Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[]

[]

[]

[]

[]

[]

[ "moving the discussion out of the review commentary as it disappears as soon as it's resolved, so it's best to discuss it in the normal comments as this is what this PR is trying to solve.\r\n\r\n-------------\r\n\r\nOh, I see - thank you for catching that. So I didn't solve the actual problem, but had a luck of hiding it under the carpet.\r\n\r\nThe problem is that the `distributed=...` is wrong here - it is currently coded to expect ddp when `distributed==True` and not dp. dp doesn't have `get_world_size()`/etc and so it fails, so should that arg be called `dpp` instead of `distributed`? But in any case the correct solution is then:\r\n```\r\n self.train_dataset.make_sortish_sampler(\r\n self.args.per_device_train_batch_size, distributed=self.args.local_rank != -1)\r\n```\r\nor re-coded to handle dp too? I don't know the initial intention - should it support `sortish_sampler` under dp or not?\r\n\r\nwe need to know whether to:\r\n\r\n1. recode `make_sortish_sampler` to support dp (can't use `get_world_size()`/etc)\r\n2. recode `make_sortish_sampler` to change its `distributed` arg to `dpp`, so that it only does the special case for dpp.\r\n\r\nAnd somewhat unrelated to the actual bug, I'd like to repeat the request at https://github.com/huggingface/transformers/issues/8822 - let's have a simple flag so that the downstream code knows which mode it is under and not via checking ranks and n_gpus which is very confusing and error-prone.", "Here is where the problem happens with dp:\r\n\r\nhttps://github.com/huggingface/transformers/blob/9995a341c9d68a9963d86c506d17330b3ad813f9/examples/seq2seq/utils.py#L361-L368\r\n\r\nSo `dist.is_available()` returns `True` under `dp`, but `dist.get_world_size()` fails, since it only works under `dpp` and requires `torch.distributed.init_process_group()` which doesn't get called under `dp`.", "In `DataParallel` mode, you don't need to do anything to your datalaoder (only in DistributedDataParallel where you need to split the batches across the various processes somehow) so you should make a regular datalaoder in that case.\r\nIn general, the only proper way to detect if you are in distributed data parallel is to look at the test `local_rank != -1` as `torch.distributed` can give you false information there. I agree it would all be much easier if the training arguments contained something that directly gives the distributed environment. ", "> In `DataParallel` mode, you don't need to do anything to your datalaoder (only in DistributedDataParallel where you need to split the batches across the various processes somehow) so you should make a regular datalaoder in that case.\r\n\r\nGreat, so then should we change the signature to make it clear ddp is wanted and not any distributed:\r\n\r\n```\r\n- def make_sortish_sampler(self, batch_size, distributed=False, shuffle=True, **kwargs):\r\n+ def make_sortish_sampler(self, batch_size, ddp=False, shuffle=True, **kwargs):\r\n```\r\n\r\nand adjust the invocations accordingly?\r\n\r\n> In general, the only proper way to detect if you are in distributed data parallel is to look at the test `local_rank != -1` as `torch.distributed` can give you false information there. I agree it would all be much easier if the training arguments contained something that directly gives the distributed environment.\r\n\r\nGreat. Should we create a feature request for that?\r\n", "I think there is a misunderstanding on the terminology: `DataParallel` is not distributed: distributed means launching several processes with the same script. The package `torch.distributed` does not return anything useful for `DataParallel` and ddp stands for *distributed* data parallel, so leaving that argument as distributed seems better to me.\r\n\r\n> Great. Should we create a feature request for that?\r\n\r\nWe can do that, yes.", "If you stick to the specific implementation, yes, dpp is the only distributed mode. But logically it doesn't make sense. DP is just as distributed as DPP, just isn't using the `torch.distributed`, so it's not a very clear distinction and will lead to such confusions all over. \r\n\r\nAs an example if you look at this function usage pattern it's mostly `dataset.make_sortish_sampler(batch_size, distributed=self.hparams.gpus > 1)` which clearly implies for any multi gpu mode (and erroneously so).", "I disagree, in the sense that code use PyTorch should stick with the PyTorch naming conventions. They chose to have a not distributed `DataParallel`, so we should honor that in our naming as well. In Distributed data parallel, you have to use a `DistributedSampler` (but not in `DataParallel`) etc. Those are all *parallel* modes (as you're training with multiple GPUs) but only one is *distributed*.", "That is a reasonable choice to follow. I'm only flagging how this leads to coding errors when a developer assumes that n_gpu> 1 == ddp. So perhaps some extra support is needed there.", "Let's see how it goes once we add the \"distributed_env\" to `TrainingArguments`!", "@sgugger, please kindly review at your convenience - I addressed all the issues you have raised - all should be good - CI failures are unrelated. Thank you!", "> Perfect, thanks a lot for humoring me and my annoying comments :-)\r\n\r\nOn the contrary, your comments were excellent and to the point. \r\n\r\nI was just slow on getting your point of view since in my mind if we solve a problem on multiple gpus it's distributed across multiple-gpus, regardless of the way it's implemented. But here distributed means distributed across multiple processes. Different semantics.", "So this is probably wrong too:\r\n\r\n```\r\n# examples/seq2seq/finetune.py: \r\nsampler = dataset.make_sortish_sampler(batch_size, distributed=self.hparams.gpus > 1)\r\n```\r\n\r\nBut that's code base on PL.\r\n\r\n@patil-suraj, may be you could have a look when you start working at this one? I suspect that it should do a different check for distributed and not check the number of gpus. Let me know if you prefer that I open a separate issue.\r\n", "Dunno how PL works.", "> Let's see how it goes once we add the \"distributed_env\" to `TrainingArguments`!\r\n\r\nAdded a feature request: https://github.com/huggingface/transformers/issues/8858", "Thank you HuggingFace Team and @stas00 , I cannot express how much I appreciate your efforts. " ]

[]

[]

[ "Hi, thanks @stas00 , I would be grateful to integrate this fix, I am currently dealing with this issue and using this script. thanks.", "Once https://github.com/huggingface/transformers/pull/8823 is merged (hopefully Monday), it will be in master, but feel free to use that branch until then.\r\n", "Awesome, I am really thankful, it helps me a lot.\n\nOn Sun, Nov 29, 2020 at 1:48 AM Stas Bekman <[email protected]>\nwrote:\n\n> Once #8823 <https://github.com/huggingface/transformers/pull/8823> is\n> merged (hopefully Monday), it will be in master, but feel free to use that\n> branch until then.\n>\n> —\n> You are receiving this because you commented.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/8822#issuecomment-735311547>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ABP4ZCDF7U67LXQTURXHFRTSSGK6TANCNFSM4UFIIMLA>\n> .\n>\n", "@rabeehk, here is a quick update - as @sgugger pointed out my fix wasn't correct if you wanted the sortish_sampler and we are trying to figure out how to fix it correctly. If all you want is to use sortish_sample with dpp only then the correct fix is most likely this:\r\n```\r\n self.train_dataset.make_sortish_sampler(\r\n self.args.per_device_train_batch_size, distributed=self.args.local_rank != -1)\r\n```\r\nplease watch the development in https://github.com/huggingface/transformers/pull/8823", "OK, the right fix has been merged into master https://github.com/huggingface/transformers/pull/8823 so just update the master and you should have it working, @rabeehk \r\n" ]

[]

[]

[ "I believe the `EncoderDecoderConfig` class has an argument `tie_encoder_decoder` which can be used to share weights between the encoder and decoder.\r\n\r\nIs this what you're looking for?\r\n\r\n@patrickvonplaten ", "Thanks @LysandreJik; it's close enough to what I'm looking for. Closing this issue.", "Hey @bayanbatn, \r\n\r\nWe are currently working on a function (see #8224) that automatically ties encoder and decoder word embeddings only in an automatic fashion...for now what one can do is to simply set it yourself via\r\n\r\n```python\r\nmodel.decoder.word_embeddings = model.encoder.word_embeddings\r\n```" ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "[https://drive.google.com/file/d/1DDIs0MsvmpJU402o1v7eM-8BS3ACbcKV/view?usp=drivesdk]() ", "#\r\nDuplicate of #" ]

[]

[]

[ "Hi @baeseongsu \r\nthanks for opening this. I am actually working on a fairly big PR to revamp these distillation scripts. I'll integrate your modifications directly there to have everything in a single place!\r\nI **hope** to do this by end of week\r\nVictor" ]

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

[]

[ "Mmm, this looks like a problem in Datasets, @lhoestq @thomwolf ?", "Could you check if the same speed differences appear if you're iterating through the `tokenized_datasets` ? \r\nFor example doing\r\n```python\r\nfor i in range(0, len(tokenized_datasets), batch_size):\r\n batch = tokenized_datasets[i:i + batch_size]\r\n```\r\nIf so, please open an issue on the Datasets repo so we can investigate", "I have run the following code for both dataset 10M rows and 100M rows, and the speed is a bit slower for the 100M dataset compared to the 10M dataset. \r\nHowever, when training BERT, I have much higher differences (e.g. in the timing above is 3s Vs. 0.25s). \r\n\r\n```python\r\nprint(\"--- Starting test for 10M batches ---\")\r\nnum_batches = 10000000\r\nbatch_size = 128\r\nimport time\r\nstart_time = time.time()\r\nfor i in range(0, num_batches, batch_size):\r\n batch = tokenized_datasets['train'][i:i + batch_size]\r\nend_time = time.time() - start_time\r\nprint(\"--- %3.3f seconds per 10M batches ---\" % (end_time))\r\n\r\n```\r\n\r\n| Number of lines in dataset | Seconds |\r\n|---|---|\r\n| 10M | 241 |\r\n| 100M | 303 |", "It looks like the major slowdown comes from somewhere else then.\r\nIt could either be from the PyTorch DataLoader but my best guess would be the PyTorch RandomSampler.\r\nFor big datasets the sampler takes a ton of RAM as mentioned in https://github.com/huggingface/datasets/issues/610#issuecomment-731725078, it could slowdown your training signficantly.\r\n\r\nCould you run the same experiment with the dataloader and the random sampler to make sure ?", "To debug I have replaced RandomSampler with SequentialSampler in Trainer class, https://github.com/huggingface/transformers/blob/master/src/transformers/trainer.py at line 374. \r\nWith SequentialSampler it works as expected, with no slower time.\r\n\r\nHow to fix RandomSampler now? ", "`RandomSampler` comes from PyTorch, so you can open an issue there. You will most likely need to implement your own random-ih sampler that goes faster than PyTorch.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "This is still an issue today. I got a 10x speed up in training on a larger dataset, by switching to the sequential sampler (due to the RandomSampler bottleneck). Monkeypatch here to switch to the seqential sampler in case it's useful:\r\n\r\n```python\r\nimport transformers.trainer as trainer\r\nfrom transformers.trainer import SequentialSampler\r\n\r\n\r\ndef sampler_monkey_patch(dataset, generator):\r\n return SequentialSampler(dataset)\r\n\r\n\r\ntrainer.RandomSampler = sampler_monkey_patch\r\n```\r\n\r\nVersions this was used with:\r\n```\r\ntransformers==4.26.1\r\npytorch==1.13.1\r\ndatasets==2.10.1\r\npython==3.9.16\r\n```\r\n\r\nTo detect if this is an issue for you, it's useful to compare the rate at which samples are processed (and gpu utilization), for a small dataset slice versus a large one.", "You can also use an IterableDataset :\r\n\r\n```python \r\ntrain_dataset = train_dataset.to_iterable_dataset()\r\n```\r\n\r\nPS : pass `num_shards=` with a factor of `num_workers` to distribute the data evenly across DataLoader workers \r\n\r\nPS2 : for distributed, see the \"Distributed\" section at https://huggingface.co/docs/datasets/use_with_pytorch", "@sgugger - this issue was closed prematurely. See the comment above where monkeypatching `Trainer` to use a `SequentialSampler` instead of a `RandomSampler` results in a 10x speedup. Could you please re-open this issue?\r\n\r\nI disagree that this is simply a PyTorch issue. Even if this was purely caused by a bug in PyTorch (which is by no means obvious or clear; see the corresponding PyTorch issue [here](https://github.com/pytorch/pytorch/issues/50689)), HF should simply not include this class in its codebase if it entails causing such a bug to users of the `Trainer` class. At the very least, there should be an official workaround." ]

[]

[]

[ "You should use [`save_to_disk`](https://huggingface.co/docs/datasets/package_reference/main_classes.html?highlight=save#datasets.DatasetDict.save_to_disk) and [`load_from_disk`](https://huggingface.co/docs/datasets/package_reference/loading_methods.html?highlight=save_to_disk#datasets.load_from_disk):\r\n\r\n```python\r\nimport datasets\r\n# Will download and preprocess the dataset\r\ndata = datasets.load_dataset(\"wikitext\", \"wikitext-2-raw-v1\")\r\n# Save it in a folder that you can copy\r\ndata.save_to_disk(\"PATH/TO/FOLDER\")\r\n\r\n# On the other machine - reload the ready to use dataset from the copied folder\r\ndata = datasets.load_from_disk(\"PATH/TO/FOLDER\")\r\n```", "thanks!" ]

[]

[]

[]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "Thanks! Looks good to me, just pinging @iliaschalkidis for information/validation", "Wow, thanks for the fix @mdermentzi. If I recall correctly, I was trying things in an initial scratch python script that was using a universal `text` variable back to back and I though it would be better to rename those in 3 different variables to make it clearer. It seems I was quite unwary... ", "No worries @iliaschalkidis! Just a small typo. ;) Thank you for publishing this model. I am so happy I can play with it for my current uni project. " ]

[]

[]

[ "A bit more context would be appreciated here" ]

[]

[]

[ "Replacing the error by a warning defeats the purpose of the check as it will make the CI pass when it should fail. We can see if we want to move it in another command, I'm just afraid it will make the failures in the CI (and what the corresponding fixes are) less understandable to the user.", "Let me think about this over the weekend, I'll try to find a solution by Monday :-)", "Awesome! I let you handle this, so I'm closing the PR." ]

