Datasets:

DanielPFlorian

/

Transformers-Github-Issues

like 0

{ "login": "fancyerii", "id": 5372812, "node_id": "MDQ6VXNlcjUzNzI4MTI=", "avatar_url": "https://avatars.githubusercontent.com/u/5372812?v=4", "gravatar_id": "", "url": "https://api.github.com/users/fancyerii", "html_url": "https://github.com/fancyerii", "followers_url": "https://api.github.com/users/fancyerii/followers", "following_url": "https://api.github.com/users/fancyerii/following{/other_user}", "gists_url": "https://api.github.com/users/fancyerii/gists{/gist_id}", "starred_url": "https://api.github.com/users/fancyerii/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/fancyerii/subscriptions", "organizations_url": "https://api.github.com/users/fancyerii/orgs", "repos_url": "https://api.github.com/users/fancyerii/repos", "events_url": "https://api.github.com/users/fancyerii/events{/privacy}", "received_events_url": "https://api.github.com/users/fancyerii/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @fancyerii, thanks for raising an issue! \r\n\r\nThis is a question best placed in our [forums](https://discuss.huggingface.co/). We try to reserve the github issues for feature requests and bug reports.", "@amyeroberts thanks. But when I post this to forum, it says \"similar content already exist\" and do not let me create new post here.", "@fancyerii Were you able to find related issues on the forum? Alternatively you can ask on [our discord](https://discord.com/invite/hugging-face-879548962464493619). ", "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/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28055/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "kwen2501", "id": 6676466, "node_id": "MDQ6VXNlcjY2NzY0NjY=", "avatar_url": "https://avatars.githubusercontent.com/u/6676466?v=4", "gravatar_id": "", "url": "https://api.github.com/users/kwen2501", "html_url": "https://github.com/kwen2501", "followers_url": "https://api.github.com/users/kwen2501/followers", "following_url": "https://api.github.com/users/kwen2501/following{/other_user}", "gists_url": "https://api.github.com/users/kwen2501/gists{/gist_id}", "starred_url": "https://api.github.com/users/kwen2501/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/kwen2501/subscriptions", "organizations_url": "https://api.github.com/users/kwen2501/orgs", "repos_url": "https://api.github.com/users/kwen2501/repos", "events_url": "https://api.github.com/users/kwen2501/events{/privacy}", "received_events_url": "https://api.github.com/users/kwen2501/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28054). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28054/reactions", "total_count": 2, "+1": 2, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28054", "html_url": "https://github.com/huggingface/transformers/pull/28054", "diff_url": "https://github.com/huggingface/transformers/pull/28054.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28054.patch", "merged_at": 1702651532000 }

{ "login": "dakinggg", "id": 43149077, "node_id": "MDQ6VXNlcjQzMTQ5MDc3", "avatar_url": "https://avatars.githubusercontent.com/u/43149077?v=4", "gravatar_id": "", "url": "https://api.github.com/users/dakinggg", "html_url": "https://github.com/dakinggg", "followers_url": "https://api.github.com/users/dakinggg/followers", "following_url": "https://api.github.com/users/dakinggg/following{/other_user}", "gists_url": "https://api.github.com/users/dakinggg/gists{/gist_id}", "starred_url": "https://api.github.com/users/dakinggg/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/dakinggg/subscriptions", "organizations_url": "https://api.github.com/users/dakinggg/orgs", "repos_url": "https://api.github.com/users/dakinggg/repos", "events_url": "https://api.github.com/users/dakinggg/events{/privacy}", "received_events_url": "https://api.github.com/users/dakinggg/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Also I thought `torch_dtype` wasn't even used when its in the config, maybe that has changed though.", "Also even specifying `torch_dtype` results in a warnings `In [21]: transformers.AutoModelForCausalLM.from_config(llamacfg, torch_dtype=torch.bfloat16, use_flash_attention_2=True)\r\nYou are attempting to use Flash Attention 2.0 without specifying a torch dtype. This might lead to unexpected behaviour`. I feel like I'm missing something fundamental about how this arg works.", "I also tried setting `config._attn_implemention` myself as a workaround, but it doesn't seem to respect that.", "Thanks a lot for this and other reports, we’ll fix it asap cc @fxmarty as well", "While you're there, it'd be good to revisit all of the hard checks in there and make sure they are truly necessary. For anyone else who ends up here, you can work around it via\r\n```\r\n def _autoset_attn_implementation_monkeypatch(\r\n cls, config, *args, **kwargs): # type: ignore\r\n config._attn_implementation = requested_attention_implementation\r\n return config\r\n\r\n PreTrainedModel._autoset_attn_implementation = classmethod(\r\n _autoset_attn_implementation_monkeypatch)\r\n```", "same issue", "in src/modeling_utils.py\r\n\r\n```python\r\n config = self._autoset_attn_implementation(\r\n config, torch_dtype=torch.get_default_dtype(), check_device_map=False\r\n )\r\n```\r\nchange to \r\n```\r\n config = self._autoset_attn_implementation(\r\n config, check_device_map=False\r\n )\r\n```", "~Same issue~\r\n\r\nEdit: I was actually loading models (w/ custom model classes) from scratch using the \\_\\_init\\_\\_ function but it looks like you should use `_from_config` instead, where you can specify torch_dtype.\r\n\r\n### Old code\r\n```\r\nconfig = AutoConfig.from_pretrained(model_name_or_path,\r\n attn_implementation=\"flash_attention_2\",\r\n torch_dtype=torch.bfloat16) # here, torch_dtype is ignored\r\nmodel = CustomPretrainedLM(config)\r\n```\r\n\r\n### New code\r\n```\r\nconfig = AutoConfig.from_pretrained(model_name_or_path, attn_implementation=\"flash_attention_2\")\r\nmodel = CustomPretrainedLM(config, torch_dtype=torch.bfloat16)\r\n```", "Same issue, but fig out a simple solution to bypass the check:\r\n\r\n```\r\ndefault_dtype = torch.get_default_dtype()\r\ntorch.set_default_dtype(torch.bfloat16)\r\nmodel = transformers.AutoModelForCausalLM.from_pretrained(\r\n model_path,\r\n attn_implementation=\"flash_attention_2\",\r\n config=config,\r\n)\r\ntorch.set_default_dtype(default_dtype)\r\n```", "```\r\n\r\nmodel = transformers.AutoModelForCausalLM.from_pretrained(\r\n model_path,\r\n attn_implementation=\"flash_attention_2\",\r\n config=config,\r\n torch_dtype=torch.bfloat16\r\n)\r\n```\r\n" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28052/reactions", "total_count": 2, "+1": 2, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "aismlv", "id": 13088690, "node_id": "MDQ6VXNlcjEzMDg4Njkw", "avatar_url": "https://avatars.githubusercontent.com/u/13088690?v=4", "gravatar_id": "", "url": "https://api.github.com/users/aismlv", "html_url": "https://github.com/aismlv", "followers_url": "https://api.github.com/users/aismlv/followers", "following_url": "https://api.github.com/users/aismlv/following{/other_user}", "gists_url": "https://api.github.com/users/aismlv/gists{/gist_id}", "starred_url": "https://api.github.com/users/aismlv/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/aismlv/subscriptions", "organizations_url": "https://api.github.com/users/aismlv/orgs", "repos_url": "https://api.github.com/users/aismlv/repos", "events_url": "https://api.github.com/users/aismlv/events{/privacy}", "received_events_url": "https://api.github.com/users/aismlv/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Per my understanding this would only be applied for models that went through the recent cache refactoring. Llava is impacted since it uses Llama and Mistral under the hood, which are part of this PR: https://github.com/huggingface/transformers/pull/26681 \r\nThe recent Mixtral has it: https://github.com/huggingface/transformers/blob/main/src/transformers/models/mixtral/modeling_mixtral.py#L863 so I would say to just keep this in mind for future models that copy from Llama/Mistral and/or uses Llama or Mistral under the hood such as Llava", "@younesbelkada Great - thanks for explaining and providing the relevant links. We're good to merge then! " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28051/reactions", "total_count": 2, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 1, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28051", "html_url": "https://github.com/huggingface/transformers/pull/28051", "diff_url": "https://github.com/huggingface/transformers/pull/28051.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28051.patch", "merged_at": 1702649120000 }

{ "login": "younesbelkada", "id": 49240599, "node_id": "MDQ6VXNlcjQ5MjQwNTk5", "avatar_url": "https://avatars.githubusercontent.com/u/49240599?v=4", "gravatar_id": "", "url": "https://api.github.com/users/younesbelkada", "html_url": "https://github.com/younesbelkada", "followers_url": "https://api.github.com/users/younesbelkada/followers", "following_url": "https://api.github.com/users/younesbelkada/following{/other_user}", "gists_url": "https://api.github.com/users/younesbelkada/gists{/gist_id}", "starred_url": "https://api.github.com/users/younesbelkada/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/younesbelkada/subscriptions", "organizations_url": "https://api.github.com/users/younesbelkada/orgs", "repos_url": "https://api.github.com/users/younesbelkada/repos", "events_url": "https://api.github.com/users/younesbelkada/events{/privacy}", "received_events_url": "https://api.github.com/users/younesbelkada/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey when can we expect this to merge ?", "Hi @rishabh063 \r\nThanks very much for your interest in this PR - hopefully quite soon, I need some time to make sure logits match then we can merge ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored.", "hey @younesbelkada any updates on this ?", "hi @rishabh063 unfortunately I had to work on more urgent things now, I will come back to this ASAP. If you have bandwith you can also take over the PR ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28050/reactions", "total_count": 4, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 4, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28050", "html_url": "https://github.com/huggingface/transformers/pull/28050", "diff_url": "https://github.com/huggingface/transformers/pull/28050.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28050.patch", "merged_at": null }

{ "login": "arnavgarg1", "id": 106701836, "node_id": "U_kgDOBlwkDA", "avatar_url": "https://avatars.githubusercontent.com/u/106701836?v=4", "gravatar_id": "", "url": "https://api.github.com/users/arnavgarg1", "html_url": "https://github.com/arnavgarg1", "followers_url": "https://api.github.com/users/arnavgarg1/followers", "following_url": "https://api.github.com/users/arnavgarg1/following{/other_user}", "gists_url": "https://api.github.com/users/arnavgarg1/gists{/gist_id}", "starred_url": "https://api.github.com/users/arnavgarg1/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/arnavgarg1/subscriptions", "organizations_url": "https://api.github.com/users/arnavgarg1/orgs", "repos_url": "https://api.github.com/users/arnavgarg1/repos", "events_url": "https://api.github.com/users/arnavgarg1/events{/privacy}", "received_events_url": "https://api.github.com/users/arnavgarg1/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @arnavgarg1 \r\nphi-1 from microsoft still uses code on the Hub feature: https://huggingface.co./microsoft/phi-1_5 \r\nIf you want to use the HF version of Phi-1 you need to use the converted checkpoints from @susnato such as : https://huggingface.co./susnato/phi-1_5_dev - I think we should transfer those weights under microsoft org with the suffix `-hf`\r\n\r\n I did not reviewed the Phi integration PR but if the keys are the same we can also open PRs on the Hub on the original repos cc @ArthurZucker ", "Hi, I have already opened PRs on the Hub for transferring weights and necessary files, [here](https://huggingface.co./microsoft/phi-1_5/discussions/62) and [here](https://huggingface.co./microsoft/phi-1/discussions/5). We need to wait for someone from the org to merge them then you can use without passing `trust_remote_code=True`.", "Thanks for the prompt response @younesbelkada and @susnato! That makes sense! Will keep on the lookout for when the weights and necessary files PR on the hub gets merged in - for now, this looks like it works nicely!\r\n\r\nThanks for the awesome work in adding support for Phi @susnato! \r\n\r\nOut of curiosity - are there plans to add support for Phi-2 as well? ", "I haven't checked phi-2 yet but If the architecture is the same as phi-1/phi-1.5, then we need to change the keys in the weights and update the config file and we will be good to go. ", "Hey @susnato, based on inspection, they seem architecturally similar, just that phi-2 is slightly bigger than phi 1.5. See this diff I created: https://www.diffchecker.com/cAspnGZ3/ \r\n\r\nI assume that we just need to update the keys and config and we should be good to go? Is this something I could potentially help with?", "We’ll be adding phi2 as well yes 😊 we asked the author if he is interested but given the community’s interest it’s a good way to go anyways ", "Great to hear @ArthurZucker! Has work already started/is there an approximate timeline for when you expect it to get added?", "ETA is probably end of next week. Gotta finish DECI, add mamba and then Phi2! Should be quite fast 🤗 ", "Hi @ArthurZucker, can I please help in adding `phi-2` in any way?\r\n\r\nAs far as I can tell, the architecture is the same as the `phi` that we have in the library(but slightly bigger). Just need to convert the weights using the existing phi script and transfer the weights. Also maybe add an integration test to make sure the logits are the same. \r\n\r\n\r\nLet me know if I could be of any help.", "Sure, as you added phi, feel free to take this! 🤗 \r\n", "Thanks for taking this @susnato! ", "Hi @ArthurZucker, thanks.\r\n\r\nJust wanted to be clear on the License part - As you can see in this [discussion](https://huggingface.co./microsoft/phi-2/discussions/4) there seems to be a problem regarding the license...Is it okay to modify the weights and then transfer them to my profile for this addition? Later once we finish transferring the code we can update the checkpoints to the official one.", "Yeah sure, the idea will be to open a PR once everything is done, making sure we don't have issues with the model type and ask the author to merge. There is for me no issue with this regarding the licence, we don't modify the weight, we modify the dictionnary that stores them / the split !", "Thanks for the clarification @ArthurZucker! I will ping you for a review when it's done.", "Hey @susnato! I saw that you created https://huggingface.co./susnato/phi-2 🥳 . Has this been verified/is it safe to use?", "Hi @arnavgarg1, the weights are messed up...they changed the modeling file on the Hub so the existing conversion script is not working properly...I will try to fix it and get it running tomorrow.", "@susnato No worries at all, let me know if there's anything I can do to help!", "@susnato do you have a draft PR already? I think this would help us follow your progress and potentially help you! 🤗 ", "Hi @arnavgarg1, I fixed some things and now `phi2` should work as expected, could you please run it from `susnato/phi-2` and let me know if it is showing expected results or not. ", "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/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28049/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "amyeroberts", "id": 22614925, "node_id": "MDQ6VXNlcjIyNjE0OTI1", "avatar_url": "https://avatars.githubusercontent.com/u/22614925?v=4", "gravatar_id": "", "url": "https://api.github.com/users/amyeroberts", "html_url": "https://github.com/amyeroberts", "followers_url": "https://api.github.com/users/amyeroberts/followers", "following_url": "https://api.github.com/users/amyeroberts/following{/other_user}", "gists_url": "https://api.github.com/users/amyeroberts/gists{/gist_id}", "starred_url": "https://api.github.com/users/amyeroberts/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/amyeroberts/subscriptions", "organizations_url": "https://api.github.com/users/amyeroberts/orgs", "repos_url": "https://api.github.com/users/amyeroberts/repos", "events_url": "https://api.github.com/users/amyeroberts/events{/privacy}", "received_events_url": "https://api.github.com/users/amyeroberts/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28048). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28048/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28048", "html_url": "https://github.com/huggingface/transformers/pull/28048", "diff_url": "https://github.com/huggingface/transformers/pull/28048.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28048.patch", "merged_at": 1702583420000 }

{ "login": "BramVanroy", "id": 2779410, "node_id": "MDQ6VXNlcjI3Nzk0MTA=", "avatar_url": "https://avatars.githubusercontent.com/u/2779410?v=4", "gravatar_id": "", "url": "https://api.github.com/users/BramVanroy", "html_url": "https://github.com/BramVanroy", "followers_url": "https://api.github.com/users/BramVanroy/followers", "following_url": "https://api.github.com/users/BramVanroy/following{/other_user}", "gists_url": "https://api.github.com/users/BramVanroy/gists{/gist_id}", "starred_url": "https://api.github.com/users/BramVanroy/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/BramVanroy/subscriptions", "organizations_url": "https://api.github.com/users/BramVanroy/orgs", "repos_url": "https://api.github.com/users/BramVanroy/repos", "events_url": "https://api.github.com/users/BramVanroy/events{/privacy}", "received_events_url": "https://api.github.com/users/BramVanroy/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Somewhat related to this, the conversational pipeline's typing and docstrings do not seem correct (which brought me to the issue above):\r\n\r\nhttps://github.com/huggingface/transformers/blob/c48787f347bd604f656c2cfff730e029c8f8c1fe/src/transformers/pipelines/conversational.py#L262-L280\r\n\r\nThe signature allows for a list of dicts (as a single conversation) but not a list of list of dicts (a batch of conversations), although `List[Conversation]` is allowed. According to the docstrings, a `List[dict]` is also not allowed - only Conversation(s). Finally, for compatibility with the pipeline call, other types of input (such as a generator or KeyDataset) should also be allowed but they are not specified.\r\n", "cc @Narsil ", "cc @Rocketknight1 for the conversational pipeline docstring part " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28047/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "Rocketknight1", "id": 12866554, "node_id": "MDQ6VXNlcjEyODY2NTU0", "avatar_url": "https://avatars.githubusercontent.com/u/12866554?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Rocketknight1", "html_url": "https://github.com/Rocketknight1", "followers_url": "https://api.github.com/users/Rocketknight1/followers", "following_url": "https://api.github.com/users/Rocketknight1/following{/other_user}", "gists_url": "https://api.github.com/users/Rocketknight1/gists{/gist_id}", "starred_url": "https://api.github.com/users/Rocketknight1/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Rocketknight1/subscriptions", "organizations_url": "https://api.github.com/users/Rocketknight1/orgs", "repos_url": "https://api.github.com/users/Rocketknight1/repos", "events_url": "https://api.github.com/users/Rocketknight1/events{/privacy}", "received_events_url": "https://api.github.com/users/Rocketknight1/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28046). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "@amyeroberts It's hard to say - it would depend exactly on how they were loading the weights. What has changed with this PR is that if you initialize a model via `from_config` **and then explicitly call its build() method**, you'll get slightly different weight names for it than you would have before. For every other case (e.g. initialize with `from_pretrained()` or implicitly call `build()` via calling the model), there should be no difference.\r\n\r\nI think for users to notice an issue they would need to:\r\n1) Load a model with `from_config`\r\n2) Explicitly build it with `build()`\r\n3) Have their own weight-loading script that they use to load weights by name into their model after 2)\r\n\r\nI think it's pretty unlikely that anyone is doing that, and even if so, they can fix it by using `build_in_name_scope()` instead." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28046/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28046", "html_url": "https://github.com/huggingface/transformers/pull/28046", "diff_url": "https://github.com/huggingface/transformers/pull/28046.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28046.patch", "merged_at": 1702575745000 }

{ "login": "wuxb45", "id": 564235, "node_id": "MDQ6VXNlcjU2NDIzNQ==", "avatar_url": "https://avatars.githubusercontent.com/u/564235?v=4", "gravatar_id": "", "url": "https://api.github.com/users/wuxb45", "html_url": "https://github.com/wuxb45", "followers_url": "https://api.github.com/users/wuxb45/followers", "following_url": "https://api.github.com/users/wuxb45/following{/other_user}", "gists_url": "https://api.github.com/users/wuxb45/gists{/gist_id}", "starred_url": "https://api.github.com/users/wuxb45/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/wuxb45/subscriptions", "organizations_url": "https://api.github.com/users/wuxb45/orgs", "repos_url": "https://api.github.com/users/wuxb45/repos", "events_url": "https://api.github.com/users/wuxb45/events{/privacy}", "received_events_url": "https://api.github.com/users/wuxb45/received_events", "type": "User", "site_admin": false }

[]

[]

