Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "Hey @ontocord,\r\n\r\nThanks a lot for opening this pull request :-) \r\n\r\nWe are very much open to adding FastSpeech2 to `transformers`! One thing that is quite important to us is that we stick as much as possible to the original implementation of the model. \r\n\r\nIMO, the easiest way to approach this would be to translate [this](TF code) to PyTorch: https://github.com/TensorSpeech/TensorFlowTTS/blob/master/tensorflow_tts/models/fastspeech2.py and add it to our lib.\r\n\r\nLet me know if it would also be ok/interesting for you to stay closer to the official code -> I'm very happy to help you get this PR merged then :-)", "Happy for you to close this PR in favor of an official implementation. Just as an aside, the G2P (GRU->GRU) code is actually based on the original impementation from the Fastspeech2 paper. But it uses https://github.com/Kyubyong/g2p which is slower than pytorh and based on Numpy. I re-wrote the G2P in pytorch based on the G2P author's notes, and retrained it so it's faster.\r\n\r\nFrom the paper: \r\n\r\n\"To alleviate the mispronunciation problem, we convert the\r\ntext sequence into the phoneme sequence (Arik et al., 2017; Wang et al., 2017; Shen et al., 2018;\r\nSun et al., 2019) with an open-source grapheme-to-phoneme tool5\r\n...\r\n5https://github.com/Kyubyong/g2p\"\r\n\r\nI think this module is really one of the things that keeps the Fastspeech2 model (and tacotron 2 and similar models) from generalizing to more languages. In theory you could just train on character level, but it's harder. DM if you want to discuss work arounds...", "Hey @ontocord @patrickvonplaten, I was wondering if there has been a followup to this PR. I'd love to see transformer TTS models like FastSpeech2 in this library and would be more than happy to help contribute if possible!", "I also think we should eventually add models like FastSpeech2 to the library. Gently ping to @anton-l here who was interested in this addition as well.", "@patrickvonplaten @anton-l Do we only add models with official weights from the paper authors? AFAIK FastSpeech2 has plenty of unofficial implementations with weights, but there is no official repository ([PwC](https://paperswithcode.com/paper/fastspeech-2-fast-and-high-quality-end-to-end)). I think we should reach out to the author (Yi Ren is on GitHub), and if that doesn't work out, consider which implementation/weights we want to port. What do you think?\r\n\r\nAlso if you'd prefer, I'll open a new issue dedicated to this discussion instead of hijacking this PR.", "I think we should definitely reach out to the original authors! Feel free to contact them :-)", "Just emailed the first author and cc'd both you and Anton! I'll keep you posted. ", "I would be interested in working on something more generic than fastspeech2 which needs a g2p module. It’s not truly end to end. \n\n> On Jan 3, 2022, at 7:55 AM, Jake Tae ***@***.***> wrote:\n> \n> \n> Just emailed the first author and cc'd both you and Anton! I'll keep you posted.\n> \n> —\n> Reply to this email directly, view it on GitHub, or unsubscribe.\n> Triage notifications on the go with GitHub Mobile for iOS or Android. \n> You are receiving this because you were mentioned.\n" ]

[]

[]

[ "@chatzich your question seems similar to this - #2323. ", "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." ]

[]

[]

[]

[]

[]

[ "Have you looked at the [run_ner](https://github.com/huggingface/transformers/blob/master/examples/token-classification/run_ner.py) example and its corresponding [notebook](https://github.com/huggingface/notebooks/blob/master/examples/token_classification.ipynb)?", "Thanks for getting back to me so quickly @sgugger. \r\nI am using what looks like the same exact version of run_ner with a different notebook. However, I don't see anything in the notebook you provided or run_ner that adds new labels. For example, if you look at the cell 8 of the notebook you linked, you see that it is only using the labels loaded with the model. What if I wanted to find addresses or some other entity type?\r\n\r\nThanks for your help!\r\nGregg\r\n\r\n", "I am confused, the labels are loaded from the dataset, not the model. If you have another dataset with other labels, the rest of the notebook will work the same way.", "Sorry, \r\nWhat I want to do is load the Bert model for NER trained from Conll2003 and use transfer learning to add addition training with new, additional data tagged with additional labels.\r\nIn the end, I want to take advantage of the existing training and add my own; teaching it to recognize more entity types.\r\n\r\nI have seen that some people seem to have done this but I haven't found the complete list of steps. ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[]

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

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "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." ]

[]

[]

[ "The text you are providing is probably too long for the model. Most Transformer models accept a sequence length of 512 tokens.\r\n\r\nWhich model did you use?", "I just used the general q-a pipeline:\r\nfrom google.colab import files\r\nuploaded = files.upload() \r\nfilename = \"New_Spark_Questions.txt\"\r\nnew_file = uploaded[filename].decode(\"utf-8\")\r\n!pip3 install sentencepiece\r\n!pip3 install git+https://github.com/huggingface/transformers\r\nquestion = \"Why should I take part in SPARK?\" \r\nfrom transformers import pipeline\r\nqa = pipeline(\"question-answering\")\r\nanswer = qa(question=question, context=new_file)\r\nprint(f\"Question: {question}\")\r\nprint(f\"Answer: '{answer['answer']}' with score {answer['score']}\")\r\nHow does the q-a pipeline decide the score?\r\nAlso how do we use a model like Bert, XLNet, etc. on a q-a pipeline\r\nDoes the input in the q-a pipeline have to be a dictionary?", "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." ]

[]

[]

[ "Have you installed the [sentencepiece](https://github.com/google/sentencepiece) library?", "> Have you installed the [sentencepiece](https://github.com/google/sentencepiece) library?\r\n\r\nYes.\r\n\r\n```\r\n!pip install transformers\r\n!pip install sentencepiece\r\n```\r\n\r\n```\r\nRequirement already satisfied: transformers in /usr/local/lib/python3.7/dist-packages (4.5.0)\r\nRequirement already satisfied: numpy>=1.17 in /usr/local/lib/python3.7/dist-packages (from transformers) (1.19.5)\r\nRequirement already satisfied: tqdm>=4.27 in /usr/local/lib/python3.7/dist-packages (from transformers) (4.41.1)\r\nRequirement already satisfied: requests in /usr/local/lib/python3.7/dist-packages (from transformers) (2.23.0)\r\nRequirement already satisfied: importlib-metadata; python_version < \"3.8\" in /usr/local/lib/python3.7/dist-packages (from transformers) (3.8.1)\r\nRequirement already satisfied: filelock in /usr/local/lib/python3.7/dist-packages (from transformers) (3.0.12)\r\nRequirement already satisfied: sacremoses in /usr/local/lib/python3.7/dist-packages (from transformers) (0.0.44)\r\nRequirement already satisfied: tokenizers<0.11,>=0.10.1 in /usr/local/lib/python3.7/dist-packages (from transformers) (0.10.2)\r\nRequirement already satisfied: regex!=2019.12.17 in /usr/local/lib/python3.7/dist-packages (from transformers) (2019.12.20)\r\nRequirement already satisfied: packaging in /usr/local/lib/python3.7/dist-packages (from transformers) (20.9)\r\nRequirement already satisfied: chardet<4,>=3.0.2 in /usr/local/lib/python3.7/dist-packages (from requests->transformers) (3.0.4)\r\nRequirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.7/dist-packages (from requests->transformers) (2020.12.5)\r\nRequirement already satisfied: urllib3!=1.25.0,!=1.25.1,<1.26,>=1.21.1 in /usr/local/lib/python3.7/dist-packages (from requests->transformers) (1.24.3)\r\nRequirement already satisfied: idna<3,>=2.5 in /usr/local/lib/python3.7/dist-packages (from requests->transformers) (2.10)\r\nRequirement already satisfied: zipp>=0.5 in /usr/local/lib/python3.7/dist-packages (from importlib-metadata; python_version < \"3.8\"->transformers) (3.4.1)\r\nRequirement already satisfied: typing-extensions>=3.6.4; python_version < \"3.8\" in /usr/local/lib/python3.7/dist-packages (from importlib-metadata; python_version < \"3.8\"->transformers) (3.7.4.3)\r\nRequirement already satisfied: six in /usr/local/lib/python3.7/dist-packages (from sacremoses->transformers) (1.15.0)\r\nRequirement already satisfied: joblib in /usr/local/lib/python3.7/dist-packages (from sacremoses->transformers) (1.0.1)\r\nRequirement already satisfied: click in /usr/local/lib/python3.7/dist-packages (from sacremoses->transformers) (7.1.2)\r\nRequirement already satisfied: pyparsing>=2.0.2 in /usr/local/lib/python3.7/dist-packages (from packaging->transformers) (2.4.7)\r\nRequirement already satisfied: sentencepiece in /usr/local/lib/python3.7/dist-packages (0.1.95)\r\n```", "I restart the runtime and it is fixed now. Seems like some strange compatibility issues with colab." ]

[]

[]

[]

[]

[]

[ "hi @GeetDsa \r\n\r\nThanks a lot for the PR. I understand the issue and IMO what should be done here is to make sure to pass the same `max_length` to the `BeamScorer` and `beam_search` instead of changing the method.\r\n\r\nThis is because the overall philosophy of `generate` is that whenever some argument is `None` its value should explicitly default to the value specified in `config`. This how all generation methods work.", "Thanks for the issue & PR @GeetDsa! I agree with @patil-suraj that we should not change the way `max_length` is set in `beam_search`.\r\n\r\nOverall, the problem IMO is actually that `BeamScorer` has a `max_length` attribute... => this shouldn't be the case IMO:\r\n- `BeamHypotheses` has a `max_length` attribute that is unused and can be removed\r\n- `BeamSearchScorer` has a `max_length` attribute that is only used for the function `finalize` => the better approach here would be too pass `max_length` as an argument to `finalize(...)` IMO\r\n\r\nThis solution will then ensure that only one `max_length` is being used and should also help to refactor out `max_length` cc @Narsil longterm.\r\n\r\nDo you want to give it a try @GeetDsa ? :-)", "> Thanks for the issue & PR @GeetDsa! I agree with @patil-suraj that we should not change the way `max_length` is set in `beam_search`.\r\n> \r\n> Overall, the problem IMO is actually that `BeamScorer` has a `max_length` attribute... => this shouldn't be the case IMO:\r\n> \r\n> * `BeamHypotheses` has a `max_length` attribute that is unused and can be removed\r\n> * `BeamSearchScorer` has a `max_length` attribute that is only used for the function `finalize` => the better approach here would be too pass `max_length` as an argument to `finalize(...)` IMO\r\n> \r\n> This solution will then ensure that only one `max_length` is being used and should also help to refactor out `max_length` cc @Narsil longterm.\r\n> \r\n> Do you want to give it a try @GeetDsa ? :-)\r\n\r\nI can give a try :)\r\n", "> BeamHypotheses has a max_length attribute that is unused and can be removed\r\n\r\nNice !\r\n\r\n> BeamSearchScorer has a max_length attribute that is only used for the function finalize => the better approach here would be too pass max_length as an argument to finalize(...) IMO\r\n\r\nSeems easier.\r\n@GeetDsa Do you think you could also add a test that reproduces your issue without your fix and that passes with the fix ? That will make backward compatibility easier to test (we're heading towards a direction to remove `max_length` as much as possible while maintaining backward compatbility)", "I have created a new pull request #11378 ; @Narsil, I think it will be little hard and time consuming for me to implement a test as I am not well-versed with it.", "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." ]