[]

[]

[ "Same as GPT-2, this would benefit from also handling padding on the left; I'll work on this in another PR." ]

[]

[]

[ "Hey @vishwa30 - what exactly do you mean by incorrect output? -> The model obviously doesn't always classify the input correctly. Since this issue doesn't seem to be a bug, could you maybe look in the forum whether people asked a similar question before and if not post this one there? :-)\r\n\r\nhttps://discuss.huggingface.co/", "@patrickvonplaten Sure! I will do that. Thanks!!" ]

[]

[]

[]

[]

[]

[ "@LysandreJik, I can totally see your point.\r\n\r\nBut this one says the same thing twice:\r\n```\r\n f\"All model checkpoint weights were used when initializing {model.__class__.__name__}.\\n\"\r\n f\"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at {pretrained_model_name_or_path}.\\n\"\r\n```\r\n\r\nThey aren't exactly the same but they say twice that there was no problem in loading the model\r\n\r\nSo as with your excellent example the following would be fitting:\r\n```\r\n\"Loading model {name} at {path}\"\r\n(any exceptions go here)\r\n\"Model loaded\"\r\n```\r\n\r\nWould you be open if changed this PR to follow this strategy?", "They don't really say the same thing, I would see the first one as:\r\n\r\n```\r\nChecking if all checkpoints weights were used in the model ...\r\nAll checkpoint weights were used.\r\n[...]\r\nChecking if all weights of the model were initialized by the checkpoint ...\r\nAll model weights are initialized.\r\n```\r\nI think both serve a purpose:\r\n1. Is the checkpoint meant for that architecture, or was it trained for another one.\r\n2. Is the model perfectly initialized from that checkpoint, or will it require some fine-tuning.", "The full current log for point 2 is:\r\n``` \r\n logger.info(\r\n f\"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at {pretrained_model_name_or_path}.\\n\"\r\n f\"If your task is similar to the task the model of the checkpoint was trained on, \"\r\n f\"you can already use {model.__class__.__name__} for predictions without further training.\"\r\n )\r\n```\r\n\r\nSo that log's line 2+3 are still there. I didn't suggest to remove those.\r\n\r\nBut I feel that this is bordering on splitting hairs (from my side), so I will just let it be. \r\n\r\nYour explanation of its purpose makes sense, @LysandreJik. I'd have just tuned it up to make it more factual and less verbose ...\r\n\r\nMy issue is that I read those logs and not all of them feel like very readable to me...\r\n\r\n\r\n" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Hey @StillKeepTry - could you maybe link the paper corresponding to your model add a small PR description? :-) That would be very helpful", "Thanks for the new PR @StillKeepTry - could you add a `test_modeling_mpnet.py` file - it would be important to test the model :-) \r\n\r\nAlso it would be amazing if you could give some context of MPNet - is there a paper, blog post, analysis, results that go along with the model? And are there pretrained weights? \r\n\r\nThanks a lot! ", "> Thanks for the new PR @StillKeepTry - could you add a `test_modeling_mpnet.py` file - it would be important to test the model :-) \n> \n> \n> \n> Also it would be amazing if you could give some context of MPNet - is there a paper, blog post, analysis, results that go along with the model? And are there pretrained weights? \n> \n> \n> \n> Thanks a lot! \n\nhttps://arxiv.org/abs/2004.09297", "> > Thanks for the new PR @StillKeepTry - could you add a `test_modeling_mpnet.py` file - it would be important to test the model :-)\r\n> > Also it would be amazing if you could give some context of MPNet - is there a paper, blog post, analysis, results that go along with the model? And are there pretrained weights?\r\n> > Thanks a lot!\r\n> \r\n> https://arxiv.org/abs/2004.09297\r\n\r\nok", "Oh and another thing to do after the merge will be to add your new model to the main README and the documentation so that people can use it! The template should give you a file for the `.rst` (or you can use `docs/model_doc/bert.rst` as an example).", "I have updated `test_modeling_mpnet.py` now.", "Hi, every reviewer. Thank you for your valuable reviews. I have fixed previous comments (like doc, format, and so on) and updated the `tokenization_mpnet.py` and `tokenization_mpnet_fast.py` by removing the inheritance. Besides, I also upload test files (`test_modeling_mpnet.py`, `test_modeling_tf_mpnet.py`) for testing, and model weights into the model hub. ", "Fantastic, thanks for working on it! Will review today.", "@patrickvonplaten Hi, are there any new comments?", "Hello!\r\n\r\nStill some comments:\r\n\r\n1. Update the inputs handling in the TF file, we have merged an update for the booleans last Friday. You can see an example in the TF BERT file if you need one.\r\n2. rebase and fix the conflicting files.\r\n3. Fix the check_code_quality test.", "I think something went wrong with the merge here :-/ Could you try to open a new PR that does not include all previous commits or fix this one? ", "> I think something went wrong with the merge here :-/ Could you try to open a new PR that does not include all previous commits or fix this one?\r\n\r\nOK, it seems something wrong when I update to the latest version. :(", "@patrickvonplaten @JetRunner @sgugger @LysandreJik @jplu The new PR is moved to [https://github.com/huggingface/transformers/pull/8971](https://github.com/huggingface/transformers/pull/8971)" ]

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

[]

[ "I think that would be a cool addition! What do you think @julien-c?", "Yes, I almost wrote something like that a while ago, so go for it 👍\r\n\r\nTo remove old weights you don't use anymore @cdpierse, we could also document a unix command to `find` files sorted by last access time and `rm` them\r\n\r\n(I think @sshleifer or @patrickvonplaten had a bash alias for this at some point?)", "@julien-c I have the PR for handling cached models pushed but I've been trying to think of some way to add a function that allows model deletion, we could use the model names and etags returned by `get_cached_models()` to select specfic model `.bin` files to delete but the problem is that it will still leave behind stray config and tokenizer files which probably isn't great. The filenames for tokenizers, configs, and models don't seem to be related so I'm not sure if they can be deleted that way. The two approaches seem to be to delete from last accessed date or just delete model `.bin` files. ", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "Hello, 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!" ]

[]

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

[ "Sorry to insist. Could you at least share your thoughts on this? @LysandreJik @patrickvonplaten ", "Is the only dependency you change `transformers`? The PyTorch version remained the same (v.1.7.0) for both `transformers` versions? If so, I'll take a deeper look this week.", "@LysandreJik In both experiments `torch==v1.7.0`. However, downgrading to `torch==v.1.6.0` (in both experiments) leads to exactly the same problem. I did `pip install transformers==v3.1.0` and `pip install transformers==v3.2.0` back and forth so that could be the only way other dependencies got updated.\r\n\r\nThank you!", "Okay, thanks for checking. I'll have a look this week.", "Hello @LysandreJik , is there any update on this issue?\r\nThank in advance!", "This issue has been stale for 1 month.", "I think I managed to solve the problem.\r\n\r\nInstead of using the environment variable `os.environ[\"CUDA_VISIBLE_DEVICES\"] = str(gpu)` to specify the GPUs one needs to provide it via `torch.device(f\"cuda:{gpu}\")`." ]

[]

[]

[ "I have simplified the code to show that it is definitely the pretrained tokenizer that breaks the execution:\r\n\r\n```\r\npython run_mlm.py \\\r\n--model_name_or_path bert-base-cased \\\r\n--train_file data.txt \\\r\n--tokenizer_name custom_tokenizer \\\r\n--output_dir output \\\r\n--do_train \\\r\n--num_train_epochs 1 \\\r\n--overwrite_output_dir\r\n\r\n```", "This seems like an issue that concerns the new mlm script, `tokenizers` and `datasets` so I'll ping the holy trinity that may have an idea where the error comes from: @sgugger @n1t0 @lhoestq ", "Looks like the tokenizer returns an empty batch of elements, which causes an `IndexError` ?", "is this issue resolved? ran into the same error.", "From recent experience, I think this might happen if no `model_max_length` is set for the tokenizer.\r\n\r\nIn the directory where your tokenizer files live, do you mind adding another file called `tokenizer_config.json`, with the following information: `{\"model_max_length\": 512}`?\r\n\r\nThank you.", "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." ]

[]

[]

[ "Failing tests seem to come from some other code (seq2seq)", "@thomwolf could you review this PR as you're the mastermind behind this code?", "@LysandreJik May I merge (failing tests and quality is linked to unrelated `finetune.py` code, I tried to rebase but it does not seem to be enough)" ]

[]

[]

[]

[]

[]

[ "Looks like you have a styling problem, could you run the command `make style` after doing a dev install with\r\n```\r\npip install -e .[dev]\r\n```\r\nin the repo?", "Oops! Done." ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "Thanks @ahmed451! \r\n\r\nFor context, please also read https://discuss.huggingface.co/t/announcement-all-model-cards-will-be-migrated-to-hf-co-model-repos/2755" ]

[]

[]

[ "Oh, looks like there is a last code-style issue to fix. Could you run `make style` on your branch? Then we can merge this.", "I don't have `make` installed 😄 , what is the style issue? Wonder what style issue can go wrong in such a simple patch. The only thing we added is `self.` in those 2 lines", "`check_code_quality` complains about `finetune.py`, but it's not modified by this patch", "Weird indeed. Will merge and fix if the issue persists." ]

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

[]

[ "Sadly not at the moment... -> we are currently thinking about how to improve the `generate()` outputs though! ", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

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

[]

[ "The pooling layer is not used during the fine-tuning if doing MLM, so gradients are not retro-propagated through that layer; the parameters are not updated.", "@LysandreJik The pooling parameters are not needed in MLM fine-tune. But usually, we use MLM to fine-tune BERT on our own corpus, then we use the saved model weight(missed pooling parameters) in downstream task.\r\nIt's unreasonable for us to random initialize the pool parameters, we should reload google's original pooling parameter(though it was not update in MLM fine-tune).", "I see, thank you for explaining! In that case, would using the `BertForPreTraining` model fit your needs? You would only need to pass the masked LM labels, not the NSP labels, but you would still have all the layers that were used for the pre-training.\r\n\r\nThis is something we had not taken into account when implementing the `add_pooling_layer` argument cc @patrickvonplaten @sgugger ", "Hi @LysandreJik,\r\n\r\nI also tried to further pre-train BERT with new, domain specific text data using the recommended run_mlm_wwm.py file, since I read a paper which outlines the benefits of this approach. I also got the warning that the Pooling Layers are not initialized from the model checkpoint. I have a few follow up questions to that:\r\n\r\n- Does that mean that the final hidden vector of the [CLS] token is randomly initialized? That would be an issue for me since I need it in my downstream application.\r\n- If the former point is true: Why is not at least the hidden vector of the source model copied?\r\n- I think to get a proper hidden vector for [CLS], NSP would be needed. If I understand your answers in issue #6330 correctly, you don't support the NSP objective due to the results of the RoBERTa paper. Does that mean there is no code for pre-training BERT in the whole huggingface library which yields meaningful final [CLS] hidden vectors?\r\n- Is there an alternative to [CLS] for downstream tasks that use sentence/document embeddings rather than token embeddings?\r\n\r\nI would really appreciate any kind of help. Thanks a lot!\r\n\r\n", "The [CLS] token was not randomly initialized. It's a token in BERT vocabulary.\r\nWe talk about Pooling Layer in [here](https://github.com/huggingface/transformers/blob/master/src/transformers/models/bert/modeling_bert.py#L609).\r\n", "Oh okay, I see. Only the weight matrix and the bias vector of that feed forward operation on the [CLS] vector are randomly initalized, not the [CLS] vector itself. I misunderstood a comment in another forum. Thanks for clarification @wlhgtc!", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "Found the source of breakage: https://github.com/huggingface/transformers/pull/8604 - I guess that PR needs more work" ]

[]

[]

[ "@mfuntowicz @avital\r\nI just fixed the bug that I was hitting. There might be other places that need this fix as well.", "Wuuhu - first Flax PR :-). This looks great to me @KristianHolsheimer - think you touched all necessary files as well! \r\n\r\n@mfuntowicz - maybe you can take a look as well", "**EDIT** I re-enabled GPU memory preallocation but set the mem fraction < 1/parallelism. That seemed to fix the tests. The problem with this is that future tests might fail if a model doesn't fit in 1/8th of the GPU memory.\r\n\r\n---\r\n\r\nThe flax tests time out. When I ran the tests locally with `pytest -n auto`, I did notice OOM issues due to preallocation of GPU memory by XLA. I addressed this in commit 6bb1f5e600cd35c712f4f980699df7735b4f59eb.\r\n\r\nOther than that, it's hard to debug the tests when there's no output.\r\n\r\nWould it be an option to run these tests single-threaded instead?", "I had the changes in another branche I'm working on, happy to merge this one and will rebase mine 👍.\r\n\r\nThanks for looking at it @KristianHolsheimer " ]

[]

[]

[]

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

[]

[ "This is weird, as the script is tested for evaluation. What does your `dev.txt` file look like?", "Dev.txt contains text in english, one sentence by line.\r\nThe PC I use has 2 graphic cards, so run_clm.py uses the 2 cards for the training, perhaps the bug occurs only when 2 or more graphic card are used for the training ?", "The script is tested on 2 GPUs as well as one. Are you sure this file contains enough text to have a least one batch during evaluation? This is the only thing I can think of for not having an eval_loss returned.", "The dev.txt file contains 46 lines, the train file contains 268263 lines.\r\n\r\nthe specifications of the PC I use :\r\n\r\n- Intel Xeon E5-2650 v4 (Broadwell, 2.20GHz)\r\n- 128 Gb ram\r\n- 2 x Nvidia GeForce GTX 1080 Ti\r\n\r\n", "Like I said, the dev file is maybe too short to provide at least one batch and return a loss. You should try with a longer dev file.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "Thanks for sharing; your filenames are wrong, they should be nested inside folders named from your model id", "> Thanks for sharing; your filenames are wrong, they should be nested inside folders named from your model id\r\n\r\nThanks Julien, I have updated the branch accordingly.", "closing in favor of #8796" ]

[]