[ "hi @wuxb45 \r\ncan you share a fully reproducible snippet?", "I have the same issue with transofrmers 4.36.1. I am using DeepSpeed framework to generate a response and face this the same error. ", "We cannot help you if you don't share a reproducible snippet. The way this part of the code works should not trigger this error because the past key values are casted to the `DynamicCache` if `use_legacy_cache`. Thus there is probably a versioning issue", "> I have the same issue with transofrmers 4.36.1. I am using DeepSpeed framework to generate a response and face this the same error.\r\n\r\nMe too. I also struggled with this problem for a long time, using deepspeed-chat to train reinforcement learning code.@liziniu", "> hi @wuxb45 can you share a fully reproducible snippet?\r\n\r\nI don't have the capacity to generate a reprod at this time. The issue was from running a code base forked from deepspeed chat's step 3. I'm sorry that I cannot provide more information now.", "I solved this by removing tensor parallel. It seems that merging perdevicetensor converted Cache to Tuple.", "> I solved this by removing tensor parallel. It seems that merging perdevicetensor converted Cache to Tuple.\r\n\r\nHI, i also faced the same issue. May I ask how you actually removed the tensor parallel if you are also using the deepspeed chat code?", "I had the same errors with `4.36.0` and `4.36.2` versions for Llama inferencing on multiple GPUs with `tensor_paralle`l package. ", "I tried version `4.36.0.dev0` . I don't have the issue. Other versions including `4.37.0.dev0` will give the AttributeError.", "Hi everyone, please let us know whenever you can share a small reproducible snippet as we can't do anything without a repro to fix the bug", "@younesbelkada \r\n\r\nYou can probably try the following code with different transformers versions to reproduce:\r\n\r\n```\r\nimport torch\r\nfrom tensor_parallel import TensorParallelPreTrainedModel\r\nfrom transformers import LlamaTokenizer, LlamaForCausalLM, GenerationConfig\r\n\r\nmodel_path = \"meta-llama/Llama-2-7b-chat-hf\"\r\nmodel = LlamaForCausalLM.from_pretrained(model_path, torch_dtype=torch.float16)\r\nmodel = TensorParallelPreTrainedModel(model, [\"cuda:0\", \"cuda:1\", \"cuda:2\", \"cuda:3\"])\r\ntokenizer = LlamaTokenizer.from_pretrained(model_path)\r\ninputs = tokenizer(\"Hi, how are you doing?\", return_tensors=\"pt\", add_special_tokens=False)\r\noutputs = tokenizer.decode(model.generate(inputs[\"input_ids\"].cuda(0), attention_mask=inputs[\"attention_mask\"].cuda(0), max_length=256)[0], add_special_tokens=False)\r\n\r\nprint(outputs)\r\n```\r\nTo me, only `4.36.0.dev0` works. After updating to new version, it won't work and I was not able to go back to the old `4.36.0.dev0` version.", "Alright this is pretty much a duplicate of #28003. We made a mistake by not advertising to test a bit more for other repos to get ready, feel free to share it on the `tensor_parallel` repo", "To me, it works with `transformers==4.34.1`\r\n\r\n", "Apparently this issue was introduced due to this this commit PR #26681 by @tomaarsen and @patrickvonplaten \r\n\r\nnext_decoder_cache should be a cache, which means it is not well initialized as a cache. Instead of a tuple , the new HF implementation pass a list of cache:\r\n\r\n```python\r\nhttps://github.com/tomaarsen/transformers/blob/ee60b1cc13e2819ef31e69952c0b6f616bd724b8/src/transformers/models/llama/modeling_llama.py#L287C45-L287C76\r\nlayer_idx: Optional[int] = None\r\n\r\n#https://github.com/tomaarsen/transformers/blob/ee60b1cc13e2819ef31e69952c0b6f616bd724b8/src/transformers/models/llama/modeling_llama.py#L355\r\npast_key_value: Optional[Cache] = None,\r\n```\r\n\r\nLayer_idx is later used by `past_key_value`, and `past_key_value` is currently replaced as list of Cache.\r\n\r\nNote the diff contains a kind of cache (for attention KV cache) which implements `to_legacy_cache`.\r\n\r\nI guess deepspeed version does not instantiate llama attention correctly or we should change the code as @fxmarty suggests:\r\n\r\n```\r\n if use_cache:\r\n\t use_legacy_cache = not isinstance(past_key_values, Cache) and past_key_values is not None\r\n\t if use_legacy_cache:\r\n\t past_key_values = DynamicCache.from_legacy_cache(past_key_values)\r\n\t elif past_key_values is None:\r\n\t past_key_values = DynamicCache()\r\n\t past_key_values_length = past_key_values.get_seq_length()\r\n```\r\n\r\n", "您好，邮箱主人会认真阅读！谢谢关注/", "I am facing a similar issue AttributeError: 'tuple' object has no attribute 'to_legacy_cache' while training Llama 7B. What is the concluded solution?", "If you did not change your version of `transformers` that is expected. Upgrading to the latest / providing a repo should help!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28045/reactions", "total_count": 3, "+1": 3, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "massabaali7", "id": 100831623, "node_id": "U_kgDOBgKRhw", "avatar_url": "https://avatars.githubusercontent.com/u/100831623?v=4", "gravatar_id": "", "url": "https://api.github.com/users/massabaali7", "html_url": "https://github.com/massabaali7", "followers_url": "https://api.github.com/users/massabaali7/followers", "following_url": "https://api.github.com/users/massabaali7/following{/other_user}", "gists_url": "https://api.github.com/users/massabaali7/gists{/gist_id}", "starred_url": "https://api.github.com/users/massabaali7/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/massabaali7/subscriptions", "organizations_url": "https://api.github.com/users/massabaali7/orgs", "repos_url": "https://api.github.com/users/massabaali7/repos", "events_url": "https://api.github.com/users/massabaali7/events{/privacy}", "received_events_url": "https://api.github.com/users/massabaali7/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @gante ", "@amyeroberts @gante did you get the chance to review my code?\r\nThanks in advance ", "Hi @massabaali7 👋 \r\n\r\nIt is my understanding that constrained beam search is not very used by the community. As such, I'm not adding more code related to it at the moment, as we have limited bandwidth to empower the community 🤗 ", "I am doing a research and publishing a paper about this. Its a very important idea proposed for backchannel insertion. ", "@massabaali7 and we'd be delighted to add it, if the published results show significant improvements. However, we do not add features while they are in a research phase -- that's why we made the whole codebase open to everyone, so you can freely experiment locally or in a forked repository 🤗 ", "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/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28044/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28044", "html_url": "https://github.com/huggingface/transformers/pull/28044", "diff_url": "https://github.com/huggingface/transformers/pull/28044.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28044.patch", "merged_at": null }

{ "login": "younesbelkada", "id": 49240599, "node_id": "MDQ6VXNlcjQ5MjQwNTk5", "avatar_url": "https://avatars.githubusercontent.com/u/49240599?v=4", "gravatar_id": "", "url": "https://api.github.com/users/younesbelkada", "html_url": "https://github.com/younesbelkada", "followers_url": "https://api.github.com/users/younesbelkada/followers", "following_url": "https://api.github.com/users/younesbelkada/following{/other_user}", "gists_url": "https://api.github.com/users/younesbelkada/gists{/gist_id}", "starred_url": "https://api.github.com/users/younesbelkada/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/younesbelkada/subscriptions", "organizations_url": "https://api.github.com/users/younesbelkada/orgs", "repos_url": "https://api.github.com/users/younesbelkada/repos", "events_url": "https://api.github.com/users/younesbelkada/events{/privacy}", "received_events_url": "https://api.github.com/users/younesbelkada/received_events", "type": "User", "site_admin": false }

[]

[]

[ "I am not 100% sure this approach is correct cc @fxmarty does this looks good to you (as you took care of the attention refactor) ?", "I'm a bit concerned about this - this is effectively a patch inside `from_pretrained` to add backwards compatibility that should have already been handled. The main question this raises for me is whether there other FA parameters/behaviours we need to check? \r\n\r\nIs it still possible to pass in both `use_flash_attention_2` and `config` to `from_pretrained`? If not, it's not clear to me from the diff how this is addressed: `use_flash_attention_2` isn't handled from the model kwargs. \r\n\r\nDidn't do a final review on the recent refactor, so might be missing something. It's also not clear to me from just this PR why passing in a config would change whether or not I can pass in `attn_implementation`.", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28043). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "Good catch\r\n\r\n@amyeroberts Even witout a fix, `use_flash_attention_2=True` along with a provided config IMO works thanks to https://github.com/huggingface/transformers/blob/050e0b44f6a63131b56d493543ab39fb7b4f20ca/src/transformers/modeling_utils.py#L1295-L1299", "cc @amyeroberts @fxmarty requesting another round of review!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28043/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28043", "html_url": "https://github.com/huggingface/transformers/pull/28043", "diff_url": "https://github.com/huggingface/transformers/pull/28043.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28043.patch", "merged_at": 1702634907000 }

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

[]

[]

[ "@patrickvonplaten Thanks for reporting. I think I know the culprit PR but indeed it shouldn't be triggered here. Looking into it 🕵️ ", "Or, at least I know about `AnnotionFormat` - I don't have any immediate ideas regarding `text_config_dict`", "~~The `text_config_dict` is from~~\r\n\r\n~~https://huggingface.co./openai/clip-vit-large-patch14/blob/main/config.json~~", "Opened a PR to address the Annotion format errors. \r\n\r\nFor the text_config errors, using the `\"runwayml/stable-diffusion-v1-5\"`, at least one of the errors is coming from using [this config](https://huggingface.co./runwayml/stable-diffusion-v1-5/blob/main/safety_checker/config.json). \r\n\r\nAs the error mentions, both `text_config_dict` and `text_config` are being used in the model config. The warning is letting you know some of the values in `\"text_config\"` won't be used e.g. `eos_token_id`, even though `eos_token_id` isn't in `text_config_dict`. This is because `text_config_dict` is used to instantiate a `CLIPTextConfig` class, which will have default values of e.g. `eos_token_id = 49407`. I think the warning itself is good - it told me exactly what to look for and flags potentially confusing behaviour. I believe the correct resolution would be to add the overwritten values to `text_config` - @ydshieh is this right? \r\n\r\n@patrickvonplaten would adding a note on how to fix be sufficient or is there more information you'd like to see? \r\n\r\n", "> I believe the correct resolution would be to add the overwritten values to text_config - @ydshieh is this right?\r\n\r\nYes, if the 2 config (dict) get the same value for a key, then no warning won't be given (for that key).\r\n(or just update the config file to not use `text_config_dict` with an updated `text_config`)\r\n\r\nWe can open Hub PRs to the most used CLIP repositories to avoid such warning.\r\n", "I have three warnings triggered by calling StableDiffusionSafetyChecker.from_pretrained():\r\n\r\n```\r\n`text_config_dict` is provided which will be used to initialize `CLIPTextConfig`. The value `text_config[\"id2label\"]` will be overriden.\r\n`text_config_dict` is provided which will be used to initialize `CLIPTextConfig`. The value `text_config[\"bos_token_id\"]` will be overriden.\r\n`text_config_dict` is provided which will be used to initialize `CLIPTextConfig`. The value `text_config[\"eos_token_id\"]` will be overriden.\r\n```\r\n\r\n", "> As the error mentions, both text_config_dict and text_config are being used in the model config. The warning is letting you know some of the values in \"text_config\" won't be used e.g. eos_token_id, even though eos_token_id isn't in text_config_dict. This is because text_config_dict is used to instantiate a CLIPTextConfig class, which will have default values of e.g. eos_token_id = 49407. I think the warning itself is good - it told me exactly what to look for and flags potentially confusing behaviour. I believe the correct resolution would be to add the overwritten values to text_config - @ydshieh is this right?\r\n\r\n> Yes, if the 2 config (dict) get the same value for a key, then no warning won't be given (for that key).\r\n(or just update the config file to not use text_config_dict with an updated text_config)\r\n\r\nI have some trouble understanding this tbh :sweat_smile: \r\n\r\nWhat exactly do we need to change here: https://huggingface.co./runwayml/stable-diffusion-v1-5/blob/main/safety_checker/config.json ? (could you maybe open a PR?) " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28042/reactions", "total_count": 1, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "peacefulotter", "id": 32218033, "node_id": "MDQ6VXNlcjMyMjE4MDMz", "avatar_url": "https://avatars.githubusercontent.com/u/32218033?v=4", "gravatar_id": "", "url": "https://api.github.com/users/peacefulotter", "html_url": "https://github.com/peacefulotter", "followers_url": "https://api.github.com/users/peacefulotter/followers", "following_url": "https://api.github.com/users/peacefulotter/following{/other_user}", "gists_url": "https://api.github.com/users/peacefulotter/gists{/gist_id}", "starred_url": "https://api.github.com/users/peacefulotter/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/peacefulotter/subscriptions", "organizations_url": "https://api.github.com/users/peacefulotter/orgs", "repos_url": "https://api.github.com/users/peacefulotter/repos", "events_url": "https://api.github.com/users/peacefulotter/events{/privacy}", "received_events_url": "https://api.github.com/users/peacefulotter/received_events", "type": "User", "site_admin": false }

[]

[]

[ "This would be better posted in the safetensors repo rather than transformers I believe? https://github.com/huggingface/safetensors", "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/main/CONTRIBUTING.md) are likely to be ignored.", "For anyone stumbling upon this: if you save the weights of a *compiled* model then you also need to load the weights inside that *compiled* model. \n\nThe repro above is actually supposed to work.." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28041/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "dwojcik92", "id": 10101471, "node_id": "MDQ6VXNlcjEwMTAxNDcx", "avatar_url": "https://avatars.githubusercontent.com/u/10101471?v=4", "gravatar_id": "", "url": "https://api.github.com/users/dwojcik92", "html_url": "https://github.com/dwojcik92", "followers_url": "https://api.github.com/users/dwojcik92/followers", "following_url": "https://api.github.com/users/dwojcik92/following{/other_user}", "gists_url": "https://api.github.com/users/dwojcik92/gists{/gist_id}", "starred_url": "https://api.github.com/users/dwojcik92/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/dwojcik92/subscriptions", "organizations_url": "https://api.github.com/users/dwojcik92/orgs", "repos_url": "https://api.github.com/users/dwojcik92/repos", "events_url": "https://api.github.com/users/dwojcik92/events{/privacy}", "received_events_url": "https://api.github.com/users/dwojcik92/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @dwojcik92, thanks for raising this issue! \r\n\r\nI'm unable to reproduce this on my side. It looks like the checkpoints have only partially downloaded in the cache. If you're not setting your cache path, the model downloads will be found under `$HOME/.cache/huggingface/hub`. Could you check there and see if the expected models are there? \r\n\r\nI'd first try running the pipeline with a small model which can be quickly downloaded to see if that works: \r\n```py\r\nfrom transformers import pipeline\r\n\r\npipe = pipeline(\"text-generation\", model=\"gpt2\")\r\n```\r\n\r\nIf this doesn't work, could you also try loading a model outside of the pipeline API: \r\n```\r\nfrom transformers import AutoModel\r\n\r\nmodel = AutoModel.from_pretrained(\"gpt2\")\r\n```", "@amyeroberts thank you for quick response! \r\nI know about cache and what I tried is to specify cache with cache_dir. I didn't helped.\r\n\r\n```python\r\nbase_model = GPTNeoXForCausalLM.from_pretrained(\r\n \"EleutherAI/pythia-6.9b-deduped-v0\",\r\n revision=\"step3000\",\r\n)\r\n```\r\nThe above command worked well while the one below failed.\r\n```python\r\n# Use a pipeline as a high-level helper\r\nfrom transformers import pipeline\r\n\r\npipe = pipeline(\"text-generation\", model=\"mistralai/Mistral-7B-v0.1\")\r\n```\r\n3 hours later and it seems that I can download all models without problem. It seems to me that the problem was with HF servers and not with `transformers`.\r\n\r\nAnyway, it would be a good idea to be able to have some verbose output here for debugging to see if the problem is with package itself or the HF servers.", "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/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28039/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "hiyouga", "id": 16256802, "node_id": "MDQ6VXNlcjE2MjU2ODAy", "avatar_url": "https://avatars.githubusercontent.com/u/16256802?v=4", "gravatar_id": "", "url": "https://api.github.com/users/hiyouga", "html_url": "https://github.com/hiyouga", "followers_url": "https://api.github.com/users/hiyouga/followers", "following_url": "https://api.github.com/users/hiyouga/following{/other_user}", "gists_url": "https://api.github.com/users/hiyouga/gists{/gist_id}", "starred_url": "https://api.github.com/users/hiyouga/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/hiyouga/subscriptions", "organizations_url": "https://api.github.com/users/hiyouga/orgs", "repos_url": "https://api.github.com/users/hiyouga/repos", "events_url": "https://api.github.com/users/hiyouga/events{/privacy}", "received_events_url": "https://api.github.com/users/hiyouga/received_events", "type": "User", "site_admin": false }

[]

[]

[ "hi @hiyouga \r\nThanks a lot for the issue! \r\nI think that you cannot pass both the config and `attn_implementation` , can you elaborate a bit on why you would like to pass the config as well as `attn_implementation` into `from_pretrained`? The canonical way to load a FA2 model is:\r\n\r\n```python\r\nimport torch\r\nfrom transformers import AutoModelForCausalLM, AutoTokenizer, LlamaForCausalLM\r\n\r\nmodel_id = \"tiiuae/falcon-7b\"\r\ntokenizer = AutoTokenizer.from_pretrained(model_id)\r\n\r\nmodel = AutoModelForCausalLM.from_pretrained(\r\n model_id, \r\n torch_dtype=torch.bfloat16, \r\n attn_implementation=\"flash_attention_2\",\r\n)\r\n```\r\n\r\nIt is also not recommended to enable FA2 through the config directly. However you can enable FA2 by passing `attn_implementation=\"flash_attention_2\"` in `from_config` methd: https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_utils.py#L1235", "Thanks for replying!\r\nI wish to modify the model config before loading the pre-trained models, such as setting `rope_scaling` and `torch_dtype`.\r\nI wonder why I could pass both `config` and `use_flash_attention_2` to `from_pretrained`", "Thanks @hiyouga \r\nIndeed, it was possible to do this before, therefore there is a regression now, I just made https://github.com/huggingface/transformers/pull/28043 which should resolve your problem" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28038/reactions", "total_count": 1, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "gante", "id": 12240844, "node_id": "MDQ6VXNlcjEyMjQwODQ0", "avatar_url": "https://avatars.githubusercontent.com/u/12240844?v=4", "gravatar_id": "", "url": "https://api.github.com/users/gante", "html_url": "https://github.com/gante", "followers_url": "https://api.github.com/users/gante/followers", "following_url": "https://api.github.com/users/gante/following{/other_user}", "gists_url": "https://api.github.com/users/gante/gists{/gist_id}", "starred_url": "https://api.github.com/users/gante/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/gante/subscriptions", "organizations_url": "https://api.github.com/users/gante/orgs", "repos_url": "https://api.github.com/users/gante/repos", "events_url": "https://api.github.com/users/gante/events{/privacy}", "received_events_url": "https://api.github.com/users/gante/received_events", "type": "User", "site_admin": false }

[]

[]

[ "This indeed seems to address what was described [here](https://github.com/huggingface/transformers/issues/27985#issuecomment-1855524359), well done!", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28037). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "As @younesbelkada mentioined and in the official site\r\n\r\n> Ampere, Ada, or Hopper GPUs (e.g., A100, RTX 3090, RTX 4090, H100). Support for Turing GPUs (T4, RTX 2080) is coming soon, please use FlashAttention 1.x for Turing GPUs for now.\r\n\r\nnothing we can't do unless we run on a different machine" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28037/reactions", "total_count": 1, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28037", "html_url": "https://github.com/huggingface/transformers/pull/28037", "diff_url": "https://github.com/huggingface/transformers/pull/28037.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28037.patch", "merged_at": 1702565565000 }

{ "login": "gante", "id": 12240844, "node_id": "MDQ6VXNlcjEyMjQwODQ0", "avatar_url": "https://avatars.githubusercontent.com/u/12240844?v=4", "gravatar_id": "", "url": "https://api.github.com/users/gante", "html_url": "https://github.com/gante", "followers_url": "https://api.github.com/users/gante/followers", "following_url": "https://api.github.com/users/gante/following{/other_user}", "gists_url": "https://api.github.com/users/gante/gists{/gist_id}", "starred_url": "https://api.github.com/users/gante/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/gante/subscriptions", "organizations_url": "https://api.github.com/users/gante/orgs", "repos_url": "https://api.github.com/users/gante/repos", "events_url": "https://api.github.com/users/gante/events{/privacy}", "received_events_url": "https://api.github.com/users/gante/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@ylacombe Re-opening as this isn't resolved by #28060 (it just makes sure our CI doesn't break)", "Alright, copying the reasons for flakiness from #28060 for trackability then!\r\n\r\n> After investigating the reasons for the test_retain_grad_hidden_states_attentions flaky failure, I realized the speech encoder attentions can be None with a non-zero probability when training=True.\r\n" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28036/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "gante", "id": 12240844, "node_id": "MDQ6VXNlcjEyMjQwODQ0", "avatar_url": "https://avatars.githubusercontent.com/u/12240844?v=4", "gravatar_id": "", "url": "https://api.github.com/users/gante", "html_url": "https://github.com/gante", "followers_url": "https://api.github.com/users/gante/followers", "following_url": "https://api.github.com/users/gante/following{/other_user}", "gists_url": "https://api.github.com/users/gante/gists{/gist_id}", "starred_url": "https://api.github.com/users/gante/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/gante/subscriptions", "organizations_url": "https://api.github.com/users/gante/orgs", "repos_url": "https://api.github.com/users/gante/repos", "events_url": "https://api.github.com/users/gante/events{/privacy}", "received_events_url": "https://api.github.com/users/gante/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28035/reactions", "total_count": 1, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 1, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28035", "html_url": "https://github.com/huggingface/transformers/pull/28035", "diff_url": "https://github.com/huggingface/transformers/pull/28035.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28035.patch", "merged_at": 1702562164000 }