[]

[]

[ "Maybe of interest to @Narsil ", "@DeveloperInProgress .\r\nSorry, there is no exhaustive list of those as of yet (as a number of them are actual exceptions raised by transformers itself)\r\n\r\nWhat I can say, is that Environment and ValueError are simply displayed as is and treated as user error (usually problem in the model configuration or inputs of the model).\r\n\r\nAny other exception is raised as a server error (and looked at regularly).\r\n\r\nAny \"unknown error\" is an error for which we can't find a good message, we try to accompany it (when it's possible) with any warnings that might have been raised earlier by transformers (for instance too long sequences make certain models crash, deep cuda errors are unusable as is, the warning is better).\r\n\r\nDoes that answer your question ?", "@Narsil gotcha" ]

[]

[]

[]

[]

[]

[ "Thank you, @bhavitvyamalik. This is excellent\r\n\r\nLet's first review what we are trying to correct.\r\n\r\nLooking under `~/.cache/huggingface/transformers/` I see:\r\n```\r\n-rw------- 1 stas stas 1.1K Oct 16 13:27 00209bab0f0b1af5ef50d4d8a2f8fb0589ec747d29d975f496d377312fc50ea7.688a102406298bdd2190bac9e0c6da7c3ac2bfa26aa40e9e07904fa\r\ne563aeec3\r\n-rw-rw-r-- 1 stas stas 158 Oct 16 13:27 00209bab0f0b1af5ef50d4d8a2f8fb0589ec747d29d975f496d377312fc50ea7.688a102406298bdd2190bac9e0c6da7c3ac2bfa26aa40e9e07904fa\r\ne563aeec3.json\r\n-rwxrwxr-x 1 stas stas 0 Oct 16 13:27 00209bab0f0b1af5ef50d4d8a2f8fb0589ec747d29d975f496d377312fc50ea7.688a102406298bdd2190bac9e0c6da7c3ac2bfa26aa40\r\ne9e07904fae563aeec3.lock*\r\n-rw------- 1 stas stas 4.9M Oct 14 11:56 002911b8e4cea0a107864f5b17f20c10f613d256e92e3c1247d6d174fbf56fe5.bf6ebaf6162cfbfbad2ce1909278a9ea1fbfe9284d318bff8bccddf\r\ndaa104205\r\n-rw-rw-r-- 1 stas stas 130 Oct 14 11:56 002911b8e4cea0a107864f5b17f20c10f613d256e92e3c1247d6d174fbf56fe5.bf6ebaf6162cfbfbad2ce1909278a9ea1fbfe9284d318bff8bccddf\r\ndaa104205.json\r\n```\r\n\r\nSo some files already have the correct perms `-rw-rw-r--`, but the others don't (`-rw-------` missing group/other perms)\r\n\r\nIf I try to get a new cached file:\r\n\r\n```\r\nPYTHONPATH=\"src\" python -c \"from transformers import AutoModel; AutoModel.from_pretrained('sshleifer/student_pegasus_xsum_16_8')\"\r\n```\r\n\r\nwe can see how the tempfile uses user-only perms while downloading it:\r\n```\r\n-rw------- 1 stas stas 246M Apr 7 10:06 tmplse9bwr1\r\n```\r\n\r\nand then your fix, adjusts the perms:\r\n\r\n```\r\n-rw-rw-r-- 1 stas stas 1.7G Apr 7 10:08 6636af980d08a3205d570f287ec5867d09d09c71d8d192861bf72e639a8c42fc.c7a07b57c0fbcb714c5b77aa08bea4f26ee23043f3c28e7c1af1153\r\na4bdfeea5\r\n-rw-rw-r-- 1 stas stas 180 Apr 7 10:08 6636af980d08a3205d570f287ec5867d09d09c71d8d192861bf72e639a8c42fc.c7a07b57c0fbcb714c5b77aa08bea4f26ee23043f3c28e7c1af1153\r\na4bdfeea5.json\r\n-rwxrwxr-x 1 stas stas 0 Apr 7 10:06 6636af980d08a3205d570f287ec5867d09d09c71d8d192861bf72e639a8c42fc.c7a07b57c0fbcb714c5b77aa08bea4f26ee23043f3c28\r\ne7c1af1153a4bdfeea5.lock*\r\n```\r\n\r\nSo this is goodness.", "There is also a recipe subclassing `NamedTemporaryFile` https://stackoverflow.com/a/44130605/9201239 so it's even more atomic. But I'm not sure how that would work with resumes. I think your way is just fine for now and if we start doing more of that we will use a subclass that fixes perms internally.\r\n", "That makes sense, moving it from `cached_path` to `get_from_cache`. Let me push your suggested changes. Yeah, even I came across this subclassing `NamedTemporaryFile` when I had to fix this for Datasets but I felt adding more such tempfiles and then using subclassing would be more beneficial.", "Any plans for asking user what file permission they want for this model?", "> Any plans for asking user what file permission they want for this model?\r\n\r\nCould you elaborate why would a user need to do that?\r\n\r\nFor shared environment this is a domain of `umask` and may be \"sticky bit\".\r\n", "When we started working on this feature for Datasets someone suggested this to us:\r\n\r\n> For example, I might start a training run, downloading a dataset. Then, a couple of days later, a collaborator using the same repository might want to use the same dataset on the same shared filesystem, but won't be able to under the default permissions.\r\n> \r\n> Being able to specify directly in the top-level load_dataset() call seems important, but an equally valid option would be to just inherit from the running user's umask (this should probably be the default anyway).\r\n> \r\n> So basically, argument that takes a custom set of permissions, and by default, use the running user's umask!\r\n\r\nSay if someone doesn't want the default running user's umask then they can specify what file permissions they want for that model. Incase they opt for this, we can avoid the umask part and directly `chmod` those permissions for the newly downloaded model. I'm not sure how useful would this be from the context from Transformers library.", "Thank you for sharing the use case, that was helpful.\r\n\r\nBut this can be solved on the unix side of things. If you want a shared directory you can set it up as such. If you need to share files with other members of the group you put them into the same unix group.\r\n\r\nIMHO, in general programs shouldn't mess with permissions directly. Other than the fix you just did which compensates for the temp facility restrictions.\r\n", "@LysandreJik, could you please have a look so that we could merge this? Thank you!", "@stas00 @bhavitvyamalik I must say that I am not familiar with the umask command, but it seems = as @LysandreJik rightfully points out in my feature request https://github.com/huggingface/transformers/issues/12169#issuecomment-861467551 - that this may solve the issue that we were having.\r\n\r\nIn brief (but please read the whole issue if you have the time): we are trying to use a single shared cache directory for all our users to prevent duplicate models. This did not work as we were running into permission errors (due to `-rw-------` as @stas00 shows). Does this PR change the behaviour of created/downloaded files so that they adhere to the permission level of the current directory? Or at least that those files are accessible by all users?\r\n\r\nThanks!", "I think, yes, this was the point of this PR. The problem is that `tempfile` forces user-only perms, so this PR restored them back to `umask`'s setting.\r\n\r\nOne other thing that helps is setting set group id bit `g+s`, which makes sub-dirs and files create under such dirs inherit the perms of the parent dir.\r\n\r\nSo your setup can be something like:\r\n\r\n```\r\nsudo find /shared/path -type d -execdir chmod g+rwxs {} \\;\r\nsudo find /shared/path -type f -execdir chmod g+rw {} \\;\r\nsudo chgrp -R shared_group_name /shared/path\r\n```\r\n\r\nwhere `/shared/path` is obvious and `shared_group_name` is the group name that all users that should have access belong to.\r\n\r\nFinally, each user having `umask 0002` or `umask 0007` in their `~/.bashrc` will make sure that the files will be read/write-able by the group on creation. `0007` is if you don't want files to be readable by others.\r\n\r\nNote that some unix programs don't respect set gid, e.g. `scp` ignores any sub-folders copied with `scp -r` and will set them to user's `umask` perms and drop setgid. But I don't think you'll be affected by it.", "Thanks, this looks promising! We currently have a \"hack\" implemented that simply watches for new file changes and on the creation of a new file, changes the permissions. Not ideal, but seeing that some colleagues use older versions of transformers in their experiments, we will have to make do for now." ]

[]

[]

[]

[]

[]

[ "@forest1988 thank you for your PR! Patrick is off for a couple of weeks, but will review once he's back. Thank you for your patience!", "@LysandreJik \r\nThank you for letting me know about that! That's totally fine!\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." ]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

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

[]

[ "Hi @tanmaylaud ,\r\n\r\nThanks for creating this issue and your interest.\r\n\r\nWe already integrated our models into hugging face :)\r\nYou can find all of them here:\r\n[https://huggingface.co/models?search=code_trans_](https://huggingface.co/models?search=code_trans_)\r\n\r\n@patrickvonplaten : FYI, we have released the paper and all the models (146 models). All the models are fully integrated into hugging face library.", "@agemagician Sorry, I didn't notice it may be because of the search terms I used. \nThanks ! \nI think it deserves a separate category instead of 'Summarization'. ", "Closed by @agemagician :)" ]

[]

[ { "login": "patrickvonplaten", "id": 23423619, "node_id": "MDQ6VXNlcjIzNDIzNjE5", "avatar_url": "https://avatars.githubusercontent.com/u/23423619?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patrickvonplaten", "html_url": "https://github.com/patrickvonplaten", "followers_url": "https://api.github.com/users/patrickvonplaten/followers", "following_url": "https://api.github.com/users/patrickvonplaten/following{/other_user}", "gists_url": "https://api.github.com/users/patrickvonplaten/gists{/gist_id}", "starred_url": "https://api.github.com/users/patrickvonplaten/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patrickvonplaten/subscriptions", "organizations_url": "https://api.github.com/users/patrickvonplaten/orgs", "repos_url": "https://api.github.com/users/patrickvonplaten/repos", "events_url": "https://api.github.com/users/patrickvonplaten/events{/privacy}", "received_events_url": "https://api.github.com/users/patrickvonplaten/received_events", "type": "User", "site_admin": false }, { "login": "patil-suraj", "id": 27137566, "node_id": "MDQ6VXNlcjI3MTM3NTY2", "avatar_url": "https://avatars.githubusercontent.com/u/27137566?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patil-suraj", "html_url": "https://github.com/patil-suraj", "followers_url": "https://api.github.com/users/patil-suraj/followers", "following_url": "https://api.github.com/users/patil-suraj/following{/other_user}", "gists_url": "https://api.github.com/users/patil-suraj/gists{/gist_id}", "starred_url": "https://api.github.com/users/patil-suraj/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patil-suraj/subscriptions", "organizations_url": "https://api.github.com/users/patil-suraj/orgs", "repos_url": "https://api.github.com/users/patil-suraj/repos", "events_url": "https://api.github.com/users/patil-suraj/events{/privacy}", "received_events_url": "https://api.github.com/users/patil-suraj/received_events", "type": "User", "site_admin": false } ]