[]

[ "After updating the run_squad.py script with a newer version of transformers, it works now!\r\n\r\nThank you!", "@TrupeshKumarPatel Seems that this is not working still. What was the actual solution to this?", "Hi @aleSuglia,\r\nhere is the updated link: https://github.com/uabinf/nlp-group-project-fall-2020-deepbiocomp/blob/main/scripts/qa_script/qa_squad_v1.ipynb , see if this help. If not then please elaborate on the error or problem that you are facing. ", "I have exactly the same error that you reported: `TypeError: TextInputSequence must be str`\r\nBy debugging, I can see that the variable `truncated_query` has a list of integers (which should be the current question's token ids). However, when you pass that to the [encode_plus](https://github.com/huggingface/transformers/blob/df2af6d8b8765b1ac2cda12d2ece09bf7240fba8/src/transformers/data/processors/squad.py#L181) method, you get the error. I guess it's because `encode_plus` expects strings and not integers. Do you have any suggestion?", "If you googled this error and you are reading this post, please do the following. When you create your tokenizer make sure that you set the flag `use_fast` to `False` like this: \r\n```python\r\nAutoTokenizer.from_pretrained(tokenizer_name, use_fast=False)\r\n```\r\n\r\nThis fixes the error. However, I wonder why there is no backward compatibility...", "Had the similar issue with the above. What @aleSuglia suggested indeed works, but the issue still persists; fast version of the tokenizer should be compatible with the previous methods. In my case, I narrowed the problem down to `InputExample`, where `text_b` can be `None`, https://github.com/huggingface/transformers/blob/447808c85f0e6d6b0aeeb07214942bf1e578f9d2/src/transformers/data/processors/utils.py#L47-L48\r\n\r\nbut the tokenizer apparently doesn't accept `None` as an input. So, I found a workaround by changing\r\n```\r\nInputExample(guid=some_id, text_a=some_text, label=some_label)\r\n-> InputExample(guid=some_id, text_a=some_text, text_b='', label=some_label)\r\n```\r\nI'm not sure this completely solves the issue though.", "Potentially related issues: https://github.com/huggingface/transformers/issues/6545 https://github.com/huggingface/transformers/issues/7735 https://github.com/huggingface/transformers/issues/7011" ]

[]

[]

[ "Maybe @julien-c can answer!", "You can check on the model hub:\r\neg. apache-2.0 models: https://huggingface.co/models?filter=license:apache-2.0\r\nmit models: https://huggingface.co/models?filter=license:mit\r\n\r\netc.", "Thanks @julien-c for the update." ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "Decided to delete \"inference: false\"" ]

[]

[]

[ "Hi @moniquebm .\r\n\r\nThe hosted inference is running `nlp = pipeline('ner', model=model, tokenizer=tokenizer, grouped_entities=True)` by default.\r\nThere is currently no way to overload it.\r\n\r\nDoes that explain the difference? ", "Hi @Narsil \r\n\r\n>Does that explain the difference?\r\n\r\nI'm afraid it doesn't explain... I've just tested nlp = pipeline('ner', model=model, tokenizer=tokenizer, grouped_entities=True) and the following (correct) result is generated programmatically:\r\n\r\n[{'entity_group': 'PUB', 'score': 0.9921221256256103, 'word': 'Tribunal de Contas da União'}, {'entity_group': 'LOC', 'score': 0.9405767321586609, 'word': 'Brasília'}, {'entity_group': 'PESSOA', 'score': 0.9840216636657715, 'word': 'Rui Barbosa'}, {'entity_group': 'ORG', 'score': 0.9529051184654236, 'word': 'Veigamed'}, {'entity_group': 'ORG', 'score': 0.8636592030525208, 'word': 'Buyerbr'}]\r\n\r\nBut in fact it seems the API does not group entities:\r\n\r\n```json\r\n[\r\n {\r\n \"entity_group\": \"LOC\",\r\n \"score\": 0.8127626776695251,\r\n \"word\": \"bras\"\r\n },\r\n {\r\n \"entity_group\": \"PESSOA\",\r\n \"score\": 0.7101765692234039,\r\n \"word\": \"rui barbosa\"\r\n },\r\n {\r\n \"entity_group\": \"ORG\",\r\n \"score\": 0.7679458856582642,\r\n \"word\": \"ve\"\r\n },\r\n {\r\n \"entity_group\": \"ORG\",\r\n \"score\": 0.45047426223754883,\r\n \"word\": \"##igamed\"\r\n },\r\n {\r\n \"entity_group\": \"ORG\",\r\n \"score\": 0.8467527627944946,\r\n \"word\": \"bu\"\r\n },\r\n {\r\n \"entity_group\": \"ORG\",\r\n \"score\": 0.6024420410394669,\r\n \"word\": \"##yerbr\"\r\n }\r\n]\r\n```\r\n\r\nThe tokens are also different. One possible explaination is that the Hosted Inference API may be using English tokenizer, but my model/code used Portuguese tokenizer from this model: https://huggingface.co/neuralmind/bert-base-portuguese-cased\r\n\r\nDoes it make sense?", "You need to update your tokenizer on your model in the hub: 'monilouise/ner_pt_br' to reflect this. The hosted inference can't know how to use a different tokenizer than the one you provide.\r\n\r\nIf you are simply using the one from `neuralmind/bert-base-portuguese-case`, you probably just download theirs, and reupload it as your own following this doc: https://huggingface.co/transformers/model_sharing.html", "May I suggest moving the discussion here: \r\nhttps://discuss.huggingface.co/c/intermediate/6\r\nAs it's not really and transformers problem but a Hub one. \r\n\r\nI am closing the issue here. Feel free to comment here to show the new location of the discussion or ping me directly on discuss.\r\n" ]

[]

[]

[]

[]

[]

[ "@jc-hou: I just tested the above script and I get the same output as you. Why did you close the issue?", "`res = model.generate(input_ids, num_beams=1, max_length=100, forced_bos_token_id=0)` solves the issue" ]

[]

[]

[]

[]

[ { "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 } ]

[ "Was fixed on master, could you try from master?\r\n\r\ncc @lhoestq @patrickvonplaten ", "Thanks @julien-c ! It worked using master. But I had this other issue:\r\n\r\n`Using custom data configuration dummy.psgs_w100.nq.no_index\r\nReusing dataset wiki_dpr (/Users/rcoutin/.cache/huggingface/datasets/wiki_dpr/dummy.psgs_w100.nq.no_index-dummy=True,with_index=False/0.0.0/14b973bf2a456087ff69c0fd34526684eed22e48e0dfce4338f9a22b965ce7c2)\r\nUsing custom data configuration dummy.psgs_w100.nq.exact\r\nReusing dataset wiki_dpr (/Users/rcoutin/.cache/huggingface/datasets/wiki_dpr/dummy.psgs_w100.nq.exact-80150455dfcf97d4/0.0.0/14b973bf2a456087ff69c0fd34526684eed22e48e0dfce4338f9a22b965ce7c2)\r\nTraceback (most recent call last):\r\n File \"/Users/rcoutin/git/examples/backup/rag.py\", line 5, in <module>\r\n model = RagTokenForGeneration.from_pretrained(\"facebook/rag-token-nq\", retriever=retriever)\r\n File \"/Users/rcoutin/git/transformers/src/transformers/modeling_utils.py\", line 947, in from_pretrained\r\n model = cls(config, *model_args, **model_kwargs)\r\n File \"/Users/rcoutin/git/transformers/src/transformers/models/rag/modeling_rag.py\", line 1009, in __init__\r\n self.rag = RagModel(config=config, question_encoder=question_encoder, generator=generator, retriever=retriever)\r\n File \"/Users/rcoutin/git/transformers/src/transformers/models/rag/modeling_rag.py\", line 487, in __init__\r\n question_encoder = AutoModel.from_config(config.question_encoder)\r\n File \"/Users/rcoutin/git/transformers/src/transformers/models/auto/modeling_auto.py\", line 615, in from_config\r\n return MODEL_MAPPING[type(config)](config)\r\n File \"/Users/rcoutin/git/transformers/src/transformers/models/dpr/modeling_dpr.py\", line 514, in __init__\r\n self.question_encoder = DPREncoder(config)\r\n File \"/Users/rcoutin/git/transformers/src/transformers/models/dpr/modeling_dpr.py\", line 155, in __init__\r\n self.bert_model = BertModel(config)\r\n File \"/Users/rcoutin/git/transformers/src/transformers/models/bert/modeling_bert.py\", line 764, in __init__\r\n self.embeddings = BertEmbeddings(config)\r\n File \"/Users/rcoutin/git/transformers/src/transformers/models/bert/modeling_bert.py\", line 181, in __init__\r\n self.position_embedding_type = config.position_embedding_type\r\nAttributeError: 'DPRConfig' object has no attribute 'position_embedding_type'`", "Not sure about this one, sorry :/ Calling the RAG gurus!", "Thanks man. I’ll try debug little more my env. Thanks!\n\nEm qua, 25 de nov de 2020 às 20:22, Julien Chaumond <\[email protected]> escreveu:\n\n> Not sure about this one, sorry :/ Calling the RAG gurus!\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/8780#issuecomment-733988235>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ABBZZ6SQTOLH4JHX3MVUUTTSRWGT7ANCNFSM4UCODMUA>\n> .\n>\n", "Looks like the issue comes from the changes of #8276\r\ncc @patrickvonplaten @LysandreJik @zhiheng-huang", "Thanks a lot for spotting the bug @racoutinho and pinpointing it @lhoestq. The PR should fix it", "I love how well maintained this repo is ❤️ \r\nJust ran into this issue yesterday, and was very surprised to see it fixed just 1 day later 👍 ", "Thank you, guys!!!! You are rock stars!!!!" ]

[]

[]

[ "PPLM is unfortunately not maintained anymore. The fix would be to pin the `transformers` version in the PPLM README.", "pinging @w4nderlust and @mimosavvy just in case", "This looks good to me, I believe the return of dict (very welcome change by the way) from the model should be the only thing breaking the pplm code, right?", "> This looks good to me, I believe the return of dict (very welcome change by the way) from the model should be the only thing breaking the pplm code, right?\r\n\r\nYeah and the named argument of model `past_key_values`. I have no issue running the provided example command in [here]( https://github.com/huggingface/transformers/tree/master/examples/text-generation/pplm) and `python run_pplm_discrim_train.py --dataset SST --epochs 1 --batch_size 8`.", "Can you run `make style` to fix the code quality check? Then we should be good for merge :-) ", "Much appreciated!" ]

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

[]

[ "@patrickvonplaten any thoughts on this? Since, I found your work on Bert2Bert very informative :)", "Hey @gulnazaki - you can use XLNet as an encoder, but not as a decoder because it'll be very difficult to add cross-attention functionality to XLNet for the decoder...", "Thanks @patrickvonplaten , I thought so. \r\nAlso, the concept of XLNet is kinda the opposite of uni-directional. \r\n\r\nI will try to increase the sequence length of GPT2 for the output sequence.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "Thanks @patrickvonplaten!\r\n\r\nAs detailed in the first post, boolean parameters cannot be set during the model call in graph mode. This is the major feature brought by this PR. I wanted to focus of TF T5 and TF Bart on a later PR once this logic is ok at least for all the others.", "There is now a better warning message.", "> You say there is a breaking change in graph mode. Does it mean that currently, both eager & graph mode can handle arguments through the configuration & through the function call? I'm unsure on where we stand on this currently.\r\n\r\nYes, both can be done, but it raises issues when through the function call in graph mode. So this PR fixes this with a better handling of this case.\r\n\r\n> It seems like the tests that would be impacted by these changes are the slow tests. Have you run the slow tests? If not, could you run the slow tensorflow tests on this PR? If you don't know how to do that, happy to show you how for next time.\r\n\r\nThis PR partially fixes these tests. Remembert that they do not pass for T5 and BART for the reasons expressed by Patrick. These models, including the saved model tests, will be fixed in same time in a PR just after this one.\r\n\r\nAlso, in a future PR I will rethink the way the attributes are handled in all the layers.", "> Yes, both can be done, but it raises issues when through the function call in graph mode. So this PR fixes this with a better handling of this case.\r\n\r\nSo right now it fails, and with this PR it also fails but with better error handling?\r\n\r\n> This PR partially fixes these tests. Remembert that they do not pass for T5 and BART for the reasons expressed by Patrick. These models, including the saved model tests, will be fixed in same time in a PR just after this one.\r\n\r\nI meant *all* the slow tests, not only the saved models with saved attentions tests. And this PR doesn't only impact the T5 and BART models, so re-running all the slow tests on this PR seems necessary.", "> So right now it fails, and with this PR it also fails but with better error handling?\r\n\r\nNo, before nothing was working in graph mode when the boolean was updated through the function call. Now, I disabled this functionality and there is no more fail, and everything works properly and as expected in eager+graph mode except T5 and BART in graph mode, which will be handled in a later PR.\r\n\r\n> I meant all the slow tests, not only the saved models with saved attentions tests. And this PR doesn't only impact the T5 and BART models, so re-running all the slow tests on this PR seems necessary.\r\n\r\nOk, I will run all of them.", "@LysandreJik All the slow tests are passing but two:\r\n\r\n- `tests/test_modeling_tf_transfo_xl.py::TFTransfoXLModelLanguageGenerationTest::test_lm_generate_transfo_xl_wt103`, I started to see that with @patrickvonplaten \r\n- `tests/test_utils_check_copies.py::CopyCheckTester::test_is_copy_consistent`, @sgugger any idea why this test don't pass anymore? Here the output:\r\n```\r\ndef test_is_copy_consistent(self):\r\n # Base copy consistency\r\n> self.check_copy_consistency(\r\n \"# Copied from transformers.models.bert.modeling_bert.BertLMPredictionHead\",\r\n \"BertLMPredictionHead\",\r\n REFERENCE_CODE + \"\\n\",\r\n )\r\n\r\ntests\\test_utils_check_copies.py:71:\r\n_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ \r\ntests\\test_utils_check_copies.py:59: in check_copy_consistency\r\n self.assertTrue(len(check_copies.is_copy_consistent(fname)) == 0)\r\nE AssertionError: False is not true\r\n```", "@LysandreJik Any other needs for this PR to be merged?", "I investigated why the `test_is_copy_consistent` test failed, that is probably because you launched your command from inside the `tests/` directory, and it has a path hardcoded to `src/transformers`, and therefore cannot find the path `tests/src/transformers`. \r\n\r\nNo issues there it seems! Reviewing a final time and merging if all is good.", "@patrickvonplaten you haven't approved this PR, do you want to give it a final look and merge if ok for you?", "> @LysandreJik All the slow tests are passing but two:\r\n> \r\n> * `tests/test_modeling_tf_transfo_xl.py::TFTransfoXLModelLanguageGenerationTest::test_lm_generate_transfo_xl_wt103`, I started to see that with @patrickvonplaten\r\n> * `tests/test_utils_check_copies.py::CopyCheckTester::test_is_copy_consistent`, @sgugger any idea why this test don't pass anymore? Here the output:\r\n> \r\n> ```\r\n> def test_is_copy_consistent(self):\r\n> # Base copy consistency\r\n> > self.check_copy_consistency(\r\n> \"# Copied from transformers.models.bert.modeling_bert.BertLMPredictionHead\",\r\n> \"BertLMPredictionHead\",\r\n> REFERENCE_CODE + \"\\n\",\r\n> )\r\n> \r\n> tests\\test_utils_check_copies.py:71:\r\n> _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ \r\n> tests\\test_utils_check_copies.py:59: in check_copy_consistency\r\n> self.assertTrue(len(check_copies.is_copy_consistent(fname)) == 0)\r\n> E AssertionError: False is not true\r\n> ```\r\n\r\nI'll investigate for `tests/test_modeling_tf_transfo_xl.py::TFTransfoXLModelLanguageGenerationTest::test_lm_generate_transfo_xl_wt103` -> thanks for pinging me on that! PR is good for me!" ]

