probably add a public API accessor?\r\n4. We can now probably remove and refactor the following code, as we now have a simpler way to get the internal model -\r\nhttps://github.com/huggingface/transformers/blob/cbe63949d76efd153a1f389f38fe9ce1287e06b0/src/transformers/trainer.py#L1635-L1651\r\n or is it used in some deep code where there is no trainer object? In which case this code won't work as it needs double unwrapping under deepspeed - `model.module.module` (since we have DDP too).\r\n \r\n I see it's used only in `floating_point_ops` which does have access to `self` so I'm not sure why it was needed in first place. Also in 2 tests, but that could be moved into the tests if need be.\r\n \r\nIf we want a general unwrap function it needs to do it recursively until there is no more `.module`.", "Let me think.\r\nFor 1, I think the `wrapped_model` should be the model, in this case, just to avoid the inconvenience of testing if `None`.\r\nFor 2, I have no strong opinion, so you can pick the version you prefer.\r\nFor 3, none of the attributes of the `Trainer` are properly documented yet. This could be added in the main docstring\r\nFor 4, yes, absolutely. This was a quick fix that was merged when I didn't get much time to do a nice solution, I though I had removed all uses of that function.\r\n\r\nCould you add the `wrapped_model` or `model_wrapped` in a separate PR? This would be easier to follow and not highjack the discussion on the deepspeed integration. We can rebase this when that PR is merged.", "> For 1, I think the `wraped_model` should be the model, in this case, just to avoid the inconvenience of testing if `None`.\r\n\r\nThen we somewhat lose information - `None` is telling us that notihng is wrapping the model. But I suppose we could achieve the same by `self.model == self.wrapped_mode` - OK, that works!\r\n\r\nThank you for the rest of the answers, @sgugger. Will integrate those and make a separate PR with wrapped_model.", "@stas00 Should it be mentioned in some README/documentation that folks can only use DeepSpeed with the PyTorch `Trainer` and not the TF `Trainer`? There's a hard dependency of using `torch.distributed` with the NCCL backend to use DeepSpeed.\r\n\r\nSecondly, what is the plan in terms of introducing DeepSpeed in the `transformers` [`setup.py`](https://github.com/huggingface/transformers/blob/master/setup.py) and the PyTorch GPU [`Dockerfile`](https://github.com/huggingface/transformers/blob/master/docker/transformers-pytorch-gpu/Dockerfile)?\r\n\r\nWhile DeepSpeed has a pip installable PyPI package, IIRC it is highly recommended that it be installed from source. Also, in order to use certain features in DeepSpeed such as 1-bit Adam, there are certain special installations to be done that do not come with the PyPI package. Will this PR support every underlying DeepSpeed feature? If not, can the scope of the initial DeepSpeed integration be defined clearly in some README/documentation, while allowing for further iterations in future to enable the utilization of more DeepSpeed features with the `transformers` `Trainer`?", "> @stas00 Should it be mentioned in some README/documentation that folks can only use DeepSpeed with the PyTorch Trainer and not the TF Trainer? There's a hard dependency of using torch.distributed with the NCCL backend to use DeepSpeed.\r\n\r\nYes, we should definitely be clear about that. thank you!\r\n\r\nAt the moment the idea is to put all the ZeRO related docs here: https://github.com/huggingface/transformers/pull/9208 (that PR covers fairscale at the moment)\r\n\r\n> Secondly, what is the plan in terms of introducing DeepSpeed in the transformers setup.py \r\n\r\nIt'll be up to users to install `deepspeed`, just like it's the case with `fairscale` or any other libraries the `transformers` core doesn't require. Currently if you use `--deepspeed` and you don't have it installed the trainer will assert with a suggestion to install that library.\r\n\r\n> and the PyTorch GPU Dockerfile?\r\n\r\nI have no idea. I don't see any reason why it can't be included. \r\n\r\nLet's do it in baby steps. First, make the support available, test it out, solve initial issues if any. Then worry about everything else?\r\n\r\n> While DeepSpeed has a pip installable PyPI package, IIRC it is highly recommended that it be installed from source. Also, in order to use certain features in DeepSpeed such as 1-bit Adam, there are certain special installations to be done that do not come with the PyPI package. Will this PR support every underlying DeepSpeed feature? If not, can the scope of the initial DeepSpeed integration be defined clearly in some README/documentation, while allowing for further iterations in future to enable the utilization of more DeepSpeed features with the transformers Trainer?\r\n\r\nAs I have shown in the example of the upcoming fairscale-support doc PR (we are waiting for fairscale to make a new pypi release before we merge it), we will document the same for DeepSpeed and address your questions. Your comments would be super-helpful for that document, so please save them for when we get to write that document. With your permission I can tag you on that future PR. Thank you.\r\n\r\nWrt to specifics let's see what ends up working out of box and what needs to be polished. I think the main issues will be bugs on the DS side. Otherwise there is a ton of features and I have only been testing a few.\r\n\r\nIf you feel inspired and are already experienced with DS it'd be awesome if you made a checklist of features and then between you and I, and anybody else who wants to contribute, test those features and check what's supported and report back to DS what is not. Since DS does most of the things on its own, I don't think there will be much to change in `transformers` Trainer once this PR is polished. I can be wrong of course.\r\n\r\n**edit**: actually there is no point waiting - I started adding notes into docs/source/training.rst in this PR. So already added a few of your comments - will need to expand those later.", "> how is the optimizer/scheduler creation handled? E.g. how is the fact we don't apply weight decay to some parameters or the proper schedule with the number of training steps handled? I don't see it in the current version of the code since deepspeed is responsible for creating the optimizer and scheduler\r\n\r\nIt wasn't, but it is now. Please have a look at what I just committed https://github.com/huggingface/transformers/pull/9211/commits/869173f62373a068c2cb932680093e7fd1d4ab78\r\n\r\n- schedulers - I was able to remap 2 `constant_with_warmup` + `linear`\r\n- optimizers - I think only Adam is possible at the moment with our cl args - but users can do whatever they want in the config file\r\n- the main logic is that if either `optimizer` or `scheduler` is configured in the config file we keep those and don't override the corresponding parts - if they are missing - we build our own config.\r\n- one thing I could use your help with, @sgugger. To use `linear` I needed `num_training_steps` early in `__init__` before distributed is set up. I tried postponing ds init for the time we have `num_training_steps` available in `train`, but it was too late - must set up dist early in `__init__` - so I added a hacky `get_num_training_steps` - please have a look if you can think of a better way to handle this chicken-n-egg problem. Thank you!", "> Left some comments. Not super happy about the `get_num_training_steps` method, but I don't see a way around it either.\r\n\r\nI agree. I tried to move `_init_deepspeed` into a later stage when we already have that figured out cleanly, but there is a chicken and egg problem there.\r\n\r\nI have made another commit to only run `get_num_training_steps` if absolutely necessary, and not by default for `--deepspeed`\r\n\r\n> Last thing I wonder is if it works okay for checkpointing (saving model/optimizer/scheduler after `save_steps` steps) and then resuming training? I.e., does DeepSpeed gets in the way of saving/reloading a model/optimzier state/scheduler state.\r\n\r\nIt's on the todo list in the OP, where I'm tracking what needs to be done. It's just a matter of doing it.\r\n", "@sgugger, as `_init_deepspeed` was born during this PR's evolution and it continues evolving and growing - are you happy with it in the main file or should it be moved into some other file, say `trainer_integrations.py`?", "@sgugger, I'd love some guidance from you to that last point\r\n \r\nWe have `deepspeed.DeepSpeedEngine.save_checkpoint` and `deepspeed.DeepSpeedEngine.load_checkpoint` ([doc](https://deepspeed.readthedocs.io/en/latest/model-checkpointing.html))\r\n\r\nSo in `_save_checkpoint` I will delegate to `deepspeed.DeepSpeedEngine.save_checkpoint` the 2 parts:\r\n```\r\n # Save model checkpoint\r\n # Save optimizer and scheduler\r\n```\r\nand leave the rest untouched, right?\r\n\r\nThen I see `_tune_save_checkpoint` - looks very similar to `_save_checkpoint` but it's much simpler - I guess doing the same here? \r\n\r\n`deepspeed.DeepSpeedEngine.save_checkpoint` saves/loads everything in one call - doesn't separate them into model/sched/optim -.\r\n\r\nAnd then on loading - there are quite a few places where this is happening. Any suggestions at how to approach this? Just match any place where there is `torch.load`?\r\n\r\nAlternatively, I can just give it a try and then you can comment on what I have missed / did wrong - if that sounds like a better use of your time.\r\n\r\nThank you!", "For the checkpointing, `_tune_save_checkpoint` is only there for Ray tune saving of checkpoints, so it's a bit different.\r\n\r\nI think it's fine to delegate to DS the saving of model/optimizer/scheduler. For the reloading, I'd go for using `torch.load` when reloading the model at the end of training but use deepspeed for loading optimzier/scheduler (it will also reload the model, but it should match the previous one).", "@sgugger, what should we do about tests - I only have a basic test that runs a full train/eval with `finetune_trainer.py` - like we have for `sharded_ddp`.\r\n1. do we want to test all the myriad of different config combinations - probably not since we can't control them anyway and perhaps it's best to do it on demand if and when we find bugs in our integration code.\r\n2. I could move it from `examples` to `tests` but then it'd be much more difficult to do a functional test I think - please correct me if I'm wrong. If there is an easy way then by all means - and `sharded_ddp` tests need to go there too then. we can also deal with this separately after concluding this PR as it's a project of its own.\r\n3. In either case - should we give it a dedicated test file/have integrations dedicated file/leave it for now as is? I think the latter.\r\n\r\nThe thing is we only pass on the configuration to DS and aren't really doing anything other than really getting out of its way, so there isn't much to test. Perhaps checkpointing should be tested as they will be part of the trainer code, which I think is the last big thing I need to integrate.\r\n\r\nThanks.\r\n", "> Could we however move it to `integrations` since the trainer code is already quite long? It only seems to use `self` to access the `args` so we could have its signature be `(args, model)`\r\n\r\nIt needs `self` for the num of training steps calculations, and I tried to only run it if it's needed (linear scheduler), so I can't pre-calculate and pass it to the `_init_deepspeed` method as it'd be inefficient for code that doesn't need it. I'm waiting for the DS team to follow up on my question about the chicken and the egg - perhaps there will be a way to avoid that hack.\r\n\r\nOtherwise I'm in total agreement with you wrt to moving it away and having only the args it needs.", "> It needs `self` for the num of training steps calculations, and I tried to only run it if it's needed (linear scheduler), so I can't pre-calculate and pass it to the `_init_deepspeed` method as it'd be inefficient for code that doesn't need it.\r\n\r\nThis is just a few math operations, I'd prefer we pass it either way to avoid the added complexity in the Trainer for everyone. \r\n\r\n> @sgugger, what should we do about tests\r\n\r\nFor now, a simple integration test seems enough (I imagine in the multi-GPU tests). It's going to be challenging to properly test all those frameworks that require multi GPU but this is a discussion for another PR. ", "> , I'd prefer we pass it either way to avoid the added complexity in the Trainer for everyone.\r\n\r\nSorry I don't know what you mean when you say that. Would you please rephrase that sentence with explicit terms - pass what/where - \"it\" being what? Thank you.", "> Sorry I don't know what you mean when you say that. Would you please rephrase that sentence with explicit terms - pass what/where - \"it\" being what? Thank you.\r\n\r\nI was speaking of the number of training steps, always pass it even if we end up not needing it because the scheduler does without.", "> I was speaking of the number of training steps, always pass it even if we end up not needing it because the scheduler does without.\r\n\r\nThank you for clarifying. \r\n\r\nI disagree. I imagine most of the time users will run their own ds_config.json with custom config, which would not use that expensively calculated number - so why introduce a repeated wasted overhead at calculating something that will never be used? (or am I exaggerating and it's not that much of an overhead?)\r\n\r\nWhy is it so bad if we pass the `trainer` object if it helps to optimize speed? I understand that since we are talking about changing `_init_deepspeed` not to be a trainer method it'd be good to make it so - but there is nothing wrong with it taking a `trainer` object as the first arg, no?\r\n\r\nI hope this whole thing might be resolved in a different, much cleaner way if I can only get help on it here https://github.com/microsoft/DeepSpeed/issues/633 - this is a hack we are doing that shouldn't be needed in first place...\r\n", "Oh, if you prefer sending it `self`, I have no objections. We do this too for some of the other integrations method.", "So our training steps calculation hack is no longer needed as @jeffra just happened to add `deepspeed.init_distributed` recently which is what we were missing, and adding which lead to a much cleaner code. yay!", "OK, this is pretty much done, just waiting for some feedback from the DeepSpeed team on the best way to handle checkpoints, but the code is in place.\r\n\r\nSo please kindly send me any final comments/suggestions/requests and let's merge it. Please re-invite old reviewers and add new ones if needed.\r\n\r\nThank you!", "deepspeed-0.3.10 has been just released by @jeffra on pypi - I verified that it works - so we are ready to merge this whenever you're happy with it.\r\n\r\nit'd be great if you tried running it too, since I think it has been only me running it, so my work is only as good as my environment is and I may not know of other culprits - e.g. I can't test with pytorch < pt-nightly since my card doesn't work with those pytorch versions.\r\n\r\nYou will need 2+ gpus to use it\r\n\r\nFirst install it:\r\n```\r\npip install deepspeed\r\n```\r\nAt the very least do the test:\r\n```\r\ncd examples/seq2seq\r\npytest -sv test_finetune_trainer.py -k deepspeed\r\n```\r\n\r\nOr if you want to fiddle with the normal run, here is what I have been using.\r\n```\r\ncd examples/seq2seq\r\nwget https://cdn-datasets.huggingface.co/translation/wmt_en_ro.tar.gz\r\ntar -xzvf wmt_en_ro.tar.gz\r\nexport BS=20; rm -r output_dir; CUDA_VISIBLE_DEVICES=0,1 PYTHONPATH=../../src USE_TF=0 deepspeed --num_gpus=2 ./finetune_trainer.py --model_name_or_path t5-small --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_predict --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 5 --n_train 100 --n_val 100 --n_test 100 --deepspeed ds_config.json --fp16 --save_steps 1\r\n```\r\n", "@sgugger, \r\n\r\n1. While working on docs I discovered that DS does its own gradient clipping (that doc was buried and I didn't see it) so I had to undo the code in the trainer that did that on behalf of DS - just skipping it\r\n2. I did a major rewrite/expansion of the docs (including the fairscale section) - so please kindly have a look. It's mainly mirroring the config logic in the integration code.\r\n3. In the docs I used consistently Trainer (upcase) to refer to HF trainer. I know you didn't like it when I did that for Issue in a different PR, let me know if you prefer it to be a lowercase trainer.\r\n\r\nWhile this PR is perfectly ready for a final review, I need to wait for https://github.com/microsoft/DeepSpeed/pull/656 to be answered before we can merge this as I'm unsure about their defaults for gradient clipping.\r\n\r\nThank you.\r\n", "> Went through the documentation and left comments. \r\n\r\nAwesome - thank you - all integrated.\r\n\r\n> On the optimizer side, it doesn't seem like DeepSpeed supports AdamW from what you're saying, so we should document the default optimizer is changed at the very beginning of the DeepSpeed section. It does change drastically the value of `weight_decay` to use.\r\n\r\nI found a way to use AdamW, thank you for catching that, @sgugger. I documented the nuances.\r\n", "I think the DeepSpeed team is on vacation, as there is no response since several days. And since I have no way of talking to anyone there, I have no way of knowing when they will be back. So I will go ahead and merge this so that others can start experimenting and then we can fix whatever needs to be fixed when I get the gradient clipping Issue answered.", "Amazing work @stas00 !" ]