[ "Hi @DapangLiu \r\n\r\nThe `resume_from_checkpoint` should work for any `PreTrainedModel` class. Even though `EncoderDecoder` model is initialized using two sepearte models when save it using `save_pretarined` it's saved as a single model and can be loaded using `from_pretrained`.\r\n\r\nMake sure that your checkpoint directory contains the saved model file, the saved optimizer and scheduler (if you also want to resume that) and it should work.", "Solved my problem. Thanks @patil-suraj !" ]

[]

[]

[ "Wow, that was quick, I just saw @stefan-it recent comment in our failing PR: https://github.com/deepset-ai/FARM/pull/727#issuecomment-814641885\r\n\r\nSo loading the XLNet Tokenizer in slow mode with\r\n```python\r\ntokenizer_orig = AutoTokenizer.from_pretrained(\"xlnet-base-cased\", use_fast=True, from_slow=True)\r\n```\r\n\r\nresolved the issue.", "Thanks for opening an issue @Timoeller. The `from_slow=True` is a bandaid over the underlying issue.\r\n\r\nWe're actively working on solving this issue, but it will take a bit of time as it's hard to do in a backwards compatible way." ]

[]

[]

[ "> encoding = tokenizer.encode_plus(\r\n> sample_txt,\r\n> max_length=32,\r\n> add_special_tokens=True, # Add '[CLS]' and '[SEP]'\r\n> return_token_type_ids=False,\r\n> pad_to_max_length=True,\r\n> return_attention_mask=True,\r\n> return_tensors='pt', # Return PyTorch tensors\r\n> )\r\n\r\n@nithinreddyy you need to add truncation=True here.", "> > encoding = tokenizer.encode_plus(\n> > sample_txt,\n> > max_length=32,\n> > add_special_tokens=True, # Add '[CLS]' and '[SEP]'\n> > return_token_type_ids=False,\n> > pad_to_max_length=True,\n> > return_attention_mask=True,\n> > return_tensors='pt', # Return PyTorch tensors\n> > )\n> \n> @nithinreddyy you need to add truncation=True here.\n\nThank you will try.", "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." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "We don't care since we are not feeding those inputs to the model, they are split afterwards: the preprocessing has two stages, first tokenize everything then create the blocks of size `block_size`. We can't truncate large inputs because then the special tokens would be added weirdly and there is no way I know of of deactivating that warning.\r\n\r\nI don't see it as being bad, like the warning when we load a checkpoint in an architecture that doesn't completely fit and some weights are randomly initialized. We can add it to the README documenting the script if you find it that scary.", "Thank you for explaining why this is done in this way, @sgugger \r\n\r\nHere are a few possible solutions:\r\n\r\n1. Perhaps the tokenizer utils shouldn't be in charge of dispensing warnings out of context? If you propose the user to ignore a warning, it's a good idea to ask ourselves why are we issuing this warning in the first place?\r\n \r\n2. Pass a new flag like `unchunked_long_input=True` which will skip the warning as now we know it's by design.\r\n\r\nAnd to explain why `1462828 > 1024` looks scary because it's not \"warning: 2048 > 1024` - ok, I'm feeding a slightly longer input - I know what I'm doing. `1462828` is much much larger and it makes no ready sense as it can be an arbitrary huge number which doesn't immediately correlate to anything. And I saw that warning when dealing with NaN, so I immediately had a question - could that somehow be connected? It's not but when you deal with an error, every warning could potentially be flagging the pointer to the cause of the error.", "Looking more, there is a `verbose` keyword argument that we could pass to this tokenization call that will then not issue the warning.", "Indeed, this:\r\n```\r\ndiff --git a/examples/language-modeling/run_clm.py b/examples/language-modeling/run_clm.py\r\nindex a49c815e2..682a9e328 100755\r\n--- a/examples/language-modeling/run_clm.py\r\n+++ b/examples/language-modeling/run_clm.py\r\n@@ -317,7 +317,9 @@ def main():\r\n text_column_name = \"text\" if \"text\" in column_names else column_names[0]\r\n\r\n def tokenize_function(examples):\r\n- return tokenizer(examples[text_column_name])\r\n+ # clm input could be much much longer than block_size - so silence the warning about size\r\n+ # mismatch - and all other warnings with it as a side-effect\r\n+ return tokenizer(examples[text_column_name], verbose=False)\r\n\r\n tokenized_datasets = datasets.map(\r\n tokenize_function,\r\n```\r\n\r\nturns it off.\r\n\r\nBut it turns off all warnings and bunch of info and error logs too.\r\n\r\nWhat is the purpose of `verbose`? I see it's used to control:\r\n* logger.info\r\n* logger.error\r\n* logger.warning - (mostly deprecation warnings)\r\n\r\nso if it's turned off it could mask some possible issues. It looks more like `silent` flag. Still not sure it should silence `logger.error`.\r\n", "Well we will have to choose one of the two ;-) I'm not really fond of adding yet another flag for silencing some things but not others so either the warning stays or we may miss other important infos.\r\n\r\nWhat we can do is add a debug flag to `run_clm` and if the flag is off, we pass `verbose=False` and skip some other things like the part where we show some random samples. If the flag is on, we leave everything as is.", "You consider this silencing, whereas I feel that this should not be happening in the first place. If a given function may receive an input that is very long by design, it should be instructed that it is so and have no qualms about it. The all-or-nothing approach to software communicating back to users is a difficult one.\r\n\r\nAt the same time I trust that if you were to design this area of the software from scratch you'd have done so considering this possibility and I hear that you just don't feel that it's important enough to warrant a special flag the way things are now.\r\n\r\nIf we are on the same page, then I'd rather not turn it all off.\r\n\r\nI think that perhaps a sort of band-aid would be to catch the warning in `run_clm.py` and if it's indeed matching that the input is too long not print it.", "Maybe the warning should be rephrased, but it should still be there in my opinion since in 95% of the cases, the user will be using the tokenzier outputs directly inside the model, this is really an uncommon use case here. So it says something not necessarily warranted in those 5% but will help the rest of the time, I think it's a good trade-off.\r\n\r\nCatching the warning in the script sounds like a good idea, to not scare a beginner.", "By no means I was suggesting to disable the warning in the general case.\r\n\r\nIt's far from trivial to capture and hide it though, since it's not a warning, but logger warning. And while it's probably possible to disable the default logger handler and then re-enable it, somehow it feels unsafe to do.\r\n\r\nHow about this:\r\n\r\n```\r\n def tokenize_function(examples):\r\n from transformers.testing_utils import CaptureLogger\r\n tok_logger = transformers.utils.logging.get_logger(\"transformers.tokenization_utils_base\")\r\n with CaptureLogger(tok_logger) as cl:\r\n output = tokenizer(examples[text_column_name])\r\n # clm input could be much much longer than block_size\r\n if \"Token indices sequence length is longer than the\" in cl.out:\r\n tok_logger.warning(\"^^^^^^^^^^^^^^^^ Please ignore the warning above - it's just a long input\")\r\n return output\r\n```\r\n\r\nSo we end up with an extra warning:\r\n\r\n```\r\n[WARNING|tokenization_utils_base.py:3143] 2021-04-07 21:09:22,144 >> Token indices sequence length is longer than the specified maximum sequence length for this model (1462828 > 1024). Running this sequence through the model will result in indexing errors\r\n[WARNING|run_clm.py:326] 2021-04-07 21:13:14,300 >> ^^^^^^^^^^^^^^^^ Please ignore the warning above - it's just a long input\r\n```\r\n\r\nunfortunately I'm not sure how to make it look as it's coming from the same file, so it's not aligned - tried to add some `^^^^^` to make it loud and clear the helper warning is about the one above and to sort of align it.\r\n\r\nNot great, but at least it explains what happens.\r\n\r\nAlso I'm not sure if it's safe to import from `testing_utils` in examples (any extra dependencies?)\r\n\r\nThoughts.\r\n", "Yeah, it's not ideal to import from `testing_utils` but I don't see any import that is not part of the standard lib there so I don't think it adds any dependence.\r\n\r\nFor the rest, I like your solution, and I don't think it's a problem if the warnings don't come from the same file. It's very clear to me." ]

[]

[ { "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 } ]

[ "Thank you for this report, @guyrosin \r\n\r\nAny idea how we could reproduce this problem? This works for me:\r\n```\r\n$ python -c \"from importlib.metadata import version; print(version('tokenizers'))\"\r\n0.10.1\r\n```\r\n\r\nI do see a different problem though. I see we now have: `\"tokenizers>=0.10.1,<0.11\"`\r\n\r\nI didn't expect a range definition, when I wrote this code, so currently it tries to do:\r\n```\r\nversion.parse('0.10.0') > version.parse('0.10.1,<0.11')\r\n```\r\nwhich is wrong - I'm surprised `version.parse` doesn't assert, it just quietly returns the same unparsed string `0.10.1,<0.11`\r\nso this definitely needs to be fixed to split by ',' and test each condition separately.\r\n\r\nActually does the problem go away if you edit `transformers/dependency_versions_table.py` to use just `\"tokenizers>=0.10.1\"`?\r\n\r\nAlso if you could add a debug print and see what's in `got_ver` and `want_ver` just before it fails to `version.parse`. I think this is the real culprit according to the trace. i.e. it's not that it can't find `tokenizers` - but it fails to parse one of the 2 version inputs.\r\n\r\nThanks.", "Could you please give a try to this PR: https://github.com/huggingface/transformers/pull/11110 and let me know if the problem goes away? Thank you.\r\n", "Thanks for the fast response @stas00! I'm glad this helped you find another bug :)\r\n\r\nI guess the problem in my case is with the tokenizers distribution. I'm getting:\r\n```\r\n$ python -c \"from importlib.metadata import version; print(version('tokenizers'))\"\r\nNone\r\n```\r\nEven after reinstalling `tokenizers`.\r\nSo trying your PR results in \"got_ver is None\" (`want_ver` is 0.10.1)\r\nNo idea how to reproduce it though :\\\r\n\r\nEdit: to make it clear: using `pkg_resources` instead of `importlib` works:\r\n```\r\n$ python -c \"import pkg_resources; print(pkg_resources.get_distribution('tokenizers').version)\"\r\n0.10.1\r\n```", "Does it help if you update/explicitly install this library:\r\n```\r\npip install importlib_metadata -U\r\n```\r\nand then retry?\r\n\r\nIt's good to know that `pkg_resources` does report the right thing. But it also has a cache which apparently reports the past state and not the current, which is the main reason it was replaced.\r\n\r\nBut sometimes the `site-packages` folder gets messed up. \r\n\r\nDoes it make any difference if you first uninstall `tokenizers` **twice** in a row and then install it (I know you said you reinstalled it, but this is slightly different)\r\n\r\nWhat if you create a new environment and try there?", "OK, it seems like my environment was corrupted indeed - there was a `tokenizers-0.9.4.dist-info` folder inside my env's `site-packages` folder... After deleting it (manually) and reinstalling `tokenizers`, everything works!\r\nThanks a lot for your help @stas00!", "Yay! Glad it worked, @guyrosin!" ]