[]

[]

[ "> which feeds the pooled output into a multilayer feedforward-network with one hidden layer and tanh activation. So this is more than only a simple linear layer.\r\n\r\nActually, the final hidden representation of the `[CLS]` token (or `<s>` token in case of RoBERTa) is not the pooled output. Applying the feedforward neural network with tanh activation on this hidden representation actually gives you the pooled output (which is a vector of size 768 in case of the base sized model). Then, after this, a linear layer called [`out_proj`](https://github.com/huggingface/transformers/blob/90d5ab3bfe8c20d9beccfe89fdfd62a8e5ac31e5/src/transformers/models/roberta/modeling_roberta.py#L1248) is used to project the pooled output of size 768 into a vector of size `num_labels`. So the documentation is still correct.\r\n\r\nFor the second question, actually BERT does the same, it is just implemented differently. In `modeling_bert.py`, they use the `pooled_output` of `BertModel`, and then apply the linear layer on top of this. This pooled output has already applied the feedforward neural network + tanh activation on top of the `[CLS]` token hidden representation, as you can see [here](https://github.com/huggingface/transformers/blob/90d5ab3bfe8c20d9beccfe89fdfd62a8e5ac31e5/examples/movement-pruning/emmental/modeling_bert_masked.py#L371). In `modeling_roberta.py`, they implement it differently: they start from the `sequence_output` (which is a tensor containing the final hidden representations of all tokens in the sequence), then get the hidden repr of the `<s>` token by typing `[:,0,:]`, then apply the feedforward nn + tanh and finally the linear projection layer.\r\n\r\nSo your confusion probably comes from the different ways in which this is implemented in BERT vs RoBERTa, and the meaning of `pooled_output`. Actually, some people use \"pooled output\" to denote the final hidden representation of the [CLS] token, but in HuggingFace transformers, this always refers to the output of a linear layer + tanh on top of this vector. ", "Thank you very much for the explanation @NielsRogge !\r\nMy confusion indeed comes from the different implementations and the meaning of \"pooled output\".\r\n\r\nSo this makes it consistent for the HuggingFace transformers library. But do you know the origin of it (now I am interested for both models)? Why is the `[CLS]` token representation transformed by a linear layer with tanh? I couldn't find any reference to tanh in the [original BERT paper](https://www.aclweb.org/anthology/N19-1423/). What they describe in section 4.1, e.g., sounds to me like there is only one linear layer on top of the [CLS] token representation. Is this a HuggingFace invention then? They don't seem to mention it in [their arXiv paper](https://arxiv.org/abs/1910.03771) either.", "Interesting question! Turns out this has [already been asked before here](https://github.com/huggingface/transformers/issues/782) and the answer by the author is [here](https://github.com/google-research/bert/issues/43#issuecomment-435980269).", "Thank you again @NielsRogge !\r\nI had only searched for issues with RoBERTa. Now it makes sense!" ]

[ { "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" } ]

[]

[ "I think it would be nice indeed.\r\n\r\nIt's been on the TO-DO list for a long time (cc @julien-c) but I think nobody's reached it yet so feel free to tackle it :)\r\n\r\nFull (and tested) backward compatibility will be paramount though", "@thomwolf Thanks for responding— I can start working on this once I get done with my current project, which is adding support for Performer-style attention ([see issue here](https://github.com/huggingface/transformers/issues/7675)).", "Seems like this issue is a bit stale but I could use it as a first issue to start contributing. Mind if I take this?", "What do you think @sgugger @patrickvonplaten?", "Not sure it's worth the time: what exactly would we gain from it apart avoiding storing all args from the init? Since we would need to convert all configurations together (for inheritance you need to go from dataclasses to dataclasses) this is work that can't be split across several PRs and needs to happen all at once.\r\n\r\nI also had trouble several times in `TrainingArguments` where the fact it's a dataclass made things harder than they should be, so we may very well lost some of the features of configs (settings params with harmonized names for instance).", "> Not sure it's worth the time: what exactly would we gain from it apart avoiding storing all args from the init? Since we would need to convert all configurations together (for inheritance you need to go from dataclasses to dataclasses) this is work that can't be split across several PRs and needs to happen all at once.\r\n> \r\n> I also had trouble several times in `TrainingArguments` where the fact it's a dataclass made things harder than they should be, so we may very well lost some of the features of configs (settings params with harmonized names for instance).\r\n\r\nI think this is a bit easier than you thought because `class` can inherit from `dataclass` and vice versa. We don't have to have a gigantic PR that changes 150+ files. We can change `PretrainedConfig` into `dataclass` first and do the rest in separate PRs in a backward compatible fashion. See the example blow: \r\n```\r\n@dataclass\r\nclass PretrainedConfig:\r\n x: int = 10\r\n y: int = 15\r\n \r\n def __post_init__(self):\r\n # for the logic that wouldn't fit in the data class constructor, we can add it here\r\n self.z = self.x+ self.y\r\n\r\n@dataclass\r\nclass AlbertConfig(PretrainedConfig):\r\n xy: int = 100\r\n\r\n# this should be backward compatible\r\nclass BertConfig(PretrainedConfig):\r\n def __init__(self, a, b, *args, **kwargs):\r\n super().__init__(*args, **kwargs)\r\n self.a = a\r\n self.b = b\r\n```\r\nBut I do agree that this is not the best ROI and quite a bit of trivial work just to shave off some boilerplate code.", "In that case, okay for me if you want to try to convert `PretrainedConfig` first. If the tests don't pass and you struggle to fix them though, don't spend too much time on it and look for another way to contribute :-)", "Hmm this is more complicated than I thought. `PretrainedConfig`, with all that logic in the constructor, properties and class methods, is too heavy of a class to be a textbook `dataclass`. You are right that I should look for something else :)" ]

[]

[]

[]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "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!" ]

[ { "id": 2627272588, "node_id": "MDU6TGFiZWwyNjI3MjcyNTg4", "url": "https://api.github.com/repos/huggingface/transformers/labels/Model%20Parallel", "name": "Model Parallel", "color": "8B66A5", "default": false, "description": "Model Parallelilsm Implementations" }, { "id": 2796628563, "node_id": "MDU6TGFiZWwyNzk2NjI4NTYz", "url": "https://api.github.com/repos/huggingface/transformers/labels/WIP", "name": "WIP", "color": "234C99", "default": false, "description": "Label your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progress" } ]