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

[]

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

[]

[]

[ "Hi @adithyaan-creator \r\n\r\nSeems like you used `pytorch-lightning` for training, `pl` prefixes every `state_dict` key with `model.` so to load the `pl` checkpoint to hf format, remove the `model.` prefix from `state_dict` keys.\r\n\r\n```python3\r\ndef remove_prefix(text: str, prefix: str):\r\n if text.startswith(prefix):\r\n return text[len(prefix) :]\r\n return text # or whatever\r\n\r\nstate_dict = {remove_prefix(k, \"model.\"): v for k, v in state_dict.items()}\r\nhf_model.load_state_dict(state_dict)\r\nhf_model.save_pretrained(save_path)\r\n```\r\nAlso, would be nice if you could post code snippet instead of screenshots :)\r\n\r\n", "Hi @patil-suraj,\r\nI tried it with loading the 'state_dict' item to the model, and getting the following error \"Unexpected key(s) in state_dict: \"decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight\". \".\r\n\r\nThe code is :: \r\ndef remove_prefix(text: str, prefix: str):\r\n if text.startswith(prefix):\r\n return text[len(prefix) :]\r\n return text # or whatever\r\n\r\nckpt = {remove_prefix(k, \"model.\"): v for k, v in ckpt['state_dict'].items()}\r\n\r\nmodel.load_state_dict(ckpt)\r\nmodel.save_pretrained(\"/content/model\")`\r\n ", "Hey @adithyaan-creator,\r\n\r\nYou can ignore this warning, see: https://github.com/huggingface/transformers/pull/9231", "@patrickvonplaten the code that was merged wasnt working out, but since the weight was unnecessary I deleted the weight layer and it worked out.\r\n'del ckpt[\"decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight\"]'\r\nThanks.", "@adithyaan-creator Hi, I am having the same issue now. Could you elaborate on how and where did you add 'del ckpt[\"decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight\"]' ? Thank you very much!" ]

[]

[]

[]

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

[]