[]

[]

[ "I think that's because of the dynamic loss scaling in FP16, which skips some optimizer steps at the beginning of training if the gradients overflowed (while it's trying to find the best loss scale). Not sure of the best way to treat this:\r\n\r\n- If we somehow manage to catch those skipped optimizer steps and delay the scheduler steps, then we won't respect the number of steps in the scheduler, leading to some wrong end learning rate\r\n- If we don't change anything, then we skip the first values of the learning rate and get that warning.\r\n\r\nSince it's impossible to know beforehand how many optimizer steps will be skipped, I don't see how to deal with that and would just go for 2, with some manual update of ` _step_count` to avoid catching the warning maybe?", "You nailed it - it's `--fp16`, removing it resolves the problem.\r\n\r\nSo the problem is here:\r\n\r\n```\r\n elif self.use_amp:\r\n self.scaler.step(self.optimizer)\r\n self.scaler.update()\r\n [...]\r\n self.lr_scheduler.step()\r\n```\r\n\r\nThis sounds like a pytorch amp issue then. I stepped through with debugger and indeed - the first time it skips optimizer step because of the overflow. \r\n\r\nI will file an issue with pytorch. https://github.com/pytorch/pytorch/issues/55585", "Yes I agree it is more on the PyTorch side to do something for that warning. Thanks for reporting it to them!", "Solved at https://github.com/huggingface/transformers/pull/11144" ]

[]

[]

[]

[ { "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" } ]

[]

[ "Yes, this is known and on the roadmap for the tokenizers: the goal is for them to use their specific config instead of the model config. Hoping to get to that in the next month or so!" ]

[]

[]

[]

[]

[]

[ "Hi @MoonshotQuest, this is because you haven't resized your token embedding matrix after adding new tokens to your vocabulary. The tokenizer can therefore generate new tokens, but the model doesn't know how to handle them. The `add_tokens` method has a [note for that](https://huggingface.co/transformers/internal/tokenization_utils.html?highlight=add_special#transformers.tokenization_utils_base.SpecialTokensMixin.add_tokens)\r\n\r\nIt was unfortunately forgotten for the `add_special_tokens` method and only put in the example, so I'm updating it.", "Hi @LysandreJik - Thanks so much for looking into it!\r\nI did check your note and I think I'm already resizing the token embedding matrix as I add my code on line 308.\r\nLine 309 (unchanged) is already: **model.resize_token_embeddings(len(tokenizer))**\r\nhttps://github.com/huggingface/transformers/blob/master/examples/language-modeling/run_clm.py\r\n\r\nThat gives the following:\r\n```\r\n special_tokens_dict = {\r\n 'bos_token': '<|startoftext|>',\r\n 'eos_token': '<|endoftext|>',\r\n 'additional_special_tokens': [\r\n \"<A>\",\r\n \"<B>\",\r\n \"<C>\",\r\n \"<D>\",\r\n \"<E>\",\r\n \"<F>\",\r\n \"<G>\",\r\n \"<H>\"\r\n ]\r\n }\r\n num_added_toks = tokenizer.add_special_tokens(special_tokens_dict)\r\n model.resize_token_embeddings(len(tokenizer))\r\n```\r\n\r\nIs there another matrix to resize? \r\n", "I am unable to reproduce your issue. In my side adding the code you mentioned to the script runs perfectly with\r\n```\r\npython examples/language-modeling/run_clm.py \\\r\n --model_type gpt2 \\\r\n --tokenizer_name gpt2 \\\r\n --dataset_name wikitext \\\r\n --dataset_config_name wikitext-2-raw-v1 \\\r\n --do_train \\\r\n --do_eval \\\r\n --output_dir ~/tmp/tst-clm \\\r\n --block_size 128 \\\r\n --max_train_samples 100 \\\r\n --overwrite_output_dir \\\r\n --no_use_fast_tokenizer\r\n``` \r\n\r\nI am using wikitext2 since I don't have access to your dataset. Note that your command contains statements that are a bit contradictory:\r\n- you are mixing a `model_name_or_path` (fine-tuning an existing model) with `model_type` (training from scratch).\r\n- you are mixiing `no_cuda` (so training on CPU) with FP16 options, which are not supported on CPU.", "Hi @sgugger thanks for trying to reproduce!\r\n\r\nI removed the GPU related args and isolated the issue to the use of my own folder & pre-trained model:\r\n`--model_type gpt2` # Works perfectly\r\n`--model_name_or_path \"models/original/\"` # Doesn't work and throw the IndexError\r\n\r\nI believe the issue is the model files I'm using in my folder `models/original/` as the pre-trained GPT2 Medium. They seem to be different than the ones downloaded and cached when using the `--model_type gpt2` argument. I only have 2 files in the folder the .bin and the .config. I would like to keep these files in a folder offline as a precaution. \r\nI pulled the files from these URL. Is there a different bin file used for fine-tuning vs a file for inferences only?\r\nhttps://huggingface.co/gpt2-medium/resolve/main/config.json\r\nhttps://s3.amazonaws.com/models.huggingface.co/bert/gpt2-medium-pytorch_model.bin\r\n\r\nI see in some other part of transformers code the following, which may suggest that different pre-trained models .bin and .config files are used for different purpose. Maybe I'm completely wrong! 😂\r\nThanks for your guidance on this.\r\n```\r\n _import_structure[\"models.gpt2\"].extend(\r\n [\r\n \"GPT2_PRETRAINED_MODEL_ARCHIVE_LIST\",\r\n \"GPT2DoubleHeadsModel\",\r\n \"GPT2ForSequenceClassification\",\r\n \"GPT2LMHeadModel\",\r\n \"GPT2Model\",\r\n \"GPT2PreTrainedModel\",\r\n \"load_tf_weights_in_gpt2\",\r\n ]\r\n )\r\n```", "Ah, this is because your checkpoint should have the resized weights: it's resized inside the script but since it's a local folder, it's also passed as a checkpoint to the Trainer later in the script, which then reloads the model from that folder without the `model.resize_token_embeddings(len(tokenizer))` this time. So you have two solutions:\r\n- either load your model, apply `model.resize_token_embeddings(len(tokenizer))` then resave it.\r\n- or remove the line that interprets the folder as a checkpoint [here](https://github.com/huggingface/transformers/blob/c6d664849bdc580cf813b2d3a555a9b33d31b33d/examples/language-modeling/run_clm.py#L414)", "Great thank you so much! @sgugger @LysandreJik \r\nThat makes sense now, I removed the line and it works perfectly. 👍\r\n\r\nI will let you know when we get closer to a launch date for our AI based game. It's going to be awesome!\r\nSorry to troll this thread but does Huggingface has a place to showcase apps made using your incredible libraries? 😊", "We have a [community page](https://huggingface.co/transformers/community.html) in the documentation, otherwise we'll be happy to help you share on social media!", "Awesome!! Take care 🤗" ]

[ { "id": 1897896961, "node_id": "MDU6TGFiZWwxODk3ODk2OTYx", "url": "https://api.github.com/repos/huggingface/transformers/labels/Migration", "name": "Migration", "color": "e99695", "default": false, "description": "" } ]

[]

[ "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." ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Really awesome! Thanks a lot Sylvain!" ]

[]