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

[ "Thank you, @alexorona!\r\n\r\nI'm still in the process of gathering info/reading up and doing some small experimentation, so will post my thoughts once I have something concrete to share.\r\n\r\nHere are some resources if someone wants to join in:\r\n\r\nAbbreviations:\r\n\r\n- MP = Model Parallelism\r\n- DP = Data Parallelism\r\n- PP = Pipeline Parallelism\r\n\r\nResources:\r\n\r\n- Parallel and Distributed Training tutorials at pytorch - a handful, starting with https://pytorch.org/tutorials/beginner/dist_overview.html\r\n\r\n- fairscale\r\n * github https://github.com/facebookresearch/fairscale\r\n * the MP part of fairscale is a fork of https://github.com/NVIDIA/Megatron-LM\r\n\r\n- ZeRO and deepspeed:\r\n * paper ZeRO: Memory Optimizations Toward Training Trillion Parameter Models https://arxiv.org/abs/1910.02054\r\n * paper ZeRO-Offload: Democratizing Billion-Scale Model Training https://arxiv.org/abs/1910.02054\r\n * detailed blog posts with diagrams: \r\n - https://www.microsoft.com/en-us/research/blog/deepspeed-extreme-scale-model-training-for-everyone/\r\n - https://www.microsoft.com/en-us/research/blog/zero-deepspeed-new-system-optimizations-enable-training-models-with-over-100-billion-parameters/\r\n - https://www.microsoft.com/en-us/research/blog/turing-nlg-a-17-billion-parameter-language-model-by-microsoft/\r\n * github https://github.com/microsoft/DeepSpeed\r\n * deepspeed examples git https://github.com/microsoft/DeepSpeedExamples\r\n * deepspeed in PL https://github.com/PyTorchLightning/pytorch-lightning/issues/817\r\n * deepspeed in PT https://github.com/pytorch/pytorch/issues/42849\r\n * discussion of the paper with visuals https://www.youtube.com/watch?v=tC01FRB0M7w\r\n\r\n- Pipeline Parallelism\r\n * DeepSpeed https://www.deepspeed.ai/tutorials/pipeline/\r\n * Fairscale https://fairscale.readthedocs.io/en/latest/api/nn/pipe.html\r\n * GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism https://arxiv.org/abs/1811.06965\r\n * PipeDream: Fast and Efficient Pipeline Parallel DNN Training https://arxiv.org/abs/1806.03377\r\n ", "Update: so we have \r\n* fairscale's sharded_ddp pretty much ready to go https://github.com/huggingface/transformers/pull/9208\r\n* and deepspeed is nicely coming along https://github.com/huggingface/transformers/pull/9211\r\n\r\nI don't have proper benchmarks yet, but I can definitely see 3-5 times less gpu ram usage! So these would be the first go-to solution when a model doesn't fit onto a single GPU.", "OK, so studying @alexorona's t5 MP implementation I think we have a few issues related to how we spread out the models across different devices. \r\n\r\nFor the purpose of this discussion let's use a simplistic approach of having just 2 GPUs (g1 and g2)\r\n\r\n@alexorona's current approach is to assume that encoder and decoder are of the same size and then split 1/2 encoder layers onto g1 and the other half onto g2. Repeat the same for decoder.\r\n\r\nThis approach has 3 issues:\r\n\r\n1. it doesn't work if encoder and decoder aren't of the same size, which is the case with many models. \r\n\r\n2. it introduces unnecessary copying of data from g1 to g2 in the middle of encoder and then again in the middle of decoder, rather than doing just one copy between end of encoder and beginning of decoder. 3 times vs 1 (in our simplistic 2-gpu example). \r\n\r\n3. it leaves out all other layers from the device map and assigns them to the first or the last device in a hardcoded way depending to where they fit better, so the user has no control over where these go.\r\n\r\nIt does make the implementation relatively simple, since we just need to move half the layers of the encoder to g1 and the other half to g2 and bring the inputs/outputs to the right devices.\r\n\r\n* Issue 1 can be fixed by providing 2 device maps - one for encoder and a different one for decoder. They would be the same if `len(encoder) == len(decoder)`. i.e. we are still using @alexorona, split-encoder and split-decoder approach.\r\n\r\n* Issue 2 can be solved again by 2 separate device maps, but the first one will map encoder - the second decoder. So there will be no splitting of the layers of encoder or decoder between separate devices. I think I may try to use this solution for Bart. \r\n```\r\nencoder_device_map > {0 => [1...6]}\r\ndecoder_device_map=> {1 => [1..6]}\r\n```\r\n(note: I'm using a non-python notation of a range here)\r\n\r\nIt will be trickier to allow overlap if the number of layers is different between encoder and decoder - say 6:9 or 6:12 - In which case it might be:\r\n```\r\nencoder_device_map > {0 => [1...6]} # 6 layer encoder\r\ndecoder_device_map=> {0 => [1..2], 1=> [3..9]} # 9 layer decoder\r\n```\r\nSo the model will need to be able to transparently handle switching layers and inputs/outputs not only through its encode/decoder layers but also from encoder to decoder - but it's quite doable.\r\n\r\nThis uneven situation would also be the case on some weird setups like mine where the gpus are of different sizes. On my setup I have one card of 8GB and another 24GB. This won't be an issue with @alexorona's current implementation.\r\n\r\n* To solve Issue 3 would be much more complicated as then almost any main layer/param can be on any device. Not sure about this one. It'd be trivial if pytorch could automatically bring inputs to the device of the params. I sent out a feeler for such possibility here https://github.com/pytorch/pytorch/issues/49961\r\n\r\nIf any of you have had a chance to think about possible solutions and some totally different ways of approaching that please share your insights.", "I was so full of hope that a simple dictionary could serve as a `device_map` for everything, but now you have shattered my blissful ignorance @stas00. But thanks so much for pointing this out! Super important! The characterization is not quite right and I think it's because you're using 2 GPUs, but the problem you identified is real. Basically both the decoder and encoder use the same map, so the first attention block of the decoder is located on the same device as the first attention block of the encoder. The performance degradation is trivial because the hand-off between GPUs when you have 8 or less is pretty efficient (when you have more, there's problems you have to work around by changing the NCCL environment variables). I thought about trying to do what you've suggested, but it meant that the `device_map` would have to get more complicated, which I was trying to avoid. However, if some of the decoder architectures have a different number of layers in the decoder than the encoder, the generalizability of the implementation will just collapse. Oh well. It was nice while it lasted.\r\n\r\nIt looks like you've really busy the last week. Responding to your comments and PRs...", "Thank you for your follow up, @alexorona. \r\n\r\nAs you're saying that from your experience the copying overhead is negligible then your current solution would work perfectly fine in some situations, like the balanced t5, but will need to be altered in others. So very likely it's this and that, rather than not this but that. i.e. no shuttered hopes.\r\n\r\nAnd if this doesn't fit in other situations it can be extended with a separate device_map for encoder and decoder. Perhaps for some models it'd be most efficient to keep the encoder on one set of devices and decoder on the other, and others shared. So that means we need to come with a way of accepting a variety of different device maps.\r\n\r\nPerhaps, we make the device_map to have two parts, but the second part (decoder) to be optional and if not passed then the first one is used for both? Then the simple solution remains mainly unchanged.\r\n\r\nMay I ask if you have used some existing implementation to model your current implementation after, and perhaps you have a list of various MP implementations so that we could study and find the most suitable way that would fit. So far I have only studied the way you approached it.\r\n\r\nThank you.\r\n\r\np.s. here are some examples of models with different encoder/decoder sizes:\r\n* https://huggingface.co/models?search=mbart_\r\n* https://huggingface.co/models?search=allenai%2Fwmt", "I have a few follow up questions, @alexorona \r\n\r\n1. on use of `torch.cuda.empty_cache()` - I guess as long as it remains in `deparallelize` it is not really going to interfere with whatever normal caching is going on. I don't think it will do what you intended it to do with an explicit `gc.collect()` as I explained in https://github.com/huggingface/transformers/pull/9354\r\n\r\n2. when do you think it's better to use this split as you implemented it (again simplifying to 2 gpus 6 layers in encoder and same in decoder):\r\n```\r\n encoder decoder\r\ngpu0 1 2 3 1 2 3\r\ngpu1 4 5 6 4 5 6\r\n```\r\nvs giving the whole gpu to one of them:\r\n```\r\n encoder decoder\r\ngpu0 1 2 3 4 5 6 \r\ngpu1 1 2 3 4 5 6\r\n```\r\n\r\nThank you!", "@alexorona I had a chance to briefly look at your approach to model-parallelism via explicit device map construction. What are your thoughts on extending this approach via the construction of a generic Megatron-style `mpu` object that implements basic methods such as `get_{model,data}_parallel_{rank,group,world_size}()`? My understanding is that DeepSpeed works with any model-parallelism approach that implements these methods (the `mpu` object needs to be passed to `deepspeed.initialize()`), it doesn't have to necessarily be a tensor-splicing approach like Megatron.\r\n\r\nWould it make sense to extend/tweak the device map approach to model-parallelism to fit within the `mpu` setup, as opposed to trying to get deepspeed's memory optimization primitives to work with the MP implementation without leveraging `mpu`?", "@alexorona, I think I found at least one culprit for needing `torch.cuda.set_device(id)` all over the place. There could be more than one culprit, but at least with pytorch-nightly I have to add it in a bunch of places if `apex.normalization.FusedLayerNorm` is used. https://github.com/NVIDIA/apex/issues/1022 If I remove its use, I don't need any `torch.cuda.set_device(id)`.\r\n\r\nOn the other hand I don't see `apex.normalization.FusedLayerNorm` is being used in either t5 or gpt2. So perhaps it's something else. I see many bug reports wrt to switching devices and some ops failing without `torch.cuda.set_device(id)` or some solid pytorch op running just before it. It sounds like a bug in some pytorch operations.\r\n", "Meanwhile I've finished porting `BartForConditionalGeneration` to MP and pretty much adopted a variation of your device_map, so it won't change much from your original design if accepted.\r\n\r\nIt supports either type of map - your split approach or the one I proposed (flat). Here are some examples:\r\n\r\n```\r\ndevice_maps_flat = {\r\n \"sshleifer/tinier_bart\": {\r\n \"encoder\": {0: [0, 1] },\r\n \"decoder\": {1: [0] },\r\n },\r\n \"sshleifer/distilbart-xsum-6-6\": {\r\n \"encoder\": {0: [0, 1, 2, 3, 4, 5] },\r\n \"decoder\": {1: [0, 1, 2, 3, 4, 5] },\r\n },\r\n}\r\n\r\n\r\ndevice_maps_split = {\r\n \"sshleifer/tinier_bart\": {\r\n \"encoder\": {0: [0],\r\n 1: [1],\r\n },\r\n \"decoder\": {1: [0] },\r\n },\r\n \"sshleifer/distilbart-xsum-6-6\": {\r\n \"encoder\": {0: [0, 1, 2],\r\n 1: [3, 4, 5],\r\n },\r\n \"decoder\": {0: [0, 1, 2],\r\n 1: [3, 4, 5],\r\n },\r\n },\r\n}\r\n```\r\n\r\nI think down the road we could support other types by simply using different keys for whatever other configuration is desired.\r\n\r\nI think eventually we will need to benchmark the different splits and see which one is more efficient. e.g. the flat approach currently suffers from the shared embeddings since they need to be constantly switched back and forth between devices!\r\n\r\nI also have much improved magical device switching functions so it should be much faster to port to MP in the future.\r\n\r\nOne other design change I will propose is to drop first/last devices and instead have `self.main_device`, so that everything happens on just one device and we only send to other devices whatever needs to be offloaded - layer/block work that is. So probably it'd mean that the main device should have less than equal number of layers/blocks assigned to it as it'll use more memory for all the inputs and outputs. I still need to polish this idea.", "We also may need to take into consideration @osalpekar's suggestion at https://github.com/pytorch/pytorch/issues/49961#issuecomment-754306157 - I haven't studied that side of things yet so can't comment at the moment. On one side it appear much more complex to setup, on the other side it might make things much easier model-side-wise. If you already familiar with that side of things please share your insights.\r\n", "And another suggestion is to potentially use Pipe Parallelism here: https://github.com/pytorch/pytorch/issues/49961#issuecomment-754326342 by @pritamdamania87\r\n\r\nThe main issue would be that it'll be enabled in pt-1.8\r\n\r\nBut @pritamdamania87 raises a super-important point - and that the current implementation doesn't take advantage of the multiple gpus, other than for their memory. So all the other gpus idle while one works, which is probably not what we want.\r\n\r\nUnless I'm missing something then this means that the current approach that we have been discussing (and released) is really a no-go. Please correct me if I'm wrong.", "Pipeline parallelism is already supported in DeepSpeed, although I haven't played around with it.\r\n\r\nhttps://www.deepspeed.ai/tutorials/pipeline/", "yes, and `fairscale` too!\r\n", "@alexorona, please have a look at this super-important comment https://github.com/pytorch/pytorch/issues/49961#issuecomment-754319348\r\nwhich I understand that `torch.cuda.set_device()` is not just for fixing bugs in some pytorch ops, but it's actually an essential tool to avoid back-n-forth copying of data which happens when `torch.cuda.set_device()` is not set to the device the ops are happening on. Ouch. I couldn't find any docs covering that culprit. \r\n\r\nWe were trying to get rid of it. Now it looks like we need to make sure we have it in every place we switch to a new device. So when switching to a new device we need:\r\n\r\n1. `torch.cuda.set_device(device)`\r\n2. `inputs.to(device)`\r\n3. `layer.to(device)`\r\n\r\n\r\n", "I was asked to share a sort of design/explanation of what we have implemented so far, so here you go (@alexorona please correct me if I have missed anything - thank you!)\r\n\r\n-------------------\r\n\r\n\r\n\r\nHere is an example of a `sshleifer/distilbart-xsum-6-6` `BartForConditionalGeneration` model:\r\n\r\n```\r\n (model): BartModel(\r\n (shared): Embedding(50264, 1024, padding_idx=1)\r\n (encoder): BartEncoder(\r\n (embed_tokens): Embedding(50264, 1024, padding_idx=1)\r\n (embed_positions): BartLearnedPositionalEmbedding(1026, 1024, padding_idx=1)\r\n (layers): ModuleList( 6 x BartEncoderLayer)\r\n (layernorm_embedding): FusedLayerNorm(torch.Size([1024]), eps=1e-05, elementwise_affine=True)\r\n )\r\n (decoder): BartDecoder(\r\n (embed_tokens): Embedding(50264, 1024, padding_idx=1)\r\n (embed_positions): BartLearnedPositionalEmbedding(1026, 1024, padding_idx=1)\r\n (layers): ModuleList( 6 x BartDecoderLayer)\r\n (layernorm_embedding): FusedLayerNorm(torch.Size([1024]), eps=1e-05, elementwise_affine=True)\r\n )\r\n )\r\n (lm_head): Linear(in_features=1024, out_features=50264, bias=False)\r\n)\r\n```\r\nNote that I collapsed the huge bulk of it and it's represented by just 2 lines that I wrote myself - it was not the output of the model dump.\r\n```\r\n (layers): ModuleList( 6 x BartEncoderLayer)\r\n (layers): ModuleList( 6 x BartDecoderLayer)\r\n```\r\nthis is some 90% of the model and that's what we want to spread out through multiple gpus.\r\n\r\n\r\n\r\n\r\n\r\n\r\nSo we have the bulk of memory used by 6 x `BartEncoderLayer` and 6 x `BartDecoderLayer`, plus some other components.\r\n\r\nFor the simplicity of the example let's say we have 2 gpus we want to split the model into.\r\n\r\nCurrently the idea is to put the 6 encoder layers on gpu 0 and the same for decoder layers but on gpu 1:\r\n```\r\ndevice_map = {\r\n \"encoder\": {0: [0, 1, 2, 3, 4, 5] },\r\n \"decoder\": {1: [0, 1, 2, 3, 4, 5] },\r\n }\r\n```\r\nor alternatively, splice each group as following:\r\n```\r\ndevice_map = {\r\n \"encoder\": {0: [0, 1, 2],\r\n 1: [3, 4, 5],\r\n },\r\n \"decoder\": {0: [0, 1, 2],\r\n 1: [3, 4, 5],\r\n },\r\n }\r\n```\r\nand the remaining non-encoder/decoder layer modules can be all on gpu 0 or grouped closer to where they are needed. We still haven't quite finalized that map.