[ "I realised that\r\n\r\n`tokenizer.model.save(\"./tokenizer\")`\r\n\r\nIs unnecessary. I've started saving only the `tokenizer.json` since this contains not only the merges and vocab but also the pipeline.\r\n\r\nAnd I noticed that `tokenizer.save_pretrained()` has a parameter `legacy_format` which defaults to True. When I set it to false it properly round trips (i.e. saves out the unified tokenizer.json rather than the model files (merges.txt and vocab.json)\r\n\r\nThe only issue now is if I create a trainer\r\n\r\n```\r\ntrainer = Trainer(\r\n model=model,\r\n tokenizer=tokenizer\r\n args=training_args,\r\n train_dataset=tokenized_datasets[\"train\"],\r\n eval_dataset=tokenized_datasets[\"validation\"] ,\r\n data_collator=data_collator,\r\n)\r\n\r\nlogger.info(\"*** Train ***\")\r\ntrainer.train(model_path=model_path)\r\ntrainer.save_model()\r\n```\r\n\r\nThe last line doesn't accept kwargs so it saves the tokenizer in the legacy format.\r\n\r\nA workaround is to do it manually ie.\r\n\r\n```\r\ntrainer = Trainer(\r\n model=model,\r\n #tokenizer=tokenizer\r\n args=training_args,\r\n train_dataset=tokenized_datasets[\"train\"],\r\n eval_dataset=tokenized_datasets[\"validation\"] ,\r\n data_collator=data_collator,\r\n)\r\n\r\nlogger.info(\"*** Train ***\")\r\ntrainer.train(model_path=model_path)\r\ntrainer.save_model()\r\ntokenizer.save_pretrained(output_dir, legacy_format=False)\r\n```\r\n\r\nOnly issue is I don't see a workaround to ensure the checkpoints are created using `legacy_format=False`\r\n\r\n\r\n\r\n", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "Indeed, will fix that notebook on Monday. Thanks for reporting!", "Note that this notebook is quite old, so you might prefer the most recent one [here](https://github.com/huggingface/notebooks/blob/master/examples/text_classification.ipynb). I fixed the issue by pinning the transformers version inside it, we will keep it for legacy reasons but not update it.", "Same problem on this notebook: [https://colab.research.google.com/drive/1-JIJlao4dI-Ilww_NnTc0rxtp-ymgDgM?usp=sharing](https://colab.research.google.com/drive/1-JIJlao4dI-Ilww_NnTc0rxtp-ymgDgM?usp=sharing).\r\nThis notebook is referenced here: [https://huggingface.co/transformers/training.html#additional-resources]().\r\nAlso, if I understand correctly, `evaluate_during_training=True` should be replaced with `evaluation_strategy='epoch'`, and then it will work fine. ", "I think this notebook is community-contributed (it's not in our repos) so it's up to the person who wrote it to port it to the most recent version of transformers." ]

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

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

[ "Doesn't that script also loads and preprocess the data? From what you're reporting, I don't interpret this as \"transformers takes a long time to load the model\" (since the line that does that takes the same time as a torch load) but as \"stuff that happens in that script before the model loading takes a lot of time\" (which is probably data preprocessing + the 3s to import transformers if TF is in your env). Or am I missing something?\r\n", "Perhaps my first post is confusing, what I did is bracketing the `torch.load` call in modeling_utils.py:\r\n```\r\n start_time = time.time()\r\n state_dict = torch.load(resolved_archive_file, map_location=\"cpu\")\r\n end_time = time.time() - start_time\r\n```\r\nSo all the other stuff isn't being measured, just the `torch.load` call. ", "Ah, I understand better. I don't think your comparison is fair: `AutoModel.from_pretrained` does two things: creating a model and filling it with the weights. From a small experiment in timing on my side, I believe all the time is spent in the model creation. So you should compare the timing of creating the model and loading the weights inside to have something that's apple to apple.", "I removed the 2nd part that was showing the same issue from a different angle, as it appears to just confuse and isn't contributing to understanding the issue at hand.\r\n\r\nThere is just `state_dict = torch.load(resolved_archive_file, map_location=\"cpu\")` call - and nothing else. On its own:\r\n`python -c \"import torch; torch.load('/hf/transformers-master/data/distill-mbart-en-ro-12-4/pytorch_model.bin')\"`\r\n it takes ~1s, the exact same call inside `modeling_utils` takes 22+ secs.", "OK, somehow I made a mistake and was taking the snapshot of startime before `model = cls(config, *model_args, **model_kwargs)` and not `torch.load()` - my apologies :( and thank you for double checking my invalid report.\r\n\r\n```\r\n import time\r\n t0 = time.time()\r\n model = cls(config, *model_args, **model_kwargs)\r\n t1 = time.time()\r\n state_dict = torch.load(resolved_archive_file, map_location=\"cpu\")\r\n t2 = time.time()\r\n print(f\"cls init { round(t1-t0, 4)}\")\r\n print(f\"load { round(t2-t1, 4)}\")\r\n import sys\r\n sys.exit(0)\r\n\r\n```\r\n\r\n```\r\ncls init 21.2055\r\nload 0.5074\r\n```\r\n\r\nSo it's setting up the model that takes so long, just as you said.\r\n\r\nCan this somehow be sped up? I was integrating deepspeed and re-running the same command repeatedly and 23 extra secs of waiting to just discover that something is off was very painful for debugging. All the failures happened at much later stages. I worked around it it by switching to a tiny model, but even that takes some secs.\r\n\r\nCan we think of a way to make an image and load it rather than rebuilding the model from scratch? So we torch.load the weights, but also cache the model image itself and load it too, rather then create it anew. It seems to be so wasteful and slow if I'm not debugging the model creation but say tuning up something in the trainer and I want the other parts to load blazingly fast and get me to the point of interest quickly. What would be the best way to approach such need?\r\n", "So doing profiling on model instantiation code it can be seen that `_init_weights` is where some 75% of that slowdown happens\r\n\r\n```\r\n ncalls tottime percall cumtime percall filename:lineno(function)\r\n 354 18.942 0.054 18.942 0.054 {method 'normal_' of 'torch._C._TensorBase' objects}\r\n 225 2.286 0.010 2.286 0.010 {method 'uniform_' of 'torch._C._TensorBase' objects}\r\n````\r\n\r\nSo we are completely wasting time doing init weights, since we are immediately replacing them. (with the exception to `SinusoidalPositionalEmbedding` which do not get loaded from the pretrained model). \r\n\r\nIf you prefer the visual version:\r\n\r\n\r\n\r\nChances are that model init needs to be made context aware and not init weights which will be immediately replaced. Thoughts?\r\n\r\nThat would make `transformers` so much faster to start! (e.g. think the model pages website which takes forever to load a model).\r\n\r\nThe profiling was done with:\r\n```\r\n# prep\r\npip install graphviz gprof2dot\r\ncat <<EOT > prog\r\nfrom transformers import AutoModelForSeq2SeqLM\r\nAutoModelForSeq2SeqLM.from_pretrained(\"sshleifer/distill-mbart-en-ro-12-4\")\r\nEOT\r\n\r\n# text profile\r\nUSE_TF=0 PYTHONPATH=src python -m cProfile -s tottime prog > profile.txt\r\nhead -10 profile.txt\r\n\r\n# visual profile\r\nUSE_TF=0 PYTHONPATH=src python -m cProfile -o profile.pstats prog\r\ngprof2dot -f pstats profile.pstats | dot -Tsvg -o callgraph.svg\r\ndisplay callgraph.svg\r\n```\r\n\r\n", "If we see a significant gain in loading time, maybe it's worth to explore a way to only apply `init_weights` on missing layers. Not sure how easy it would be to implement it though...\r\n\r\nMaybe a `init_weights` function arg in `__init__` might make sense:\r\n```python\r\nmodel = cls(config, init_weights=False, *model_args, **model_kwargs) # don't call init_weights, but initialize all weights to zero because it's much faster\r\n# load weights into model and get missing layers\r\n# init missing layers\r\n```", "Yeah Patrick's suggestion is probably the best, though I'm not sure it can easily be achieved in the current API. Note that this is only one slowdown at the beginning of training, so I don't think this should be high priority.", "I totally get it that it's not high priority, since most people don't care for a slow start when they run it non-stop for hours - it only affects people who need a quick start - which is the case when debugging something or as I suggested the demo function on the model pages which takes a really long time to load.\r\n\r\nIn the case of BART, its deterministic segments do the init internally, so it's enough to just monkeypatch as a proof of concept:\r\n\r\n```\r\n # modeling_utils.py::from_pretrained\r\n init_weights_orig = PreTrainedModel.init_weights\r\n def init_weights_pretrained(self):\r\n # self.apply(self._init_weights)\r\n if self.config.pruned_heads: self.prune_heads(self.config.pruned_heads)\r\n self.tie_weights()\r\n \r\n PreTrainedModel.init_weights = init_weights_pretrained\r\n model = cls(config, *model_args, **model_kwargs)\r\n PreTrainedModel.init_weights = init_weights_orig\r\n\r\n```\r\nand this command:\r\n```\r\nPYTHONPATH=../../src USE_TF=0 time python -c 'from transformers import AutoModelForSeq2SeqLM; AutoModelForSeq2SeqLM.from_pretrained(\"sshleifer/distill-mbart-en-ro-12-4\")'\r\n```\r\ngoes from 25sec to 8secs. The instantiation goes from 22 secs to 5 secs.\r\n\r\nThere are few `uniform_` calls left which account for 2.3 extra secs, which if shaves off we should be down to 2-3 secs (from 22!).\r\n\r\nI quickly checked that the core functions normally - same scores - well, I did just one finetune_trainer run.\r\n\r\nOne way is to solve this as @patrickvonplaten suggested, and I'm also thinking of changing the design a bit. So that each model has a normal `init_weights` and `init_weights_pretrained` - then it's very clear to the developer what goes where and then simply invoke one or the other depending on the context. And then it's just a matter of choosing how to signal the context.\r\n", "I don't see how you could have an `init_weights_pretrained`: it depends on the checkpoint you pass: if you pass the checkpoint of a `BertModel` to `BertForMaskedLM`, you just have one bias to initialize (if weights are tied). But if you pass a checkpoint of a `BertForMaskedLM` checkpoint then you have nothing to initialize. And the same holds for every variant (which would have different specific weights to initialize in case of a pretrained model) so I don't really see how you can do this API-wise.\r\n\r\nThe only way I see through it is to allow the `init_weights` to get the list of model parameters to randomly initialize, but since we use the `apply` method afterward (and rely on it to get modules inside each model specific `_init_weights` method) I don't see how to use it properly. It would probably require some clever recursive method.\r\n\r\nAgain, lots of headaches and possibilities for errors for an end result that doesn't strike me as high priority.\r\n\r\n> it only affects people who need a quick start - which is the case when debugging something or as I suggested the demo function on the model pages which takes a really long time to load.\r\n\r\nIt doesn't take 25 seconds on a tiny model, only a big one. So I'd suggest debugging on a tiny model :-)", "Thank you both for entertaining possible approaches and suggesting that you are not quite seeing a smooth solution. I just don't know enough about all of it, so I'm surely missing on cases I haven't thought of, but somehow in my mind it looks simple. The devil is in the details.\r\n\r\n> It doesn't take 25 seconds on a tiny model, only a big one. So I'd suggest debugging on a tiny model :-)\r\n\r\nUnfortunately the tiny model approach doesn't work with debugging OOM in deepspeed, as its configuration correlates to the model size. I guess it's not special to deepspeed at all. So the tiny model trick works for checking mechanics (i.e. that the code compiles), but isn't helpful for OOM debug.", "@patrickvonplaten, @sgugger, @LysandreJik - could we please revisit this - working on making t5-11b train was painful - it was taking really really really long time to init the model, just to drop it and replace with pre-trained weights. Transformers is mainly about pre-trained models, so perhaps this can be made somehow configurable? \r\n\r\nWe know when a pretrained model is loaded, so why not propagate that information and let the model know it's being loaded in pre-trained mode, so that it could skip any weight inits that are going to be replaced anyway?\r\n\r\nAnd while we are at it, I don't suppose there is a way to involve more than one CPU core in loading the model? I guess that would be a question for pytorch.\r\n\r\nThank you!", "I'm happy to add such a featurue. It should be feasible to only initialize those layers that are not in the saved `.pt` file.", "Indeed, this would be a welcome feature, big models aren't going away.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "@patrickvonplaten, I should probably work on it - since it doesn't seem like you will have time any time soon.", "It's on my To-Do List, but still don't think, I'll be able to take a look within the next 2,3 weeks - sorry :-/ If you find some time for this, it would be great", "Ihave finetuned a longformer encoder decoder model, and trying to convert it into an api but model takes too long to load that api throws a not responding error.\r\n\r\nKindly if anyone can guide me on how can I reduce the time for the model to load.\r\nThank You in advance.", "Hello @AyeshaSarwar,\r\n\r\ncould you please use the forum: https://discuss.huggingface.co/ instead for such questions? We don't support Flask compatibility in `transformers`. Please keep in mind that the issues are mainly used for issues related to just `transformers`.\r\n\r\nThanks", "Im on the same boat as @stas00 . I understand that the code need to maintain a wider compatibility across the oceans of models, but people needs a working workaround before an elegant solution born into reality. I believe as huggingface slowly graduating from pure research field, more and more people are being hurt by the tremendous model initialization time. \r\nHoping for a change", "@DeXtmL, this thread is 2 years old - the particular problem I raised in this Issue has been solved a long time ago. The model is no longer being init'ed twice.\r\n\r\nIf you feel something is still slow please start a new Issue.\r\n\r\nthank you." ]