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

[ "Hi @sboparen \r\n\r\nRight now the caching is implemented such that when `past_key_values` are passed current token length must be 1.\r\nThis is due to the local attention layer which uses dynamic block length. This is a known limitation and I'm working on it at the moment.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Unstale", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "This does not appear to occur with `BertTokenizerFast`, `GPT2TokenizerFast`, or `RobertaTokenizerFast`.", "This bug was introduced in `4.3.0`, I do not encounter this issue when running with `transformers==4.2.2` ", "Hi @johnmccain! This is unfortunately a known issue, and an issue on which we're working. It's a complex issue as it requires updating the `tokenizer.json` on the hub, which needs to be done very carefully in order to not break backwards compatibility for users.\r\n\r\nHowever, I would argue that it's the tokenizer loaded directly from the hub that has wrong offsets. If we look at the following:\r\n```py\r\nfrom transformers import XLMRobertaTokenizerFast\r\n\r\ncheckpoint = \"xlm-roberta-base\"\r\n\r\ntokenizer = XLMRobertaTokenizerFast.from_pretrained(checkpoint)\r\ntokenizer.save_pretrained(checkpoint + \"_local\")\r\ntokenizer_loaded = XLMRobertaTokenizerFast.from_pretrained(checkpoint + \"_local\")\r\n\r\nprint(tokenizer('hello world son', return_offsets_mapping=True))\r\nprint(tokenizer('hello world son').encodings[0].tokens)\r\nprint(tokenizer_loaded('hello world son', return_offsets_mapping=True))\r\n```\r\nWe'll obtain the following:\r\n```\r\n{'input_ids': [0, 33600, 31, 8999, 775, 2], 'attention_mask': [1, 1, 1, 1, 1, 1], 'offset_mapping': [(0, 0), (0, 4), (4, 5), (6, 11), (12, 15), (0, 0)]}\r\n['<s>', '▁hell', 'o', '▁world', '▁son', '</s>']\r\n{'input_ids': [0, 33600, 31, 8999, 775, 2], 'attention_mask': [1, 1, 1, 1, 1, 1], 'offset_mapping': [(0, 0), (0, 4), (4, 5), (5, 11), (11, 15), (0, 0)]}\r\n```\r\n\r\nI would argue that these (the second example) are the true offsets:\r\n```\r\n[ '<s>', '▁hell', 'o', '▁world', '▁son', '</s>']\r\n[(0, 0), (0, 4), (4, 5), (5, 11), (11, 15), (0, 0)]\r\n```\r\n\r\n---\r\n\r\n### Solving this\r\n\r\nWe are aware of this issue and are working on solving it. You should have been getting the wrong offsets in version v4.2.2, because this was an offset bug fixed in a recent `tokenizers` version; hence why the tokenizer file on the hub gives you the wrong offsets, while the conversion using the recent `tokenizers` yields accurate offsets.\r\n\r\n In the meantime, you can work around this by forcing your `XLMRobertaTokenizerFast` to proceed to a conversion from the slow tokenizer, which will correctly instantiate your fast tokenizer. You can do so with the `from_slow` flag:\r\n\r\n```\r\ntokenizer = XLMRobertaTokenizerFast.from_pretrained(\"xlm-roberta-base\", from_slow=True)\r\n```\r\n\r\nPlease let me know if this fixes your issue.", "Hey @LysandreJik, thanks for the quick reply! Glad to know y'all are already aware of this and working on a solution.\r\n\r\nI'm not sure that I agree that the tokenizer from the hub is wrong here. It seems to me that the difference between the offsets from the tokenizer loaded via the hub vs. the tokenizer loaded from disk is the inclusion of leading spaces in the offset mappings. My understanding of the offset mappings is that they indicate the indices in the original string that the token represents, excluding the whitespace preceding the token. \r\n\r\nTokenizers other than `XLMRobertaTokenizerFast` are consistent with that behavior and do not include leading whitespace in their token offsets (whether loaded from the hub or from disk).\r\n\r\nExample:\r\n\r\n```python\r\nfrom transformers import BertTokenizerFast, XLMRobertaTokenizerFast, RobertaTokenizerFast\r\nxlm_roberta_tokenizer = XLMRobertaTokenizerFast.from_pretrained('xlm-roberta-base')\r\nxlm_roberta_tokenizer.save_pretrained('test')\r\nxlm_roberta_tokenizer_loaded = XLMRobertaTokenizerFast.from_pretrained('test')\r\nroberta_tokenizer = RobertaTokenizerFast.from_pretrained('roberta-base')\r\nbert_tokenizer = BertTokenizerFast.from_pretrained('bert-base-uncased')\r\n\r\ntest_str = 'hello world son'\r\n\r\nprint('\\n', 'XLMRobertaTokenizerFast from hub')\r\ntokenized = xlm_roberta_tokenizer(test_str, return_offsets_mapping=True)\r\nfor (offset_start, offset_end), token in zip(tokenized.offset_mapping, tokenized.encodings[0].tokens):\r\n print(repr(token), repr(test_str[offset_start:offset_end]))\r\n \r\nprint('\\n', 'XLMRobertaTokenizerFast from disk')\r\ntokenized = xlm_roberta_tokenizer_loaded(test_str, return_offsets_mapping=True)\r\nfor (offset_start, offset_end), token in zip(tokenized.offset_mapping, tokenized.encodings[0].tokens):\r\n print(repr(token), repr(test_str[offset_start:offset_end]))\r\n \r\nprint('\\n', 'RobertaTokenizerFast from hub')\r\ntokenized = roberta_tokenizer(test_str, return_offsets_mapping=True)\r\nfor (offset_start, offset_end), token in zip(tokenized.offset_mapping, tokenized.encodings[0].tokens):\r\n print(repr(token), repr(test_str[offset_start:offset_end]))\r\n\r\nprint('\\n', 'BertTokenizerFast from hub')\r\ntokenized = bert_tokenizer(test_str, return_offsets_mapping=True)\r\nfor (offset_start, offset_end), token in zip(tokenized.offset_mapping, tokenized.encodings[0].tokens):\r\n print(repr(token), repr(test_str[offset_start:offset_end]))\r\n```\r\n\r\nOutput:\r\n```\r\n XLMRobertaTokenizerFast from hub\r\n'<s>' ''\r\n'▁hell' 'hell'\r\n'o' 'o'\r\n'▁world' 'world'\r\n'▁son' 'son'\r\n'</s>' ''\r\n\r\n XLMRobertaTokenizerFast from disk\r\n'<s>' ''\r\n'▁hell' 'hell'\r\n'o' 'o'\r\n'▁world' ' world'\r\n'▁son' ' son'\r\n'</s>' ''\r\n\r\n RobertaTokenizerFast from hub\r\n'<s>' ''\r\n'hello' 'hello'\r\n'Ġworld' 'world'\r\n'Ġson' 'son'\r\n'</s>' ''\r\n\r\n BertTokenizerFast from hub\r\n'[CLS]' ''\r\n'hello' 'hello'\r\n'world' 'world'\r\n'son' 'son'\r\n'[SEP]' ''\r\n```\r\n", "> My understanding of the offset mappings is that they indicate the indices in the original string that the token represents, excluding the whitespace preceding the token.\r\n\r\nThe offset mapping indicates the indices in the original string that the token represents. If the token contains any leading whitespace, then by default it will be part of the offset mapping.\r\n\r\nThe Roberta tokenizer is an exception as it uses an option available only for the byte-level tokenizers which allows to trim the offsets to avoid including the leading whitespace. (cf https://github.com/huggingface/transformers/blob/11505fa139f6eb4896d7f1d181da1e026878d489/src/transformers/convert_slow_tokenizer.py#L293)\r\n\r\nFor more info about the bug @LysandreJik mentioned, you can have a look at https://github.com/huggingface/transformers/issues/9637. The fact that the leading whitespace is not included when you load your tokenizer from the hub is indeed a bug, and it leads to wrong offset mappings in many situations.", "Thanks for the context! Closing as a duplicate of #9637" ]

[]

[]

[ "Hi,\r\n\r\nCan you please ask this question on the [forum](https://discuss.huggingface.co/) rather than here? For example, [this comment](https://discuss.huggingface.co/t/what-are-some-recommended-pretrained-models-for-extracting-semantic-feature-on-single-sentence/2698/3?) might help you already.\r\n\r\nThe authors of HuggingFace like to keep Github issues for bugs/feature requests.\r\n\r\nThank you!\r\n", "Thank you, closing this." ]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Hi @dupuisIRT \r\n\r\nI'm not sure if this qualifies as an issue with the model, I think you should post this on the [forum](https://discuss.huggingface.co/) instead. I am not sure if I can help with this.", "Thanks @patil-suraj, you are right, I will do it.\r\n" ]

[]

[]

[ "Thankyou @sgugger for guiding me at every step. ", "Thank *you* for all your work on this :-)", "One last step: could you make a PR to update the README of this folder to explain how to use the new scripts you added (you can take inspiration in the other folders).", "> One last step: could you make a PR to update the README of this folder to explain how to use the new scripts you added (you can take inspiration in the other folders).\r\n\r\nSure will do that . " ]

[]

[]

[]

[]

[]

[ "Save problem with #10947 but it is still unsolved. Can anyone of HuggingFace's authors please help us out on this issue?\r\n", "The issue you are mentioning concerns TensorFlow models which you are not using so there is no link. In your case, tehre is a problem with your dataset (this is where the error arises from the stack trace you copied) but you are using a class that is not maintained anymore, so you should switch to preprocessing with the datasets library like in [this notebook](https://github.com/huggingface/notebooks/blob/master/examples/language_modeling.ipynb).", "Thanks for the heads-up @sgugger! As a personal favour, can you please convey via the internal channels to update their colab notebooks? [this notebook](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/01_how_to_train.ipynb) is the first search result for training LM from scratch and it was updated a year ago - which makes it difficult for anyone to judge what support for classes has been dropped.", "Yes, creating an updated version of this notebook is on our roadmap.", "So I tried doing the exact same way - the LM does converge but when fine-tuning, I am wracked with weird errors - https://stackoverflow.com/questions/67004233/typeerror-zeros-like-argument-input-when-fine-tuning-on-mlm if anyone has any idea or suggestions, can they please convey it to me? I am completely out of ideas.", "You should use the [forums](https://discuss.huggingface.co/) instead of stack overflow, there will be more people to answer your questions there.", "@sgugger I have posted this in the [forum](https://discuss.huggingface.co/t/error-when-fine-tuning-pretrained-masked-language-model/5386/6) but there is a discrepancy - using `datasets` API again for fine-tuning leads to the same error. I tried with quite different variants of tokenization functions, but this error persists. \r\n\r\nIs there another new way to fine-tune models?", "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 also try to git pull the updated codes and run the official command:\r\n`python run_generation.py --model_type ctrl --model_name ctrl --temperature 0 --repetition 1.2`\r\n\r\nERROR:\r\n\r\n> 04/07/2021 10:22:08 - WARNING - main - device: cpu, n_gpu: 0, 16-bits training: False\r\n> 04/07/2021 10:23:25 - INFO - main - Namespace(device=device(type='cpu'), fp16=False, k=0, length=20, model_name_or_path='ctrl', model_type='ctrl', n_gpu=0, no_cuda=False, num_return_sequences=1, p=0.9, padding_text='', prefix='', prompt='', repetition_penalty=1.2, seed=42, stop_token=None, temperature=0.0, xlm_language='')\r\n> Model prompt >>> Links Hello, my dog is cute\r\n> 2021-04-07 10:24:00.280690: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'libcudart.so.11.0'; dlerror: libcudart.so.11.0: cannot open shared object file: No such file or directory\r\n> 2021-04-07 10:24:00.280732: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.\r\n> Traceback (most recent call last):\r\n> File \"run_generation.py\", line 290, in\r\n> main()\r\n> File \"run_generation.py\", line 231, in main\r\n> preprocessed_prompt_text = prepare_input(args, model, tokenizer, prompt_text)\r\n> File \"run_generation.py\", line 92, in prepare_ctrl_input\r\n> encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False)\r\n> File \"/usr/local/lib/python3.6/dist-packages/transformers/tokenization_utils_base.py\", line 2032, in encode\r\n> **kwargs,\r\n> File \"/usr/local/lib/python3.6/dist-packages/transformers/tokenization_utils_base.py\", line 2357, in encode_plus\r\n> **kwargs,\r\n> File \"/usr/local/lib/python3.6/dist-packages/transformers/tokenization_utils.py\", line 460, in _encode_plus\r\n> verbose=verbose,\r\n> File \"/usr/local/lib/python3.6/dist-packages/transformers/tokenization_utils_base.py\", line 2794, in prepare_for_model\r\n> return_attention_mask=return_attention_mask,\r\n> File \"/usr/local/lib/python3.6/dist-packages/transformers/tokenization_utils_base.py\", line 2593, in pad\r\n> f\"type of {first_element} unknown: {type(first_element)}. \"\r\n> ValueError: type of None unknown: <class 'NoneType'>. Should be one of a python, numpy, pytorch or tensorflow object.", "Hello, I'm having a hard time reproducing. When running your last example, I get an error because you've specified a temperature of 0 whereas it should be strictly positive, but that's it.\r\n\r\nFrom your error I believe it might be due to a setup issue. Can you fill in the environment information requested as asked in the template? `transformers-cli env`\r\n\r\nOr do you have a colab handy with a reproducer? Thank you!", "Oh, I've tweaked a few things and I now have a reproducer! Looking into it.", "Hey @yananchen1989, there was an issue with the vocabulary, I've patched it in [`huggingface@c03b9b`](https://huggingface.co/ctrl/commit/c03b9b1e9f9269e4ec848bc341bb71acab8d18c7). Can you let me know if you manage to use the CTRL model now?", "@LysandreJik \r\nHi ,\r\nhere is the env info:\r\n\r\n> \r\n> - `transformers` version: 4.6.0.dev0\r\n> - Platform: Linux-4.9.2-3.0.0.std7b.el7.5.x86_64-x86_64-with-Ubuntu-18.04-bionic\r\n> - Python version: 3.6.9\r\n> - PyTorch version (GPU?): 1.7.1+cu101 (True)\r\n> - Tensorflow version (GPU?): 2.3.1 (True)\r\n> - Using GPU in script?: <fill in>\r\n> - Using distributed or parallel set-up in script?: <fill in>", "And another machine's env. (actually I run the codes in both the two machines to check if there is something wrong in the env rather than the code itself)\r\n\r\n> - `transformers` version: 4.4.2\r\n> - Platform: Linux-4.9.2-3.0.0.std7b.el7.5.x86_64-x86_64-with-debian-buster-sid\r\n> - Python version: 3.7.10\r\n> - PyTorch version (GPU?): 1.8.1+cu102 (True)\r\n> - Tensorflow version (GPU?): 2.3.1 (True)\r\n> - Using GPU in script?: <fill in>\r\n> - Using distributed or parallel set-up in script?: <fill in>", "After reinstallation by the command: \r\n`pip install git+https://github.com/huggingface/transformers`\r\n\r\nThe env info:\r\n\r\n> - `transformers` version: 4.6.0.dev0\r\n> - Platform: Linux-4.9.2-3.0.0.std7b.el7.5.x86_64-x86_64-with-Ubuntu-18.04-bionic\r\n> - Python version: 3.6.9\r\n> - PyTorch version (GPU?): 1.5.0 (False)\r\n> - Tensorflow version (GPU?): 2.4.1 (False)\r\n> - Using GPU in script?: <fill in>\r\n> - Using distributed or parallel set-up in script?: <fill in>\r\n> \r\n\r\n> In [5]: model_name='ctrl'\r\n> \r\n> In [6]: from transformers import pipeline\r\n> ...: model = pipeline(\"text-generation\", model=model_name, device=0) #\r\n> ...:\r\n> \r\n> In [7]: results = model('Links Hello, my dog is cute ', max_length=250, \\\r\n> ...: repetition_penalty=1.2, \\\r\n> ...: do_sample=True, top_p=0.9, top_k=0, num_return_sequences=5)\r\n> \r\n> In [8]: results\r\n> Out[8]:\r\n> [{'generated_text': 'Links Hello, my dog is cute and amazing thanks to my little girl’s consistent loving. After over 25 years, of honing her character even in a parking lot behind the grocery store. It seems like every time we take our beloved Ava out, something else will change. So I’ll ask you two questions: What kind of place did she grow up in? What type of houses does your wonderful girl live in?? And also – how do all these changes look from when this picture was taken in 2014 but as I talked about earlier, it still holds true today. It’s worth noting that none of those homes are condos. There aren’t many empty houses, so things should be manageable. \\n \\n When I got home there were eight pets at home, six dog owners, three cats, one bunny. Nothing gets shipped or made for us. Before unloading them into their crates with everything on hand – furniture carriers, leashes, harnesses, treats, flashlights, supplies, food bowls, crates and such, etc., – between each person trying its best to figure out if it’s fair to invite more animals in, I helped unload some stuff ourselves … either by heaving boxes off two successive helpers, or helping friends heave in excess to build a makeshift bed inside the trailer. All too often, once we’ve cleaned up, left the mobile units set free, loaded them onto our pickup truck, we’re expected to head'},\r\n\r\nThe bug is fixed, thanks. @LysandreJik ", "@LysandreJik What is the default temperature have you set ?\r\n\r\nI check the official CTRL git, https://github.com/salesforce/ctrl \r\nit set the temperature to zero.", "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." ]