\r\n\r\nOf course, other models may have more or less layers and they don't have to have the same number of layers in encoder and decoder.\r\n\r\n\r\nNow that we have the map, we can place different layers/blocks on different devices\r\n\r\n\r\nA simplified explanation would be with the usual drawing of the deep nn (random blocks in this example)\r\n```\r\nblocks | [blk] ... [blk 2] | [blk 3] ... [blk 5] | [blk 6] ... [blk 7] | [head]\r\ndevices | 0 | 1 | 2 | 0\r\n```\r\n\r\nImplementation details:\r\n\r\n1. create model\r\n2. `model.parallelize()`: run through the model's layers and remap them to specific devices as defined by the device map by simply runnin `to(device)`\r\n3. inside `forward` we switch inputs to the same device as the layer's params using a handy wrapper I shared here: https://github.com/pytorch/pytorch/issues/49961#issuecomment-753441248\r\n4. some outputs need to be brought back to the device where the logic of the main program happens (e.g. beam search) \r\n\r\nComplications:\r\n\r\n* shared embeds are a performance issue - we have to switch them back and forth between different devices.\r\n* because some layers have params on different devices the developer has to explicitly choose which device to switch input to\r\n* looks like we may need to sort out that `torch.cuda.set_device()` which apparently is needed too - sometimes to cover for bugs in pytorch, other times for performance - I haven't figured it out yet, I opened an issue:\r\nhttps://github.com/pytorch/pytorch/issues/50112\r\n* beam search works extremely slow with this approach - 10x slowdown.\r\n\r\nTo port a model one needs to apply the device map (stage 2 above) and then gradually deal with wrong device errors, by remapping the inputs to the devices of the params of the layer. Alex was doing each variable manually, which is a huge pain. I automated this process (it's in 2 PRs that haven't been merged yet, the Bart PR has a smarter function)\r\n\r\nTransitions:\r\n\r\n- Alex defined first/last devices to work with. In Bart MP I shifted to a different mapping where everything happens on main_device (say 0), and we only ever switch devices for those stacks of encoder/decoder layers that repeat, but all the helping params remain on device 0, which greatly simplifies things.\r\n\r\n- So when we pass data to the parallelized model we `.to(main_device)` and most of the layers are already on the main_device, so now we only need to switch devices when the stacks end. So if you take the following map:\r\n\r\n```\r\ndevice_map = {\r\n \"encoder\": {0: [0, 1, 2, 3, 4, 5] },\r\n \"decoder\": {1: [0, 1, 2, 3, 4, 5] },\r\n }\r\n```\r\n\r\nHere one only need to change devices twice\r\n1. once when switching between encoder.5 and encoder.0 and\r\n2. once more when returning from forward of decoder.5,\r\n\r\nbut of course, since the user may choose to split them vertically as so:\r\n\r\n```\r\ndevice_map = {\r\n \"encoder\": {0: [0, 1, 2],\r\n 1: [3, 4, 5],\r\n },\r\n \"decoder\": {0: [0, 1, 2],\r\n 1: [3, 4, 5],\r\n },\r\n }\r\n```\r\n\r\nthere will be more switches here.\r\n\r\nSo with the automation of switching `forward` input to the desired device it's only a few surprises that one has to resolve, since each model has some unexpected needs.\r\n\r\nOverall, with the great foundation @alexorona laid out and with a bit of the automation I added the implementation is solid and would work just fine for those who can afford idling gpus.\r\n\r\nWhat we need to figure out next is how these idling gpus will co-operate with all the other great components we have been working on (fairscale/deepspeed/pytorch pipelines/etc.)\r\n", "Great recap @stas00 ", "update: I made t5 work with HF trainer and --model_parallel in eval mode https://github.com/huggingface/transformers/pull/9323 - needed to copy the outputs back to the first device - it's more or less fine in the training stage (it worked in the first place), **but w/ beam search size 4 it's 10x slower on eval w/ MP than w/o MP** - it gets hit badly by the back-n-forth data copying.", "The more I'm reading on various Parallelization strategies the more I see how confusing the terminology is.\r\n\r\nWhat's most call Model Parallel (MP) should probably be called \"Model Distributed\" - since all we are doing here is splitting the model across several GPUs, as such \"Model Distributed\" is a much closer to reality term.\r\n\r\nNext comes Pipeline Parallelism (PP) - where we split the mini-batch into micro-batches and feed into Model Parallel / Model Distributed, so that while a GPU that completed its `forward` idles waiting for other GPUs to compute their chunks of layers of the model and backprop, it can start on a new input. It is a Pipeline for sure, is this parallel though - I have a hard time calling it Parallel, since all the ops are sequential still.\r\n\r\nIt's much easier to understand this by studying this diagram from the [GPipe paper](https://ai.googleblog.com/2019/03/introducing-gpipe-open-source-library.html)\r\n\r\n\r\n\r\nThis diagram makes it very clear why what we have implemented is what it calls a a naive MP, and you can see the huge idling with 4 GPUs.\r\n\r\nIt then shows how it tries to resolve this idling problem with Pipeline. There is still idling but less so.\r\n\r\nIt also misrepresents the length of time forward and backward paths take. From asking the experts in general backward is ~2x slower than forward. But as I was corrected on slack, the length of the bubble is about the same regardless of their execution speed. (Thanks @deepakn94)\r\n\r\nAnd Deepak also stressed out that since with PP there is a splitting into micro-batches, the effective batch size has to be big enough, otherwise PP will be idling too - so it requires experimentation to find a good batch size.\r\n\r\nBottom line, PP is an improved version of MP, according to my current understanding. I'm still still researching. \r\n\r\nI think the real Parallelization is the [ZeRO paper](https://arxiv.org/abs/1910.02054) where Sharding/Partitioning is done and then it's truly parallel processing, but I'm still trying to understand what exactly is going on there. (Need to find a good diagram visually showing what it does) Grr, I see others use sharding/partitioning as a replacement for parallelism... so confusing. \r\n\r\nI updated https://github.com/huggingface/transformers/issues/8771#issuecomment-733224520 with resources on PP and next need to try to convert perhaps t5 to PP and see how it works in practice. There will be issues to overcome due to BN and tied weights.", "@deepakn94 helped me to finally grasp ZeRO-powered data parallelism, as it's described on this diagram from this [blog post](https://www.microsoft.com/en-us/research/blog/zero-deepspeed-new-system-optimizations-enable-training-models-with-over-100-billion-parameters/)\r\n\r\n\r\nSo it's quite simple conceptually, this is just your usual DataParallel (DP), except, instead of replicating the full model params, gradients and optimizer states, each gpu stores only a slice of it. And then at run-time when the full layer params are needed just for the given layer, all gpus sync to give each other parts that they miss - this is it.\r\n\r\nConsider this simple model with 3 layers and each layer has 3 params:\r\n```\r\nLa | Lb | Lc\r\n---|----|---\r\na0 | b0 | c0\r\na1 | b1 | c1\r\na2 | b2 | c2\r\n```\r\nLx being the layer and we have 3 layers, and ax being the weights - 3 weights\r\n\r\nIf we have 3 GPUs, the Sharded DDP (= Zero DP) splits the model onto 3 GPUs like so:\r\n\r\n```\r\nGPU0:\r\nLa | Lb | Lc\r\n---|----|---\r\na0 | b0 | c0\r\n\r\nGPU1:\r\nLa | Lb | Lc\r\n---|----|---\r\na1 | b1 | c1\r\n\r\nGPU2:\r\nLa | Lb | Lc\r\n---|----|---\r\na2 | b2 | c2\r\n```\r\n\r\nIn a way this is horizontal slicing, if you imagine the typical DNN diagram. Vertical slicing is where one puts whole layer-groups on different GPUs. But it's just the starting point.\r\n\r\nNow each of these GPUs will get the usual mini-batch as it works in DP:\r\n```\r\nx0 => GPU0\r\nx1 => GPU1\r\nx2 => GPU2\r\n```\r\n\r\nThe inputs are unmodified - they think they are going to be processed by the normal model.\r\n\r\nSo the inputs first hit the first layer La.\r\n\r\nLet's focus just on GPU0: x0 needs a0, a1, a2 params to do its forward path, but GPU0 has only a0 - so what it does is it gets sent a1 from GPU1 and a2 from GPU2. Now the forward step can happen.\r\n\r\nIn parallel GPU1 gets mini-batch x1 and it only has a1, but needs a0 and a2 params, so it gets those from GPU0 and GPU2. \r\n\r\nSame happens to GPU2 that gets input x2. It gets a0 and a1 from GPU0 and GPU1. \r\n\r\nAs soon as the calculation is done, the data that is no longer needed gets dropped - it's only used during the calculation.\r\n\r\nThe same is repeated at every other stage.\r\n\r\nAnd the whole larger thing is repeated for layer Lb, then Lc forward-wise, and then backward Lc -> Lb -> La.\r\n\r\nTo me this sounds like an efficient group backpacking weight distribution strategy:\r\n\r\n1. person A carries the tent\r\n2. person B carries the stove\r\n3. person C carries the entertainment system\r\n\r\nNow each night they all share what they have with others and get from others what the don't have, and in the morning they pack up their allocated type of gear and continue on their way. This is Sharded DDP / Zero DP.\r\n\r\nCompare this strategy to the simple one where each person has to carry their own tent, stove and entertainment system, which would be far more inefficient. This is DataParallel in pytorch.\r\n\r\nAnd I think pretty much everywhere I read Sharded == Partitioned, so I think those are synonyms in the context of distributed models.\r\n\r\n", "**edit: 2021-02-15: Note that `finetune_trainer.py` was moved to `examples/legacy/seq2seq/`, and there is a new script `run_seq2seq.py` that took over `finetune_trainer.py`, you will find transition notes [here](https://github.com/huggingface/transformers/issues/10036)**\r\n\r\nThe simplest way to quickly reproduce the following is to switch to the transformers sha of the time this was posted, that is:\r\n\r\n```\r\ngit clone https://github.com/huggingface/transformers\r\ncd transformers\r\ngit checkout 7e662e6a3be0ece4 \r\n```\r\n\r\n--------------\r\n\r\nThe amazing discovery of the day is DeepSpeed's [Zero-Offload](https://www.deepspeed.ai/tutorials/zero-offload/). ZeRO-Offload is a ZeRO optimization that offloads the optimizer memory and computation from the GPU to the host CPU. \r\n\r\nYou can use DeepSpeed with a single GPU and train with huge models that won't normally fit onto a single GPU.\r\n\r\nFirst let's try to finetune the huge `t5-3b` with a 24GB rtx-3090:\r\n```\r\nexport BS=1; rm -r output_dir; CUDA_VISIBLE_DEVICES=0 PYTHONPATH=../../src USE_TF=0 ./finetune_trainer.py \\\r\n--model_name_or_path t5-3b --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval \\\r\n--do_predict --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 \\\r\n--logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 \\\r\n--overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate \\\r\n--eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 \\\r\n--val_max_target_length 128 --warmup_steps 5 --n_train 60 --n_val 10 --n_test 10 --fp16\r\n```\r\nNo cookie, even with BS=1\r\n```\r\nRuntimeError: CUDA out of memory. Tried to allocate 64.00 MiB (GPU 0; 23.70 GiB total capacity; 21.37 GiB already allocated; 45.69 MiB free; 22.05 GiB reserved in total by PyTorch)\r\n```\r\n\r\nNow update your `transformers` to master, then install deepspeed:\r\n```\r\npip install deepspeed\r\n```\r\n\r\nand let's try again:\r\n```\r\nexport BS=20; rm -r output_dir; CUDA_VISIBLE_DEVICES=0 PYTHONPATH=../../src USE_TF=0 deepspeed --num_gpus=1 \\\r\n./finetune_trainer.py --model_name_or_path t5-3b --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro \\\r\n--do_eval --do_predict --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 \\\r\n--logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 \\\r\n--overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate \\\r\n--eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 \\\r\n--val_max_target_length 128 --warmup_steps 5 --n_train 60 --n_val 10 --n_test 10 --deepspeed ds_config_1gpu.json --fp16\r\n```\r\net voila! we get a BS=20 trained just fine. I can probably push BS even further. It OOMed at BS=30.\r\n```\r\n2021-01-12 19:06:31 | INFO | __main__ | train_n_objs = 60\r\n2021-01-12 19:06:31 | INFO | __main__ | train_runtime = 8.8511\r\n2021-01-12 19:06:35 | INFO | __main__ | val_n_objs = 10\r\n2021-01-12 19:06:35 | INFO | __main__ | val_runtime = 3.5329\r\n2021-01-12 19:06:39 | INFO | __main__ | test_n_objs = 10\r\n2021-01-12 19:06:39 | INFO | __main__ | test_runtime = 4.1123\r\n```\r\n\r\nAmazing!\r\n\r\nImportant note - I used `CUDA_VISIBLE_DEVICES=0` to single out one gpu, but deepspeed has a bug now where it ignores that env var, so it'll be using the first GPU instead. microsoft/DeepSpeed#662 But hoping it will get fixed eventually.\r\n\r\nThe config file `ds_config_1gpu.json` is:\r\n```\r\n{\r\n \"fp16\": {\r\n \"enabled\": true,\r\n \"loss_scale\": 0,\r\n \"loss_scale_window\": 1000,\r\n \"hysteresis\": 2,\r\n \"min_loss_scale\": 1\r\n },\r\n\r\n \"zero_optimization\": {\r\n \"stage\": 2,\r\n \"allgather_partitions\": true,\r\n \"allgather_bucket_size\": 2e8,\r\n \"reduce_scatter\": true,\r\n \"reduce_bucket_size\": 2e8,\r\n \"overlap_comm\": true,\r\n \"contiguous_gradients\": true,\r\n \"cpu_offload\": true\r\n },\r\n\r\n \"optimizer\": {\r\n \"type\": \"Adam\",\r\n \"params\": {\r\n \"adam_w_mode\": true,\r\n \"lr\": 3e-5,\r\n \"betas\": [ 0.9, 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\r\n```\r\n\r\nI had to lower the ZeRO buffers from the default 5e8 to 2e8, otherwise it was OOM'ing even on BS=1.\r\n\r\n**important**: DeepSpeed made some changes in the non-released version as of this writing and so the above config won't work anymore. It dropped `adam_w_mode` and added a proper `AdamW` optimizer (it was always there, but just not exposed normally), so replace that section with:\r\n\r\n```\r\n \"optimizer\": {\r\n \"type\": \"AdamW\",\r\n \"params\": {\r\n \"lr\": 3e-5,\r\n \"betas\": [ 0.9, 0.999 ],\r\n \"eps\": 1e-8,\r\n \"weight_decay\": 3e-7\r\n }\r\n },\r\n```\r\n\r\nAnd it's not optimized yet, I just found at least one config that worked for this simple proof-of-concept test.\r\n\r\nGo and check it out!\r\n\r\n**edit:** I was asked about RAM usage for this task, it was 71GB peak, I re-run the same command as above with:\r\n`/usr/bin/time -v ` before `deepspeed` and got:\r\n\r\n```\r\n User time (seconds): 117.12\r\n System time (seconds): 53.46\r\n Percent of CPU this job got: 122%\r\n Elapsed (wall clock) time (h:mm:ss or m:ss): 2:19.38\r\n Average shared text size (kbytes): 0\r\n Average unshared data size (kbytes): 0\r\n Average stack size (kbytes): 0\r\n Average total size (kbytes): 0\r\n Maximum resident set size (kbytes): 70907544\r\n Average resident set size (kbytes): 0\r\n Major (requiring I/O) page faults: 3245\r\n Minor (reclaiming a frame) page faults: 31346864\r\n Voluntary context switches: 16348\r\n Involuntary context switches: 52489\r\n Swaps: 0\r\n File system inputs: 1402864\r\n File system outputs: 11143504\r\n Socket messages sent: 0\r\n Socket messages received: 0\r\n Signals delivered: 0\r\n Page size (bytes): 4096\r\n Exit status: 0\r\n```\r\n\r\nSo the peak RSS entry is 71GB:\r\n```\r\n Maximum resident set size (kbytes): 70907544\r\n```\r\n\r\nThe doc is here: https://huggingface.co/transformers/master/main_classes/trainer.html#deepspeed\r\nAnd it's already slightly outdated - I need to modify it to cover that it works with single GPUs too!\r\n\r\n@alexorona, I think you'd be super-happy about this one.\r\n\r\np.s. if you need to setup the dir and the data, first do:\r\n```\r\ngit clone https://github.com/huggingface/transformers/\r\ncd transformers/\r\ncd examples/seq2seq\r\nwget https://cdn-datasets.huggingface.co/translation/wmt_en_ro.tar.gz\r\ntar -xzvf wmt_en_ro.tar.gz\r\n```\r\nbefore running any of the above scripts.\r\n\r\nOh, and I'm on pytorch-nightly since that's the only version that works at the moment with rtx-3090.", "**edit: 2021-02-15: Note that `finetune_trainer.py` was moved to `examples/legacy/seq2seq/`, and there is a new script `run_seq2seq.py` that took over `finetune_trainer.py`, you will find the transition notes [here](https://github.com/huggingface/transformers/issues/10036)**\r\n\r\nThe simplest way to quickly reproduce the following is to switch to the transformers sha of the time this was posted, that is:\r\n\r\n```\r\ngit clone https://github.com/huggingface/transformers\r\ncd transformers\r\ngit checkout 7e662e6a3be0ece4 \r\n```\r\n\r\n--------------\r\n\r\nOK and to finish the day here are some benchmarks - thank you @sgugger for letting me run those on your machine with dual titan rtx.