[]

[]

[ "Hey @Jess0-0,\r\n\r\nYou can load TF checkpoints into PT and vice-versa via:\r\n\r\n`XLMRoberta.from_pretrained(...., from_tf=True)`\r\n\r\nor \r\n\r\n`TFXLMRoberta.from_pretrained(...., from_pt=True)`.\r\n\r\nSince `XLMRobertaForSequenceClassification` is just an alias of `RobertaForSequenceClassification` there should be no problem in doing \r\n\r\n```python\r\nTFRobertaForSequenceClassification.from_pretrained(\"<your/path/to/saved/xlm/roberta/pytorch/dir>\", from_pt=True)\r\n```" ]

[]

[]

[]

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

[]

[ "I don't agree on an always-document-everything basis, as the documentation page of each model is already quite long (which may be the reason not a lot of people seem to be reading them...) In the case of generate, since there is no link to the part where the generate method is documented, we could add the documentation of the generate method there, but in the case of tokenizers for instance, I prefer our approach where we document just the main methods in the subclasses and point to the superclass for less important ones.\r\nWhich again also goes in the direction of documenting generate in the models that implement it, so why not go with it. I don't think it should be harder than adding `generate` to the members field, though sphinx might not like `GenerationMixin`. If anyone wants to give it a try, I'll happily review a PR.", "`generate` was just an example, so let's not solve just a sub-case if possible.\r\n\r\nAs mentioned in OP if you feel that including full docs is too much - a short entry that links to the super class or mixin's method is a satisfactory solution, but the user shouldn't hunt for where that entry might be elsewhere. \r\n\r\nAlso for methods such as `T5ForConditionalGeneration.generate` mentioned in free prose it'd be awesome to have it linked to the right doc entry. I understand it should happen automatically with sphinx if there is an actual entry to link to. \r\n\r\n> the documentation page of each model is already quite long (which may be the reason not a lot of people seem to be reading them...)\r\n\r\nI don't know what you mean by \"not a lot of people seem to be reading them\" - do you imply that users ask a lot of questions that are already answered by the existing documentation or do you have some other way to measure the \"not a lot\" part?\r\n\r\n", "> I don't know what you mean by \"not a lot of people seem to be reading them\" - do you imply that users ask a lot of questions that are already answered by the existing documentation\r\n\r\nYes, I was implying exactly that. For the general rule, I'd keep it to: main parts of the API of a class should be documented in which class (within reasons) and more minor parts should be documented once in the superclass/mixin with a link from all the subclasses (like what is done for `Tokenizer`)\r\n", "I agree the process is long an feels like a lot of unnecessary steps.", ">> I don't know what you mean by \"not a lot of people seem to be reading them\" - do you imply that users ask a lot of questions that are already answered by the existing documentation\r\n> Yes, I was implying exactly that. \r\n\r\nDo we have a way to reach out and ask why users don't per-use docs before asking questions? \r\n\r\nIs this:\r\n\r\n* an issue of the quality/readability/navigationability of the docs\r\n* users are just lazy (as a virtue)\r\n* users don't know that there are docs to search through\r\n* the search engine doesn't have the smarts to show most relevant info and shows too many irrelevant hits? many docs search engines are pretty crappy in my experience. It's usually the best to use google with: \r\n\r\n`\"my search query\" site:https://huggingface.co/transformers/`\r\n\r\nto find the best information. (swap in whatever other docs site you need, I wasn't singling out transformers, it was just an example)\r\n\r\n> For the general rule, I'd keep it to: main parts of the API of a class should be documented in which class (within reasons) and more minor parts should be documented once in the superclass/mixin with a link from all the subclasses (like what is done for Tokenizer)\r\n\r\nThat works 100% for me. \r\n\r\nWhat needs to be done to make this happen? Definitely no rush.", "I don't know whether this helps, but when I managed a huge open source project many years back I trained our users to per-use docs by almost never replying with the repeat information they requested but with a direct link to where it was in the documentation. Over time it was very clear to the community that the information is available in the docs and less and less questions were asked and more and more one line answers with the link to the right info were posted by various users of that community.\r\n\r\nWhen one answers in clear text it signals users that the information is in your brain and not outside of it.\r\n\r\nThat is just how my experience had been, YMMV.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "@sgugger, how do we resolve this - please let me know if I can be of help.", "I think we should agree to disagree on this point. I didn't see anything in the last survey that showed a strong desire from the user to have the documentation of each model page be even longer, quite the opposite, so I stand by what I said with the generic methods like `generate` be documented in the base classes unless they have some model-specific behavior that needs to override that documentation. It's the same for the tokenizer encode and call methods.", "I think there are two different possibilities here:\r\n1. duplicating content - you disagree with - OK\r\n2. having a placeholder with a link to the main location of this entry\r\n\r\nFor example, when documentation refers to `T5ForConditionalGeneration.generate` it could:\r\n1. link directly to the general `generate` doc entry\r\n2. link to an entry `generate` in the t5 doc which will link to the main location of this entry\r\n\r\nWhy make the user work extra hard searching for something, when it can be automated and get the user what they need in 1 or 2 quick clicks.\r\n\r\nI'm not asking for a hypothetical nice-to-have feature, I often find myself frustrated when I can't quickly link to a method when I address someone's Issue and have to search for it. So it's a very selfish request. And I'm willing to work for it, since it'll save my time and minimize frustration where it doesn't have to happen.", "> For example, when documentation refers to T5ForConditionalGeneration.generate it could:\r\n> 1. link directly to the general generate doc entry\r\n\r\nI am not sure I know how to do that in sphinx, but if it's possible, I'm all for it!\r\n\r\n> 2. link to an entry generate in the t5 doc which will link to the main location of this entry\r\n\r\nThis one will require writing a custom docstring for the generate method of T5. Probably more possible and can be automated with a decorator to use on the classes with a generate method I think.", "Great. I will research it then and report back if I find a way. \r\n\r\nThank you for your feedback, @sgugger " ]