{ "login": "u7122029", "id": 111028268, "node_id": "U_kgDOBp4oLA", "avatar_url": "https://avatars.githubusercontent.com/u/111028268?v=4", "gravatar_id": "", "url": "https://api.github.com/users/u7122029", "html_url": "https://github.com/u7122029", "followers_url": "https://api.github.com/users/u7122029/followers", "following_url": "https://api.github.com/users/u7122029/following{/other_user}", "gists_url": "https://api.github.com/users/u7122029/gists{/gist_id}", "starred_url": "https://api.github.com/users/u7122029/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/u7122029/subscriptions", "organizations_url": "https://api.github.com/users/u7122029/orgs", "repos_url": "https://api.github.com/users/u7122029/repos", "events_url": "https://api.github.com/users/u7122029/events{/privacy}", "received_events_url": "https://api.github.com/users/u7122029/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @u7122029, thanks for raising this issue and taking the time to write up all of these details - it really helps us. \r\n\r\nIndeed, this doesn't seem like a desired behaviour. From the linked issue and these comments - [[1](https://github.com/huggingface/transformers/issues/25024#issuecomment-1648023318)], [[2](https://github.com/huggingface/transformers/issues/25024#issuecomment-1649544696)] - it seems this is a known issue which requires refactoring of the BLIP model and to convert the weights. \r\n\r\n@younesbelkada Is there any update on the progress of this? Are there weights on the hub which are compatible which can be used instead for the code examples? \r\n", "Until @younesbelkada responds, could I please have some guidance on how I can fix this through a pull request?", "Hi @u7122029 \r\nThanks very much for the issue, I have been discussing with @amyeroberts offline about this problem\r\nThere is a huge confusion around the `BlipModel` class, and it is not possible to make that class compatible with weights that are on the Hub as it will involve many breaking changes, e.g. `BlipForConditionalGeneration` has `text_decoder` whereas `BlipModel` has `text_model`. `BlipModel` also exposes those attributes:\r\n```python\r\n self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias=False)\r\n self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias=False)\r\n self.logit_scale = nn.Parameter(torch.tensor(self.config.logit_scale_init_value))\r\n```\r\nwhich do not exist in the origjnal blip-model: https://github.com/salesforce/BLIP/blob/main/models/blip.py#L39-L43 - the HF BlipModel class is unfortunately just a copy-pasta from `CLIPModel` that shouldn’t be designed as I designed it at first place :confused:\r\nIn the near future, we will remove that class from the docs and deprecate `BlipModel`\r\n\r\nIf your intent is to retrieve text / vision logits from BLIP, I believe that you can still use `BlipForConditionalGeneration` without any problem. ", "Hi @younesbelkada, great to hear from you. I'd like to use the BLIP model to get probabilities on each text output. For example, if I gave a list of text outputs such as `[\"an image of a cat\", \"an image of a dog\"]`, and I gave an image with a cat in it, then I would expect a high logit for `\"an image of a cat\"` and a lower logit for `\"an image of a dog\"`, which is pretty much what `BLIPModel` does.\r\n\r\nI gave the `BlipForConditionalGeneration` class a spin, and please forgive me if I sound like a noob, but I couldn't understand how I could use the `logits` from the `output` variable (as per https://huggingface.co./docs/transformers/model_doc/blip#transformers.BlipForConditionalGeneration) to achieve what I wanted to do. Could I please have some advice for this?", "Hi @u7122029 \r\nThank you very much \r\nIf I understood correctly your usecase, I think that `BlipForImageTextRetrieval` might do the trick for you. \r\n\r\nconsider the snippet below:\r\n```python\r\nfrom PIL import Image\r\nimport requests\r\nfrom transformers import AutoProcessor, BlipForImageTextRetrieval\r\n\r\nmodel = BlipForImageTextRetrieval.from_pretrained(\"Salesforce/blip-itm-base-coco\")\r\nprocessor = AutoProcessor.from_pretrained(\"Salesforce/blip-itm-base-coco\")\r\n\r\nurl = \"http://images.cocodataset.org/val2017/000000039769.jpg\"\r\nimage = Image.open(requests.get(url, stream=True).raw)\r\ntext = \"an image of two cats sleeping on a couch with tv remotes\"\r\n\r\ninputs = processor(images=image, text=text, return_tensors=\"pt\")\r\noutputs = model(**inputs)\r\nprint(outputs.itm_score.softmax(-1))\r\n>>> tensor([[0.0045, 0.9955]]) # Label 1 means match\r\n```\r\n\r\n`outputs.itm_score` gives a tensor of size batch_size, 2 - the first label corresponds to the case where the image did not matched the text, the second label means it has matched. \r\nFor the image below:\r\n\r\n\r\n\r\nIt gave a matched score of `0.9955` for the prompt `\"an image of two cats sleeping on a couch with tv remotes\"` and a matched score of `0.0633` for the prompt `\"an image of two dogs sleeping on a couch with tv remotes\"`", "I think this is an interesting idea, but I want to consider more than just binary yes/no image-text matching, i.e: generate a probability vector corresponding to more than 2 different text labels. For example, CIFAR-10 has 10 different labels, and I would pass a list `[\"an image of an airplane\", \"an image of an automobile\", ..., \"an image of a truck\"]` which has 10 labels, not 2. Furthermore, wouldn't entry 0 in your code snippet output correspond to \"*not* an image of two cats sleeping on a couch with tv remotes\", which is much more general than \"an image of two dogs sleeping on a couch with tv remotes\"?", "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/main/CONTRIBUTING.md) are likely to be ignored.", "> I think this is an interesting idea, but I want to consider more than just binary yes/no image-text matching, i.e: generate a probability vector corresponding to more than 2 different text labels. For example, CIFAR-10 has 10 different labels, and I would pass a list `[\"an image of an airplane\", \"an image of an automobile\", ..., \"an image of a truck\"]` which has 10 labels, not 2. Furthermore, wouldn't entry 0 in your code snippet output correspond to \"_not_ an image of two cats sleeping on a couch with tv remotes\", which is much more general than \"an image of two dogs sleeping on a couch with tv remotes\"?\r\n\r\n@younesbelkada @amyeroberts Hi guys, could I please have this confirmed?\r\n\r\nHave there also been any updates to the `BLIPModel`? Like I said before: I'd like to fix it if you can't at the moment, but I'm completely new to contributing to this repo so I would at least like some advice to do so.\r\n\r\n\r\n\r\n", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28034/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "DopeorNope-Lee", "id": 86828497, "node_id": "MDQ6VXNlcjg2ODI4NDk3", "avatar_url": "https://avatars.githubusercontent.com/u/86828497?v=4", "gravatar_id": "", "url": "https://api.github.com/users/DopeorNope-Lee", "html_url": "https://github.com/DopeorNope-Lee", "followers_url": "https://api.github.com/users/DopeorNope-Lee/followers", "following_url": "https://api.github.com/users/DopeorNope-Lee/following{/other_user}", "gists_url": "https://api.github.com/users/DopeorNope-Lee/gists{/gist_id}", "starred_url": "https://api.github.com/users/DopeorNope-Lee/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/DopeorNope-Lee/subscriptions", "organizations_url": "https://api.github.com/users/DopeorNope-Lee/orgs", "repos_url": "https://api.github.com/users/DopeorNope-Lee/repos", "events_url": "https://api.github.com/users/DopeorNope-Lee/events{/privacy}", "received_events_url": "https://api.github.com/users/DopeorNope-Lee/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @DopeorNope-Lee, thanks for raising this issue! \r\n\r\nCould you share your running environment: run `transformers-cli env` in the terminal and copy-paste the output? ", "Use version `flash_attn==2.3.4`\r\nand on the axo cfg `flash_attention_2: true`", "Hi @fblgit, thanks for providing this information. To be able to debug, we'll information about the transformers version and any relevant libraries (obtained by running `transformers-cli env`). \r\n\r\nCould you also provide a minimal reproducer? The link provided in the issue description just goes directly to a repo, rather than a specific script. We get many issues and PRs, so that we can get to all of them in a timely manner, we need you to help us help you. In this case, we'll need the smallest amount of code that can be run to reproduce the error. ", "same error.", "> Hi @DopeorNope-Lee, thanks for raising this issue!\r\n> \r\n> Could you share your running environment: run `transformers-cli env` in the terminal and copy-paste the output?\r\n\r\nSame error, and I also cannot successfully install the flash_attn2. The transformer-cli env is:\r\n**- `transformers` version: 4.32.0.dev0\r\n- Platform: Linux-5.15.0-1042-azure-x86_64-with-glibc2.17\r\n- Python version: 3.8.13\r\n- Huggingface_hub version: 0.20.3\r\n- Safetensors version: 0.4.2\r\n- Accelerate version: 0.21.0\r\n- Accelerate config: not found\r\n- PyTorch version (GPU?): 1.12.1+rocm5.2.3 (True)\r\n- Tensorflow version (GPU?): not installed (NA)\r\n- Flax version (CPU?/GPU?/TPU?): not installed (NA)\r\n- Jax version: not installed\r\n- JaxLib version: not installed\r\n- Using GPU in script?: <fill in>\r\n- Using distributed or parallel set-up in script?: <fill in>**\r\nCould I know how to fix it? ", "Hi @SUNJIMENG, thanks for providing the environment information. Could you also provide a minimal code snippet to reproduce the issue? \r\n\r\nFlash Attention was added for mistral in v4.34 and for mixstral in v4.36. I'd first suggest upgrading to the latest version of transformers `pip install -U transformers` and re-running your code. ", "@fblgit from his advice I can fix the error.\r\n\r\n@SUNJIMENG could you remove your flash-attn and reinstall with `pip install -U flash-attn`?", "Facing same issue, Followed\r\n\r\n```\r\npip install -U transformers\r\npip install -U flash-attn\r\n```\r\n\r\nStill same error", "@BakingBrains If you're also experiencing this error, could you provide your environment information - run`transformers-cli env` - and a snippet we can run to reproduce the error? ", "@amyeroberts the problem was with the CUDA drivers, I re-installed, now it's fine.\r\nThank you." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28033/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "younesbelkada", "id": 49240599, "node_id": "MDQ6VXNlcjQ5MjQwNTk5", "avatar_url": "https://avatars.githubusercontent.com/u/49240599?v=4", "gravatar_id": "", "url": "https://api.github.com/users/younesbelkada", "html_url": "https://github.com/younesbelkada", "followers_url": "https://api.github.com/users/younesbelkada/followers", "following_url": "https://api.github.com/users/younesbelkada/following{/other_user}", "gists_url": "https://api.github.com/users/younesbelkada/gists{/gist_id}", "starred_url": "https://api.github.com/users/younesbelkada/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/younesbelkada/subscriptions", "organizations_url": "https://api.github.com/users/younesbelkada/orgs", "repos_url": "https://api.github.com/users/younesbelkada/repos", "events_url": "https://api.github.com/users/younesbelkada/events{/privacy}", "received_events_url": "https://api.github.com/users/younesbelkada/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @amyeroberts , I agree it is quite hacky, let me take some time to further investigate and provide a proper fix", "@younesbelkada I saw no other issue except here onwards `batch_index, non_attended_tokens = torch.where(first_layer_past_key_value == 0)`\r\n\r\n`first_layer_past_key_value.shape[1]` is bigger than `extended_attention_mask.shape[1]`, so it should be expected that an index could be present in `non_attended_tokens` that is larger than `extended_attention_mask.shape[1]`.\r\n\r\nThat is why filtering out made sense to me at-least. To avoid this hack, I see that either `first_layer_past_key_value` or `extended_attention_mask` is the root of the problem.\r\n\r\n", "hey - facing a similar issue: it seems to appear when both the inputs and generated outputs are long enough, hence different behaviour for different images. One way to replicate is:\r\n\r\n```\r\nimport requests\r\nfrom PIL import Image\r\n\r\nimport torch\r\nfrom transformers import AutoProcessor, LlavaForConditionalGeneration\r\n\r\nmodel_id = \"llava-hf/llava-1.5-7b-hf\"\r\n\r\nk = 200\r\nuser_prompt = \"Describe the image:?\\n\" * k\r\nprompt = f\"USER: <image>\\n{user_prompt}ASSISTANT:\"\r\nimage_file = \"http://images.cocodataset.org/val2017/000000039769.jpg\"\r\n\r\nmodel = LlavaForConditionalGeneration.from_pretrained(\r\n model_id, \r\n torch_dtype=torch.float16, \r\n).to(0)\r\n\r\nprocessor = AutoProcessor.from_pretrained(model_id)\r\n\r\nraw_image = Image.open(requests.get(image_file, stream=True).raw)\r\ninputs = processor(prompt, raw_image, return_tensors='pt').to(0, torch.float16)\r\n\r\nprint(k, inputs['input_ids'].size())\r\noutput = model.generate(**inputs, max_new_tokens=200, do_sample=False)\r\nprint(k, output.size())\r\nprint(processor.decode(output[0][-100:], skip_special_tokens=True))\r\n```\r\n\r\nRunning on A10G on current main\r\n\r\n```\r\n!CUDA_LAUNCH_BLOCKING=1 python test_llava.py\r\n\r\n2023-12-18 14:20:27.237501: E tensorflow/compiler/xla/stream_executor/cuda/cuda_dnn.cc:9342] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered\r\n2023-12-18 14:20:27.237560: E tensorflow/compiler/xla/stream_executor/cuda/cuda_fft.cc:609] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered\r\n2023-12-18 14:20:27.237608: E tensorflow/compiler/xla/stream_executor/cuda/cuda_blas.cc:1518] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered\r\n2023-12-18 14:20:27.244438: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.\r\nTo enable the following instructions: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.\r\nLoading checkpoint shards: 100%|██████████████████| 3/3 [00:01<00:00, 1.59it/s]\r\nSpecial tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.\r\n200 torch.Size([1, 1412])\r\n/databricks/python/lib/python3.10/site-packages/torch/nn/modules/conv.py:459: UserWarning: Applied workaround for CuDNN issue, install nvrtc.so (Triggered internally at ../aten/src/ATen/native/cudnn/Conv_v8.cpp:80.)\r\n return F.conv2d(input, weight, bias, self.stride,\r\n\r\n\r\n--- added some prints\r\nExtended attention mask: (1, 575); \r\nAttention mask: (1, 1413); \r\nfirst_layer_past_key_value (1, 1987); \r\nTarget seqlen: 1988; \r\nBatch index: tensor([0], device='cuda:0'); \r\nNon attended tokens: tensor([1881], device='cuda:0')\r\n---\r\n\r\n\r\n../aten/src/ATen/native/cuda/IndexKernel.cu:92: operator(): block: [0,0,0], thread: [0,0,0] Assertion `index >= -sizes[i] && index < sizes[i] && \"index out of bounds\"` failed.\r\nTraceback (most recent call last):\r\n File \"\", line 25, in <module>\r\n output = model.generate(**inputs, max_new_tokens=200, do_sample=False)\r\n File \"/databricks/python/lib/python3.10/site-packages/torch/utils/_contextlib.py\", line 115, in decorate_context\r\n return func(*args, **kwargs)\r\n File \"/local_disk0/.ephemeral_nfs/envs/pythonEnv-e7088835-8718-43c0-b531-9c937824ca9c/lib/python3.10/site-packages/transformers/generation/utils.py\", line 1731, in generate\r\n return self.greedy_search(\r\n File \"/local_disk0/.ephemeral_nfs/envs/pythonEnv-e7088835-8718-43c0-b531-9c937824ca9c/lib/python3.10/site-packages/transformers/generation/utils.py\", line 2592, in greedy_search\r\n outputs = self(\r\n File \"/databricks/python/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1501, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/local_disk0/.ephemeral_nfs/envs/pythonEnv-e7088835-8718-43c0-b531-9c937824ca9c/lib/python3.10/site-packages/transformers/models/llava/modeling_llava.py\", line 439, in forward\r\n extended_attention_mask[batch_index, non_attended_tokens] = 0\r\nRuntimeError: CUDA error: device-side assert triggered\r\nCompile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.\r\n```\r\n\r\nSo if I understand correctly we should be masking the resulting `attention_mask` instead of `extended_attention_mask`?\r\n\r\nAlso, how do these zeros appear in `first_layer_past_key_value` in the first place?\r\n\r\n", "Thanks for the reproducer, I'll try to run some experiments on my end\r\n\r\n> Also, how do these zeros appear in first_layer_past_key_value in the first place?\r\n\r\nBecause the extended hidden states are initialized with all zeros , hence on the first layer they should stay un-touched so the first past kv cache should remain all zeros in the places where you have padd tokens", "@adilzhan-ismailov-depop Thanks for sharing the example. I would also agree that it has something to do with length of the generated output for a certain image. I am not sure if it has anything to do with input length, as the three prompts I tried, the different between prompt length was not much.\r\n\r\n@younesbelkada To answer your questions:\r\n\"Do you use one image per prompt? Are the prompts you use long? Can you somehow reproduce it with an image that you can find on the internet?\"\r\n\r\nYes, I used one image per prompt. Three different prompts were used in different runs. Smallest prompt was 7 words and the biggest one was 17 words. I can try to find more example images on the internet for which the same error is thrown, if needed.", "Hi @gullalc \r\n\r\n> I can try to find more example images on the internet for which the same error is thrown, if needed.\r\n\r\nYes that would be really great, thanks !", "> Because the extended hidden states are initialized with all zeros , hence on the first layer they should stay un-touched so the first past kv cache should remain all zeros in the places where you have padd tokens\r\n\r\nThanks - but I think in the example with batch size of one we shouldn't have any pad tokens? We can reproduce the error by adding a padding token to any input manually though:\r\n\r\n```\r\nimport requests\r\nfrom PIL import Image\r\n\r\nimport torch\r\nfrom transformers import AutoProcessor, LlavaForConditionalGeneration\r\n\r\nmodel_id = \"llava-hf/llava-1.5-7b-hf\"\r\n\r\nprompt = f\"USER: <image>\\nDescribe the image:?\\nASSISTANT:\"\r\nimage_file = \"http://images.cocodataset.org/val2017/000000039769.jpg\"\r\n\r\nmodel = LlavaForConditionalGeneration.from_pretrained(\r\n model_id, \r\n torch_dtype=torch.float16, \r\n device_map='auto'\r\n)\r\n\r\nprocessor = AutoProcessor.from_pretrained(model_id)\r\n\r\ndevice = 'cuda'\r\n\r\nraw_image = Image.open(requests.get(image_file, stream=True).raw)\r\ninputs = processor(prompt, raw_image, return_tensors='pt').to(device)\r\n\r\n# add padding token manually\r\npad_token_id = processor.tokenizer.pad_token_id # 32001\r\ninputs['input_ids'] = torch.hstack([torch.ones((1, 1), dtype=torch.int64, device=device) * pad_token_id, inputs['input_ids']])\r\ninputs['attention_mask'] = torch.hstack([torch.zeros((1, 1), dtype=torch.int64, device=device), inputs['input_ids']])\r\n\r\noutput = model.generate(**inputs, max_new_tokens=200, do_sample=False)\r\n```\r\n\r\nThis fails because since we have an image, padding token is not at the first position, so we fail the first time we create extended_attention_mask and try to index it\r\n\r\nSo why does this happen without padding tokens, and likelihood is higher with longer inputs? This is likely to do with half-precision. If you run this experiment you can see that this happens for float16 much more frequently than for float32:\r\n\r\n```\r\nimport torch\r\nimport altair as alt\r\nimport pandas as pd\r\nfrom tqdm.auto import tqdm\r\n\r\n# Function to run the experiment\r\ndef run_experiment(dtype, num_runs=1000):\r\n lengths = []\r\n for _ in tqdm(range(num_runs), desc=f\"Running with dtype={dtype}\"):\r\n random_tensor = torch.rand((2000, 1), dtype=dtype) # num. of elements in the original example\r\n batch_index, non_attended_tokens = torch.where(random_tensor == 0)\r\n lengths.append(len(batch_index))\r\n return lengths\r\n\r\n# Running the experiments\r\nlengths_float16 = run_experiment(torch.float16)\r\nlengths_float32 = run_experiment(torch.float32)\r\n\r\n# Creating a DataFrame for visualization\r\ndf = pd.DataFrame({\r\n \"Length\": lengths_float16 + lengths_float32,\r\n \"Dtype\": [\"float16\"] * len(lengths_float16) + [\"float32\"] * len(lengths_float32)\r\n})\r\n\r\n# Plotting the results\r\nchart = alt.Chart(df).mark_bar().encode(\r\n x=alt.X('Length:Q', title=\"Num. of zero entries in 2000-element array\"),\r\n y='count(Length):Q',\r\n column='Dtype:N'\r\n).properties(\r\n width=220,\r\n height=200\r\n)\r\n\r\nchart\r\n```\r\n\r\n\r\nIn practical terms I think it's ok, but maybe there is a more elegant way to identify non-attended tokens. The logic that handles the attention mask is still an issue though in case we have real padding tokens in the batch", "Hi :)\r\nThanks for investigating this! Just to let you know I'm facing the same issue when using images from the German split of the XM3600 dataset with a batch size > 1.\r\n\r\nHere is some log extract:\r\n```python\r\n File \"/home/XXX/miniforge3/envs/lmmm/lib/python3.11/site-packages/lightning/pytorch/loops/evaluation_loop.py\", line 134, in run\r\n self._evaluation_step(batch, batch_idx, dataloader_idx, dataloader_iter)\r\n File \"/homeXXX/miniforge3/envs/lmmm/lib/python3.11/site-packages/lightning/pytorch/loops/evaluation_loop.py\", line 391, in _evaluation_step\r\n output = call._call_strategy_hook(trainer, hook_name, *step_args)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home/XXX/miniforge3/envs/lmmm/lib/python3.11/site-packages/lightning/pytorch/trainer/call.py\", line 309, in _call_strategy_hook\r\n output = fn(*args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^\r\n File \"/home/XXX/miniforge3/envs/lmmm/lib/python3.11/site-packages/lightning/pytorch/strategies/strategy.py\", line 416, in test_step\r\n return self.lightning_module.test_step(*args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home/XXX/gitrepos/lmmm/lmmm/model/mixins/evaluation.py\", line 201, in test_step\r\n return self._in_text_image_out_text(\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home/XXX/gitrepos/lmmm/lmmm/model/mixins/evaluation.py\", line 101, in _in_text_image_out_text\r\n _, pred_text = self.generate(\r\n ^^^^^^^^^^^^^^\r\n File \"/home/XXX/gitrepos/lmmm/lmmm/model/lit_llava.py\", line 73, in generate\r\n generated_ids: torch.Tensor = self.model.generate(\r\n ^^^^^^^^^^^^^^^^^^^^\r\n File \"/home/XXX/miniforge3/envs/lmmm/lib/python3.11/site-packages/torch/utils/_contextlib.py\", line 115, in decorate_context\r\n return func(*args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home/XXX/miniforge3/envs/lmmm/lib/python3.11/site-packages/transformers/generation/utils.py\", line 1718, in generate\r\n return self.greedy_search(\r\n ^^^^^^^^^^^^^^^^^^^\r\n File \"/home/XXX/miniforge3/envs/lmmm/lib/python3.11/site-packages/transformers/generation/utils.py\", line 2579, in greedy_search\r\n outputs = self(\r\n ^^^^^\r\n File \"/home/XXX/miniforge3/envs/lmmm/lib/python3.11/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home/XXX/miniforge3/envs/lmmm/lib/python3.11/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home/XXX/miniforge3/envs/lmmm/lib/python3.11/site-packages/transformers/models/llava/modeling_llava.py\", line 428, in forward\r\n extended_attention_mask[batch_index, non_attended_tokens] = 0\r\n ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\nRuntimeError: CUDA error: device-side assert triggered\r\nCompile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.\r\n```\r\n\r\nThe fix introduced in this PR fixes the issue though", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28032). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "Note I will try to fix the SDPA regression (for users that perform multi-image & multi-prompt such as https://github.com/huggingface/transformers/issues/28184) in a separate PR , meanwhile users can always use the model with `attn_implementation=\"eager\"` to revert the previous behaviour. ", "Thanks a lot for the review! Tests are passing on my VM which is a 2xT4 with the same pytorch & bnb version as the docker image we use! Merging ! 🚀 \r\nThanks to all contributors for the insightful discussion and the fix! ", "Have the same problem and this is very helpful. Thanks!", "For anyone that wants to use this fix before the next release:\r\n```bash\r\npip install -U git+https://github.com/huggingface/transformers.git\r\n```", "I tried this pr but it still gives me an error.\r\nI think the core issue is \r\n```python\r\nfirst_layer_past_key_value.size(-1) > extended_attention_mask.size(-1)\r\n# May induce index errors\r\n```\r\n, which hasn't been addressed.\r\nBut I really don't quite understand what this piece of code is doing, the fix I think should look like this, but I don't know if that's correct:\r\n```python\r\nextended_attention_mask = torch.ones(\r\n (attention_mask.shape[0], target_seqlen - attention_mask.shape[1]),\r\n dtype=attention_mask.dtype,\r\n device=attention_mask.device,\r\n)\r\n\r\n# Zero-out the places where we don't need to attend\r\nattention_mask = torch.cat((attention_mask, extended_attention_mask), dim=1)\r\nattention_mask[batch_index, non_attended_tokens] = 0\r\n# attention_mask.size() = first_layer_past_key_value.size() + 1\r\n```\r\nor\r\n\r\n```python\r\nextended_attention_mask = torch.ones(\r\n (attention_mask.shape[0], target_seqlen - attention_mask.shape[1]),\r\n dtype=attention_mask.dtype,\r\n device=attention_mask.device,\r\n)\r\n\r\nvalid_indices = non_attended_tokens >= attention_mask.size(-1)\r\nnew_batch_index = batch_index[valid_indices]\r\nnew_non_attended_tokens = non_attended_tokens[valid_indices]\r\n\r\n# Zero-out the places where we don't need to attend\r\nextended_attention_mask[new_batch_index, new_non_attended_tokens - attention_mask.size(-1)] = 0\r\n```\r\nmight be better", "Thanks @NicholasCao !\r\nCan you file a separate issue for this and tag me? If you can also provide a reproducer it would be great. This PR that got merged fixes the issue explained here: https://github.com/huggingface/transformers/pull/28032#issuecomment-1860650043 which hopefully should cover most of the issues related with llava and index errors", "@NicholasCao I managed to repro your issue that seems to happen in the case one passes a custom past key value, which is the case for AWQ. It should be fixed in https://github.com/huggingface/transformers/pull/28239", "I'm not using awq, I'm having this problem when I'm batch inference images, it's harder to reproduce to find the specific image", "@NicholasCao #28239 should solve it, let me know if the PR fixes your issue", "thx, it works", "Thanks @NicholasCao !" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28032/reactions", "total_count": 2, "+1": 2, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28032", "html_url": "https://github.com/huggingface/transformers/pull/28032", "diff_url": "https://github.com/huggingface/transformers/pull/28032.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28032.patch", "merged_at": 1703263658000 }