\r\n\r\nLet's start with the results table:\r\n\r\n\r\n| Method | max BS | train time | eval time |\r\n|---------------------------|--------|------------|-----------|\r\n| baseline | 16 | 30.9458 | 56.3310 |\r\n| fp16 | 20 | 21.4943 | 53.4675 |\r\n| sharded_ddp | 30 | 25.9085 | 47.5589 |\r\n| sharded_ddp+fp16 | 30 | 17.3838 | 45.6593 |\r\n| deepspeed w/o cpu offload | 40 | **10.4007** | 34.9289 |\r\n| deepspeed w/ cpu offload | **50** | 20.9706 | **32.1409** |\r\n\r\nBaseline + data setup was:\r\n```\r\ngit clone https://github.com/huggingface/transformers/\r\ncd transformers/\r\ncd examples/seq2seq\r\nwget https://cdn-datasets.huggingface.co/translation/wmt_en_ro.tar.gz\r\ntar -xzvf wmt_en_ro.tar.gz\r\nexport BS=16; rm -r output_dir; PYTHONPATH=../../src USE_TF=0 python -m torch.distributed.launch \\\r\n--nproc_per_node=2 ./finetune_trainer.py --model_name_or_path t5-large --output_dir output_dir \\\r\n--adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_train --evaluation_strategy=steps --freeze_embeds \\\r\n--label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 \\\r\n--max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS \\\r\n--per_device_train_batch_size $BS --predict_with_generate --eval_steps 25000 --sortish_sampler \\\r\n--task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 500 \\\r\n--n_train 2000 --n_val 500\r\n```\r\n\r\nNotes:\r\n\r\n- We are doing a small train=2000, eval=500 items to do the comparisons. Eval does by default beam search size=4, so it's slower than training with the same number of samples, that's why I used 4x less eval items\r\n- task: translation\r\n- model: t5-large\r\n- We have 2x 24GB GPUs\r\n- DeepSpeed wasn't really designed for evaluation according to its developers but you can see it rocks there too.\r\n\r\nResults: Well, Deepspeed beats all solutions that were compared - it's much faster and can fit much bigger batches into the given hardware. as you can see from the previous post https://github.com/huggingface/transformers/issues/8771#issuecomment-759176685 - the cpu offloading while is slower on training it can fit more into your hardware. and it's the winner for eval!\r\n\r\nNote: these benchmarks aren't perfect as they take a lot of time to handle you can see that BS numbers are pretty rounded - surely they can be somewhat bigger and speed somewhat better as a result, so I'm sure both sharded ddp and deepspeed can be optimized further.\r\n\r\nBut that's a good start. As both sharded ddp and deepspeed are now in master https://huggingface.co/transformers/master/main_classes/trainer.html#trainer-integrations please go ahead and do your own benchmarks.\r\n\r\nAnd now the raw results - sorry it's not markdown'ed:\r\n\r\n```\r\n\r\n# setup\r\n\r\nconda install -y pytorch==1.7.1 torchvision cudatoolkit=10.2 -c pytorch\r\npip install deepspeed fairscale\r\n\r\n# versions\r\n\r\nPyTorch version: 1.7.1\r\nIs debug build: False\r\nCUDA used to build PyTorch: 10.2\r\nROCM used to build PyTorch: N/A\r\n\r\nOS: Ubuntu 20.04.1 LTS (x86_64)\r\nGCC version: (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0\r\nClang version: 10.0.0-4ubuntu1\r\nCMake version: version 3.16.3\r\n\r\nPython version: 3.8 (64-bit runtime)\r\nIs CUDA available: True\r\nCUDA runtime version: 10.0.130\r\nGPU models and configuration:\r\nGPU 0: TITAN RTX\r\nGPU 1: TITAN RTX\r\n\r\nNvidia driver version: 450.102.04\r\ncuDNN version: Probably one of the following:\r\n/usr/local/cuda-10.2/targets/x86_64-linux/lib/libcudnn.so.7.6.5\r\n\r\ntransformers_version\": \"4.2.0dev0\", (master)\r\n\r\n# baseline\r\n\r\n\r\nmax that I could fit was BS=16\r\n\r\nexport BS=16; rm -r output_dir; PYTHONPATH=../../src USE_TF=0 python -m torch.distributed.launch --nproc_per_node=2 ./finetune_trainer.py --model_name_or_path t5-large --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 500 --n_train 2000 --n_val 500\r\n\r\n\r\n01/13/2021 05:31:19 - INFO - __main__ - train_runtime = 30.9458\r\n01/13/2021 05:32:15 - INFO - __main__ - val_bleu = 25.8269\r\n01/13/2021 05:32:15 - INFO - __main__ - val_runtime = 56.331\r\n\r\n# w/ --fp16\r\n\r\ncould fit BS=20\r\n\r\nexport BS=20; rm -r output_dir; PYTHONPATH=../../src USE_TF=0 python -m torch.distributed.launch --nproc_per_node=2 ./finetune_trainer.py --model_name_or_path t5-large --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 500 --n_train 2000 --n_val 500 --fp16\r\n\r\n01/13/2021 05:33:49 - INFO - __main__ - train_runtime = 21.4943\r\n01/13/2021 05:34:42 - INFO - __main__ - val_bleu = 25.7895\r\n01/13/2021 05:34:42 - INFO - __main__ - val_runtime = 53.4675\r\n\r\n\r\n------------------------------------------------\r\n\r\n# w/ --sharded_ddp\r\n\r\nto compare with BS=20\r\n\r\nexport BS=20; rm -r output_dir; PYTHONPATH=../../src USE_TF=0 python -m torch.distributed.launch --nproc_per_node=2 ./finetune_trainer.py --model_name_or_path t5-large --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 500 --n_train 2000 --n_val 500 --sharded_ddp\r\n\r\n\r\n01/13/2021 06:26:11 - INFO - __main__ - train_runtime = 28.9404\r\n01/13/2021 05:36:16 - INFO - __main__ - val_bleu = 25.7201\r\n01/13/2021 05:36:16 - INFO - __main__ - val_runtime = 55.0909\r\n\r\nbut can fit more now, so same with BS=30\r\n\r\nexport BS=30; rm -r output_dir; PYTHONPATH=../../src USE_TF=0 python -m torch.distributed.launch --nproc_per_node=2 ./finetune_trainer.py --model_name_or_path t5-large --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 500 --n_train 2000 --n_val 500 --sharded_ddp\r\n\r\n01/13/2021 06:28:02 - INFO - __main__ - train_runtime = 25.9085\r\n01/13/2021 05:39:08 - INFO - __main__ - val_bleu = 25.7178\r\n01/13/2021 05:39:08 - INFO - __main__ - val_runtime = 47.5589\r\n\r\n\r\n# w/ --sharded_ddp --fp16\r\n\r\nexport BS=20; rm -r output_dir; PYTHONPATH=../../src USE_TF=0 python -m torch.distributed.launch --nproc_per_node=2 ./finetune_trainer.py --model_name_or_path t5-large --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 500 --n_train 2000 --n_val 500 --sharded_ddp --fp16\r\n\r\n01/13/2021 06:29:08 - INFO - __main__ - train_runtime = 21.4775\r\n01/13/2021 05:41:39 - INFO - __main__ - val_bleu = 25.7162\r\n01/13/2021 05:41:39 - INFO - __main__ - val_runtime = 53.2397\r\n\r\nbut can fit more now, so same with BS=30\r\n\r\n01/13/2021 06:30:03 - INFO - __main__ - train_runtime = 17.3838\r\n01/13/2021 05:43:56 - INFO - __main__ - val_bleu = 25.7314\r\n01/13/2021 05:43:56 - INFO - __main__ - val_runtime = 45.6593\r\n\r\n# w/ --deepspeed ds_config.json (stage 2 w/o cpu offloading)\r\n\r\nI changed the config file to:\r\n\r\n \"cpu_offload\": false\r\n\r\nexport BS=40; rm -r output_dir; PYTHONPATH=../../src USE_TF=0 deepspeed ./finetune_trainer.py --model_name_or_path t5-large --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 500 --n_train 2000 --n_val 500 --deepspeed ds_config.json\r\n\r\n01/13/2021 06:32:35 - INFO - __main__ - train_runtime = 10.4007\r\n01/13/2021 06:33:10 - INFO - __main__ - val_bleu = 25.9687\r\n01/13/2021 06:33:10 - INFO - __main__ - val_runtime = 34.9289\r\n\r\n\r\n# w/ --deepspeed ds_config.json (stage 2 w/ cpu offloading)\r\n\r\nif we lower the buffers to `1.5e8` and enable cpu offloading:\r\n\r\n \"allgather_bucket_size\": 1.5e8,\r\n \"reduce_bucket_size\": 1.5e8,\r\n \"cpu_offload\": true\r\n\r\nwe can get to BS=50!\r\n\r\nBS=50 rm -r output_dir; PYTHONPATH=../../src USE_TF=0 deepspeed ./finetune_trainer.py --model_name_or_path t5-large --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 500 --n_train 2000 --n_val 500 --deepspeed ds_config.json\r\n\r\n01/13/2021 06:40:51 - INFO - __main__ - train_runtime = 20.9706\r\n01/13/2021 06:41:23 - INFO - __main__ - val_bleu = 25.9244\r\n01/13/2021 06:41:23 - INFO - __main__ - val_runtime = 32.1409\r\n\r\nI'm pretty sure if the buffers are even smaller it could do even higher BS. But it's late and I'm going to sleep.\r\n\r\n```\r\n\r\nHere is the config file that was used for deepspeed: https://github.com/huggingface/transformers/blob/69ed36063a732c37fdf72c605c65ebb5b2e85f44/examples/seq2seq/ds_config.json\r\n", "Whoah! ZeRO stage 1: sharded optimizer has been just merged into pytorch! https://github.com/pytorch/pytorch/pull/46750\r\nWith complements of @blefaudeux and the FairScale and DeepSpeed teams!\r\n\r\nPipeline too: https://github.com/pytorch/pytorch/tree/master/torch/distributed/pipeline\r\n\r\nAnd more coming later: https://github.com/pytorch/pytorch/issues/42849\r\n", "> Whoah! ZeRO stage 1: sharded optimizer has been just merged into pytorch! [pytorch/pytorch#46750](https://github.com/pytorch/pytorch/pull/46750)\r\n> With complements of @blefaudeux and the FairScale and DeepSpeed teams!\r\n> \r\n> Pipeline too: https://github.com/pytorch/pytorch/tree/master/torch/distributed/pipeline\r\n> \r\n> And more coming later: [pytorch/pytorch#42849](https://github.com/pytorch/pytorch/issues/42849)\r\n\r\nthanks ! the whole fairscale suite will take a little more time, so it's good that HF is integrated already, the work will not be lost. Great [blog post](https://github.com/huggingface/blog/pull/71) also, and thanks for the numbers ! Some improvements planned over time speed wise within fairscale/shardedddp which should trickle down automatically, thinking for instance about the experimental optimizers in pytorch which flatten the params or better bucketing for the reduce part", "These are great news, @blefaudeux! Thank you for sharing.\r\n\r\nI hope you create a page on github with such news, so it'd be easy to keep abreast of the speed improvements and to appraise users of the need to update to this or that version if they want certain improvements/speed ups. \r\n\r\nIf it's not too much trouble that is.\r\n\r\np.s. my fantasy is that there will be a ZeRO Central, where updates from the all collaborating ZeRO implementations get posted.\r\n\r\ne.g. DeepSpeed just released a new paper: https://arxiv.org/abs/1910.02054 - this would have been a great candidate for such sharing.", "This is very impressive work!\r\n\r\nFrom the perspective of an end-user doing seq2seq (e.g. T5), running the above examples for T5-11B (both sharded_ddp and deepspeed) doesn't appear to be performing complete model parallelism (or, at least, I am getting OOM errors on a machine with four A100-SXM4-40GBs, Python 3.7, pull of HF ~4.3.0 master from yesterday, CUDA 11.0, Pytorch 1.7.1, DeepSpeed compiled from source with the A100 8.0 arch enabled for the A100, BS=1). I understand from the blog post this is likely because sharding is only currently implemented for the optimizer and gradients, but not the model parameters? Is there an interim suggestion for easily running these large models in 4.3? It looks like there's currently confusion since --model_parallel was removed in 4.2 (and some confusion about how to run large models using the /examples/ now, e.g. #9243 ) ", "> This is very impressive work!\r\n\r\nTotally agree. Those both teams and the inventors of ZeRO are awesome!\r\n \r\n> From the perspective of an end-user doing seq2seq (e.g. T5), running the above examples for T5-11B (both sharded_ddp and deepspeed)\r\n\r\none of them - not both. Will send a PR to block such attempts. https://github.com/huggingface/transformers/pull/9712/\r\n\r\nDeepSpeed already does sharded ddp. Slowly, slowly we will get a better understanding and better documentation.\r\n\r\n> doesn't appear to be performing complete model parallelism (or, at least, I am getting OOM errors on a machine with four A100-SXM4-40GBs, Python 3.7, pull of HF ~4.3.0 master from yesterday, CUDA 11.0, Pytorch 1.7.1, DeepSpeed compiled from source with the A100 8.0 arch enabled for the A100, BS=1). I understand from the blog post this is likely because sharding is only currently implemented for the optimizer and gradients, but not the model parameters?\r\n\r\nThat's correct. Not yet.\r\n\r\n* With fairscale you get sharding or optim/grads.\r\n* With deepspeed you get all that, plus cpu-offload, plus better memory management.\r\n\r\nWe would need to have Pipeline parallelism working to support 2D parallelism, which probably should fit t5-11b onto 4 gpus. I'm working on this at the moment, but run into multiple limitations of the PP implementations https://github.com/pytorch/pytorch/pull/50693 and https://github.com/microsoft/DeepSpeed/pull/659.\r\n\r\nIn any case please update your master as I merged a bug fix some 6 hours ago, but I don't think it'd make any difference to your situation.\r\n\r\n> Is there an interim suggestion for easily running these large models in 4.3? It looks like there's currently confusion since --model_parallel was removed in 4.2 (and some confusion about how to run large models using the /examples/ now, e.g. #9243 )\r\n\r\nThe `--model_parallel` flag was half-baked so it was removed until the day we actually have something solid in place. but you can still use model parallelism.\r\n\r\nWhat you can do now is to activate our naive model parallelism, which I think may just fit the 45GB model over 4x 40GB GPUs. See: https://huggingface.co/transformers/model_doc/t5.html?highlight=parallel#transformers.T5EncoderModel.parallelize\r\nWe currently have t5, gpt2 and (unmerged bart pr) with this version of naive MP.\r\n\r\nBut it's going to be slow, see: https://github.com/huggingface/transformers/issues/8771#issuecomment-758250421 because 3 out of 4 gpus will be idling at any given moment. Basically, you will have a speed of a single gpu, with extra slowdown due to data being copied between gpus back and forth. We need to get PP working to overcome this.", "@PeterAJansen As Stas points out, you should use the model parallelism implementation from 4.1.0. You'll likely need somewhere around 256 GB total GPU memory to train t5-11b with max 512 input tokens and 320 GB for 1024 tokens (so p4 instance in AWS). \r\n\r\nIn 4.1.0, there's only a few changes to the code you'd need to do to accomplish this: 1) set `train_args = TrainingArguments(model_parallel = True) `and, 2) after loading the model, call `model.parallelize()` (no arguments needed -- custom device map won't help you with t5-11b). @stas00, can you confirm what the procedure is for >= 4.2.0? I haven't been able to keep up with the changes with the move. ", "@alexorona, as the doc [goes](https://huggingface.co/transformers/model_doc/t5.html?highlight=parallel#transformers.T5EncoderModel.parallelize), in the current master incarnation all you need to do is to call:\r\n```\r\nmodel.parallelize()\r\n```\r\nbefore you do the training. This then sets:\r\n\r\n`self.is_model_parallel` to `True` and the trainer does the same thing it was doing when `--model_parallel` was used. It just does it smarter now and no longer requires an extra flag.\r\n\r\nThe new logic is:\r\n```\r\n if hasattr(model, \"is_parallelizable\") and model.is_parallelizable and model.model_parallel:\r\n self.is_model_parallel = True\r\n else:\r\n self.is_model_parallel = False\r\n```\r\n\r\nThe reason `--model_parallel` was removed is because it exposed that flag to all example scripts, but the scripts like `finetune_trainer.py` weren't synced, so as a user would run `finetune_trainer.py --model_parallel` nothing would happen, that's why just that flag was removed.\r\n\r\nBut nothing else changed from your original implementation API-wise, @alexorona. The PRs I proposed which would change the device map have been parked for now.\r\n\r\nWe may re-add this flag in the future once the scripts will be able to activate MP internally.", "@stas00 The flag was exposed because `TrainingArguments` would automatically increase the batch size if more than one GPU was detected (it would default to model parallelism behavior), thus defeating the purpose of model parallelism. Did you change that behavior?", "As I was trying to convey we found a way to do the exact same thing without needing an extra flag. That's it's enough to run \r\n`model.parallelize()` right after creating the model, for the trainer to do the right thing.\r\n\r\n> TrainingArguments would automatically increase the batch size if more than one GPU was detected \r\n\r\nAs you can see it forces it to appear as having just 1 gpu, so no DP will be activated.\r\n\r\nhttps://github.com/huggingface/transformers/blob/7acfa95afb8194f8f9c1f4d2c6028224dbed35a2/src/transformers/trainer.py#L285-L290\r\n\r\nPlease let me know if we have missed anything in the re-shuffle.\r\n\r\n" ]