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

[]

[ "Is there any update on this? \r\n\r\nSeems like it allows dynamic batching. ", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "As the trace suggests, the error seem to come from the `_reorder_cache` method in `generation_utils.py`. Since the model is parallelized among multiple devices, it fails since the device of `beam_idx` and `layer_past` don't match for all layers.\r\n\r\nI just tried to modify line 229 in `generation_utils.py` to:\r\n```\r\nreturn tuple(layer_past.index_select(1, beam_idx.to(layer_past.device)) for layer_past in past)\r\n```\r\nwhich seems to work. \r\nI'm happy to file a PR with this change if you approve. Please let me know if there is anything I should be aware of, or pay extra attention to.", "FWIW, this fix doesn't currently work for T5, as the fix to `_reorder_cache` is not reflected in the `modeling_t5.py` file. Following the above, changing [this line](https://github.com/huggingface/transformers/blob/fa84540e98a6af309c3007f64def5011db775a70/src/transformers/models/t5/modeling_t5.py#L1679) to `layer_past_state.index_select(0, beam_idx.to(layer_past_state.device)),` appears to fix it.\r\n\r\n@patrickvonplaten ", "@OyvindTafjord - would you mind opening a new PR for it? :-)" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "cc @sgugger @patrickvonplaten @LysandreJik @lhoestq ", "Nice, good to merge then!", "Awesome, thank you so much for the reviews @lhoestq @patrickvonplaten -- happy holidays!", "Thanks guys!", "@amogkam @patrickvonplaten I need some help to implement an end-to-end retrieval training feature for the rag with Ray.\r\n\r\nHow can I run document encoding and indexing with an updated doc-encoder (context encoder network that kept frozen in the original RAG) using a Ray actor separated from the main training process? \r\n\r\nHow can I access the document index inside Ray actors during the training incase I want to update the index, say in every 5000 steps. \r\n\r\n", "@shamanez could you open a new issue to track this?", "@richardliaw \r\n\r\nI have already opened one a few weeks ago. Please refer to this [issue](https://github.com/huggingface/transformers/issues/9646)\r\n\r\nI added a new issue explaining the exact problem in [this](https://github.com/huggingface/transformers/issues/10135)" ]

[]

[]

[ "may I ask for one more bit, while you're at it - sorting the metrics before printing them out?\r\n\r\nI added this already into the final json file writing, but it'd make it easier to read the info logs.\r\n\r\nNow we have:\r\n\r\n```\r\n2020-12-18 11:51:55 | INFO | __main__ | val_loss = 368.4116\r\n2020-12-18 11:51:55 | INFO | __main__ | val_bleu = 26.3465\r\n2020-12-18 11:51:55 | INFO | __main__ | val_gen_len = 31.2\r\n2020-12-18 11:51:55 | INFO | __main__ | val_runtime = 22.1214\r\n2020-12-18 11:51:55 | INFO | __main__ | val_samples_per_second = 9.041\r\n2020-12-18 11:51:55 | INFO | __main__ | epoch = 1.0\r\n2020-12-18 11:51:55 | INFO | __main__ | val_n_objs = 200\r\n```\r\n\r\nas compared to sorted:\r\n```\r\n2020-12-18 11:51:55 | INFO | __main__ | epoch = 1.0\r\n2020-12-18 11:51:55 | INFO | __main__ | val_bleu = 26.3465\r\n2020-12-18 11:51:55 | INFO | __main__ | val_gen_len = 31.2\r\n2020-12-18 11:51:55 | INFO | __main__ | val_loss = 368.4116\r\n2020-12-18 11:51:55 | INFO | __main__ | val_n_objs = 200\r\n2020-12-18 11:51:55 | INFO | __main__ | val_runtime = 22.1214\r\n2020-12-18 11:51:55 | INFO | __main__ | val_samples_per_second = 9.041\r\n```\r\n\r\nor I could make a PR later if it's too unrelated... I probably should do that and not waste your time.\r\n\r\nThanks." ]

[]

[]

[ "Please try\r\n`_, cls_hs = self.bert(sent_id, attention_mask=mask)`\r\nto\r\n`_, cls_hs = self.bert(sent_id, attention_mask=mask)[:2]`\r\nor\r\n`_, cls_hs = self.bert(sent_id, attention_mask=mask, return_dict=False)`", "It fixes the issue. Thanks!", "> Please try\r\n> `_, cls_hs = self.bert(sent_id, attention_mask=mask)`\r\n> to\r\n> `_, cls_hs = self.bert(sent_id, attention_mask=mask)[:2]`\r\n> or\r\n> `_, cls_hs = self.bert(sent_id, attention_mask=mask, return_dict=False)`\r\n\r\nYou the man! I also got stuck in this issue for an hour, and your solution just fixes it perfectly! \r\nThanks man!" ]

[]

[]

[ "Looking into it!\r\n\r\nAn easy fix for now would be the following:\r\n\r\n```python\r\nfrom transformers import AutoTokenizer, AutoModel\r\n\r\nlocal_dir = 'mpnet-model/'\r\n\r\nmodel = AutoModel.from_pretrained('microsoft/mpnet-base')\r\ntokenizer = AutoTokenizer.from_pretrained('microsoft/mpnet-base')\r\nmodel.save_pretrained(local_dir)\r\ntokenizer.save_pretrained(local_dir)\r\n\r\n#Load tokenizer and model from dir\r\nmodel = AutoModel.from_pretrained(local_dir)\r\n\r\n# The following command will throw an exception\r\ntokenizer = AutoTokenizer.from_pretrained(local_dir, use_fast=False)\r\n```\r\n\r\nbut that's more of a hack than really solving the underlying problem as it just loads the slow tokenizer. Will check what's going on!\r\n", "The PR linked to the issue adds the required converter so that the above code:\r\n\r\n```python\r\nfrom transformers import AutoTokenizer, AutoModel\r\n\r\nlocal_dir = 'mpnet-model/'\r\n\r\nmodel = AutoModel.from_pretrained('microsoft/mpnet-base')\r\ntokenizer = AutoTokenizer.from_pretrained('microsoft/mpnet-base')\r\nmodel.save_pretrained(local_dir)\r\ntokenizer.save_pretrained(local_dir)\r\n\r\n#Load tokenizer and model from dir\r\nmodel = AutoModel.from_pretrained(local_dir)\r\n\r\n# The following command will throw an exception\r\ntokenizer = AutoTokenizer.from_pretrained(local_dir)\r\n```\r\n\r\nshould work after merging. It's a bit weird that one can use a FastTokenizer if the model id is `microsoft/mpnet-base` but not if the model is serialized and loaded again...we should maybe think of a better way to prevent such issues in the future. Maybe just not allow one to add a \"FastTokenizer\" class without adding the corresponding converter? @sgugger @LysandreJik ", "Great, thanks for the quick fix." ]

[]

[]