{ "login": "younesbelkada", "id": 49240599, "node_id": "MDQ6VXNlcjQ5MjQwNTk5", "avatar_url": "https://avatars.githubusercontent.com/u/49240599?v=4", "gravatar_id": "", "url": "https://api.github.com/users/younesbelkada", "html_url": "https://github.com/younesbelkada", "followers_url": "https://api.github.com/users/younesbelkada/followers", "following_url": "https://api.github.com/users/younesbelkada/following{/other_user}", "gists_url": "https://api.github.com/users/younesbelkada/gists{/gist_id}", "starred_url": "https://api.github.com/users/younesbelkada/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/younesbelkada/subscriptions", "organizations_url": "https://api.github.com/users/younesbelkada/orgs", "repos_url": "https://api.github.com/users/younesbelkada/repos", "events_url": "https://api.github.com/users/younesbelkada/events{/privacy}", "received_events_url": "https://api.github.com/users/younesbelkada/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28031/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28031", "html_url": "https://github.com/huggingface/transformers/pull/28031", "diff_url": "https://github.com/huggingface/transformers/pull/28031.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28031.patch", "merged_at": 1702552785000 }

{ "login": "gante", "id": 12240844, "node_id": "MDQ6VXNlcjEyMjQwODQ0", "avatar_url": "https://avatars.githubusercontent.com/u/12240844?v=4", "gravatar_id": "", "url": "https://api.github.com/users/gante", "html_url": "https://github.com/gante", "followers_url": "https://api.github.com/users/gante/followers", "following_url": "https://api.github.com/users/gante/following{/other_user}", "gists_url": "https://api.github.com/users/gante/gists{/gist_id}", "starred_url": "https://api.github.com/users/gante/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/gante/subscriptions", "organizations_url": "https://api.github.com/users/gante/orgs", "repos_url": "https://api.github.com/users/gante/repos", "events_url": "https://api.github.com/users/gante/events{/privacy}", "received_events_url": "https://api.github.com/users/gante/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28030). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "@amyeroberts the failing test is also failing in the daily CI (i.e. unrelated to this PR, as it doesn't depend on assisted generation), and I can't reproduce it on my end 🤔 ", "@amyeroberts I can't merge due to the failing test (which is also failing on `main` 👀 ). Would you be able to merge?", "(the test is flaky 👉 https://github.com/huggingface/transformers/pull/28035)", "@gante In this case we can merge :) \r\n\r\nedit: note this was discussed offline as the reason for failing tests was identified and confirmed as independent from this PR " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28030/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28030", "html_url": "https://github.com/huggingface/transformers/pull/28030", "diff_url": "https://github.com/huggingface/transformers/pull/28030.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28030.patch", "merged_at": 1702560674000 }

{ "login": "ydshieh", "id": 2521628, "node_id": "MDQ6VXNlcjI1MjE2Mjg=", "avatar_url": "https://avatars.githubusercontent.com/u/2521628?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ydshieh", "html_url": "https://github.com/ydshieh", "followers_url": "https://api.github.com/users/ydshieh/followers", "following_url": "https://api.github.com/users/ydshieh/following{/other_user}", "gists_url": "https://api.github.com/users/ydshieh/gists{/gist_id}", "starred_url": "https://api.github.com/users/ydshieh/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ydshieh/subscriptions", "organizations_url": "https://api.github.com/users/ydshieh/orgs", "repos_url": "https://api.github.com/users/ydshieh/repos", "events_url": "https://api.github.com/users/ydshieh/events{/privacy}", "received_events_url": "https://api.github.com/users/ydshieh/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28029/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28029", "html_url": "https://github.com/huggingface/transformers/pull/28029", "diff_url": "https://github.com/huggingface/transformers/pull/28029.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28029.patch", "merged_at": 1702554240000 }

{ "login": "moufuyu", "id": 62968285, "node_id": "MDQ6VXNlcjYyOTY4Mjg1", "avatar_url": "https://avatars.githubusercontent.com/u/62968285?v=4", "gravatar_id": "", "url": "https://api.github.com/users/moufuyu", "html_url": "https://github.com/moufuyu", "followers_url": "https://api.github.com/users/moufuyu/followers", "following_url": "https://api.github.com/users/moufuyu/following{/other_user}", "gists_url": "https://api.github.com/users/moufuyu/gists{/gist_id}", "starred_url": "https://api.github.com/users/moufuyu/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/moufuyu/subscriptions", "organizations_url": "https://api.github.com/users/moufuyu/orgs", "repos_url": "https://api.github.com/users/moufuyu/repos", "events_url": "https://api.github.com/users/moufuyu/events{/privacy}", "received_events_url": "https://api.github.com/users/moufuyu/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @SunMarc ", "Hi @moufuyu \r\nThanks for the issue and for your interest in using this feature! \r\nCurrently autoawq on pypi is broken with transformers 4.36.0, I fixed the issue on AutoAWQ side with https://github.com/casper-hansen/AutoAWQ/pull/244 but there is no release yet. To fix your issue you can either\r\n- downgrade transformers to 4.35.2 `pip install -U transformers==4.35.0`\r\n- install auto-awq from main branch", "cc @casper-hansen do you think it would make sense to do a patch release to include https://github.com/casper-hansen/AutoAWQ/pull/244 ? 🙏 Currently users cannot perform fused modules + auto-awq, they either need to compile autoawq from source or use 4.35.2", "Users should be on 4.35.2 if they want to use AutoAWQ. Unfortunately, transformers broke AutoAWQ quite deeply with the latest changes. AutoAWQ was broken in 4/5 of the last minor releases of transformers, so bear with us while we fix it. At the same time, Mixtral was released which is a time-consuming task to support. So users will have to wait until the next version of AutoAWQ unless transformers could add backward compatibility.", "We'll make a patch as soon as the fixes are merged! 🤗 ", "Thank you for your comments and your efforts to fix this.\r\n\r\nI ran again with transformers==4.35.2 and got another error.\r\nSince I am using the AWQ model (\"TheBloke/Mistral-7B-OpenOrca-AWQ\") I guess this error is not expected.\r\n```\r\n---------------------------------------------------------------------------\r\nValueError Traceback (most recent call last)\r\nCell In[7], line 13\r\n 5 torch_device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\r\n 7 quantization_config = AwqConfig(\r\n 8 bits=4,\r\n 9 do_fuse=True,\r\n 10 fuse_max_seq_len=512,\r\n 11 )\r\n---> 13 model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=quantization_config).to(0)\r\n 14 tokenizer = AutoTokenizer.from_pretrained(model_id)\r\n 16 streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)\r\n\r\nFile ~/anaconda3/envs/pytorch_p310/lib/python3.10/site-packages/transformers/models/auto/auto_factory.py:566, in _BaseAutoModelClass.from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs)\r\n 564 elif type(config) in cls._model_mapping.keys():\r\n 565 model_class = _get_model_class(config, cls._model_mapping)\r\n--> 566 return model_class.from_pretrained(\r\n 567 pretrained_model_name_or_path, *model_args, config=config, **hub_kwargs, **kwargs\r\n 568 )\r\n 569 raise ValueError(\r\n 570 f\"Unrecognized configuration class {config.__class__} for this kind of AutoModel: {cls.__name__}.\\n\"\r\n 571 f\"Model type should be one of {', '.join(c.__name__ for c in cls._model_mapping.keys())}.\"\r\n 572 )\r\n\r\nFile ~/anaconda3/envs/pytorch_p310/lib/python3.10/site-packages/transformers/modeling_utils.py:2685, in PreTrainedModel.from_pretrained(cls, pretrained_model_name_or_path, config, cache_dir, ignore_mismatched_sizes, force_download, local_files_only, token, revision, use_safetensors, *model_args, **kwargs)\r\n 2680 quantization_method_from_args = getattr(\r\n 2681 quantization_config, \"quant_method\", QuantizationMethod.BITS_AND_BYTES\r\n 2682 )\r\n 2684 if quantization_method_from_args == QuantizationMethod.AWQ:\r\n-> 2685 raise ValueError(\r\n 2686 \"You cannot pass an `AwqConfig` when loading a model as you can only use AWQ models\"\r\n 2687 \" for inference. To quantize transformers models with AWQ algorithm, please refer to our\"\r\n 2688 \" quantization docs: https://huggingface.co./docs/transformers/main_classes/quantization \"\r\n 2689 )\r\n 2691 if quantization_config is None and (load_in_8bit or load_in_4bit):\r\n 2692 quantization_method_from_args = QuantizationMethod.BITS_AND_BYTES\r\n\r\nValueError: You cannot pass an `AwqConfig` when loading a model as you can only use AWQ models for inference. To quantize transformers models with AWQ algorithm, please refer to our quantization docs: https://huggingface.co./docs/transformers/main_classes/quantization \r\n```", "Hi @moufuyu \r\nThanks very much for your patience, and sorry I was wrong, actually fused modules was introduced after 4.35.2\r\nCan you try with:\r\n```bash\r\npip install -U git+https://github.com/huggingface/transformers.git@fdb85be40fa255c015819e711c15117c2aaa5101\r\n```\r\nOnce autoawq makes a release you'll be able to switch to transformers>=4.36.0", "I have confirmed that the fused module is working properly.\r\nThank you very much for your kind attention.", "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/main/CONTRIBUTING.md) are likely to be ignored.", "Closing this as it seems resolved! Now everything should work fine under latest autoawq package together with transformers main" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28028/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "tszdanger", "id": 35394351, "node_id": "MDQ6VXNlcjM1Mzk0MzUx", "avatar_url": "https://avatars.githubusercontent.com/u/35394351?v=4", "gravatar_id": "", "url": "https://api.github.com/users/tszdanger", "html_url": "https://github.com/tszdanger", "followers_url": "https://api.github.com/users/tszdanger/followers", "following_url": "https://api.github.com/users/tszdanger/following{/other_user}", "gists_url": "https://api.github.com/users/tszdanger/gists{/gist_id}", "starred_url": "https://api.github.com/users/tszdanger/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/tszdanger/subscriptions", "organizations_url": "https://api.github.com/users/tszdanger/orgs", "repos_url": "https://api.github.com/users/tszdanger/repos", "events_url": "https://api.github.com/users/tszdanger/events{/privacy}", "received_events_url": "https://api.github.com/users/tszdanger/received_events", "type": "User", "site_admin": false }

[ { "id": 5616426447, "node_id": "LA_kwDOCUB6oc8AAAABTsPdzw", "url": "https://api.github.com/repos/huggingface/transformers/labels/solved", "name": "solved", "color": "B1D6DC", "default": false, "description": "" } ]

[]

[ "See this PR: https://github.com/huggingface/transformers/pull/28009\r\nIt can help overcome this issue.", "Glad to see this issue has been fixed :) ", "> See this PR: #28009 It can help overcome this issue.\r\n\r\nit not worked!", "@jiezhangGt could you provide a reproduer / make sure you are using the latest version of main? " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28027/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "ArthurZucker", "id": 48595927, "node_id": "MDQ6VXNlcjQ4NTk1OTI3", "avatar_url": "https://avatars.githubusercontent.com/u/48595927?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ArthurZucker", "html_url": "https://github.com/ArthurZucker", "followers_url": "https://api.github.com/users/ArthurZucker/followers", "following_url": "https://api.github.com/users/ArthurZucker/following{/other_user}", "gists_url": "https://api.github.com/users/ArthurZucker/gists{/gist_id}", "starred_url": "https://api.github.com/users/ArthurZucker/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ArthurZucker/subscriptions", "organizations_url": "https://api.github.com/users/ArthurZucker/orgs", "repos_url": "https://api.github.com/users/ArthurZucker/repos", "events_url": "https://api.github.com/users/ArthurZucker/events{/privacy}", "received_events_url": "https://api.github.com/users/ArthurZucker/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28026/reactions", "total_count": 1, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 1, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28026", "html_url": "https://github.com/huggingface/transformers/pull/28026", "diff_url": "https://github.com/huggingface/transformers/pull/28026.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28026.patch", "merged_at": 1702539970000 }