[]

[]

[ "Good idea!\r\n\r\nThis could be done for tokenizers as well, no? (the `from_pretrained` for tokenizers is in `tokenization_utils_base.py`)", "I have extended the same modifications to the tokenizers, as suggested by @thomwolf , and to auto classes too." ]

[]

[]

[ "Hi @eladsegal ! \r\n\r\nThis is indeed interesting. The FRCNN config should match the exact setting used in the original demo. I will say, however, everything is finicky, but has been tested + should work.\r\n\r\nCould you let me know if you:\r\n- modified any of the settings in the config\r\n- provide a link to the script script you are using so I can see the exact changes you have made\r\n- If also possible, plot the bounding boxes on the original prediction so that I can see how the bounding boxes are different (I think that may be pretty telling)\r\n\r\nI think I should be able to figure out what may be happening if you could let me know the above!", "Thanks for the response, @eltoto1219!\r\nThe 33% I got previously was a mistake, not sure what change I made that caused it. \r\nI started from scratch now, and the results are a lot closer to the original.\r\nI've made the following changes:\r\n- https://github.com/huggingface/transformers/blob/master/examples/lxmert/utils.py#L552 - do this only if the image is from a URL (https://github.com/huggingface/transformers/issues/8333) (improves accuracy on GQA by ~4 points)\r\n\r\nAnd changes so extracting_data.py will work for batch size larger than 1:\r\n- https://github.com/huggingface/transformers/blob/master/examples/lxmert/extracting_data.py#L95 - changed to if len(batch) > 0\r\n\r\n- https://github.com/huggingface/transformers/blob/master/examples/lxmert/modeling_frcnn.py#L42 - added .view(-1, 1, 1) to allow broadcasting\r\n\r\nFor `extracting_data.py` I made a few more changes to be able to extract only for GQA's testdev images (so it will be a lot faster to run).\r\nYou can reproduce everything with the code in:\r\nhttps://github.com/eladsegal/gqa_lxmert\r\nThere's also a notebook there showing the different bounding boxes.\r\n\r\nFor some reason, inference with different batch sizes in the FRCNN model results in different features for the same image.\r\nI got the following accuracies for GQA for the following different batch sizes:\r\n6: 57.15%\r\n2: 57.64%\r\n1: 58.54%\r\nUsing features from the original LXMERT repo results in 59.29%.", "Of course! I am am real glad you were able to catch the color format error for images downloaded via a url (It would have taken me forever to find something like that)! I will have to fix that as soon as possible + do some more testing. \r\n\r\nI find the different downstream lxmert accuracies on GQA when different batch sizes are used for feature extraction really interesting aswell. In the original repo, the extraction was set up so that one image went through the faster frcnn at a time. I am thinking something may be getting unordered when using multiple images at once, so I will take a look. I think it would also be worth pointing out that there are multiple releases of visual genome images (from 2014 and 2016). If downloading directly from https://visualgenome.org/ there appears to be quite a few corrupt images. \r\n\r\nOn the other hand, all visual genome images can be downloaded from https://cs.stanford.edu/people/dorarad/gqa/about.html (which I am assuming is the latest version). That may in part be why the accuracy is lower. (GQA train + val split come from visual genome. GQA testdev split comes from the COCO test split) There may also be some very small changes to some extraction hyper-parameters (most likely the NMS threshold for the post processing of bounding boxes) which may have also resulted in slightly different inaccuracies. \r\n\r\nI'll go ahead and extract features across the different versions of visual genome + gqa, and compare element-wise with the features from the original lxmert repo and see if anything is any different, and if so, how different. It should take me a couple of days, but I will get back to you by then!", "Thank you, I really appreciate this!\r\n\r\n>I think it would also be worth pointing out that there are multiple releases of visual genome images (from 2014 and 2016). If downloading directly from https://visualgenome.org/ there appears to be quite a few corrupt images.\r\n> \r\n> On the other hand, all visual genome images can be downloaded from https://cs.stanford.edu/people/dorarad/gqa/about.html (which I am assuming is the latest version). That may in part be why the accuracy is lower. (GQA train + val split come from visual genome. GQA testdev split comes from the COCO test split) \r\n\r\nNot sure about this as a possible reason for the lower accuracy, as the images I used were downloaded from GQA's website, and I only did comparison on the testdev split, and no additional training at all.\r\n\r\n> There may also be some very small changes to some extraction hyper-parameters (most likely the NMS threshold for the post processing of bounding boxes) which may have also resulted in slightly different inaccuracies.\r\n\r\nSounds like a very reasonable explanation for the different and features and the small accuracy difference!\r\n\r\n\r\n", "@eltoto1219 The issue with batch-wise extraction is known:\r\nhttps://github.com/airsplay/py-bottom-up-attention/issues/3#issuecomment-624240642", "Hey @eladsegal !\r\n\r\nI stumbled upon that actually on Friday myself too. I rewrote the extraction script to only allow frcnn extraction for one image at a time. \r\n\r\nI think that the discrepancy in accuracy comes from the fact that Hao actually used a caffe-based frcnn pretrained model trained specifically to predict 36 images, while the pytorch one here was pretrained to predict 10-100 images. That and the potential of slightly different NMS thresholds. \r\n\r\nAslong as the batch-size is 1, the feature quality is most likely the same, however, if we are finetuning with the features from the model used in the aforementioned script, those technically wont be the exact same features used to pretrain lxmert, they should still get the job done if your okay with being ~1% lower than the reported accuracy.\r\n\r\n\r\nHere is the link to the fixed script:\r\n\r\nhttps://drive.google.com/file/d/1er2axVyGj8eW84QBGrV0dqTmKbxyS8F7/view?usp=sharing\r\n", "Thank you very much @eltoto1219, this has been extremely helpful!", "Why [`examples/lxmert/`](https://github.com/huggingface/transformers/blob/master/examples/lxmert/) no longer exists?", "This was an error, it has been put back a few days ago. Sorry for the inconvenience." ]

[]

[]

[]

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

[]

[ "Indeed, not being able to pass tokenizer arguments is a limiting factor of the pipelines. We're working on pipelines v2 (cc @mfuntowicz) which will allow such arguments to be passed.\r\n\r\nIn the meantime, we would definitely welcome a PR offering this functionality - but it would have to be agnostic to the argument, not specific to the truncation strategy.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

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

[]

[ "Hey @singhsidhukuldeep, \r\n\r\nCould you maybe upload your weights to a colab so that I can reproduce the error or upload your weights to the model hub and give me a path to it? \r\n\r\nThis way, I can reproduce the error and solve it :-) \r\n\r\nThanks a lot!", "Hi @patrickvonplaten ,\r\n\r\nI tried this\r\n```Python\r\npt_model = TFBertForPreTraining.from_pretrained(model_output_location, from_pt=True)\r\n\r\nprint(\"\\n\\n>>> Saving HuggingFace to tensorflow(pb)\")\r\ntf.saved_model.save(pt_model,TF_model_output_location)\r\n```\r\n\r\nand it worked, but I am not able to understand the limitations here!", "If you want to convert from <del>TF to PT</del> PT to TF this is exactly how you should to it...", "@patrickvonplaten I am looking to convert PyTorch to TF!", "Sorry I meant PT to TF -> your approach is correct here.", "Got it! Thanks for the help. \r\n\r\nOne last thing, this gives a `*.h5` file (weights only)\r\nIs there a way to get `*.pb` file with structure and weights?", "Hi @singhsidhukuldeep, I tried this, it did save the model in assets, variables, and saved_model.pb format, but couldn't get any .h5 file that I need, am I missing something?\r\n\r\nP.S. I am trying to convert a standard config BERT MaskedLM model\r\n\r\n> Hi @patrickvonplaten ,\r\n> \r\n> I tried this\r\n> \r\n> ```python\r\n> pt_model = TFBertForPreTraining.from_pretrained(model_output_location, from_pt=True)\r\n> \r\n> print(\"\\n\\n>>> Saving HuggingFace to tensorflow(pb)\")\r\n> tf.saved_model.save(pt_model,TF_model_output_location)\r\n> ```\r\n> \r\n> and it worked, but I am not able to understand the limitations here!\r\n\r\n", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "The errors are due to connection errors." ]

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

[]

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

[]

[]

[ "Will update the switch to fast tokenizers when a decision has been made @thomwolf @n1t0 ", "@sgugger please let me know if this is what you had in mind." ]

[]

[]

[ "I also tried to use automodel for this, I needed to modify the T5Config, I called it like this with automodel. thank you for your help. I need to make this work with not pretrained model. \r\n\r\n` model = AutoModel.from_config(config)\r\n`\r\nI am getting this error:\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"finetune_t5_trainer.py\", line 236, in <module>\r\n main()\r\n File \"finetune_t5_trainer.py\", line 89, in main\r\n model = AutoModel.from_config(config)\r\n File \"/idiap/user/rkarimi/libs/anaconda3/envs/internship/lib/python3.7/site-packages/transformers/modeling_auto.py\", line 604, in from_config\r\n config.__class__, cls.__name__, \", \".join(c.__name__ for c in MODEL_MAPPING.keys())\r\nValueError: Unrecognized configuration class <class 'seq2seq.models.t5.configuration_t5.T5Config'> for this kind of AutoModel: AutoModel.\r\nModel type should be one of RetriBertConfig, T5Config, DistilBertConfig, AlbertConfig, CamembertConfig, XLMRobertaConfig, BartConfig, LongformerConfig, RobertaConfig, LayoutLMConfig, SqueezeBertConfig, BertConfig, OpenAIGPTConfig, GPT2Config, MobileBertConfig, TransfoXLConfig, XLNetConfig, FlaubertConfig, FSMTConfig, XLMConfig, CTRLConfig, ElectraConfig, ReformerConfig, FunnelConfig, LxmertConfig, BertGenerationConfig, DebertaConfig, DPRConfig, XLMProphetNetConfig, ProphetNetConfig.\r\n\r\n\r\n\r\n```", "solved with model = T5ForConditionalGeneration(config=config) syntax has been changed from 3.5.0 to 3.5.1 thanks " ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[]