[ "I haven't reviewed in detail yet, but just looking at the API with the number of things to change for ALBERT (and in terms of line of code) is a hard pass for me. Overriding the resize method as was done before was way easier, this adds too much complexity.", "I understand that it is a big update. Nevertheless, the way it was done before didn't worked and was quite buggy (the tests basically was testing almost nothing) and to make the resizing properly working, these changes are necessary.", "In all cases I'm open to any suggestion that will reduce the number of changes :)", "@sgugger @LysandreJik I tried a new approach for the resizing that reduce a lot the changes in each model implementation, it is even much shorter than what we currently have in master. I have done my test only on ALBERT for now, can you recheck that file and let me know what you think about it.", "Ok I will clarify this a bit more:\r\n\r\n1. The fist most important issue in the current implementation is that the resizing is not graph compilation+ execution compliant because of the usage of `numpy` and `tensor.numpy()` calls and then not usable in such cases.\r\n2. The naming was depending of where the build was coming from, that's why we needed the `get_prefix_bias_name` for the bias and the manual build of the embeddings names. Which was a temporary fix, because it is very error prone, because the naming depends of several other things that are not taken into account into this manual build.\r\n3. Resizing was not working for some models, such as BART which doesn't work properly and raises an error. (Proof that the tests was not testing everything)\r\n4. The current resizing has two issues when resizing when we instantiate a model from scratch: either it raises an attribute error because the model is not fully built (weights not instantiated) and get a wrong naming, and then if we save the model with this wrong naming we cannot save/load it properly because the naming doesn't correspond to the current architecture.\r\n5. All the weights names across the models don't share the same names sometimes `embeddings.word_embeddings`, sometimes `shared.weight`, sometimes `lm_head.bias`, sometimes `lm_loss.bias`, sometimes `mlm.bias`, sometimes `mlm.predictions.bias` and many other ways...\r\n\r\nAs stated in #8657 this was just a temporary fix to go to the quickest way to wait for the real rework.\r\n\r\nThis PR aims to solve all these issues and bring something more generic and less error prone.", "I personally never understood that #8657 was a quick fix that was needing another PR afterwards. We cannot operate by adding new methods in one release then breaking them or deleting them in the next so the works that was done in #8657 needs to be built upon not destroyed (and please, say in bold next time you are just making a quick fix as I would never have approved #8657 to be merged had I known...)\r\n\r\nSo before we review this, the following need to be addressed:\r\n- `get_input_embeddings` needs to keep the same return type\r\n- `get_output_embeddings` needs to keep the same return type\r\n- `get_output_layer_with_bias` can't disappear\r\n- `get_prefix_bias_name` can't disappear\r\n\r\nThis is annoying but this is why we usually don't merge a half-baked fix introducing new APIs, we can't break that after.", "We can keep this for the next major release.\r\n\r\nWhat you ask is doable but will make the codebase more complicated. I will rework this.", "I have just done the following restore:\r\n\r\n- `get_input_embeddings` still returns a layer\r\n- `get_output_embeddings` still returns a layer\r\n- `get_output_layer_with_bias` is back\r\n- `get_prefix_bias_name` is back\r\n\r\nThe old and new approach was much more compliant than I thought so it was easier to restore what @sgugger asked, and now there should be zero breaking change. Really sorry for the misunderstanding, I will clearer next time.", "@sgugger I should have addressed all your comments :)\r\n\r\n> which makes me think there is a better way to code the default in modeling_tf_utils.\r\n\r\nShare your thoughts 😉", "Thanks @LysandreJik! For the tying test look in the `_get_resized_lm_head_decoder()` method. Unless you mean adding a test in `test_modeling_tf_common` ?", "I mean I'm not seeing a test that checks `get_input_embeddings() == get_output_embeddings()` when weights are tied, but I may be missing something here. \r\n\r\nI know these two generally point to the same tensors, but no always, do they?", "> I mean I'm not seeing a test that checks get_input_embeddings() == get_output_embeddings() when weights are tied, but I may be missing something here.\r\n\r\nYes, there is a test for this, line 884 in `modeling_tf_utils`.\r\n\r\n> I know these two generally point to the same tensors, but no always, do they?\r\n\r\nYes they always point to the same tensor when they equals, 100% sure.", "I should have addressed all the comments.", "Ah yeah, we probably need a rebase here since TF-serving just got merged :-/", "Arf good point!", "@patrickvonplaten the test `TFLEDModelTest::test_pt_tf_model_equivalence` seems very flaky, it looks like that it randomly pass/fail. ", "Good to merge for me now :)", "> @patrickvonplaten the test `TFLEDModelTest::test_pt_tf_model_equivalence` seems very flaky, it looks like that it randomly pass/fail.\r\n\r\nJust fixed it: https://github.com/huggingface/transformers/pull/9459", "@sgugger any objection to merge this PR?" ]

[]

[]

[ "Could you try putting `--block_size=512` in your command to see if it changes something?", "@LysandreJik Thank you it solved the issue! " ]

[]

[]

[ "I encountered a similar issue, in a context where `transformers` was not imported. I've [reported](https://github.com/pytorch/pytorch/issues/50858) the issue to the PyTorch project.", "I too have similar problem when running unittests for a Python package on travis, using Python 3.9. Specifically, all unit tests run fine, but the final outcome is segfault. See: https://travis-ci.org/github/LoryPack/abcpy/builds/755321073\r\n\r\n", "This is probably fixed by https://github.com/pytorch/pytorch/pull/50998" ]

[]

[]

[ "Hello! Yes, feel free to. However, you seem to have based yourself off of a `tf` branch that is very old?", "@LysandreJik Yeah sorry. Should I raise a new issue so that it can be changed accordingly?", "If you want to contribute this model, I invite you to base yourself off of the `master` branch and create a new branch from there. \r\n\r\nOnce you your branch, you can leverage the [template scripts](https://github.com/huggingface/transformers/tree/master/templates/adding_a_new_model) which should help you by creating a new model and adding it everywhere it should be added; you'll only have to update the created files.", "Reading the [contributing guide](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) will also be very helpful.", "great effort and contribution @ravi03071991 ", "@LysandreJik Sure. Will follow the guidelines. Thank you." ]

[]

[]

[ "One last detail: could you run `make style` on your branch? There seems to be some bad formatting.", "> One last detail: could you run `make style` on your branch? There seems to be some bad formatting.\r\n\r\nI did, but let me recheck it" ]

[]

[]

[ "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 realized that I didn't follow [instructions](https://github.com/pytorch/xla/blob/master/README.md) properly.", "@redrussianarmy What`s the problem? I have the same issue but can not find a solution." ]

[]

[]

[]

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

[]

[ "hi @SagarPalyal not sure what you mean here. by ckpt file do you mean a saved checkpoint ?", "Yes you are right. This is saved checkpoint file only.", "One thing to note: `pytorch_model.bin` is the name of the weights file, and every torch HF model will have that file.\r\n\r\nTo load the checkpoint simply pass the path for the checkpoint instead of `model_path`", "Hey, @SagarPalyal did you solve the problem? I met the same problem. 'from_pretrained' function needs pytorch model file, not tensorflow model file. I don't know how to convert a costumed pegasus tensorflow model to pytorch model. Anyone knows??? ", "I am not able to figure out how to convert .ckpt file to pytorch model file but in your case if you have tensorflow model file then you can use parameter from_tf=True", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[]

[]

[]

[ "`python utils/check_copies.py --fix_and_overwrite` should be run to make the `check_code_quality` test pass", "> `python utils/check_copies.py --fix_and_overwrite` should be run to make the `check_code_quality` test pass\r\n\r\nYeah, did that, for some reason, it's not working for `RobertaEmbeddings`, the code is the same as that of `BertEmbeddings`", "> > `python utils/check_copies.py --fix_and_overwrite` should be run to make the `check_code_quality` test pass\r\n> \r\n> Yeah, did that, for some reason, it's not working for `RobertaEmbeddings`, the code is the same as that of `BertEmbeddings`\r\n\r\nIf Roberta has to be different feel free to remove the copy statement", "I just merged a PR that made `cache` related tests a bit more aggressive: https://github.com/huggingface/transformers/pull/9256. It would be awesome if you could run your PR on the `EncoderDecoderModel` slow tests to make sure the cache doesn't change the results.", "Merging !" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

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

[]