{ "login": "rangehow", "id": 88258534, "node_id": "MDQ6VXNlcjg4MjU4NTM0", "avatar_url": "https://avatars.githubusercontent.com/u/88258534?v=4", "gravatar_id": "", "url": "https://api.github.com/users/rangehow", "html_url": "https://github.com/rangehow", "followers_url": "https://api.github.com/users/rangehow/followers", "following_url": "https://api.github.com/users/rangehow/following{/other_user}", "gists_url": "https://api.github.com/users/rangehow/gists{/gist_id}", "starred_url": "https://api.github.com/users/rangehow/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/rangehow/subscriptions", "organizations_url": "https://api.github.com/users/rangehow/orgs", "repos_url": "https://api.github.com/users/rangehow/repos", "events_url": "https://api.github.com/users/rangehow/events{/privacy}", "received_events_url": "https://api.github.com/users/rangehow/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @rangehow, thanks for raising this issue! \r\n\r\nAre you sure the weights save in `dir_to_bert_ckpt` aren't just gaussian initialized with bias at 0 and that's what's being loaded in? If I make my own dummy model, following the pattern in the example, the weights for the `self.bert` layer are loaded in, as expected, from the checkpoint and the weights for `self.llama` and `self.lm_head` are randomly initialized, as expected.\r\n\r\nRunning: \r\n\r\n```py\r\nfrom transformers import BertModel, LlamaModel, LlamaConfig, LlamaPreTrainedModel\r\nfrom torch import nn\r\n\r\nclass Foo(LlamaPreTrainedModel):\r\n def __init_(self, config, *model_args, **model_kwargs):\r\n super().__init__(config)\r\n self.llama = LlamaModel(config)\r\n self.bert = BertModel.from_pretrained('bert-base-uncased')\r\n self.lm_head = nn.Linear(config.hidden_size, config.vocab_size)\r\n self.post_init()\r\n\r\nconfig = LlamaConfig()\r\nmodel_0 = Foo(config)\r\nmodel_1 = Foo(config)\r\n\r\nsame_params = set()\r\ndifferent_params = set()\r\nfor ((name_0, param_0), (name_1, param_1)) in zip(model_0.named_parameters(), model_1.named_parameters()):\r\n assert name_0 == name_1\r\n if not (param_0 == param_1).all():\r\n different_params.add(name_0)\r\n else:\r\n same_params.add(name_0)\r\n``` \r\n\r\n<details><summary>Same params</summary>{'bert.embeddings.LayerNorm.bias',\r\n 'bert.embeddings.LayerNorm.weight',\r\n 'bert.embeddings.position_embeddings.weight',\r\n 'bert.embeddings.token_type_embeddings.weight',\r\n 'bert.embeddings.word_embeddings.weight',\r\n 'bert.encoder.layer.0.attention.output.LayerNorm.bias',\r\n 'bert.encoder.layer.0.attention.output.LayerNorm.weight',\r\n 'bert.encoder.layer.0.attention.output.dense.bias',\r\n 'bert.encoder.layer.0.attention.output.dense.weight',\r\n 'bert.encoder.layer.0.attention.self.key.bias',\r\n 'bert.encoder.layer.0.attention.self.key.weight',\r\n 'bert.encoder.layer.0.attention.self.query.bias',\r\n 'bert.encoder.layer.0.attention.self.query.weight',\r\n 'bert.encoder.layer.0.attention.self.value.bias',\r\n 'bert.encoder.layer.0.attention.self.value.weight',\r\n 'bert.encoder.layer.0.intermediate.dense.bias',\r\n 'bert.encoder.layer.0.intermediate.dense.weight',\r\n 'bert.encoder.layer.0.output.LayerNorm.bias',\r\n 'bert.encoder.layer.0.output.LayerNorm.weight',\r\n 'bert.encoder.layer.0.output.dense.bias',\r\n 'bert.encoder.layer.0.output.dense.weight',\r\n 'bert.encoder.layer.1.attention.output.LayerNorm.bias',\r\n 'bert.encoder.layer.1.attention.output.LayerNorm.weight',\r\n 'bert.encoder.layer.1.attention.output.dense.bias',\r\n 'bert.encoder.layer.1.attention.output.dense.weight',\r\n 'bert.encoder.layer.1.attention.self.key.bias',\r\n 'bert.encoder.layer.1.attention.self.key.weight',\r\n 'bert.encoder.layer.1.attention.self.query.bias',\r\n 'bert.encoder.layer.1.attention.self.query.weight',\r\n 'bert.encoder.layer.1.attention.self.value.bias',\r\n 'bert.encoder.layer.1.attention.self.value.weight',\r\n 'bert.encoder.layer.1.intermediate.dense.bias',\r\n 'bert.encoder.layer.1.intermediate.dense.weight',\r\n 'bert.encoder.layer.1.output.LayerNorm.bias',\r\n 'bert.encoder.layer.1.output.LayerNorm.weight',\r\n 'bert.encoder.layer.1.output.dense.bias',\r\n 'bert.encoder.layer.1.output.dense.weight',\r\n 'bert.encoder.layer.10.attention.output.LayerNorm.bias',\r\n 'bert.encoder.layer.10.attention.output.LayerNorm.weight',\r\n 'bert.encoder.layer.10.attention.output.dense.bias',\r\n 'bert.encoder.layer.10.attention.output.dense.weight',\r\n 'bert.encoder.layer.10.attention.self.key.bias',\r\n 'bert.encoder.layer.10.attention.self.key.weight',\r\n 'bert.encoder.layer.10.attention.self.query.bias',\r\n 'bert.encoder.layer.10.attention.self.query.weight',\r\n 'bert.encoder.layer.10.attention.self.value.bias',\r\n 'bert.encoder.layer.10.attention.self.value.weight',\r\n 'bert.encoder.layer.10.intermediate.dense.bias',\r\n 'bert.encoder.layer.10.intermediate.dense.weight',\r\n 'bert.encoder.layer.10.output.LayerNorm.bias',\r\n 'bert.encoder.layer.10.output.LayerNorm.weight',\r\n 'bert.encoder.layer.10.output.dense.bias',\r\n 'bert.encoder.layer.10.output.dense.weight',\r\n 'bert.encoder.layer.11.attention.output.LayerNorm.bias',\r\n 'bert.encoder.layer.11.attention.output.LayerNorm.weight',\r\n 'bert.encoder.layer.11.attention.output.dense.bias',\r\n 'bert.encoder.layer.11.attention.output.dense.weight',\r\n 'bert.encoder.layer.11.attention.self.key.bias',\r\n 'bert.encoder.layer.11.attention.self.key.weight',\r\n 'bert.encoder.layer.11.attention.self.query.bias',\r\n 'bert.encoder.layer.11.attention.self.query.weight',\r\n 'bert.encoder.layer.11.attention.self.value.bias',\r\n 'bert.encoder.layer.11.attention.self.value.weight',\r\n 'bert.encoder.layer.11.intermediate.dense.bias',\r\n 'bert.encoder.layer.11.intermediate.dense.weight',\r\n 'bert.encoder.layer.11.output.LayerNorm.bias',\r\n 'bert.encoder.layer.11.output.LayerNorm.weight',\r\n 'bert.encoder.layer.11.output.dense.bias',\r\n 'bert.encoder.layer.11.output.dense.weight',\r\n 'bert.encoder.layer.2.attention.output.LayerNorm.bias',\r\n 'bert.encoder.layer.2.attention.output.LayerNorm.weight',\r\n 'bert.encoder.layer.2.attention.output.dense.bias',\r\n 'bert.encoder.layer.2.attention.output.dense.weight',\r\n 'bert.encoder.layer.2.attention.self.key.bias',\r\n 'bert.encoder.layer.2.attention.self.key.weight',\r\n 'bert.encoder.layer.2.attention.self.query.bias',\r\n 'bert.encoder.layer.2.attention.self.query.weight',\r\n 'bert.encoder.layer.2.attention.self.value.bias',\r\n 'bert.encoder.layer.2.attention.self.value.weight',\r\n 'bert.encoder.layer.2.intermediate.dense.bias',\r\n 'bert.encoder.layer.2.intermediate.dense.weight',\r\n 'bert.encoder.layer.2.output.LayerNorm.bias',\r\n 'bert.encoder.layer.2.output.LayerNorm.weight',\r\n 'bert.encoder.layer.2.output.dense.bias',\r\n 'bert.encoder.layer.2.output.dense.weight',\r\n 'bert.encoder.layer.3.attention.output.LayerNorm.bias',\r\n 'bert.encoder.layer.3.attention.output.LayerNorm.weight',\r\n 'bert.encoder.layer.3.attention.output.dense.bias',\r\n 'bert.encoder.layer.3.attention.output.dense.weight',\r\n 'bert.encoder.layer.3.attention.self.key.bias',\r\n 'bert.encoder.layer.3.attention.self.key.weight',\r\n 'bert.encoder.layer.3.attention.self.query.bias',\r\n 'bert.encoder.layer.3.attention.self.query.weight',\r\n 'bert.encoder.layer.3.attention.self.value.bias',\r\n 'bert.encoder.layer.3.attention.self.value.weight',\r\n 'bert.encoder.layer.3.intermediate.dense.bias',\r\n 'bert.encoder.layer.3.intermediate.dense.weight',\r\n 'bert.encoder.layer.3.output.LayerNorm.bias',\r\n 'bert.encoder.layer.3.output.LayerNorm.weight',\r\n 'bert.encoder.layer.3.output.dense.bias',\r\n 'bert.encoder.layer.3.output.dense.weight',\r\n 'bert.encoder.layer.4.attention.output.LayerNorm.bias',\r\n 'bert.encoder.layer.4.attention.output.LayerNorm.weight',\r\n 'bert.encoder.layer.4.attention.output.dense.bias',\r\n 'bert.encoder.layer.4.attention.output.dense.weight',\r\n 'bert.encoder.layer.4.attention.self.key.bias',\r\n 'bert.encoder.layer.4.attention.self.key.weight',\r\n 'bert.encoder.layer.4.attention.self.query.bias',\r\n 'bert.encoder.layer.4.attention.self.query.weight',\r\n 'bert.encoder.layer.4.attention.self.value.bias',\r\n 'bert.encoder.layer.4.attention.self.value.weight',\r\n 'bert.encoder.layer.4.intermediate.dense.bias',\r\n 'bert.encoder.layer.4.intermediate.dense.weight',\r\n 'bert.encoder.layer.4.output.LayerNorm.bias',\r\n 'bert.encoder.layer.4.output.LayerNorm.weight',\r\n 'bert.encoder.layer.4.output.dense.bias',\r\n 'bert.encoder.layer.4.output.dense.weight',\r\n 'bert.encoder.layer.5.attention.output.LayerNorm.bias',\r\n 'bert.encoder.layer.5.attention.output.LayerNorm.weight',\r\n 'bert.encoder.layer.5.attention.output.dense.bias',\r\n 'bert.encoder.layer.5.attention.output.dense.weight',\r\n 'bert.encoder.layer.5.attention.self.key.bias',\r\n 'bert.encoder.layer.5.attention.self.key.weight',\r\n 'bert.encoder.layer.5.attention.self.query.bias',\r\n 'bert.encoder.layer.5.attention.self.query.weight',\r\n 'bert.encoder.layer.5.attention.self.value.bias',\r\n 'bert.encoder.layer.5.attention.self.value.weight',\r\n 'bert.encoder.layer.5.intermediate.dense.bias',\r\n 'bert.encoder.layer.5.intermediate.dense.weight',\r\n 'bert.encoder.layer.5.output.LayerNorm.bias',\r\n 'bert.encoder.layer.5.output.LayerNorm.weight',\r\n 'bert.encoder.layer.5.output.dense.bias',\r\n 'bert.encoder.layer.5.output.dense.weight',\r\n 'bert.encoder.layer.6.attention.output.LayerNorm.bias',\r\n 'bert.encoder.layer.6.attention.output.LayerNorm.weight',\r\n 'bert.encoder.layer.6.attention.output.dense.bias',\r\n 'bert.encoder.layer.6.attention.output.dense.weight',\r\n 'bert.encoder.layer.6.attention.self.key.bias',\r\n 'bert.encoder.layer.6.attention.self.key.weight',\r\n 'bert.encoder.layer.6.attention.self.query.bias',\r\n 'bert.encoder.layer.6.attention.self.query.weight',\r\n 'bert.encoder.layer.6.attention.self.value.bias',\r\n 'bert.encoder.layer.6.attention.self.value.weight',\r\n 'bert.encoder.layer.6.intermediate.dense.bias',\r\n 'bert.encoder.layer.6.intermediate.dense.weight',\r\n 'bert.encoder.layer.6.output.LayerNorm.bias',\r\n 'bert.encoder.layer.6.output.LayerNorm.weight',\r\n 'bert.encoder.layer.6.output.dense.bias',\r\n 'bert.encoder.layer.6.output.dense.weight',\r\n 'bert.encoder.layer.7.attention.output.LayerNorm.bias',\r\n 'bert.encoder.layer.7.attention.output.LayerNorm.weight',\r\n 'bert.encoder.layer.7.attention.output.dense.bias',\r\n 'bert.encoder.layer.7.attention.output.dense.weight',\r\n 'bert.encoder.layer.7.attention.self.key.bias',\r\n 'bert.encoder.layer.7.attention.self.key.weight',\r\n 'bert.encoder.layer.7.attention.self.query.bias',\r\n 'bert.encoder.layer.7.attention.self.query.weight',\r\n 'bert.encoder.layer.7.attention.self.value.bias',\r\n 'bert.encoder.layer.7.attention.self.value.weight',\r\n 'bert.encoder.layer.7.intermediate.dense.bias',\r\n 'bert.encoder.layer.7.intermediate.dense.weight',\r\n 'bert.encoder.layer.7.output.LayerNorm.bias',\r\n 'bert.encoder.layer.7.output.LayerNorm.weight',\r\n 'bert.encoder.layer.7.output.dense.bias',\r\n 'bert.encoder.layer.7.output.dense.weight',\r\n 'bert.encoder.layer.8.attention.output.LayerNorm.bias',\r\n 'bert.encoder.layer.8.attention.output.LayerNorm.weight',\r\n 'bert.encoder.layer.8.attention.output.dense.bias',\r\n 'bert.encoder.layer.8.attention.output.dense.weight',\r\n 'bert.encoder.layer.8.attention.self.key.bias',\r\n 'bert.encoder.layer.8.attention.self.key.weight',\r\n 'bert.encoder.layer.8.attention.self.query.bias',\r\n 'bert.encoder.layer.8.attention.self.query.weight',\r\n 'bert.encoder.layer.8.attention.self.value.bias',\r\n 'bert.encoder.layer.8.attention.self.value.weight',\r\n 'bert.encoder.layer.8.intermediate.dense.bias',\r\n 'bert.encoder.layer.8.intermediate.dense.weight',\r\n 'bert.encoder.layer.8.output.LayerNorm.bias',\r\n 'bert.encoder.layer.8.output.LayerNorm.weight',\r\n 'bert.encoder.layer.8.output.dense.bias',\r\n 'bert.encoder.layer.8.output.dense.weight',\r\n 'bert.encoder.layer.9.attention.output.LayerNorm.bias',\r\n 'bert.encoder.layer.9.attention.output.LayerNorm.weight',\r\n 'bert.encoder.layer.9.attention.output.dense.bias',\r\n 'bert.encoder.layer.9.attention.output.dense.weight',\r\n 'bert.encoder.layer.9.attention.self.key.bias',\r\n 'bert.encoder.layer.9.attention.self.key.weight',\r\n 'bert.encoder.layer.9.attention.self.query.bias',\r\n 'bert.encoder.layer.9.attention.self.query.weight',\r\n 'bert.encoder.layer.9.attention.self.value.bias',\r\n 'bert.encoder.layer.9.attention.self.value.weight',\r\n 'bert.encoder.layer.9.intermediate.dense.bias',\r\n 'bert.encoder.layer.9.intermediate.dense.weight',\r\n 'bert.encoder.layer.9.output.LayerNorm.bias',\r\n 'bert.encoder.layer.9.output.LayerNorm.weight',\r\n 'bert.encoder.layer.9.output.dense.bias',\r\n 'bert.encoder.layer.9.output.dense.weight',\r\n 'bert.pooler.dense.bias',\r\n 'bert.pooler.dense.weight',\r\n 'llama.layers.0.input_layernorm.weight',\r\n 'llama.layers.0.post_attention_layernorm.weight',\r\n 'llama.layers.1.input_layernorm.weight',\r\n 'llama.layers.1.post_attention_layernorm.weight',\r\n 'llama.layers.10.input_layernorm.weight',\r\n 'llama.layers.10.post_attention_layernorm.weight',\r\n 'llama.layers.11.input_layernorm.weight',\r\n 'llama.layers.11.post_attention_layernorm.weight',\r\n 'llama.layers.12.input_layernorm.weight',\r\n 'llama.layers.12.post_attention_layernorm.weight',\r\n 'llama.layers.13.input_layernorm.weight',\r\n 'llama.layers.13.post_attention_layernorm.weight',\r\n 'llama.layers.14.input_layernorm.weight',\r\n 'llama.layers.14.post_attention_layernorm.weight',\r\n 'llama.layers.15.input_layernorm.weight',\r\n 'llama.layers.15.post_attention_layernorm.weight',\r\n 'llama.layers.16.input_layernorm.weight',\r\n 'llama.layers.16.post_attention_layernorm.weight',\r\n 'llama.layers.17.input_layernorm.weight',\r\n 'llama.layers.17.post_attention_layernorm.weight',\r\n 'llama.layers.18.input_layernorm.weight',\r\n 'llama.layers.18.post_attention_layernorm.weight',\r\n 'llama.layers.19.input_layernorm.weight',\r\n 'llama.layers.19.post_attention_layernorm.weight',\r\n 'llama.layers.2.input_layernorm.weight',\r\n 'llama.layers.2.post_attention_layernorm.weight',\r\n 'llama.layers.20.input_layernorm.weight',\r\n 'llama.layers.20.post_attention_layernorm.weight',\r\n 'llama.layers.21.input_layernorm.weight',\r\n 'llama.layers.21.post_attention_layernorm.weight',\r\n 'llama.layers.22.input_layernorm.weight',\r\n 'llama.layers.22.post_attention_layernorm.weight',\r\n 'llama.layers.23.input_layernorm.weight',\r\n 'llama.layers.23.post_attention_layernorm.weight',\r\n 'llama.layers.24.input_layernorm.weight',\r\n 'llama.layers.24.post_attention_layernorm.weight',\r\n 'llama.layers.25.input_layernorm.weight',\r\n 'llama.layers.25.post_attention_layernorm.weight',\r\n 'llama.layers.26.input_layernorm.weight',\r\n 'llama.layers.26.post_attention_layernorm.weight',\r\n 'llama.layers.27.input_layernorm.weight',\r\n 'llama.layers.27.post_attention_layernorm.weight',\r\n 'llama.layers.28.input_layernorm.weight',\r\n 'llama.layers.28.post_attention_layernorm.weight',\r\n 'llama.layers.29.input_layernorm.weight',\r\n 'llama.layers.29.post_attention_layernorm.weight',\r\n 'llama.layers.3.input_layernorm.weight',\r\n 'llama.layers.3.post_attention_layernorm.weight',\r\n 'llama.layers.30.input_layernorm.weight',\r\n 'llama.layers.30.post_attention_layernorm.weight',\r\n 'llama.layers.31.input_layernorm.weight',\r\n 'llama.layers.31.post_attention_layernorm.weight',\r\n 'llama.layers.4.input_layernorm.weight',\r\n 'llama.layers.4.post_attention_layernorm.weight',\r\n 'llama.layers.5.input_layernorm.weight',\r\n 'llama.layers.5.post_attention_layernorm.weight',\r\n 'llama.layers.6.input_layernorm.weight',\r\n 'llama.layers.6.post_attention_layernorm.weight',\r\n 'llama.layers.7.input_layernorm.weight',\r\n 'llama.layers.7.post_attention_layernorm.weight',\r\n 'llama.layers.8.input_layernorm.weight',\r\n 'llama.layers.8.post_attention_layernorm.weight',\r\n 'llama.layers.9.input_layernorm.weight',\r\n 'llama.layers.9.post_attention_layernorm.weight',\r\n 'llama.norm.weight'}</details>\r\n\r\n\r\n<details><summary>Different params</summary>{'llama.embed_tokens.weight',\r\n 'llama.layers.0.mlp.down_proj.weight',\r\n 'llama.layers.0.mlp.gate_proj.weight',\r\n 'llama.layers.0.mlp.up_proj.weight',\r\n 'llama.layers.0.self_attn.k_proj.weight',\r\n 'llama.layers.0.self_attn.o_proj.weight',\r\n 'llama.layers.0.self_attn.q_proj.weight',\r\n 'llama.layers.0.self_attn.v_proj.weight',\r\n 'llama.layers.1.mlp.down_proj.weight',\r\n 'llama.layers.1.mlp.gate_proj.weight',\r\n 'llama.layers.1.mlp.up_proj.weight',\r\n 'llama.layers.1.self_attn.k_proj.weight',\r\n 'llama.layers.1.self_attn.o_proj.weight',\r\n 'llama.layers.1.self_attn.q_proj.weight',\r\n 'llama.layers.1.self_attn.v_proj.weight',\r\n 'llama.layers.10.mlp.down_proj.weight',\r\n 'llama.layers.10.mlp.gate_proj.weight',\r\n 'llama.layers.10.mlp.up_proj.weight',\r\n 'llama.layers.10.self_attn.k_proj.weight',\r\n 'llama.layers.10.self_attn.o_proj.weight',\r\n 'llama.layers.10.self_attn.q_proj.weight',\r\n 'llama.layers.10.self_attn.v_proj.weight',\r\n 'llama.layers.11.mlp.down_proj.weight',\r\n 'llama.layers.11.mlp.gate_proj.weight',\r\n 'llama.layers.11.mlp.up_proj.weight',\r\n 'llama.layers.11.self_attn.k_proj.weight',\r\n 'llama.layers.11.self_attn.o_proj.weight',\r\n 'llama.layers.11.self_attn.q_proj.weight',\r\n 'llama.layers.11.self_attn.v_proj.weight',\r\n 'llama.layers.12.mlp.down_proj.weight',\r\n 'llama.layers.12.mlp.gate_proj.weight',\r\n 'llama.layers.12.mlp.up_proj.weight',\r\n 'llama.layers.12.self_attn.k_proj.weight',\r\n 'llama.layers.12.self_attn.o_proj.weight',\r\n 'llama.layers.12.self_attn.q_proj.weight',\r\n 'llama.layers.12.self_attn.v_proj.weight',\r\n 'llama.layers.13.mlp.down_proj.weight',\r\n 'llama.layers.13.mlp.gate_proj.weight',\r\n 'llama.layers.13.mlp.up_proj.weight',\r\n 'llama.layers.13.self_attn.k_proj.weight',\r\n 'llama.layers.13.self_attn.o_proj.weight',\r\n 'llama.layers.13.self_attn.q_proj.weight',\r\n 'llama.layers.13.self_attn.v_proj.weight',\r\n 'llama.layers.14.mlp.down_proj.weight',\r\n 'llama.layers.14.mlp.gate_proj.weight',\r\n 'llama.layers.14.mlp.up_proj.weight',\r\n 'llama.layers.14.self_attn.k_proj.weight',\r\n 'llama.layers.14.self_attn.o_proj.weight',\r\n 'llama.layers.14.self_attn.q_proj.weight',\r\n 'llama.layers.14.self_attn.v_proj.weight',\r\n 'llama.layers.15.mlp.down_proj.weight',\r\n 'llama.layers.15.mlp.gate_proj.weight',\r\n 'llama.layers.15.mlp.up_proj.weight',\r\n 'llama.layers.15.self_attn.k_proj.weight',\r\n 'llama.layers.15.self_attn.o_proj.weight',\r\n 'llama.layers.15.self_attn.q_proj.weight',\r\n 'llama.layers.15.self_attn.v_proj.weight',\r\n 'llama.layers.16.mlp.down_proj.weight',\r\n 'llama.layers.16.mlp.gate_proj.weight',\r\n 'llama.layers.16.mlp.up_proj.weight',\r\n 'llama.layers.16.self_attn.k_proj.weight',\r\n 'llama.layers.16.self_attn.o_proj.weight',\r\n 'llama.layers.16.self_attn.q_proj.weight',\r\n 'llama.layers.16.self_attn.v_proj.weight',\r\n 'llama.layers.17.mlp.down_proj.weight',\r\n 'llama.layers.17.mlp.gate_proj.weight',\r\n 'llama.layers.17.mlp.up_proj.weight',\r\n 'llama.layers.17.self_attn.k_proj.weight',\r\n 'llama.layers.17.self_attn.o_proj.weight',\r\n 'llama.layers.17.self_attn.q_proj.weight',\r\n 'llama.layers.17.self_attn.v_proj.weight',\r\n 'llama.layers.18.mlp.down_proj.weight',\r\n 'llama.layers.18.mlp.gate_proj.weight',\r\n 'llama.layers.18.mlp.up_proj.weight',\r\n 'llama.layers.18.self_attn.k_proj.weight',\r\n 'llama.layers.18.self_attn.o_proj.weight',\r\n 'llama.layers.18.self_attn.q_proj.weight',\r\n 'llama.layers.18.self_attn.v_proj.weight',\r\n 'llama.layers.19.mlp.down_proj.weight',\r\n 'llama.layers.19.mlp.gate_proj.weight',\r\n 'llama.layers.19.mlp.up_proj.weight',\r\n 'llama.layers.19.self_attn.k_proj.weight',\r\n 'llama.layers.19.self_attn.o_proj.weight',\r\n 'llama.layers.19.self_attn.q_proj.weight',\r\n 'llama.layers.19.self_attn.v_proj.weight',\r\n 'llama.layers.2.mlp.down_proj.weight',\r\n 'llama.layers.2.mlp.gate_proj.weight',\r\n 'llama.layers.2.mlp.up_proj.weight',\r\n 'llama.layers.2.self_attn.k_proj.weight',\r\n 'llama.layers.2.self_attn.o_proj.weight',\r\n 'llama.layers.2.self_attn.q_proj.weight',\r\n 'llama.layers.2.self_attn.v_proj.weight',\r\n 'llama.layers.20.mlp.down_proj.weight',\r\n 'llama.layers.20.mlp.gate_proj.weight',\r\n 'llama.layers.20.mlp.up_proj.weight',\r\n 'llama.layers.20.self_attn.k_proj.weight',\r\n 'llama.layers.20.self_attn.o_proj.weight',\r\n 'llama.layers.20.self_attn.q_proj.weight',\r\n 'llama.layers.20.self_attn.v_proj.weight',\r\n 'llama.layers.21.mlp.down_proj.weight',\r\n 'llama.layers.21.mlp.gate_proj.weight',\r\n 'llama.layers.21.mlp.up_proj.weight',\r\n 'llama.layers.21.self_attn.k_proj.weight',\r\n 'llama.layers.21.self_attn.o_proj.weight',\r\n 'llama.layers.21.self_attn.q_proj.weight',\r\n 'llama.layers.21.self_attn.v_proj.weight',\r\n 'llama.layers.22.mlp.down_proj.weight',\r\n 'llama.layers.22.mlp.gate_proj.weight',\r\n 'llama.layers.22.mlp.up_proj.weight',\r\n 'llama.layers.22.self_attn.k_proj.weight',\r\n 'llama.layers.22.self_attn.o_proj.weight',\r\n 'llama.layers.22.self_attn.q_proj.weight',\r\n 'llama.layers.22.self_attn.v_proj.weight',\r\n 'llama.layers.23.mlp.down_proj.weight',\r\n 'llama.layers.23.mlp.gate_proj.weight',\r\n 'llama.layers.23.mlp.up_proj.weight',\r\n 'llama.layers.23.self_attn.k_proj.weight',\r\n 'llama.layers.23.self_attn.o_proj.weight',\r\n 'llama.layers.23.self_attn.q_proj.weight',\r\n 'llama.layers.23.self_attn.v_proj.weight',\r\n 'llama.layers.24.mlp.down_proj.weight',\r\n 'llama.layers.24.mlp.gate_proj.weight',\r\n 'llama.layers.24.mlp.up_proj.weight',\r\n 'llama.layers.24.self_attn.k_proj.weight',\r\n 'llama.layers.24.self_attn.o_proj.weight',\r\n 'llama.layers.24.self_attn.q_proj.weight',\r\n 'llama.layers.24.self_attn.v_proj.weight',\r\n 'llama.layers.25.mlp.down_proj.weight',\r\n 'llama.layers.25.mlp.gate_proj.weight',\r\n 'llama.layers.25.mlp.up_proj.weight',\r\n 'llama.layers.25.self_attn.k_proj.weight',\r\n 'llama.layers.25.self_attn.o_proj.weight',\r\n 'llama.layers.25.self_attn.q_proj.weight',\r\n 'llama.layers.25.self_attn.v_proj.weight',\r\n 'llama.layers.26.mlp.down_proj.weight',\r\n 'llama.layers.26.mlp.gate_proj.weight',\r\n 'llama.layers.26.mlp.up_proj.weight',\r\n 'llama.layers.26.self_attn.k_proj.weight',\r\n 'llama.layers.26.self_attn.o_proj.weight',\r\n 'llama.layers.26.self_attn.q_proj.weight',\r\n 'llama.layers.26.self_attn.v_proj.weight',\r\n 'llama.layers.27.mlp.down_proj.weight',\r\n 'llama.layers.27.mlp.gate_proj.weight',\r\n 'llama.layers.27.mlp.up_proj.weight',\r\n 'llama.layers.27.self_attn.k_proj.weight',\r\n 'llama.layers.27.self_attn.o_proj.weight',\r\n 'llama.layers.27.self_attn.q_proj.weight',\r\n 'llama.layers.27.self_attn.v_proj.weight',\r\n 'llama.layers.28.mlp.down_proj.weight',\r\n 'llama.layers.28.mlp.gate_proj.weight',\r\n 'llama.layers.28.mlp.up_proj.weight',\r\n 'llama.layers.28.self_attn.k_proj.weight',\r\n 'llama.layers.28.self_attn.o_proj.weight',\r\n 'llama.layers.28.self_attn.q_proj.weight',\r\n 'llama.layers.28.self_attn.v_proj.weight',\r\n 'llama.layers.29.mlp.down_proj.weight',\r\n 'llama.layers.29.mlp.gate_proj.weight',\r\n 'llama.layers.29.mlp.up_proj.weight',\r\n 'llama.layers.29.self_attn.k_proj.weight',\r\n 'llama.layers.29.self_attn.o_proj.weight',\r\n 'llama.layers.29.self_attn.q_proj.weight',\r\n 'llama.layers.29.self_attn.v_proj.weight',\r\n 'llama.layers.3.mlp.down_proj.weight',\r\n 'llama.layers.3.mlp.gate_proj.weight',\r\n 'llama.layers.3.mlp.up_proj.weight',\r\n 'llama.layers.3.self_attn.k_proj.weight',\r\n 'llama.layers.3.self_attn.o_proj.weight',\r\n 'llama.layers.3.self_attn.q_proj.weight',\r\n 'llama.layers.3.self_attn.v_proj.weight',\r\n 'llama.layers.30.mlp.down_proj.weight',\r\n 'llama.layers.30.mlp.gate_proj.weight',\r\n 'llama.layers.30.mlp.up_proj.weight',\r\n 'llama.layers.30.self_attn.k_proj.weight',\r\n 'llama.layers.30.self_attn.o_proj.weight',\r\n 'llama.layers.30.self_attn.q_proj.weight',\r\n 'llama.layers.30.self_attn.v_proj.weight',\r\n 'llama.layers.31.mlp.down_proj.weight',\r\n 'llama.layers.31.mlp.gate_proj.weight',\r\n 'llama.layers.31.mlp.up_proj.weight',\r\n 'llama.layers.31.self_attn.k_proj.weight',\r\n 'llama.layers.31.self_attn.o_proj.weight',\r\n 'llama.layers.31.self_attn.q_proj.weight',\r\n 'llama.layers.31.self_attn.v_proj.weight',\r\n 'llama.layers.4.mlp.down_proj.weight',\r\n 'llama.layers.4.mlp.gate_proj.weight',\r\n 'llama.layers.4.mlp.up_proj.weight',\r\n 'llama.layers.4.self_attn.k_proj.weight',\r\n 'llama.layers.4.self_attn.o_proj.weight',\r\n 'llama.layers.4.self_attn.q_proj.weight',\r\n 'llama.layers.4.self_attn.v_proj.weight',\r\n 'llama.layers.5.mlp.down_proj.weight',\r\n 'llama.layers.5.mlp.gate_proj.weight',\r\n 'llama.layers.5.mlp.up_proj.weight',\r\n 'llama.layers.5.self_attn.k_proj.weight',\r\n 'llama.layers.5.self_attn.o_proj.weight',\r\n 'llama.layers.5.self_attn.q_proj.weight',\r\n 'llama.layers.5.self_attn.v_proj.weight',\r\n 'llama.layers.6.mlp.down_proj.weight',\r\n 'llama.layers.6.mlp.gate_proj.weight',\r\n 'llama.layers.6.mlp.up_proj.weight',\r\n 'llama.layers.6.self_attn.k_proj.weight',\r\n 'llama.layers.6.self_attn.o_proj.weight',\r\n 'llama.layers.6.self_attn.q_proj.weight',\r\n 'llama.layers.6.self_attn.v_proj.weight',\r\n 'llama.layers.7.mlp.down_proj.weight',\r\n 'llama.layers.7.mlp.gate_proj.weight',\r\n 'llama.layers.7.mlp.up_proj.weight',\r\n 'llama.layers.7.self_attn.k_proj.weight',\r\n 'llama.layers.7.self_attn.o_proj.weight',\r\n 'llama.layers.7.self_attn.q_proj.weight',\r\n 'llama.layers.7.self_attn.v_proj.weight',\r\n 'llama.layers.8.mlp.down_proj.weight',\r\n 'llama.layers.8.mlp.gate_proj.weight',\r\n 'llama.layers.8.mlp.up_proj.weight',\r\n 'llama.layers.8.self_attn.k_proj.weight',\r\n 'llama.layers.8.self_attn.o_proj.weight',\r\n 'llama.layers.8.self_attn.q_proj.weight',\r\n 'llama.layers.8.self_attn.v_proj.weight',\r\n 'llama.layers.9.mlp.down_proj.weight',\r\n 'llama.layers.9.mlp.gate_proj.weight',\r\n 'llama.layers.9.mlp.up_proj.weight',\r\n 'llama.layers.9.self_attn.k_proj.weight',\r\n 'llama.layers.9.self_attn.o_proj.weight',\r\n 'llama.layers.9.self_attn.q_proj.weight',\r\n 'llama.layers.9.self_attn.v_proj.weight',\r\n 'lm_head.weight'}</details>\r\n\r\n\r\n\r\n\r\nNote: \r\n* Your class should inherit from `LlamaPreTrainedModel`, however I tried with both `LlamaPreTrainedModel` and `LlamaForCausalLM` and it works\r\n* Layernorm params are the same, even for randomly initialized layers as they're [initially set to 1 for the weight and 0 for the bias](https://pytorch.org/docs/stable/generated/torch.nn.LayerNorm.html)", "> Hi @rangehow, thanks for raising this issue!\r\n> \r\n> Are you sure the weights save in `dir_to_bert_ckpt` aren't just gaussian initialized with bias at 0 and that's what's being loaded in? If I make my own dummy model, following the pattern in the example, the weights for the `self.bert` layer are loaded in, as expected, from the checkpoint and the weights for `self.llama` and `self.lm_head` are randomly initialized, as expected.\r\n> \r\n> Running:\r\n> \r\n> ```python\r\n> from transformers import BertModel, LlamaModel, LlamaConfig, LlamaPreTrainedModel\r\n> from torch import nn\r\n> \r\n> class Foo(LlamaPreTrainedModel):\r\n> def __init_(self, config, *model_args, **model_kwargs):\r\n> super().__init__(config)\r\n> self.llama = LlamaModel(config)\r\n> self.bert = BertModel.from_pretrained('bert-base-uncased')\r\n> self.lm_head = nn.Linear(config.hidden_size, config.vocab_size)\r\n> self.post_init()\r\n> \r\n> config = LlamaConfig()\r\n> model_0 = Foo(config)\r\n> model_1 = Foo(config)\r\n> \r\n> same_params = set()\r\n> different_params = set()\r\n> for ((name_0, param_0), (name_1, param_1)) in zip(model_0.named_parameters(), model_1.named_parameters()):\r\n> assert name_0 == name_1\r\n> if not (param_0 == param_1).all():\r\n> different_params.add(name_0)\r\n> else:\r\n> same_params.add(name_0)\r\n> ```\r\n> \r\n> Same params\r\n> Different params\r\n> Note:\r\n> \r\n> * Your class should inherit from `LlamaPreTrainedModel`, however I tried with both `LlamaPreTrainedModel` and `LlamaForCausalLM` and it works\r\n> * Layernorm params are the same, even for randomly initialized layers as they're [initially set to 1 for the weight and 0 for the bias](https://pytorch.org/docs/stable/generated/torch.nn.LayerNorm.html)\r\n\r\nSorry for the late reply and thank you for your patient help！ The first question that ckpt incorrect saved magically disappeared after I updated Transformers without editing any code (The difference I can observe is save_pretrained method default save safetensor instead pytorch file). \r\nI still have some doubts：\r\n1. Is my approach an appropriate way to combine 2 pretrained model together so that I can use transformers trainer? \r\n2. **How can I save the strange combined model after training since they have two totally different tokenizer.** The solution I can think of now is after I initialize c class and save it ，I can delete the ckpt in bert folder and change the class c code from bert.from_pretrained(bert_folder) to load config and self.bert(BertConfig) . Now I have to keep a c_folder (contains llama tokenizer, combined weights of bert and llama, llama config) and bert_folder( bert_tokenizer and bert_config). This doesn't seem like the normal form of a transformers model.\r\n\r\nI am trying to jointly training retriever (a bert like model) and llm together using transformers's trainer .Not sure if other members of the community has tried this too, because RAG is really popular now : )", "@rangehow Thanks for updating about what made things work on your end. \r\n\r\nRegarding your other questions, these are best placed in our [forums](https://discuss.huggingface.co/). We try to reserve the github issues for feature requests and bug reports. You'll be able to have in-depth discussions and find out what other members of the community have tried there. \r\n\r\nIn general:\r\n1. It depends what you mean by \"appropriate\". You can certainly combine different modules into another `PretraindeModel` subclass. However, there's a few things to be aware of: \r\n* Using `from_pretrained` means the same checkpoint will be loaded everytime. So, if you train the model, these weight will update alongside the others in the model. When you reload, the weights in the `self.llama` model will expecting values from the fine-tuned bert model. Unless you freeze the weights in `self.bert` you'll likely encounter performance issues. \r\n* The `PretrainedModel` class you choose as the parent class controls how the model weights are initialized. If you use `from_pretrained` then the pretrained weights are loaded in, but if you use the config and `LlamaPreTrainedModel` then you will end up initializing bert's weights with Llama's logic.\r\n2. If they have two different tokenizers, then why combine them into one model? If you need both tokenizers, you can look into combining them with a processor class\r\n\r\n\r\n\r\n" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28025/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "oraluben", "id": 5031346, "node_id": "MDQ6VXNlcjUwMzEzNDY=", "avatar_url": "https://avatars.githubusercontent.com/u/5031346?v=4", "gravatar_id": "", "url": "https://api.github.com/users/oraluben", "html_url": "https://github.com/oraluben", "followers_url": "https://api.github.com/users/oraluben/followers", "following_url": "https://api.github.com/users/oraluben/following{/other_user}", "gists_url": "https://api.github.com/users/oraluben/gists{/gist_id}", "starred_url": "https://api.github.com/users/oraluben/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/oraluben/subscriptions", "organizations_url": "https://api.github.com/users/oraluben/orgs", "repos_url": "https://api.github.com/users/oraluben/repos", "events_url": "https://api.github.com/users/oraluben/events{/privacy}", "received_events_url": "https://api.github.com/users/oraluben/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@oraluben `_flash_attn_2_enabled` does not exist anymore in 4.36. Can you try to load your model with `model = LlamaForCausalLM.from_config(config, attention_implementation=\"flash_attention_2\")` and could you `print(model)` as well?", "Fixed by https://github.com/huggingface/transformers/pull/28031" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28024/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "geoffreyangus", "id": 29719151, "node_id": "MDQ6VXNlcjI5NzE5MTUx", "avatar_url": "https://avatars.githubusercontent.com/u/29719151?v=4", "gravatar_id": "", "url": "https://api.github.com/users/geoffreyangus", "html_url": "https://github.com/geoffreyangus", "followers_url": "https://api.github.com/users/geoffreyangus/followers", "following_url": "https://api.github.com/users/geoffreyangus/following{/other_user}", "gists_url": "https://api.github.com/users/geoffreyangus/gists{/gist_id}", "starred_url": "https://api.github.com/users/geoffreyangus/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/geoffreyangus/subscriptions", "organizations_url": "https://api.github.com/users/geoffreyangus/orgs", "repos_url": "https://api.github.com/users/geoffreyangus/repos", "events_url": "https://api.github.com/users/geoffreyangus/events{/privacy}", "received_events_url": "https://api.github.com/users/geoffreyangus/received_events", "type": "User", "site_admin": false }