[]

[]

[ "duplicated" ]

[]

[]

[ "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!", "ok. Created [issue](https://discuss.huggingface.co/t/using-rag-with-local-documents/5326) in [forum](https://discuss.huggingface.co)\r\n\r\nThanks @LysandreJik " ]

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

[ "Hey @patil-suraj, I've worked on your very helpful comments in #8387 and implemented them in this PR. Apart from a few formatting changes which I still need to work on the main test that is still failing is `python3 -m pytest tests/test_modeling_distilbert.py::DistilBertModelTest::test_for_causal_lm_decoder`. \r\n\r\nI'm still debugging this but it seems that the `cross_attention_outputs` aren't being produced when the model is run as a decoder because the `encoder_hidden_states` are empty. I'll try and find out why that's the case.", "Hey @patil-suraj. Managed to fix the few tests that were failing and now `make tests` passes for all tests. I'm still seeing a few failed tests in the CI but the error messages don't seem to be related to the DistilBert model.", "Hi @KMFODA , yes the errors don't look related to this PR. Could you please rebase your branch and push again ? That should fix the CI issue. Let me know if it doesn't.", "Done! Thanks for the suggestion @patil-suraj. PR now passes all checks. :)", "Thanks for the review @patil-suraj. Made all the changes as per your comments.", "No problem at all @patil-suraj. Once again I really appreciate the comments and help. Hopefully with this latest push your comments should be addressed :).", "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.", "Hey @patil-suraj. Is there anything else I should change in the latest commit?", "@KMFODA @patil-suraj Wondering if there's any reason this PR is not merged yet? It's been a few months. What is the usual turnaround time one should expect for PRs?", "Hey @ghlai9665. I'm not sure what the usual turnaround time is but I'm aware the team are quite busy with multiple other projects. I'll make sure to look at anything @patil-suraj comes back with soon as.", "Hey @KMFODA ! Sorry for being super slow here. Let me know if you still have time to work on this, I'll help with the merge conflicts and any other changes. Thanks :) ", "Hey @patil-suraj, no problems at all. Yes happy to work on it still. Do you see any conflicts that would not be resolved by a rebase?", "@KMFODA , Thanks a lot!\r\nIt would be awesome if you could rebase and resolve the conflicts. The PR is almost finished, will take another look once the conflicts are resolved. Let me know if you need any help with it!", "Hi @patil-suraj , apologies for the delay. Was stuck in loops trying to ensure all tests passed and then rebasing. Should be good for review now." ]

[]

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

[ "Are you sure it's properly installed? This code should indeed work if `tensoboard` is installed in the environment in which you execute it. You can try to force the `TensorBoard` integration by adding `report_to=[\"tensorboard\"]` in your `TrainingArguments`.", "@sgugger My bad, somehow missed tensorboard installation. Apologies for the inconvenience. Closing the issue." ]

[]

[]

[ "Will do it tomorrow. Wanted to add it to `transformers` to be included in the next release" ]

[]

[]

[ "Hi,\r\n\r\nCan you ask this question on the [forum](https://discuss.huggingface.co/) instead of here? There are plenty of people active there, and will help you with this question.\r\n\r\nGithub issues are more for bugs/feature requests, not training-related questions.\r\n\r\nThanks! ", "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." ]

[]

[]

[]

[]

[]

[ "@merleyc Instead of using pytorch_pretrained_bert try using transformers:- `from transformers import BertTokenizer`.", "Thanks a lot, @frankhart2018! It works." ]

[]

[]

[]

[]

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

[]

[]

[]

[ "I fix the quantization bug through changing the matrix multiply process as below\r\nthe old version\r\n```python\r\ncontext_layer = context_layer.permute(0, 2, 1, 3).contiguous()\r\n\r\n# Should find a better way to do this\r\nw = (\r\n self.dense.weight.t()\r\n .view(self.num_attention_heads, self.attention_head_size, self.hidden_size)\r\n .to(context_layer.dtype)\r\n)\r\nb = self.dense.bias.to(context_layer.dtype)\r\n\r\nprojected_context_layer = torch.einsum(\"bfnd,ndh->bfh\", context_layer, w) + b\r\n```\r\nmy version\r\n```python\r\nb,f,n,d = context_layer.size()[0],context_layer.size()[1],context_layer.size()[2],context_layer.size()[3]\r\ncontext_layer = context_layer.view(b,f,n*d)\r\nprojected_context_layer = self.dense(context_layer)\r\n``` " ]

[]

[]