[ "Could you run the other examples script to check? On my end they result in a roughly x2 speedup but I'm not on pytorch-nightly.", "Any recommendations and specific command lines that you use? \r\n\r\nI have been using the finetune and other scripts in seq2seq as a go-to scripts for testing, so I'm not quite experienced with the others. \r\n\r\nThank you!", "So on a finetune_trainer setup `--fp16` is slower than w/o it - on either amp or apex, so that elimination the caching concern. (w/ pytorch nightly)\r\n\r\nCaveat: The following tests aren't ideal since I have 1 fast and 1 slow card, but they should be consistent since the overall speed is always at the slowest card (with the exception of single gpu tests), so it's like having 2 slow cards.\r\n\r\n### DDP\r\n\r\n```\r\n# baseline w/ --fp16\r\nexport BS=4; rm -r output_dir; CUDA_VISIBLE_DEVICES=0,1 PYTHONPATH=../../src USE_TF=0 python -m torch.distributed.launch --nproc_per_node=2 --master_port=9910 ./finetune_trainer.py --model_name_or_path sshleifer/distill-mbart-en-ro-12-4 --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_train --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_train_batch_size $BS --sortish_sampler --src_lang en_XX --task translation --tgt_lang ro_RO --val_max_target_length 128 --warmup_steps 500 --n_train 500\r\n\r\n2020-12-17 16:04:18 | INFO | __main__ | train_runtime = 27.9693\r\n\r\n# --fp16 --fp16_backend apex\r\n\r\n2020-12-17 16:01:14 | INFO | __main__ | train_runtime = 30.0469\r\n\r\n# --fp16 --fp16_backend amp\r\n\r\n2020-12-17 16:06:41 | INFO | __main__ | train_runtime = 29.5368\r\n\r\n```\r\n\r\n### DP\r\n\r\nDP setup is about the same correlation - but runs twice as slow! (that looks wrong too!)\r\n\r\n```\r\n\r\n# baseline (no --fp16)\r\nexport BS=4; rm -r output_dir; CUDA_VISIBLE_DEVICES=0,1 PYTHONPATH=../../src USE_TF=0 python ./finetune_trainer.py --model_name_or_path sshleifer/distill-mbart-en-ro-12-4 --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_train --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_train_batch_size $BS --sortish_sampler --src_lang en_XX --task translation --tgt_lang ro_RO --val_max_target_length 128 --warmup_steps 500 --n_train 500\r\n\r\n2020-12-17 16:16:09 | INFO | __main__ | train_runtime = 56.7522\r\n\r\n# --fp16 --fp16_backend apex\r\n\r\n2020-12-17 16:14:26 | INFO | __main__ | train_runtime = 59.4309\r\n\r\n# --fp16 --fp16_backend amp\r\n\r\n2020-12-17 16:12:18 | INFO | __main__ | train_runtime = 58.4406\r\n```\r\n\r\n\r\n\r\n### Single GPU (gtx-1070 slowest)\r\n\r\nno improvement either with fp16\r\n\r\n```\r\n# baseline (no --fp16)\r\nexport BS=4; rm -r output_dir; CUDA_VISIBLE_DEVICES=1 PYTHONPATH=../../src USE_TF=0 python ./finetune_trainer.py --model_name_or_path sshleifer/distill-mbart-en-ro-12-4 --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_train --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_train_batch_size $BS --sortish_sampler --src_lang en_XX --task translation --tgt_lang ro_RO --val_max_target_length 128 --warmup_steps 500 --n_train 500\r\n\r\n2020-12-17 16:26:10 | INFO | __main__ | train_runtime = 24.6995\r\n\r\n# --fp16 --fp16_backend apex\r\n\r\n2020-12-17 16:27:26 | INFO | __main__ | train_runtime = 28.6601\r\n\r\n# --fp16 --fp16_backend amp\r\n\r\n2020-12-17 16:28:37 | INFO | __main__ | train_runtime = 27.9687\r\n\r\n```\r\n\r\n\r\n### Single GPU (rtx-3090 fastest)\r\n\r\nno improvement either with fp16\r\n\r\n```\r\n# baseline (no --fp16)\r\nexport BS=4; rm -r output_dir; CUDA_VISIBLE_DEVICES=0 PYTHONPATH=../../src USE_TF=0 python ./finetune_trainer.py --model_name_or_path sshleifer/distill-mbart-en-ro-12-4 --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_train --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_train_batch_size $BS --sortish_sampler --src_lang en_XX --task translation --tgt_lang ro_RO --val_max_target_length 128 --warmup_steps 500 --n_train 500\r\n\r\n2020-12-17 16:22:33 | INFO | __main__ | train_runtime = 11.2534\r\n\r\n# --fp16 --fp16_backend apex\r\n\r\n2020-12-17 16:19:41 | INFO | __main__ | train_runtime = 14.4828\r\n\r\n# --fp16 --fp16_backend amp\r\n\r\n2020-12-17 16:21:15 | INFO | __main__ | train_runtime = 11.7265\r\n```\r\n\r\nrtx-3040 is so much faster than gtx-1070.\r\n\r\nAlso wrt single gpu test: slow card 100%, fast one 75% - so the trainer is not fast enough feeding data in for the latter.\r\n", "> Any recommendations and specific command lines that you use?\r\n\r\nNormally the first command indicated in the README of each example folder should work well :-)", "> Normally the first command indicated in the README of each example folder should work well :-)\r\n\r\nThe one I tried doesn't report speed. `text-classification/run_glue.py`. I'm not sure what method you were using to detect speed improvements.", "Thanks for adding the speed metrics into the core trainer, with https://github.com/huggingface/transformers/pull/9198 I run benchmarks for run_glue in text-classification, the results are terrible speed-wise just like with the finetune_trainer - terrible as in all those things that are supposed to speed things up, slow things down instead:\r\n\r\n```\r\n# baseline\r\nrm -r output_dir; PYTHONPATH=../../src USE_TF=0 python -m torch.distributed.launch --nproc_per_node=2 run_glue.py --model_name_or_path bert-base-cased --task_name MRPC --do_train --do_eval --max_seq_length 128 --per_device_train_batch_size 32 --learning_rate 2e-5 --num_train_epochs 3.0 --output_dir output_dir\r\n\r\n12/18/2020 13:29:01 - INFO - __main__ - train_runtime = 97.4012\r\n\r\n# --fp16\r\n\r\n12/18/2020 13:46:09 - INFO - __main__ - train_runtime = 109.1225\r\n\r\n# --sharded_ddp\r\n\r\n12/18/2020 13:53:18 - INFO - __main__ - train_runtime = 103.1887\r\n\r\n# --fp16 --sharded_ddp\r\n\r\n12/18/2020 13:50:57 - INFO - __main__ - train_runtime = 113.9132\r\n```\r\n\r\nthis is all w/ pt-nightly, since I have to use it to run rtx-3090 card\r\n\r\n", "In that case, it looks linked to your setup/env somehow. Maybe support for the RTX-3090 is not fully fledged? On my setup and env FP16 speeds up by a factor of x2 roughly for this script. Will test the sharded DDP variants on Monday.", "if you could do the above 4 runs (https://github.com/huggingface/transformers/issues/9179#issuecomment-748338332) for comparison that would be great! I need to know if something is off with my setup since I'm trying to eval deepspeed. \r\n\r\ncuda-11.2 is out so hoping to get a normal support for rtx-3090 from pytorch really soon now.", "Hi \r\nthe command @stas00 mentioend above does not work for me, I am using last version of huggingface, could you tell me which version you used to test? please see my bug here https://github.com/huggingface/transformers/issues/9215 ", "@rabeehk: https://github.com/huggingface/transformers/issues/9156#issuecomment-748501582", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "Probably should change CI to do:\r\n```\r\npip install fairscale --no-build-isolation\r\n```\r\nwhich is much faster then the new pip system, as it doesn't need to fetch dependent packages\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." ]

[]

[]

[ "I made a request for a quicker/simpler binary: https://github.com/facebookresearch/fairscale/issues/264\r\n\r\nAnd added an issue to track this so we won't forget: https://github.com/huggingface/transformers/issues/9178\r\n" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Cleaner version here: https://github.com/huggingface/transformers/pull/9251" ]

[]

[]

[ "Hey @dan-i, \r\n\r\n1) You should mask the labels of the first token it's not enough to set the attention_mask to 0 -> `labels=torch.tensor([-100, ...])`\r\n2) You have to change the `position_ids` to ensure the expected behavior in GPT2, *i.e.* make sure that in the second case the position_ids are [0, 0, 1, 2] instead of [0, 1, 2, 3]", "thank patrick. seems to work if you just do labels to -100 where the pads are without mask and position. at least it returns same result as individual sentence scores. for anybody else that needs gpt2 batch sentence perplexity scores.\r\n\r\n```\r\ndef calc_loss(sentences,inputs,logits):\r\n # calc loss from logits returned from model\r\n lines_len = torch.sum(inputs['attention_mask'], dim=1) # quick way to get sent len. mask isn't used for anything else\r\n for line_ind in range(len(sentences)): \r\n log_prob = 0.0\r\n for token_ind in range(lines_len[line_ind] - 1): \r\n token_prob = F.softmax(logits[line_ind, token_ind], dim=0)\r\n token_id = inputs['input_ids'][line_ind, token_ind + 1]\r\n log_prob += torch.log(token_prob[token_id])\r\n sentloss=abs(log_prob)\r\n toks=lines_len[line_ind]\r\n loss=sentloss/(toks-1)\r\n print(sentences[line_ind],' loss=',loss.item(),'toks=',toks.item(),'sentloss=',sentloss.item())\r\n\r\nsentences=[\"Hello, my\",\"Hello, my dog\",\"Hello, my dog is a dog\"]\r\ntokenizer.pad_token = tokenizer.eos_token # set pad to eos\r\ninputs = tokenizer(sentences, return_tensors=\"pt\", padding=True) # right side padding as usual\r\n# mask of label_id's\r\nlabel_ids=inputs['input_ids'].clone()\r\nlabel_ids[label_ids==tokenizer.encode(tokenizer.pad_token)[0]] = -100\r\n# one shot score the batch\r\nwith torch.no_grad():\r\n model.eval()\r\n loss, logits = model(input_ids=inputs['input_ids'], labels=label_ids)[:2]\r\n# calc loss per sentence\r\ncalc_loss(sentences,inputs,logits)\r\n```\r\noutputs:\r\nHello, my loss= 3.069786548614502 toks= 3 sentloss= 6.139573097229004\r\nHello, my dog loss= 4.247798442840576 toks= 4 sentloss= 12.74339485168457\r\nHello, my dog is a dog loss= 3.671175003051758 toks= 7 sentloss= 22.027050018310547\r\n" ]

[ { "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" } ]

[]