[]

{ "login": "younesbelkada", "id": 49240599, "node_id": "MDQ6VXNlcjQ5MjQwNTk5", "avatar_url": "https://avatars.githubusercontent.com/u/49240599?v=4", "gravatar_id": "", "url": "https://api.github.com/users/younesbelkada", "html_url": "https://github.com/younesbelkada", "followers_url": "https://api.github.com/users/younesbelkada/followers", "following_url": "https://api.github.com/users/younesbelkada/following{/other_user}", "gists_url": "https://api.github.com/users/younesbelkada/gists{/gist_id}", "starred_url": "https://api.github.com/users/younesbelkada/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/younesbelkada/subscriptions", "organizations_url": "https://api.github.com/users/younesbelkada/orgs", "repos_url": "https://api.github.com/users/younesbelkada/repos", "events_url": "https://api.github.com/users/younesbelkada/events{/privacy}", "received_events_url": "https://api.github.com/users/younesbelkada/received_events", "type": "User", "site_admin": false }

[ { "login": "younesbelkada", "id": 49240599, "node_id": "MDQ6VXNlcjQ5MjQwNTk5", "avatar_url": "https://avatars.githubusercontent.com/u/49240599?v=4", "gravatar_id": "", "url": "https://api.github.com/users/younesbelkada", "html_url": "https://github.com/younesbelkada", "followers_url": "https://api.github.com/users/younesbelkada/followers", "following_url": "https://api.github.com/users/younesbelkada/following{/other_user}", "gists_url": "https://api.github.com/users/younesbelkada/gists{/gist_id}", "starred_url": "https://api.github.com/users/younesbelkada/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/younesbelkada/subscriptions", "organizations_url": "https://api.github.com/users/younesbelkada/orgs", "repos_url": "https://api.github.com/users/younesbelkada/repos", "events_url": "https://api.github.com/users/younesbelkada/events{/privacy}", "received_events_url": "https://api.github.com/users/younesbelkada/received_events", "type": "User", "site_admin": false } ]

[ "hi @geoffreyangus \r\nthanks for the clean reproducer, I can confirm #28031 fixes the bug", "Ah, @younesbelkada were you able to try with Mixtral as well (the commented out MODEL_ID in the reproducer)? It still seems to be broken on my side.", "Hi @geoffreyangus \r\nThanks, indeed, just made #28061" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28023/reactions", "total_count": 1, "+1": 0, "-1": 0, "laugh": 1, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "getao", "id": 12735658, "node_id": "MDQ6VXNlcjEyNzM1NjU4", "avatar_url": "https://avatars.githubusercontent.com/u/12735658?v=4", "gravatar_id": "", "url": "https://api.github.com/users/getao", "html_url": "https://github.com/getao", "followers_url": "https://api.github.com/users/getao/followers", "following_url": "https://api.github.com/users/getao/following{/other_user}", "gists_url": "https://api.github.com/users/getao/gists{/gist_id}", "starred_url": "https://api.github.com/users/getao/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/getao/subscriptions", "organizations_url": "https://api.github.com/users/getao/orgs", "repos_url": "https://api.github.com/users/getao/repos", "events_url": "https://api.github.com/users/getao/events{/privacy}", "received_events_url": "https://api.github.com/users/getao/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Could you share a minimal reproducer that shows that there is no improvements / how you are enabling it? ", "> Could you share a minimal reproducer that shows that there is no improvements / how you are enabling it?\r\n\r\n```\r\n\r\ndef main():\r\n parser = transformers.HfArgumentParser((ModelArguments, DataArguments, TrainingArguments))\r\n model_args, data_args, training_args = parser.parse_args_into_dataclasses()\r\n\r\n print(model_args)\r\n print(data_args)\r\n print(training_args)\r\n\r\n data_prefix = data_args.data_path\r\n \r\n train_file = f\"{data_prefix}.train.json\"\r\n eval_file = f\"{data_prefix}.eval.json\"\r\n \r\n dataset = load_dataset(\"json\", data_files={\"train\": train_file, \"eval\": eval_file})\r\n train_dataset = dataset[\"train\"]\r\n eval_dataset = dataset[\"eval\"]\r\n tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path)\r\n train_dataset = train_dataset.map(tokenize_function, batched=True, fn_kwargs={\"tokenizer\": tokenizer, \"max_seq_length\": data_args.max_seq_length)\r\n eval_dataset = eval_dataset.map(tokenize_function, batched=True, fn_kwargs={\"tokenizer\": tokenizer, \"max_seq_length\": data_args.max_seq_length)\r\n\r\n model_download_flag = False\r\n while model_download_flag is False: \r\n try:\r\n model = AutoModelForCausalLM.from_pretrained(model_args.model_name_or_path, torch_dtype=torch.float16 if training_args.bf16 is False else torch.bfloat16, use_flash_attention_2=True, resume_download=True)\r\n model_download_flag = True\r\n except:\r\n print(\"Downloading incomplete... Retry\")\r\n\r\n train_model(model, train_dataset, eval_dataset, training_args)\r\n\r\nmain()\r\n\r\n```\r\n\r\nAbove is the training script for training with llama-2.\r\n\r\nThe command line is:\r\n\r\n```\r\ntorchrun --nproc-per-node=4 baseline.py --adam_beta2 0.95 --adam_epsilon 1e-6 --num_train_epochs $epoch --per_device_train_batch_size $batch --per_device_eval_batch_size $batch --gradient_accumulation_steps 1 \\\r\n --learning_rate $lr --seed $seed --data_seed $seed **--gradient_checkpointing** --logging_steps 10 --save_strategy 'no' --evaluation_strategy 'steps' --eval_steps $eval_steps --save_steps $save_steps --bf16 --max_seq_length $max_seq_len --output_dir $OUT_DIR --logging_dir $OUT_DIR/logs --data_path $INPUT_DATA | tee $OUT_DIR/train.log\r\n\r\n```", "Sorry, you're gonna have to give more details about what the `train_model` function does. If you don't use the official trainer / training script then maybe you are just not enabling it. `model.enable_gradient_checkpointing()` should be called. \r\n", "> model.enable_gradient_checkpointing()\r\n\r\nI used the official trainer:\r\n\r\n```\r\ndef train_model(model, train_dataset, eval_dataset, training_args, data_collator=None):\r\n\r\n last_checkpoint = None\r\n if os.path.isdir(training_args.output_dir):\r\n last_checkpoint = get_last_checkpoint(training_args.output_dir)\r\n if last_checkpoint is not None and training_args.resume_from_checkpoint is None:\r\n print(\r\n f\"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change \"\r\n \"the `--output_dir` or add `--overwrite_output_dir` to train from scratch.\"\r\n )\r\n\r\n trainer = Trainer(\r\n model=model,\r\n args=training_args,\r\n train_dataset=train_dataset,\r\n eval_dataset=eval_dataset,\r\n data_collator=data_collator\r\n )\r\n\r\n checkpoint = None\r\n if training_args.resume_from_checkpoint is not None:\r\n checkpoint = training_args.resume_from_checkpoint\r\n elif last_checkpoint is not None:\r\n checkpoint = last_checkpoint\r\n\r\n print(f\"Loaded from the checkpoint: {checkpoint}\")\r\n\r\n train_result = trainer.train(resume_from_checkpoint=checkpoint)\r\n \r\n trainer.save_model()\r\n trainer.log_metrics(\"train\", train_result.metrics)\r\n metrics = trainer.evaluate()\r\n trainer.log_metrics(\"eval\", metrics)\r\n trainer.save_metrics(\"eval\", metrics)\r\n```\r\n\r\nOr, do you mean that I still need to add ```model.enable_gradient_checkpointing()``` even if I use the default official Trainer?", "Not necessarily no, the code here https://github.com/huggingface/transformers/blob/04e67503d804c5f51a667e2db8e817c1d744048b/src/transformers/trainer.py#L1650 should automatically enable it if the arg is properly passed. As it is one of the most important aspect w.r.t training with peft, this is heavily tested. \r\nCould you make sure that it is set by printing the `training_args`? ", "> Not necessarily no, the code here\r\n> \r\n> https://github.com/huggingface/transformers/blob/04e67503d804c5f51a667e2db8e817c1d744048b/src/transformers/trainer.py#L1650\r\n> \r\n> should automatically enable it if the arg is properly passed. As it is one of the most important aspect w.r.t training with peft, this is heavily tested.\r\n> Could you make sure that it is set by printing the `training_args`?\r\n\r\nSure, I checked this and found it was enabled by printing training_args:\r\n\r\ngradient_checkpointing=True,\r\ngradient_checkpointing_kwargs=None,\r\n\r\nAlso, I have observed some logs showing the gradient_checkpointing is enabled before the training progress bar appears:\r\n\r\n`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\r\n\r\nUnfortunately, no GPU memory saving is observed.", "I am probably missing something so pinging @younesbelkada here! 🤗 \r\n", "hi @getao \r\nThanks for the issue! I don't think training should be faster with GC - do you observe the same behavioru without FA?", "> hi @getao Thanks for the issue! I don't think training should be faster with GC - do you observe the same behavioru without FA?\r\n\r\nHi @younesbelkada , I didn't say that GC helps training become faster. GC is thought to save memory but at a cost of slowing down training. What I said is that I found enabling GC didn't save memory and didn't slow down training, which is weird to me.\r\n\r\nBTW, What is FA?", "I see thanks for explaining, OK! \r\nSorry I meant Flash Attention, I wonder if it is not a weird interaction between Flash Attention and GC", "> I see thanks for explaining, OK! Sorry I meant Flash Attention, I wonder if it is not a weird interaction between Flash Attention and GC\r\n\r\nOh, I see. I'll take a look at it without Flash attention.", "Regarding potential weird interaction between Flash Attention and GC -- The HF gradient checkpointing will cause inefficient redundant computation when using flash attention. TL;DR: it (computationally equivalent) computes a redundant FA forward.\r\n\r\nThe reason is the flash attention backward kernel will recompute the Q^tK for rematerialization, which is similar to the recomputation in the backward process of the original HF gradient checkpointing. \r\n\r\nIf you're interested, we implemented a more efficient flash attention friendly gradient checkpointing in [FastCkpt](https://github.com/RulinShao/FastCkpt/tree/main) to mitigate this issue, where you just need to pip install fastckpt and import a monkey patch to accelerate your training by saving one flash attention forward in every layer compared with the HF gradient checkpointing.", "> Regarding potential weird interaction between Flash Attention and GC -- The HF gradient checkpointing will cause inefficient redundant computation when using flash attention. TL;DR: it (computationally equivalent) computes a redundant FA forward.\r\n> \r\n> The reason is the flash attention backward kernel will recompute the Q^tK for rematerialization, which is similar to the recomputation in the backward process of the original HF gradient checkpointing.\r\n> \r\n> If you're interested, we implemented a more efficient flash attention friendly gradient checkpointing in [FastCkpt](https://github.com/RulinShao/FastCkpt/tree/main) to mitigate this issue, where you just need to pip install fastckpt and import a monkey patch to accelerate your training by saving one flash attention forward in every layer compared with the HF gradient checkpointing.\r\n\r\nThank you for your answer. It makes sense for the speed issue. However, does it make sense that the GPU memory is not saved?", "Hello, just an update.\r\n\r\nI find that --gradient_checkpointing is highly useful to save memory when it is used with deepspeed along with the deepspeed config file. \r\n\r\nHowever, if I use it in torchrun (without any specified config), it doesn't work to save memory. I wonder if there is anything that I missed for properly enabling gradient checkpointing.\r\n\r\nAny help would be appreciated.", "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/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28022/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "lhallee", "id": 72926928, "node_id": "MDQ6VXNlcjcyOTI2OTI4", "avatar_url": "https://avatars.githubusercontent.com/u/72926928?v=4", "gravatar_id": "", "url": "https://api.github.com/users/lhallee", "html_url": "https://github.com/lhallee", "followers_url": "https://api.github.com/users/lhallee/followers", "following_url": "https://api.github.com/users/lhallee/following{/other_user}", "gists_url": "https://api.github.com/users/lhallee/gists{/gist_id}", "starred_url": "https://api.github.com/users/lhallee/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/lhallee/subscriptions", "organizations_url": "https://api.github.com/users/lhallee/orgs", "repos_url": "https://api.github.com/users/lhallee/repos", "events_url": "https://api.github.com/users/lhallee/events{/privacy}", "received_events_url": "https://api.github.com/users/lhallee/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Sorry could you either show the issue or detail where you had a problem? The computation is different because the output shape are also different, the routing mecanism is also different. 🤗 ", "Sure! @ArthurZucker\r\n\r\nHere's the current loss function for convenience\r\n```\r\ndef load_balancing_loss_func(gate_logits: torch.Tensor, num_experts: torch.Tensor = None, top_k=2) -> float:\r\n r\"\"\"\r\n Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.\r\n\r\n See Switch Transformer (https://arxiv.org/abs/2101.03961) for more details. This function implements the loss\r\n function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between\r\n experts is too unbalanced.\r\n\r\n Args:\r\n gate_logits (Union[`torch.Tensor`, Tuple[torch.Tensor]):\r\n Logits from the `gate`, should be a tuple of tensors. Shape: [batch_size, seqeunce_length, num_experts].\r\n num_experts (`int`, *optional*):\r\n Number of experts\r\n\r\n Returns:\r\n The auxiliary loss.\r\n \"\"\"\r\n if gate_logits is None:\r\n return 0\r\n\r\n if isinstance(gate_logits, tuple):\r\n # cat along the layers?\r\n gate_logits = torch.cat(gate_logits, dim=0)\r\n\r\n routing_weights, selected_experts = torch.topk(gate_logits, top_k, dim=-1)\r\n routing_weights = routing_weights.softmax(dim=-1)\r\n\r\n # cast the expert indices to int64, otherwise one-hot encoding will fail\r\n if selected_experts.dtype != torch.int64:\r\n selected_experts = selected_experts.to(torch.int64)\r\n\r\n if len(selected_experts.shape) == 2:\r\n selected_experts = selected_experts.unsqueeze(2)\r\n\r\n expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)\r\n\r\n # For a given token, determine if it was routed to a given expert.\r\n expert_mask = torch.max(expert_mask, axis=-2).values\r\n\r\n # cast to float32 otherwise mean will fail\r\n expert_mask = expert_mask.to(torch.float32)\r\n tokens_per_group_and_expert = torch.mean(expert_mask, axis=-2)\r\n\r\n router_prob_per_group_and_expert = torch.mean(routing_weights, axis=-1)\r\n return torch.mean(tokens_per_group_and_expert * router_prob_per_group_and_expert.unsqueeze(-1)) * (num_experts**2)\r\n```\r\n\r\nAn example\r\n\r\n```\r\nnum_hidden_layers=30\r\nbatch_size = 16\r\nseq_len = 32\r\nnum_experts = 8\r\ngate_logits = tuple(torch.randn(batch_size, seq_len, num_experts) for _ in range(num_hidden_layers))\r\nload_balancing_loss_func(gate_logits=gate_logits, num_experts=num_experts)\r\n```\r\nShape error\r\n\r\n```\r\n---------------------------------------------------------------------------\r\nRuntimeError Traceback (most recent call last)\r\n[c:\\Users\\Logan](file:///C:/Users/Logan) Hallee\\Desktop\\MOE-PLM\\moesm_testing.ipynb Cell 13 line 6\r\n [3](vscode-notebook-cell:/c%3A/Users/Logan%20Hallee/Desktop/MOE-PLM/moesm_testing.ipynb#X15sZmlsZQ%3D%3D?line=2) num_experts = 8\r\n [5](vscode-notebook-cell:/c%3A/Users/Logan%20Hallee/Desktop/MOE-PLM/moesm_testing.ipynb#X15sZmlsZQ%3D%3D?line=4) gate_logits = tuple(torch.randn(batch_size, seq_len, num_experts) for _ in range(30))\r\n----> [6](vscode-notebook-cell:/c%3A/Users/Logan%20Hallee/Desktop/MOE-PLM/moesm_testing.ipynb#X15sZmlsZQ%3D%3D?line=5) load_balancing_loss_func(gate_logits=gate_logits, num_experts=8)\r\n\r\n[c:\\Users\\Logan](file:///C:/Users/Logan) Hallee\\Desktop\\MOE-PLM\\moesm_testing.ipynb Cell 13 line 4\r\n [42](vscode-notebook-cell:/c%3A/Users/Logan%20Hallee/Desktop/MOE-PLM/moesm_testing.ipynb#X15sZmlsZQ%3D%3D?line=41) tokens_per_group_and_expert = torch.mean(expert_mask, axis=-2)\r\n [44](vscode-notebook-cell:/c%3A/Users/Logan%20Hallee/Desktop/MOE-PLM/moesm_testing.ipynb#X15sZmlsZQ%3D%3D?line=43) router_prob_per_group_and_expert = torch.mean(routing_weights, axis=-1)\r\n---> [45](vscode-notebook-cell:/c%3A/Users/Logan%20Hallee/Desktop/MOE-PLM/moesm_testing.ipynb#X15sZmlsZQ%3D%3D?line=44) return torch.mean(tokens_per_group_and_expert * router_prob_per_group_and_expert.unsqueeze(-1)) * (num_experts**2)\r\n\r\nRuntimeError: The size of tensor a (480) must match the size of tensor b (32) at non-singleton dimension 1\r\n```\r\n ", "The loss is made to be used with the outputs of the model, which merge batch and sequence length 😉", "It looks like the documentation is wrong then. Could you clarify where the merge happens and the correct shape of the input?", "Hello~ does this function \"load_balancing_loss_func\" really work？ It always output a constant for me.", "> Hello~ does this function \"load_balancing_loss_func\" really work？ It always output a constant for me.\r\n\r\nSame to me, and the grad norm is 0. @ArthurZucker ", "Thanks all for the feedback I'll check it and update the doc with an example! \r\nThe merge happens in the forward of the `MixtralSparseMoeBlock` here: https://github.com/huggingface/transformers/blob/cfd3e8d1e05e11b12bf50efb90691a4ad1f68926/src/transformers/models/mixtral/modeling_mixtral.py#L706", "> Thanks all for the feedback I'll check it and update the doc with an example! The merge happens in the forward of the `MixtralSparseMoeBlock` here: https://github.com/huggingface/transformers/blob/cfd3e8d1e05e11b12bf50efb90691a4ad1f68926/src/transformers/models/mixtral/modeling_mixtral.py#L706\r\n\r\nHi, have you fix the constant loss problem ?", "Yes, #28115 fixes this sorry everyone ! 🤗 " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28021/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "Lyken17", "id": 7783214, "node_id": "MDQ6VXNlcjc3ODMyMTQ=", "avatar_url": "https://avatars.githubusercontent.com/u/7783214?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Lyken17", "html_url": "https://github.com/Lyken17", "followers_url": "https://api.github.com/users/Lyken17/followers", "following_url": "https://api.github.com/users/Lyken17/following{/other_user}", "gists_url": "https://api.github.com/users/Lyken17/gists{/gist_id}", "starred_url": "https://api.github.com/users/Lyken17/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Lyken17/subscriptions", "organizations_url": "https://api.github.com/users/Lyken17/orgs", "repos_url": "https://api.github.com/users/Lyken17/repos", "events_url": "https://api.github.com/users/Lyken17/events{/privacy}", "received_events_url": "https://api.github.com/users/Lyken17/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@amyeroberts I have revised the docs, please have a check." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28020/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28020", "html_url": "https://github.com/huggingface/transformers/pull/28020", "diff_url": "https://github.com/huggingface/transformers/pull/28020.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28020.patch", "merged_at": 1702660191000 }