[ "Thank you for the report, @LouisCastricato \r\n\r\nI think it's pretty safe to take DeepSpeed out of the equation for now, since as you're saying the problem is due to mixed precision, so let's deal with AMP first.\r\n\r\nHow was GPT-Neo pre-trained?\r\n\r\nIs it by chance the case that GPT-Neo was pre-trained with `bfloat16` like t5/mt5 (https://github.com/huggingface/transformers/pull/10956) or was it pre-trained in fp32?", "The 1.3b model was pretrained on TPUs in mesh-tf using fp16. ", "You mean mixed precision fp16, correct?\r\n\r\nAs I haven't used mesh-tf - what would be the equivalent of this setup in the pytorch land? Since if we find the exact equivalent and the model was ported correctly and is used under the same setup - this problem shouldn't exist. does it make sense? \r\n\r\nSo let's find out what is different here (assuming the porting was done correctly).", "OK so `bf16` and not `fp16` - a very important difference. thank you for this correctlion, @leogao2 \r\n\r\nI just wrote about it today: https://discuss.huggingface.co/t/mixed-precision-for-bfloat16-pretrained-models/5315\r\n\r\nI will try to look at it tomorrow, this is probably the same story as t5/mt5 then.", "It'd help to save time if you had a ready way to reproduce the problem, I tried:\r\n\r\n```\r\nexport BS=1; rm -rf /tmp/test-clm; PYTHONPATH=src USE_TF=0 CUDA_VISIBLE_DEVICES=0 python examples/language-modeling/run_clm.py \\\r\n --model_name_or_path EleutherAI/gpt-neo-1.3B \\\r\n --dataset_name wikitext \\\r\n --dataset_config_name wikitext-2-raw-v1 \\\r\n --do_train \\\r\n --max_train_samples 1 \\\r\n --per_device_train_batch_size $BS \\\r\n --output_dir /tmp/test-clm \\\r\n --block_size 128 \\\r\n --logging_steps 1 \r\n```\r\n\r\nIt hardly fits onto a 24GB card with a tiny block size, and fp16 OOMs right away. \r\n\r\nI don't suppose you have a smaller model to experiment with?\r\n\r\nStraightforward `generate` in full fp16 seems to work fine on a single sample to a full `max_length`, so this is good.\r\n\r\n```\r\nfrom transformers import GPTNeoForCausalLM, GPT2Tokenizer\r\nmodel = GPTNeoForCausalLM.from_pretrained(\"EleutherAI/gpt-neo-1.3B\")\r\ntokenizer = GPT2Tokenizer.from_pretrained(\"EleutherAI/gpt-neo-1.3B\")\r\n\r\nprompt = \"In a shocking finding, scientists discovered a herd of unicorns living in a remote, \" \\\r\n \"previously unexplored valley, in the Andes Mountains. Even more surprising to the \" \\\r\n \"researchers was the fact that the unicorns spoke perfect English.\"\r\n\r\ninputs = tokenizer(prompt, return_tensors=\"pt\")\r\ninput_ids = inputs.input_ids.to(\"cuda:0\")\r\nmodel = model.half().to(\"cuda:0\")\r\n\r\ngen_tokens = model.generate(input_ids. do_sample=True, temperature=0.9, max_length=2048,)\r\ngen_text = tokenizer.batch_decode(gen_tokens)[0]\r\nprint(gen_text)\r\n```\r\n\r\nThanks.", "We are working on producing a minimal example for you currently. After checking our internal documents we realized that 1.3b is bfp16 where as 2.7b is fp32", "If you need an A100 to test on, let us know. ", "Hi! As we're doing a few changes to the implementation to make it cleaner over in https://github.com/huggingface/transformers/pull/10985, we ran a quick training to ensure that the model could still train.\r\n\r\nWe leveraged @Xirider's script detailed in https://github.com/Xirider/finetune-gpt2xl in order to fine-tune the 1.3B checkpoint, and we did see a decrease in the loss over this small sample:\r\n\r\n\r\nWe didn't investigate further, but this allows to fine-tune the 1.3B variant on a single V100 GPU.\r\n\r\ncc @patil-suraj ", "That was sequence length 2048?", "> That was sequence length 2048?\r\n\r\nIt's 1024 on wikitext ", "Thanks for pointing that out, it was 1024. The tokenizer configurations on the hub were ill-configured to have a `model_max_length` set to `1024`, I've updated them to have the correct `2048`.\r\n\r\nI added a `--block_size=2048` parameter, see below the training loss:\r\n\r\n\r\n\r\nIt is slightly higher, but isn't a NaN!", "Hm... Maybe our project is just cursed then. Thanks for the pointer, I'll go through installations and see if anything is weird.", "I ran the fine-tuning on the recent branch so I thought this might be it; but I just tested on `master` and I don't get any NaNs either. \r\nDon't hesitate to tell us if we can help further.", "I'm running this on 24GB rtx-3090 and while it's not converging it's not getting NaNs:\r\n\r\n```\r\ngit clone https://github.com/huggingface/transformers\r\ncd transformers\r\ngit clone finetune-gpt2xl\r\nrm -rf output_dir; PYTHONPATH=src USE_TF=0 deepspeed --num_gpus=1 examples/language-modeling/run_clm.py \\\r\n--deepspeed finetune-gpt2xl/ds_config_gptneo.json \\\r\n--model_name_or_path EleutherAI/gpt-neo-1.3B \\\r\n--train_file finetune-gpt2xl/train.csv \\\r\n--validation_file finetune-gpt2xl/validation.csv \\\r\n--do_train \\\r\n--do_eval \\\r\n--fp16 \\\r\n--overwrite_cache \\\r\n--evaluation_strategy=\"steps\" \\\r\n--output_dir output_dir \\\r\n--num_train_epochs 1 \\\r\n--eval_steps 15 \\\r\n--gradient_accumulation_steps 2 \\\r\n--per_device_train_batch_size 1 \\\r\n--use_fast_tokenizer False \\\r\n--learning_rate 5e-06 \\\r\n--warmup_steps 10 --logging_steps 5 --block_size 2048\r\n```", "It looks like the version of DeepSpeed we are running (0.3.11) prevents us from running that example on our hardware. We are in the process of updating DeepSpeed to a newer version (>0.3.12) so that it is not caught by [line 287 of `integrations.py`](https://github.com/huggingface/transformers/blob/master/src/transformers/integrations.py#L287).", "I'm able to reproduce `loss=Nan` while testing deepspeed zero-3 with this:\r\n```\r\ngit clone https://github.com/huggingface/transformers\r\ncd transformers\r\ngit clone https://github.com/Xirider/finetune-gpt2xl\r\n# create finetune-gpt2xl/ds_config_gptneo_zero3.json as shown below\r\nBS=2; rm -rf output_dir; PYTHONPATH=src USE_TF=0 deepspeed --num_gpus=2 examples/language-modeling/run_clm.py \\\r\n--deepspeed finetune-gpt2xl/ds_config_gptneo_zero3.json --model_name_or_path EleutherAI/gpt-neo-1.3B \\\r\n--train_file finetune-gpt2xl/train.csv --validation_file finetune-gpt2xl/validation.csv --do_train --do_eval --fp16 \\\r\n--overwrite_cache --evaluation_strategy=\"steps\" --output_dir output_dir --num_train_epochs 1 --eval_steps 15 \\\r\n--gradient_accumulation_steps 2 --per_device_train_batch_size $BS --per_device_train_batch_size $BS \\\r\n--use_fast_tokenizer False --learning_rate 9e-06 --warmup_steps 100 --logging_steps 5 --block_size 1048\r\n```\r\n\r\naround step 24/174.\r\n\r\nexcept I'm using 2 uncommitted branches mentioned in https://github.com/huggingface/transformers/issues/11044\r\n\r\nI will try to reduce it something smaller.\r\n\r\np.s. for reproducibility purpose here is the config I used: `finetune-gpt2xl/ds_config_gptneo_zero3.json`\r\n```\r\n{\r\n \"fp16\": {\r\n \"enabled\": true,\r\n \"loss_scale\": 0,\r\n \"loss_scale_window\": 1000,\r\n \"initial_scale_power\": 16,\r\n \"hysteresis\": 2,\r\n \"min_loss_scale\": 1\r\n },\r\n\r\n \"zero_optimization\": {\r\n \"stage\": 3,\r\n \"cpu_offload\": true,\r\n \"cpu_offload_params\": true,\r\n \"cpu_offload_use_pin_memory\" : true,\r\n \"overlap_comm\": true,\r\n \"contiguous_gradients\": true,\r\n \"sub_group_size\": 1e14,\r\n \"reduce_bucket_size\": 0,\r\n \"stage3_prefetch_bucket_size\": 0,\r\n \"stage3_param_persistence_threshold\": 0,\r\n \"stage3_max_live_parameters\": 1e9,\r\n \"stage3_max_reuse_distance\": 1e9,\r\n \"stage3_gather_fp16_weights_on_model_save\": true\r\n },\r\n\r\n \"optimizer\": {\r\n \"type\": \"AdamW\",\r\n \"params\": {\r\n \"lr\": 3e-5,\r\n \"betas\": [0.8, 0.999],\r\n \"eps\": 1e-8,\r\n \"weight_decay\": 3e-7\r\n }\r\n },\r\n\r\n \"scheduler\": {\r\n \"type\": \"WarmupLR\",\r\n \"params\": {\r\n \"warmup_min_lr\": 0,\r\n \"warmup_max_lr\": 3e-5,\r\n \"warmup_num_steps\": 500\r\n }\r\n },\r\n\r\n \"steps_per_print\": 2000,\r\n \"wall_clock_breakdown\": false\r\n}\r\n```", "Thank you!", "Does that happen with zero-2?", "Oddly enough it's fine with zero-2 in this particular setup, but the configurations aren't the same so we aren't comparing the same things.\r\n\r\nBut also if I do the same zero-3 training on one gpu there no nan either.\r\n\r\nBut that doesn't matter, as long as we have a way to reproduce nans it's good enough to start working on understanding the cause and then fixing it.\r\n\r\n@samyam from DeepSpeed suggested an idea to try, so I'm going to go back to the mt5 which gets a NaN on the very first step and experiment with it first, since it's much faster than dealing with step 21 of this really heavy model. And then if it works will come back to gpt-neo.\r\n\r\nIf meanwhile you find a much faster way to get to NaNs that would be helpful. \r\n", "Also I don't know if this is somehow related, but this message looks alarming:\r\n\r\n> [WARNING|tokenization_utils_base.py:3136] 2021-04-07 00:05:14,268 >> Token indices sequence length is longer than the specified maximum sequence length for this model (1462828 > 2048). Running this sequence through the model will result in indexing errors\r\n\r\nI think there is a bug somewhere, but it might be unrelated.\r\n\r\n**edit**: I verified this is just a misplaced warning, not a problem", "@stas00 an update:\r\n\r\nWe were able to run your code on both 125M and 1.3B models without issue. The loss goes down, we get Shakespearean language, all is good.\r\n\r\nUnfortunately, we cannot use your code for our task. We are seeking to train a dual objective model with two complete different datasets. We have two datasets that we are mashing together and trying to train via contrastive loss. Unfortunately, it appears that using the HF trainer class makes that more or less impossible.\r\n\r\nIs there a better way to do the pipelining, so we can evade whatever the bug we are running into is? We tried to run it on sequence length 1024, but it ended up eventually going to NaN anyways after a thousand steps or so.\r\n\r\n", "> ... we can evade whatever the bug we are running into is?\r\n\r\nThe `NaN`s appearance in this particular situation is not caused by a bug in either `transformers` or `deepspeed`.\r\n\r\nThe model was trained in one numerical range and you're trying to run it in a different range that it wasn't trained for - there is not too much that can be done here.\r\n\r\nIt's the same problem for any bfloat16-pretrained model. Which includes t5/mt5/pegasus to name a few.\r\n\r\nThe fine-tuning/inference should be done in the same environment it was trained in or an environment that the model can numerically translate to. This is not the case with bf16 vs fp16 - please refer to my commentary at https://discuss.huggingface.co/t/mixed-precision-for-bfloat16-pretrained-models/5315\r\n\r\nWhat we are trying to do now is to find a workaround that will not provide a full mixed precision regime, but a partial one. For that we need to find which operations are safe to run in fp16 and which aren't. And unfortunately as you can see some of these \"runaways\" happen after thousands of steps.\r\n\r\n> We were able to run your code on both 125M and 1.3B models without issue.\r\n\r\nOh, fantastic! Just discovered https://huggingface.co/EleutherAI/gpt-neo-125M after you mentioned it - it'd be much easier to debug with. Thank you for that!\r\n\r\nWhich \"your code\" are you referring to? Trainer + run_clm.py?\r\n\r\nI hear you that the HF Trainer is not suitable for your task. But if you have your own Trainer that works, why won't you use that instead? On other words how can we support you in this situation?", "What I'm going to do next is:\r\n\r\n* [ ] Try to see if deepspeed can be run in fp32 - apparently it was never tried by anyone since until now mixed fp16 just worked.\r\n\r\n\r\n**This proved to be not possible at the moment**\r\n\r\n* [ ] Try to use the suggestion from samyam to only do fp16 matmul in ff layers with pre-scaling and post-unscaling since that's where the bulk of the processing happens\r\n* [ ] try to find a short example I can reproduce gpt-neo Nans with, because if I have to wait 10+min before it nans, it will be a very difficult going.\r\n\r\n", "The dual objective code we refer to can be found here: https://github.com/EleutherAI/visual-grounding\r\n\r\nAnd ok sounds good. The offer for A100s still stands btw, fp32 might be a nightmare on an RTX 3090. ", "> And ok sounds good. The offer for A100s still stands btw, fp32 might be a nightmare on an RTX 3090.\r\n\r\nThank you for your offer, @LouisCastricato - You're very likely correct - I may take you up on that offer at a later time. I'm not planning on finetuning gpt-neo in fp32 on rtx-3090, but just to test that deepspeed can even run in fp32 on a small model. Because if it works you could at least do that.\r\n\r\n> The dual objective code we refer to can be found here: https://github.com/EleutherAI/visual-grounding\r\n\r\nYes, but I'm not sure what to do with this information. My guess is that you developed your own trainer and you're trying to integrate deepspeed into it and are running into issues? What is it specifically that you need to move forward with your project or what is blocking you?", "Oh apologies. I shared it so that you could see the configuration we're using. I think I might have accidentally deleted that part though (bigs thumbs and touchscreens)\r\n\r\n\r\nYes, we're trying to integrate DeepSpeed directly with our training code. Both [ds_config.json](https://github.com/EleutherAI/visual-grounding/blob/main/Training/ds_config.json) and [amp_config.json](https://github.com/EleutherAI/visual-grounding/blob/main/Training/amp_config.json) produce the same NaN error strictly on autoregressive batches- before the forward step. We have not seen the NaN error on the backwards step. \r\n\r\nTherefore, since we do not see it on the other component of our dual objective (in this case is Google's WIT dataset) which has sequence lengths at most 128 tokens. We can see NaNs beginning to appear at sequence length 768 and once we get to 2048 its every batch that has NaNs.", "Thank you for clarifying that, @LouisCastricato \r\n\r\nUnderstood. I will have to get to know this model as I have never worked with it. So I will comment once I had a chance to sit with it after I install all kinds of debug hooks into it.\r\n\r\nwrt your config, it looks good. \r\n\r\n```\r\n \"allgather_bucket_size\": 50000000,\r\n \"reduce_bucket_size\": 50000000,\r\n```\r\nthese might be too small for an efficient operation. You want these to be in 2e8 to 5e8 range according to Samyam.\r\n\r\nI also recommend you switch to the e-notation - it's too easy to miss a zero. In zero3 they have a param with 14 zeros!\r\n\r\nYou may want to enable cpu-offload if you have extra RAM.\r\n\r\nOtherwise there isn't that much to configure in zero-2. There is a lot more to tune up in zero-3.\r\n\r\nAs I mentioned a few comments up, Deepspeed makes an efficient use of hardware, but if the model itself is an issue there is not much that changing Deepspeed configuration can do.", "Hi,\r\n\r\nI was curious if there was any update on this?", "I was busy working on the DeepSpeed ZeRO-3 integration with `transformers` release, so I haven't had a chance to research this issue yet.\r\n\r\nIf I knew it was a quick fix I'd have done it right away, but this kind of a problem is a long process so I need to have uninterrupted time to work on it. Moreover, fixing it in AMP won't necessarily fix it in DeepSpeed (but it'd surely help).\r\n\r\nI started working on the checklist, I'm aware that this is a big problem for you guys and I thought that perhaps at least you could run DeepSpeed in fp32, but, alas, currently it's not possible - you can disable `fp16`, but there are a lot of hardcoded `half()` calls in ZeRO-3 code so it basically ignores the fp16 setting at the moment and just does everything in fp16.\r\n\r\nI doubt the DeepSpeed developers will take care of this any time soon as they have no resources to do so, so if you want to help that could be one task that might help to move things forward a bit - making Deepspeed work with fp32. Then the next stage would be to optimize the parts that can be done in fp16 w/o overflow leaving most of it in fp32. Samyam suggested the matmuls in FF layers would be the best part to do in fp16 as I mentioned some comments earlier.\r\n\r\nJust to give you an idea, the process goes like this: I find something that doesn't work or is missing for `transformers` needs, I file an Issue, nothing happens for awhile, and since I need this for integration I just go and implement it, with the DeepSpeed team's guidance. If we don't do the work it will eventually happen, but that eventual might be in a very long time. \r\n\r\nLet's hope they manage to expand their team with the recent job openings they posted and have more resources to support the projects that integrate their work.\r\n\r\nI also asked them and all the models they have been working with were trained in mixed fp16 precision, so had no reason to sort out bfloat16 (yet).\r\n\r\nSo priorities-wise, will having full fp32-support be useful to you or not really?", "I've been experimenting with mt5, since the overflow is on the first step there. And now trying to see if adding a penalizing factor for large logits might help solve this problem.\r\n\r\n@LouisCastricato, would you like to try this possible workaround? https://github.com/huggingface/transformers/pull/10956#issuecomment-820712267\r\nwith may be a little bit more aggressive `z_loss` - I see the extra penalty is much much smaller on gpt-neo.\r\n\r\n-----------------------\r\n\r\nSo since we released the integration of DeepSpeed ZeRO-3 I no longer can reproduce the overflow, my deepspeed experiments all succeed - no overflow. For example I run:\r\n```\r\nBS=2; rm -rf output_dir; PYTHONPATH=src USE_TF=0 deepspeed --num_gpus=2 \\\r\nexamples/language-modeling/run_clm.py --deepspeed finetune-gpt2xl/ds_config_gptneo_zero3.json \\\r\n--model_name_or_path EleutherAI/gpt-neo-1.3B --train_file finetune-gpt2xl/train.csv \\\r\n--validation_file finetune-gpt2xl/validation.csv --do_train --do_eval --fp16 --overwrite_cache \\\r\n--evaluation_strategy=\"steps\" --output_dir output_dir --num_train_epochs 1 --eval_steps 15 \\\r\n--gradient_accumulation_steps 2 --per_device_train_batch_size $BS --per_device_train_batch_size $BS \\\r\n--use_fast_tokenizer False --learning_rate 9e-06 --warmup_steps 100 --logging_steps 5 \\\r\n--block_size 2048\r\n```\r\nwith `finetune-gpt2xl/ds_config_gptneo_zero3.json` as pasted in https://github.com/huggingface/transformers/issues/11076#issuecomment-814433061\r\n\r\nCan you give me a setup that I can reproduce the overflow with?" ]