[ "Hi @chris-tng -- thanks for trying out Flax in HF Transformers!\r\n\r\nA quick comment on `nn.compact` and `setup` (I work on Flax) -- indeed if you want access to submodules such as for transfer learning then using `setup` is the way to go. I skimmed the PR and see that you use `setup` in some places and `nn.compact` in others? I'm curious whether you found `nn.compact` more useful in particular settings.\r\n\r\nIndeed `setup` is more similar to the PyTorch style (though you still get shape inference if you use `nn.compact` in modules that don't have submodules). `nn.compact` is nice if you want to use loops or conditionals to define submodules based on hyperparameters, and some people also prefer how it \"co-locates\" their submodule definitions and usage. But ultimate it's somewhat a matter of preference.\r\n\r\n(Please do let us know whatever other thoughts or questions on Flax on our discussion board: https://github.com/google/flax/discussions)\r\n\r\nHappy holidays and new year!", "> shape\r\n\r\nHey @avital,\r\n\r\nThanks a lot for your input here! That's very useful. Most of the main contributors to Transformers are on holiday at the moment and this is a rather big design decision to make going forward with Flax, so I think we'll have to wait here until early January until everybody is back (@sgugger, @LysandreJik, @mfuntowicz) \r\n\r\nHappy holiday to you as well :-) ", "Hi @avital ,\r\n\r\nApology for my delayed response. I appreciate your great work on Flax. Regarding the use of `setup()` and `nn.compact`, personally I find `setup` works better for testing submodules. This is useful for converting and debugging the module (and submodules). For instance, I can create a model/module with many submodules:\r\n\r\n```python\r\nclass Dummy(nn.Module):\r\n\r\n def setup(self):\r\n self.submodule1 = nn.Dense(10)\r\n self.submodule2 = MyLayerNorm()\r\n\r\n def __call__(self):\r\n # do something here\r\n```\r\n\r\nAfter loading model weights from a dict, I can access/debug submodule by simply accessing the attribute: `dummy.submodule1`, `dummy.submodule2`. From this, I can debug forward pass, check model weights of invididual submodule.\r\n\r\nShameless plug, I wrote a blog post about porting huggingface pytorch model to flax, [here](https://chris-tng.github.io/blog/transformer/nlp/flax/pytorch/jax/2020/12/16/tips-flax.html). I'm a new Flax user so please correct me if I'm missing anything.\r\n\r\nHappy holiday and happy new year to everyone 🎄 🍾 ", "Hey @chris-tng, \r\n\r\nsorry to had you wait for this long. I'll solve the merge conflicts in your PR and then use your PR to change the `@nn.compact` to `setup` in all other flax models as well so that we have a common standard now. Since most of our users are used to the \"PyTorch\" style and I only see advantages for our library philosophy:\r\n\r\n- We base most design decisions on the PyTorch style\r\n- We prefer slightly less compact readable code over the slightly more \"magic\" functionalities that might reduce code\r\n- To me the are no real downsides to using `setup`", "Intermediate state is saved here: #9484 will push to this PR on Monday the latest", "Hey @chris-tng,\r\n\r\nI noticed that we will probably have to wait a bit to get this merged: https://github.com/google/flax/pull/683 to be able to continue the PR. Will keep you up-to-date :-)", "Hi folks, sorry for the delay with the new-year shuffle and school shutdown.\r\n\r\ngoogle/flax#683 required a bit more conversation and updating some other codebases but now it's merged! If you have a moment, please take a look and see if it helps unblock progress. We'll release Flax 0.4.0 soon, but installing from GitHub now is the way to go.", "Hey, sorry for barging in\r\nI was needing a small BERT-like model in Jax and so I've recently updated this in a local branch to work with the Flax refactoring that makes checkpoints directly compatible with PyTorch (plus fixing some other issues that had gone through the cracks)\r\nShould I push directly to this branch or make a new PR from my fork? Also should I wait #11364 and update my code accordingly?", "> Hey, sorry for barging in\r\n> I was needing a small BERT-like model in Jax and so I've recently updated this in a local branch to work with the Flax refactoring that makes checkpoints directly compatible with PyTorch (plus fixing some other issues that had gone through the cracks)\r\n> Should I push directly to this branch or make a new PR from my fork? Also should I wait #11364 and update my code accordingly?\r\n\r\nHey @CoderPat,\r\n\r\nIt would be great if you could wait until #11364 is merged (should be done in the next 2 days). The PR fixes a couple of bugs :-)", "No problem @patrickvonplaten! Also regarding git logistics, is it better to ask @chris-tng for permission to push directly to his branch?", "> No problem @patrickvonplaten! Also regarding git logistics, is it better to ask @chris-tng for permission to push directly to his branch?\r\n\r\nI think it's alright to copy past the code that is still useful and open a new branch, if you'd like to add Electra :-). On the branch we should then give credit to @chris-tng , but since the PR is quite old now I think he would be fine if we close this one and open a new one (Please let me know if this is not the case @chris-tng :-)) . #11364 should be the last refactor before the \"fundamental\" Flax design is finished.", "Just to confirm @patrickvonplaten , the flax refactor is merged and the structure should be stable enough that I can work on implementing Electra right?", "Exactly @CoderPat - very much looking forward to your PR :-)" ]

[]

[]

[ "**Update:** I slightly changed my script based on the [text_classification example notebook](https://github.com/huggingface/notebooks/blob/master/examples/text_classification.ipynb) and now it works. Don't know what made the difference, but in case someone else has a similar problem, I recommend following the steps in this script. " ]

[]

[]

[ "An alternative is to remove all content, but the previous state where one variable as some keys but the others not leads to failures.", "Yes, we really need to remove all of this @LysandreJik @thomwolf ", "We do, but right now we still use them in the tests", "It was a bit confusing to me which to add in the constants and which not.. 😅 ", "@sgugger one more small fix you can add is in tapas.rst under \"Usage: inference\", there should be one more space:\r\n\r\n\r\n\r\n(Sorry, if I see more things I make a new PR myself :p)", "Pushing your fix on master directly Niels" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "I think @LysandreJik will know more what to do for the logging. Also, we usually ask questions like this on the [forum](https://discuss.huggingface.co/) and keep issues for bugs/feature requests.", "If you want to output everything to a file, you can redirect the standard output/standard error to a file. We use the official `logging` package for logs, and they have a full chapter en redirecting to a file: https://docs.python.org/3/howto/logging.html#logging-to-a-file", "It is also possible to redirect the output from the script (without modifying anything). I use the following code to both have the output on my screen (console) and a copy in my log files (with stderr having the python logging logs):\r\n\r\n```bash\r\npython run_your_script.py \\\r\n > >(tee -a stdout.log) \\\r\n 2> >(tee -a stderr.log >&2)\r\n```\r\n\r\n(Found this gem on StackOverflow somewhere.)", "@Querela and @LysandreJik, thanks a lot for the help. closing the issues." ]

[]

[]

[ "Hi, note that 'suicide' and ' suicide' are indeed different tokens with respect to SentencePieces, which is used by Roberta.\r\n\r\nSentencePieces keeps white space intact. So we expect different tokens from 'suicide' and ' suicide'. \r\n\r\nI assume in the old transformers 2.x, strip was applied in the tokenizer before it was passed to SentencePiece, which is incorrect. ", "Thank you for letting me know. Regardless, the above creates some pretty big inconsistency issues and for example I end up with a model trained with 2.x, when used to predict in an env with transformers >= 3.x to have major discrepancy in predicting accuracy.\r\nThis should be at least flagged somewhere and there should be some \"backward compatibility\" option here.", "You could call `text.strip()` before passing it to the tokenizer. ", "Sorry, just to clarify. This doesn't have to do on how I can personally solve my issue with a hack. Of course I can do that or simply use the same transformers version when training/predicting. \r\nThis has to do on how a production system might get affected by this, in my view, major discrepancy. Tokenization behaviour shouldn't change \"silently\" across versions.\r\nThink of the following case: a product trained to do sentence similarity. The model is trained to understand similarity of sentences, thus tokens. Now if you happily change tokenization across versions, you end up with a model that won't work properly when you upgrade transformers library. That should not happen, or at least get flagged somewhere.", "The old tokenization was buggy and incorrect (based on the info here). I don't see a point to add an option to enable back buggy / incorrect behavior of a component. That would be quite bad to have flags that re-enable all kind of buggy behavior you had in old versions. \r\n\r\nIf you need stable results in production then there is no other way than to stay at one version or otherwise have sufficient tests to ensure the functionality of your system when you upgrade your framework. \r\n\r\nIt is the same with every framework, that between major versions that there can be significant changes that can impact your software. \r\n\r\n(Just my personal opinion, I don't work for huggingface. I don't know if the bug in the old tokenizer was known and fixed (then it is likely mentioned in the release notes) or if the fix was just a by product by some other commit.)\r\n\r\n\r\nIt appears to be fixed here\r\nhttps://github.com/huggingface/transformers/pull/5287\r\n\r\nAnd was mentioned in the subsequent release ", "Thank you this slipped my attention, it was nevertheless a pretty significant issue. Apparently it's connected to that: https://github.com/huggingface/transformers/issues/5256\r\nI guess this should be expected with a framewrok developing so quickly!", "@ypapanik I am a bit surprised that it makes such a big difference for your use case. You wouldn't expect this as the only difference is a potential white space at the beginning of the text.\r\n\r\nMy roberta models changed performance in downstream tasks only slightly between the old tokenizer and the newer tokenizer (performance improved like 0.1 percentage points actually)." ]

[]

[]

[ "@patrickvonplaten @LysandreJik @julien-c @zhiheng-huang", "Hey @hadaev8,\r\n\r\nThanks for your PR. It would be awesome if you could provide a code snippet to this PR that shows a case where the current implementation of BERT's relative positional embeddings would fail. Hope it's fine to tag the original author here: @zhiheng-huang for some discussion. ", "@patrickvonplaten \r\nColab notebook\r\nhttps://colab.research.google.com/drive/1bAwwNMbh27JW0H6uWG9rZgyjz_b0eB1M?usp=sharing\r\n", "@patrickvonplaten \r\nAny update?", "Hey @hadaev8,\r\n\r\nI'm having a hard time deciding here without the opinion of the official author @zhiheng-huang . @zhiheng-huang it would be great if you could leave a comment here :-)", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "@patrickvonplaten \r\nWell, in current state model with rel pos just would not work as expected eg process longer input.", "@hadaev8, \r\n\r\nAlright, let's try to fix it together! It would be awesome if you could post a reproducible code snippet here so that we can see when an error arises", "@patrickvonplaten \r\nHere it doesnt work\r\nhttps://colab.research.google.com/drive/1bAwwNMbh27JW0H6uWG9rZgyjz_b0eB1M?usp=sharing\r\nHere it work\r\nhttps://colab.research.google.com/drive/1OIHjR2kVndDzwro7n_SgkpJG4YwjgGQY?usp=sharing", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]