{ "login": "ylacombe", "id": 52246514, "node_id": "MDQ6VXNlcjUyMjQ2NTE0", "avatar_url": "https://avatars.githubusercontent.com/u/52246514?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ylacombe", "html_url": "https://github.com/ylacombe", "followers_url": "https://api.github.com/users/ylacombe/followers", "following_url": "https://api.github.com/users/ylacombe/following{/other_user}", "gists_url": "https://api.github.com/users/ylacombe/gists{/gist_id}", "starred_url": "https://api.github.com/users/ylacombe/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ylacombe/subscriptions", "organizations_url": "https://api.github.com/users/ylacombe/orgs", "repos_url": "https://api.github.com/users/ylacombe/repos", "events_url": "https://api.github.com/users/ylacombe/events{/privacy}", "received_events_url": "https://api.github.com/users/ylacombe/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Thanks for the suggestion @amyeroberts ! I've integrated them and will merge when the CI is green!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28019/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28019", "html_url": "https://github.com/huggingface/transformers/pull/28019", "diff_url": "https://github.com/huggingface/transformers/pull/28019.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28019.patch", "merged_at": 1702565024000 }

{ "login": "SunMarc", "id": 57196510, "node_id": "MDQ6VXNlcjU3MTk2NTEw", "avatar_url": "https://avatars.githubusercontent.com/u/57196510?v=4", "gravatar_id": "", "url": "https://api.github.com/users/SunMarc", "html_url": "https://github.com/SunMarc", "followers_url": "https://api.github.com/users/SunMarc/followers", "following_url": "https://api.github.com/users/SunMarc/following{/other_user}", "gists_url": "https://api.github.com/users/SunMarc/gists{/gist_id}", "starred_url": "https://api.github.com/users/SunMarc/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/SunMarc/subscriptions", "organizations_url": "https://api.github.com/users/SunMarc/orgs", "repos_url": "https://api.github.com/users/SunMarc/repos", "events_url": "https://api.github.com/users/SunMarc/events{/privacy}", "received_events_url": "https://api.github.com/users/SunMarc/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@SunMarc Do you have permissions to merge? If not, I can merge this in if it's good to go", "I'll merge it ! thx for the reminder " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28018/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28018", "html_url": "https://github.com/huggingface/transformers/pull/28018", "diff_url": "https://github.com/huggingface/transformers/pull/28018.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28018.patch", "merged_at": 1705335775000 }

{ "login": "tomaarsen", "id": 37621491, "node_id": "MDQ6VXNlcjM3NjIxNDkx", "avatar_url": "https://avatars.githubusercontent.com/u/37621491?v=4", "gravatar_id": "", "url": "https://api.github.com/users/tomaarsen", "html_url": "https://github.com/tomaarsen", "followers_url": "https://api.github.com/users/tomaarsen/followers", "following_url": "https://api.github.com/users/tomaarsen/following{/other_user}", "gists_url": "https://api.github.com/users/tomaarsen/gists{/gist_id}", "starred_url": "https://api.github.com/users/tomaarsen/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/tomaarsen/subscriptions", "organizations_url": "https://api.github.com/users/tomaarsen/orgs", "repos_url": "https://api.github.com/users/tomaarsen/repos", "events_url": "https://api.github.com/users/tomaarsen/events{/privacy}", "received_events_url": "https://api.github.com/users/tomaarsen/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@amyeroberts This should be a simple merge; but I'd like to leave it to a core maintainer as I'm not 100% sure whether non-core maintainers should be merging on `transformers`.", "We usually let HF people merge their own PRs 😉 Could you rebase on main before merging? ", "Good to know!\r\n\r\nRebased & force pushed :) ", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28017). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28017/reactions", "total_count": 1, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 1, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28017", "html_url": "https://github.com/huggingface/transformers/pull/28017", "diff_url": "https://github.com/huggingface/transformers/pull/28017.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28017.patch", "merged_at": 1705309683000 }

{ "login": "stevhliu", "id": 59462357, "node_id": "MDQ6VXNlcjU5NDYyMzU3", "avatar_url": "https://avatars.githubusercontent.com/u/59462357?v=4", "gravatar_id": "", "url": "https://api.github.com/users/stevhliu", "html_url": "https://github.com/stevhliu", "followers_url": "https://api.github.com/users/stevhliu/followers", "following_url": "https://api.github.com/users/stevhliu/following{/other_user}", "gists_url": "https://api.github.com/users/stevhliu/gists{/gist_id}", "starred_url": "https://api.github.com/users/stevhliu/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/stevhliu/subscriptions", "organizations_url": "https://api.github.com/users/stevhliu/orgs", "repos_url": "https://api.github.com/users/stevhliu/repos", "events_url": "https://api.github.com/users/stevhliu/events{/privacy}", "received_events_url": "https://api.github.com/users/stevhliu/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28016/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28016", "html_url": "https://github.com/huggingface/transformers/pull/28016", "diff_url": "https://github.com/huggingface/transformers/pull/28016.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28016.patch", "merged_at": 1702675050000 }

{ "login": "upperwal", "id": 5246435, "node_id": "MDQ6VXNlcjUyNDY0MzU=", "avatar_url": "https://avatars.githubusercontent.com/u/5246435?v=4", "gravatar_id": "", "url": "https://api.github.com/users/upperwal", "html_url": "https://github.com/upperwal", "followers_url": "https://api.github.com/users/upperwal/followers", "following_url": "https://api.github.com/users/upperwal/following{/other_user}", "gists_url": "https://api.github.com/users/upperwal/gists{/gist_id}", "starred_url": "https://api.github.com/users/upperwal/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/upperwal/subscriptions", "organizations_url": "https://api.github.com/users/upperwal/orgs", "repos_url": "https://api.github.com/users/upperwal/repos", "events_url": "https://api.github.com/users/upperwal/events{/privacy}", "received_events_url": "https://api.github.com/users/upperwal/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@upperwal see https://github.com/huggingface/transformers/pull/28009", "@thundergolfer Oh, awesome! Will close this issue. " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28015/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "jeddobson", "id": 11461294, "node_id": "MDQ6VXNlcjExNDYxMjk0", "avatar_url": "https://avatars.githubusercontent.com/u/11461294?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jeddobson", "html_url": "https://github.com/jeddobson", "followers_url": "https://api.github.com/users/jeddobson/followers", "following_url": "https://api.github.com/users/jeddobson/following{/other_user}", "gists_url": "https://api.github.com/users/jeddobson/gists{/gist_id}", "starred_url": "https://api.github.com/users/jeddobson/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jeddobson/subscriptions", "organizations_url": "https://api.github.com/users/jeddobson/orgs", "repos_url": "https://api.github.com/users/jeddobson/repos", "events_url": "https://api.github.com/users/jeddobson/events{/privacy}", "received_events_url": "https://api.github.com/users/jeddobson/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28014). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28014/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28014", "html_url": "https://github.com/huggingface/transformers/pull/28014", "diff_url": "https://github.com/huggingface/transformers/pull/28014.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28014.patch", "merged_at": 1702565523000 }

{ "login": "ShoufaChen", "id": 28682908, "node_id": "MDQ6VXNlcjI4NjgyOTA4", "avatar_url": "https://avatars.githubusercontent.com/u/28682908?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ShoufaChen", "html_url": "https://github.com/ShoufaChen", "followers_url": "https://api.github.com/users/ShoufaChen/followers", "following_url": "https://api.github.com/users/ShoufaChen/following{/other_user}", "gists_url": "https://api.github.com/users/ShoufaChen/gists{/gist_id}", "starred_url": "https://api.github.com/users/ShoufaChen/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ShoufaChen/subscriptions", "organizations_url": "https://api.github.com/users/ShoufaChen/orgs", "repos_url": "https://api.github.com/users/ShoufaChen/repos", "events_url": "https://api.github.com/users/ShoufaChen/events{/privacy}", "received_events_url": "https://api.github.com/users/ShoufaChen/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @ShoufaChen \r\nThanks for the issue,\r\nIn contrast to the official implementation, we decided to padd the language model's vocabulary size to introduce the new `<image>` token. Instead of adding 2, which is in fact sufficient, we decided to go for 64 for performance reasons as explained in this document: https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html#requirements-tc shared by @ArthurZucker here: https://huggingface.co./llava-hf/llava-1.5-7b-hf/discussions/5 \r\nLet me know if this makes sense to you", "Thanks for your detailed reply", "Thank you very much @ShoufaChen !" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28013/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "sanchit-gandhi", "id": 93869735, "node_id": "U_kgDOBZhWpw", "avatar_url": "https://avatars.githubusercontent.com/u/93869735?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sanchit-gandhi", "html_url": "https://github.com/sanchit-gandhi", "followers_url": "https://api.github.com/users/sanchit-gandhi/followers", "following_url": "https://api.github.com/users/sanchit-gandhi/following{/other_user}", "gists_url": "https://api.github.com/users/sanchit-gandhi/gists{/gist_id}", "starred_url": "https://api.github.com/users/sanchit-gandhi/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sanchit-gandhi/subscriptions", "organizations_url": "https://api.github.com/users/sanchit-gandhi/orgs", "repos_url": "https://api.github.com/users/sanchit-gandhi/repos", "events_url": "https://api.github.com/users/sanchit-gandhi/events{/privacy}", "received_events_url": "https://api.github.com/users/sanchit-gandhi/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28012/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28012", "html_url": "https://github.com/huggingface/transformers/pull/28012", "diff_url": "https://github.com/huggingface/transformers/pull/28012.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28012.patch", "merged_at": 1702637839000 }

{ "login": "ArthurZucker", "id": 48595927, "node_id": "MDQ6VXNlcjQ4NTk1OTI3", "avatar_url": "https://avatars.githubusercontent.com/u/48595927?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ArthurZucker", "html_url": "https://github.com/ArthurZucker", "followers_url": "https://api.github.com/users/ArthurZucker/followers", "following_url": "https://api.github.com/users/ArthurZucker/following{/other_user}", "gists_url": "https://api.github.com/users/ArthurZucker/gists{/gist_id}", "starred_url": "https://api.github.com/users/ArthurZucker/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ArthurZucker/subscriptions", "organizations_url": "https://api.github.com/users/ArthurZucker/orgs", "repos_url": "https://api.github.com/users/ArthurZucker/repos", "events_url": "https://api.github.com/users/ArthurZucker/events{/privacy}", "received_events_url": "https://api.github.com/users/ArthurZucker/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@ArthurZucker is there any plans to include this fix in patch release?", "I don't think so! But the next release will have it" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28011/reactions", "total_count": 1, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 1 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28011", "html_url": "https://github.com/huggingface/transformers/pull/28011", "diff_url": "https://github.com/huggingface/transformers/pull/28011.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28011.patch", "merged_at": 1702533064000 }

{ "login": "ArthurZucker", "id": 48595927, "node_id": "MDQ6VXNlcjQ4NTk1OTI3", "avatar_url": "https://avatars.githubusercontent.com/u/48595927?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ArthurZucker", "html_url": "https://github.com/ArthurZucker", "followers_url": "https://api.github.com/users/ArthurZucker/followers", "following_url": "https://api.github.com/users/ArthurZucker/following{/other_user}", "gists_url": "https://api.github.com/users/ArthurZucker/gists{/gist_id}", "starred_url": "https://api.github.com/users/ArthurZucker/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ArthurZucker/subscriptions", "organizations_url": "https://api.github.com/users/ArthurZucker/orgs", "repos_url": "https://api.github.com/users/ArthurZucker/repos", "events_url": "https://api.github.com/users/ArthurZucker/events{/privacy}", "received_events_url": "https://api.github.com/users/ArthurZucker/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28010). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "Just wanted to say this would be hugely helpful for us over at https://github.com/probcomp/hfppl !", "Likewise the ability to not include an extra SPIECE_UNDERLINE / Llama token 29871 when encoding a word with a space in front ( ` <word>`) would be huge for https://github.com/EleutherAI/lm-evaluation-harness !", "will make it for next release I hope! ", "Would love to see this as well! Thanks for working on this.", "Failing test is unrelated 😉 " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28010/reactions", "total_count": 1, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28010", "html_url": "https://github.com/huggingface/transformers/pull/28010", "diff_url": "https://github.com/huggingface/transformers/pull/28010.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28010.patch", "merged_at": 1708429832000 }