[]

[]

[ "> This is great, it's a very clean implementation! However, there are a few things remaining:\r\n> \r\n> * It should be in the BigBird `__init__.py` file\r\n> * It should be in the main `__init__.py` file\r\n> * It should be in the `AutoTokenizer` mapping\r\n> * It should be tested!\r\n\r\nYes, this was just to get things rolling...I would need help with the files referenced in the code (the json files)", "@LysandreJik can you help in resolving this error as you might have seen it before:\r\nException: Error while attempting to build Precompiled normalizer: Cannot parse precompiled_charsmap\r\nIt seems I would require to compile the files first right ? using the spm_compiler?\r\n", "When do you face this error? I only see an error because the tokenizer isn't documented", "> When do you face this error? I only see an error because the tokenizer isn't documented\r\n\r\nThis error occurs when I am writing tests haven't pushed the tests yet because off the error", "Do you have some code with which we can reproduce the error? I've not seen this error before si I can't really help. Maybe @n1t0's seen this error in the past, but having a reproducer would definitely help a long way.", "Looks like a malformed `tokenizer.json` file, but can't say for sure. As @LysandreJik said, a reproducer would help a lot!", "> Looks like a malformed `tokenizer.json` file, but can't say for sure. As @LysandreJik said, a reproducer would help a lot!\r\n\r\n@LysandreJik @n1t0 I have pushed the code so that the tests on circleci would show the error.\r\nHere:\r\nhttps://app.circleci.com/pipelines/github/huggingface/transformers/21981/workflows/e3d357cf-a783-4b65-8f6f-fa1f15bdce1e/jobs/193105", "Wonderful! I'll take a look at it, thanks for pushing!", "@tanmaylaud, Anthony just pushed a commit on your branch which should fix the issue you're mentioning.\r\n\r\nThe rest of the failing tests seem to be due to a mismatch between how you've configured the special tokens within converter vs how they're configured in the initialisartion + `build_inputs_with_special_tokens` methods within your implementation.\r\n\r\nMake sure to pull the branch before continuing to work! ", "> @tanmaylaud, Anthony just pushed a commit on your branch which should fix the issue you're mentioning.\r\n> \r\n> The rest of the failing tests seem to be due to a mismatch between how you've configured the special tokens within converter vs how they're configured in the initialisartion + `build_inputs_with_special_tokens` methods within your implementation.\r\n> \r\n> Make sure to pull the branch before continuing to work!\r\n\r\nThanks @LysandreJik @n1t0 for the update, I am not able to understand the error in one of the tests:\r\nKeyError: 'token_type_ids'\r\nAll slow BERT tokenizers should have this right? ", "I guess the class for the fast tokenizer should specify the available input names just like the slow here: https://github.com/huggingface/transformers/blob/master/src/transformers/models/big_bird/tokenization_big_bird.py#L83", "@n1t0 @LysandreJik could you please help me debug this one test case that is failing due to the difference in tokens being generated. I suspect it is due to non-english Unicode chars but I am not able to understand the problem. \r\n\r\n 57 passed\r\n 1 failed\r\n - tests/test_tokenization_common.py:2146 BigBirdTokenizationTest.test_tokenization_python_rust_equals\r\n 7 skipped\r\n", "Hey @tanmaylaud, sorry for the delay. I'll take a look at the failing test case in a few hours.", "This seems to be out of my depth. @n1t0, maybe you have an idea of what's happening off the top of your head:\r\n\r\nThis code sample:\r\n```py\r\nfrom transformers import AlbertTokenizer, AlbertTokenizerFast\r\nfrom transformers import BigBirdTokenizer, BigBirdTokenizerFast\r\n\r\ntext = \"This text is included to make sure Unicode is handled properly: 力加勝北区ᴵᴺᵀᵃছজটডণত\"\r\n\r\nap = AlbertTokenizer.from_pretrained(\"albert-base-v1\")\r\nar = AlbertTokenizerFast.from_pretrained(\"albert-base-v1\")\r\nbp = BigBirdTokenizer.from_pretrained(\"google/bigbird-roberta-base\")\r\nbr = BigBirdTokenizerFast.from_pretrained(\"google/bigbird-roberta-base\")\r\n\r\nprint(ap.tokenize(text))\r\nprint(ar.tokenize(text))\r\nprint(bp.tokenize(text))\r\nprint(br.tokenize(text))\r\n```\r\n\r\nReturns this:\r\n\r\n```py\r\n['▁this', '▁text', '▁is', '▁included', '▁to', '▁make', '▁sure', '▁unicode', '▁is', '▁handled', '▁properly', ':', '▁', '力加勝北区', 'in', 'ta', 'ছজটডণত']\r\n['▁this', '▁text', '▁is', '▁included', '▁to', '▁make', '▁sure', '▁unicode', '▁is', '▁handled', '▁properly', ':', '▁', '力加勝北区', 'in', 'ta', 'ছজটডণত']\r\n['▁This', '▁text', '▁is', '▁included', '▁to', '▁make', '▁sure', '▁Unicode', '▁is', '▁handled', '▁properly', ':', '▁', '力加勝北区ᴵᴺᵀᵃছজটডণত']\r\n['▁This', '▁text', '▁is', '▁included', '▁to', '▁make', '▁sure', '▁Unicode', '▁is', '▁handled', '▁properly', ':', '▁', '力加勝北区', 'IN', 'Ta', 'ছজটডণত']\r\n```\r\n\r\nAlso, the initialization of both BigBird tokenizer yields the following warning:\r\n```\r\nnormalizer.cc(50) LOG(INFO) precompiled_charsmap is empty. use identity normalization.\r\n```\r\n\r\nHave you ever seen this before?", "Thank you very much for the repro @LysandreJik. It helped me find the issue a lot faster than I would have without it.\r\n\r\nI believe the last failing tests should be fixed by merging master", "@LysandreJik @n1t0 thanks for the help! All tests are ok now. " ]

[]

[]

[]

[]

[]

[]

[]

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

[ "@sgugger I fixed an edge case in the `_collate_batch` method and added tests for that, but if it's ok I'd like to close this PR and open a fresh one.", "Closing in favor of #11128" ]

[]

[]

[]

[]

[]

[]