{ "login": "muellerzr", "id": 7831895, "node_id": "MDQ6VXNlcjc4MzE4OTU=", "avatar_url": "https://avatars.githubusercontent.com/u/7831895?v=4", "gravatar_id": "", "url": "https://api.github.com/users/muellerzr", "html_url": "https://github.com/muellerzr", "followers_url": "https://api.github.com/users/muellerzr/followers", "following_url": "https://api.github.com/users/muellerzr/following{/other_user}", "gists_url": "https://api.github.com/users/muellerzr/gists{/gist_id}", "starred_url": "https://api.github.com/users/muellerzr/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/muellerzr/subscriptions", "organizations_url": "https://api.github.com/users/muellerzr/orgs", "repos_url": "https://api.github.com/users/muellerzr/repos", "events_url": "https://api.github.com/users/muellerzr/events{/privacy}", "received_events_url": "https://api.github.com/users/muellerzr/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Somewhat of an aside, but there's no guarantee that a previous writer has created the directory before this point: https://github.com/huggingface/transformers/pull/28009/files#diff-ed55888e6665791fe92cc8fc0c499da54f4ace6738551cd9a2591881cda076deR2379\r\n\r\nI've seen recently that a process entering this function can skip past save operations which would create the directory and arrive at this point before another process (the 'main') has a chance to create the directory. \r\n\r\n---\r\n\r\nAlso, is `should_save` only ever `True` for a single process, the main process? If so then it's a misnomer. It's documented as: \r\n\r\n> Whether or not the current process should write to disk, e.g., to save models and checkpoints.\r\n\r\nBut in a multi-GPU scenario, multiple processes participate in disk writing against the checkpoint directory. \r\n\r\nPS. Sorry for the bug! I didn't test my original change on multi-GPU. ", "@thundergolfer re: \r\n\r\n> Whether or not the current process should write to disk, e.g., to save models and checkpoints.\r\n\r\nYes. You'll find that it's used sparingly during saving of the weights, but the internal check is that we're on process 0", "@thundergolfer re;\r\n\r\n> I've seen recently that a process entering this function can skip past save operations which would create the directory and arrive at this point before another process (the 'main') has a chance to create the directory.\r\n\r\nCan you give an example so I can contextualize the logic in the code to see where we need to fix/make the directory instead?\r\n\r\nMy best guess is you don't have a model?\r\n\r\nI can put an `os.mkdir` there instead as an option, but would be good for us to be able to write a test for it. ", "> Can you give an example...\r\n\r\nIt was happening in multi-GPU scenario when using the [axolotl](https://github.com/OpenAccess-AI-Collective/axolotl) framework. Figured it was because a non-main process was skipping all the model saving steps and doing the unconditional save of state before main process created the directory. Can look to contribute a failing test which would motivate the change 👍 ", "That would be great @thundergolfer. I'm hesitant to do too much in this one PR, since this addresses the main issue for base users of the Trainer. A follow up (possibly not part of this patch?) where we can look in-depth at it for axolotl and ensure that it doesn't break other parts would be good. ", " It works. Thanks for the quick fix. @muellerzr ", "it didn't fix the issue when multi-node training. \r\ntransformer 4.36.1\r\n'''\r\nFileNotFoundError: [Errno 2] No such file or directory: '*/tmp-checkpoint-14696' -> '*/checkpoint-14696'\r\n'''", "@dumpmemory do you have the full stack-trace and perhaps a small reproduction script? \r\n\r\nI think there's still a race condition where checking for directory existence is not atomic with the rename attempt. If it's possible that the dir has already been moved, should just attempt the rename and catch the potential exception.\r\n\r\nhttps://github.com/huggingface/transformers/pull/28009/files#diff-ed55888e6665791fe92cc8fc0c499da54f4ace6738551cd9a2591881cda076deR2390\r\n\r\nAlso it seems simpler to have only the `.should_save == True` (main process) do the rename. The rename can only succeed once, and only the main process should perform it.", "> @dumpmemory do you have the full stack-trace and perhaps a small reproduction script?\r\n> \r\n> I think there's still a race condition where checking for directory existence is not atomic with the rename attempt. If it's possible that the dir has already been moved, should just attempt the rename and catch the potential exception.\r\n> \r\n> https://github.com/huggingface/transformers/pull/28009/files#diff-ed55888e6665791fe92cc8fc0c499da54f4ace6738551cd9a2591881cda076deR2390\r\n> \r\n> Also it seems simpler to have only the `.should_save == True` (main process) do the rename. The rename can only succeed once, and only the main process should perform it.\r\n\r\nI am training with multi-gpu setting and shared file system. so each node's rank 0 process try to do the rename which is a race condition. ", "it didn't fix the issue when multi-node training.\r\ntransformer 4.36.2\r\n'''\r\nFileNotFoundError: [Errno 2] No such file or directory: '/tmp-checkpoint-5' -> '/checkpoint-5'", "> it didn't fix the issue when multi-node training. transformer 4.36.2 ''' FileNotFoundError: [Errno 2] No such file or directory: '/tmp-checkpoint-5' -> '/checkpoint-5'\r\n\r\npls check the main branch. it might be the issue of nfs" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28009/reactions", "total_count": 1, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28009", "html_url": "https://github.com/huggingface/transformers/pull/28009", "diff_url": "https://github.com/huggingface/transformers/pull/28009.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28009.patch", "merged_at": 1702487850000 }

{ "login": "wcy1122", "id": 31536861, "node_id": "MDQ6VXNlcjMxNTM2ODYx", "avatar_url": "https://avatars.githubusercontent.com/u/31536861?v=4", "gravatar_id": "", "url": "https://api.github.com/users/wcy1122", "html_url": "https://github.com/wcy1122", "followers_url": "https://api.github.com/users/wcy1122/followers", "following_url": "https://api.github.com/users/wcy1122/following{/other_user}", "gists_url": "https://api.github.com/users/wcy1122/gists{/gist_id}", "starred_url": "https://api.github.com/users/wcy1122/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/wcy1122/subscriptions", "organizations_url": "https://api.github.com/users/wcy1122/orgs", "repos_url": "https://api.github.com/users/wcy1122/repos", "events_url": "https://api.github.com/users/wcy1122/events{/privacy}", "received_events_url": "https://api.github.com/users/wcy1122/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The input images are not necessary homogenous (unlike a tensor with batch dimension along which the elements are the same format).\r\n\r\nYou can definitely use `datasets` batch processing feature to speedup the task.\r\n\r\ncc @amyeroberts to see if they has any further suggestion/comment.", "Hi @wcy1122, thanks for raising this feature request! \r\n\r\nAs @ydshieh highlights, because the input images can be of different sizes and formats, and making them batchable may or may-not happen depending on the processor's configuration. Using `map` alongside `datasets` is a great way to parallize this process.\r\n\r\nThe image processors are, admittedly and unfortunately, very slow. Their principle purpose is to make it as easy as possible to go from an image to inputs which can be fed to the model i.e. a user can quickly test and get a prediction. \r\n\r\nPart of the reason for not engineering a highly-performant image processing library is that many great libraries already exist. You'll notice that in our training examples for vision models, [we use torchvision](https://github.com/huggingface/transformers/blob/3060899be51fe1a96b12de97376f2e2b8315bc4c/examples/flax/vision/run_image_classification.py#L338) for this very reason.", "Get it. Thanks for the reply and suggestion. " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28008/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "Snarkdoof", "id": 5370689, "node_id": "MDQ6VXNlcjUzNzA2ODk=", "avatar_url": "https://avatars.githubusercontent.com/u/5370689?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Snarkdoof", "html_url": "https://github.com/Snarkdoof", "followers_url": "https://api.github.com/users/Snarkdoof/followers", "following_url": "https://api.github.com/users/Snarkdoof/following{/other_user}", "gists_url": "https://api.github.com/users/Snarkdoof/gists{/gist_id}", "starred_url": "https://api.github.com/users/Snarkdoof/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Snarkdoof/subscriptions", "organizations_url": "https://api.github.com/users/Snarkdoof/orgs", "repos_url": "https://api.github.com/users/Snarkdoof/repos", "events_url": "https://api.github.com/users/Snarkdoof/events{/privacy}", "received_events_url": "https://api.github.com/users/Snarkdoof/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @gante ", "I believe there's already an active PR for this: https://github.com/huggingface/transformers/pull/26699 However, the PR might need a little more work as it is not a polished solution yet (at least that's what I think).", "Fixed by https://github.com/huggingface/transformers/pull/28114 👍 " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28007/reactions", "total_count": 1, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "fxmarty", "id": 9808326, "node_id": "MDQ6VXNlcjk4MDgzMjY=", "avatar_url": "https://avatars.githubusercontent.com/u/9808326?v=4", "gravatar_id": "", "url": "https://api.github.com/users/fxmarty", "html_url": "https://github.com/fxmarty", "followers_url": "https://api.github.com/users/fxmarty/followers", "following_url": "https://api.github.com/users/fxmarty/following{/other_user}", "gists_url": "https://api.github.com/users/fxmarty/gists{/gist_id}", "starred_url": "https://api.github.com/users/fxmarty/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/fxmarty/subscriptions", "organizations_url": "https://api.github.com/users/fxmarty/orgs", "repos_url": "https://api.github.com/users/fxmarty/repos", "events_url": "https://api.github.com/users/fxmarty/events{/privacy}", "received_events_url": "https://api.github.com/users/fxmarty/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@fxmarty Are you happy for me to merge? ", "sure @amyeroberts !" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28006/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28006", "html_url": "https://github.com/huggingface/transformers/pull/28006", "diff_url": "https://github.com/huggingface/transformers/pull/28006.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28006.patch", "merged_at": 1705421444000 }

{ "login": "fxmarty", "id": 9808326, "node_id": "MDQ6VXNlcjk4MDgzMjY=", "avatar_url": "https://avatars.githubusercontent.com/u/9808326?v=4", "gravatar_id": "", "url": "https://api.github.com/users/fxmarty", "html_url": "https://github.com/fxmarty", "followers_url": "https://api.github.com/users/fxmarty/followers", "following_url": "https://api.github.com/users/fxmarty/following{/other_user}", "gists_url": "https://api.github.com/users/fxmarty/gists{/gist_id}", "starred_url": "https://api.github.com/users/fxmarty/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/fxmarty/subscriptions", "organizations_url": "https://api.github.com/users/fxmarty/orgs", "repos_url": "https://api.github.com/users/fxmarty/repos", "events_url": "https://api.github.com/users/fxmarty/events{/privacy}", "received_events_url": "https://api.github.com/users/fxmarty/received_events", "type": "User", "site_admin": false }

[ { "id": 6126880899, "node_id": "LA_kwDOCUB6oc8AAAABbTDIgw", "url": "https://api.github.com/repos/huggingface/transformers/labels/contributions-welcome", "name": "contributions-welcome", "color": "F99E09", "default": false, "description": "" } ]

[]

[ "Hi @fxmarty I can take a look at this issue. Of I can ask questions if necessary. Or has anyone taken it already?", "does someone know if longT5 and all T5 models are blocked by bias support in flash attention ?\r\n\r\nhttps://github.com/Dao-AILab/flash-attention/pull/617", "Hi @davidan5 are you working on the implementation?", "@ENate I was trying to understand the status and have an estimation of the code change to see if I can contribute.", "I see.", "I'm interested in taking a look at this for the Mistral model if that's still needed. Otherwise, please let me know if there are any other models that still need some work. Thanks", "Is LongT5 still open?", "Mistral is already covered! LongT5 if it is like T5 and has attention bias that might not be supported", "Oh yea, looks like you added support for Mistral/Mixtral last month.\r\n\r\nIt doesn't seem to be supported for BERT yet (I think someone else is working on FA2 but not SDPA), so I'll take a crack at it. It looks like there is a config for relative position embeddings for BERT, so I'll just have it fallback to the original attention for configs using relative position embeddings.\r\n\r\n@ArthurZucker - Please let me know if you know if someone else is already working on SDPA for BERT and I can look for something else to do. Thanks!", "Not sure anyone is working on that but bert is already so small that I doubt it will have a lot of impact on perf! " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28005/reactions", "total_count": 7, "+1": 0, "-1": 0, "laugh": 0, "hooray": 3, "confused": 0, "heart": 0, "rocket": 4, "eyes": 0 }

{ "login": "Merchaoui", "id": 80455763, "node_id": "MDQ6VXNlcjgwNDU1NzYz", "avatar_url": "https://avatars.githubusercontent.com/u/80455763?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Merchaoui", "html_url": "https://github.com/Merchaoui", "followers_url": "https://api.github.com/users/Merchaoui/followers", "following_url": "https://api.github.com/users/Merchaoui/following{/other_user}", "gists_url": "https://api.github.com/users/Merchaoui/gists{/gist_id}", "starred_url": "https://api.github.com/users/Merchaoui/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Merchaoui/subscriptions", "organizations_url": "https://api.github.com/users/Merchaoui/orgs", "repos_url": "https://api.github.com/users/Merchaoui/repos", "events_url": "https://api.github.com/users/Merchaoui/events{/privacy}", "received_events_url": "https://api.github.com/users/Merchaoui/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @Merchaoui, thanks for raising this issue! \r\n\r\nLoading the model with `ignore_mismatched_sizes=True` is the correct way to go i.e.:\r\n\r\n```py\r\nmodel = LayoutLMv2ForTokenClassification.from_pretrained(\r\n \"./layoutxlm_reduced/microsoft/layoutxlm-base\",\r\n num_labels=len(labels),\r\n ignore_mismatched_sizes=True\r\n)\r\n```\r\n\r\nThe other error shouldn't be happening though. Could you share a minimal code snippet we could run which reproduces the error when training? \r\n\r\nCould you also share the error message with the full traceback? \r\n", "Thank you so much, I just figured it out. When I reduced the model, I removed unnecessary tokens from LayoutXLMTokenizerFast and for the training I was processing with LayoutXLMTokenizer. Now it´s working thanks", "@Merchaoui Great - glad to hear it's working! Thanks for following up on the issue to confirm :) ", "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/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28004/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "VasilGeorgiev39", "id": 149842188, "node_id": "U_kgDOCO5pDA", "avatar_url": "https://avatars.githubusercontent.com/u/149842188?v=4", "gravatar_id": "", "url": "https://api.github.com/users/VasilGeorgiev39", "html_url": "https://github.com/VasilGeorgiev39", "followers_url": "https://api.github.com/users/VasilGeorgiev39/followers", "following_url": "https://api.github.com/users/VasilGeorgiev39/following{/other_user}", "gists_url": "https://api.github.com/users/VasilGeorgiev39/gists{/gist_id}", "starred_url": "https://api.github.com/users/VasilGeorgiev39/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/VasilGeorgiev39/subscriptions", "organizations_url": "https://api.github.com/users/VasilGeorgiev39/orgs", "repos_url": "https://api.github.com/users/VasilGeorgiev39/repos", "events_url": "https://api.github.com/users/VasilGeorgiev39/events{/privacy}", "received_events_url": "https://api.github.com/users/VasilGeorgiev39/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Thank you @VasilGeorgiev39, a quick fix is to load your transformers model with\r\n\r\n```python\r\nimport transformers\r\nimport tensor_parallel as tp\r\n\r\ntokenizer = transformers.AutoTokenizer.from_pretrained(\"meta-llama/Llama-2-13b-chat-hf\")\r\nmodel = transformers.AutoModelForCausalLM.from_pretrained(\"meta-llama/Llama-2-13b-chat-hf\", attn_implementation=\"eager\")\r\n\r\nmodelp = tp.tensor_parallel(model)\r\n```\r\n\r\nIt looks like `tp.tensor_parallel` is overloading Transformers classes and does not have `_supports_sdpa = True`, which results in this error. I'll see what we can do.\r\n\r\nFor reference: https://huggingface.co./docs/transformers/perf_infer_gpu_one#flashattention-and-memory-efficient-attention-through-pytorchs-scaleddotproductattention", "I am having the same issue with `Llama-2-70b-chat-hf model`. Tried to add `attn_implementation=\"eager`, but got the following new \"no attribute 'to_legacy_cache'\" errors:\r\n\r\n```\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 1527, in _call_impl [1/1003]\r\n return forward_call(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/tensor_parallel/pretrained_model.py\", line 76, in forward\r\n return self.wrapped_model(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/tensor_parallel/tensor_parallel.py\", line 159, in forward\r\n return parallel_apply(self.module_shards, inputs, kwargs_tup, self.devices)[self.output_device_index]\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/parallel/parallel_apply.py\", line 110, in parallel_apply\r\n output.reraise()\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/_utils.py\", line 694, in reraise\r\n raise exception\r\nAttributeError: Caught AttributeError in replica 0 on device 0.\r\nOriginal Traceback (most recent call last):\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/parallel/parallel_apply.py\", line 85, in _worker\r\n output = module(*input, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/llama/modeling_llama.py\", line 1174, in forward\r\n outputs = self.model(\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/usr/local/lib/python3.8/dist-packages/transformers/models/llama/modeling_llama.py\", line 1086, in forward\r\n next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache\r\nAttributeError: 'tuple' object has no attribute 'to_legacy_cache'\r\n```\r\n", "cc @gante as well! ", "@taozhang9527 since `tensor_parallel` overloads the models, it is possible their classes are not compatible with the models that got the new cache format in `transformers` v4.36. I'd open an issue in `tensor_parallel` :)" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28003/reactions", "total_count": 1, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "ElsebaiyMohamed", "id": 77920008, "node_id": "MDQ6VXNlcjc3OTIwMDA4", "avatar_url": "https://avatars.githubusercontent.com/u/77920008?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ElsebaiyMohamed", "html_url": "https://github.com/ElsebaiyMohamed", "followers_url": "https://api.github.com/users/ElsebaiyMohamed/followers", "following_url": "https://api.github.com/users/ElsebaiyMohamed/following{/other_user}", "gists_url": "https://api.github.com/users/ElsebaiyMohamed/gists{/gist_id}", "starred_url": "https://api.github.com/users/ElsebaiyMohamed/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ElsebaiyMohamed/subscriptions", "organizations_url": "https://api.github.com/users/ElsebaiyMohamed/orgs", "repos_url": "https://api.github.com/users/ElsebaiyMohamed/repos", "events_url": "https://api.github.com/users/ElsebaiyMohamed/events{/privacy}", "received_events_url": "https://api.github.com/users/ElsebaiyMohamed/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @ElsebaiyMohamed, thanks for raising this issue and providing details on the error + a snippet. Could you also provide information about the running environment: run `transformers-cli env` in the terminal and copy-paste the output? ", "Hi @amyeroberts ,\r\nApologies for the delayed response! 🙏 Life threw a curveball, but I'm back on track. Thanks for your patience!\r\n\r\nRegarding your request, here's the output of `transformers-cli env`:\r\n\r\n```bash\r\ntransformers version: 4.36.0\r\nPlatform: Linux-5.15.133+-x86_64-with-glibc2.35\r\nPython version: 3.10.12\r\nHuggingface_hub version: 0.19.4\r\nSafetensors version: 0.4.1\r\nAccelerate version: 0.25.0\r\nAccelerate config: \tnot found\r\nPyTorch version (GPU?): 2.0.0 (True)\r\nTensorflow version (GPU?): 2.13.0 (True)\r\nFlax version (CPU?/GPU?/TPU?): 0.7.5 (gpu)\r\nJax version: 0.4.21\r\nJaxLib version: 0.4.21\r\nUsing GPU in script?: yes\r\nUsing distributed or parallel set-up in script?: no\r\n```\r\n\r\nLet me know if there's anything else I can help you with.", "@ElsebaiyMohamed Great - thanks for providing this info! \r\n\r\ncc @sanchit-gandhi @ylacombe " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28002/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "SeeknnDestroy", "id": 44926076, "node_id": "MDQ6VXNlcjQ0OTI2MDc2", "avatar_url": "https://avatars.githubusercontent.com/u/44926076?v=4", "gravatar_id": "", "url": "https://api.github.com/users/SeeknnDestroy", "html_url": "https://github.com/SeeknnDestroy", "followers_url": "https://api.github.com/users/SeeknnDestroy/followers", "following_url": "https://api.github.com/users/SeeknnDestroy/following{/other_user}", "gists_url": "https://api.github.com/users/SeeknnDestroy/gists{/gist_id}", "starred_url": "https://api.github.com/users/SeeknnDestroy/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/SeeknnDestroy/subscriptions", "organizations_url": "https://api.github.com/users/SeeknnDestroy/orgs", "repos_url": "https://api.github.com/users/SeeknnDestroy/repos", "events_url": "https://api.github.com/users/SeeknnDestroy/events{/privacy}", "received_events_url": "https://api.github.com/users/SeeknnDestroy/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @SeeknnDestroy, thanks for raising an issue! \r\n\r\nThere's three parts to the issue being raised. \r\n\r\nWith regards to the error message, this is because the deprecated argument `max_length` is used for that checkpoint's [generation config](https://huggingface.co./distil-whisper/distil-large-v2/blob/c204f3c76ec464a0ab9bcfd19afa0add93f69983/generation_config.json#L127). @sanchit-gandhi \r\n\r\nThe second is about the transcription behaviour. @sanchit-gandhi is best placed to answer about the recommended way to treat long audio files. \r\n\r\nThe final point is how to configure the model behaviour using inference endpoints, which I'll defer to @philschmid :) ", "Whisper has a receptive field of 30s. For long-form transcription (>30s audio), we need to enable \"chunking\" to transcribe chunks of 30s audios incrementally, and then \"stitch\" the resulting transcriptions together at the boundaries. You can see how to run this in Python here: https://huggingface.co./distil-whisper/distil-large-v2#long-form-transcription\r\nIt's quite simple by passing one extra line to the pipeline: `chunk_length_s=15`\r\n\r\nI'll leave @philschmid to advise on how to integrate this into your endpoint!", "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/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28001/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "fxmarty", "id": 9808326, "node_id": "MDQ6VXNlcjk4MDgzMjY=", "avatar_url": "https://avatars.githubusercontent.com/u/9808326?v=4", "gravatar_id": "", "url": "https://api.github.com/users/fxmarty", "html_url": "https://github.com/fxmarty", "followers_url": "https://api.github.com/users/fxmarty/followers", "following_url": "https://api.github.com/users/fxmarty/following{/other_user}", "gists_url": "https://api.github.com/users/fxmarty/gists{/gist_id}", "starred_url": "https://api.github.com/users/fxmarty/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/fxmarty/subscriptions", "organizations_url": "https://api.github.com/users/fxmarty/orgs", "repos_url": "https://api.github.com/users/fxmarty/repos", "events_url": "https://api.github.com/users/fxmarty/events{/privacy}", "received_events_url": "https://api.github.com/users/fxmarty/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @fxmarty - thanks for raising this! \r\n\r\nTo help with debugging - has this been observed with other checkpoints or only the tiny random ones for testing? " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28000/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "ArthurZucker", "id": 48595927, "node_id": "MDQ6VXNlcjQ4NTk1OTI3", "avatar_url": "https://avatars.githubusercontent.com/u/48595927?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ArthurZucker", "html_url": "https://github.com/ArthurZucker", "followers_url": "https://api.github.com/users/ArthurZucker/followers", "following_url": "https://api.github.com/users/ArthurZucker/following{/other_user}", "gists_url": "https://api.github.com/users/ArthurZucker/gists{/gist_id}", "starred_url": "https://api.github.com/users/ArthurZucker/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ArthurZucker/subscriptions", "organizations_url": "https://api.github.com/users/ArthurZucker/orgs", "repos_url": "https://api.github.com/users/ArthurZucker/repos", "events_url": "https://api.github.com/users/ArthurZucker/events{/privacy}", "received_events_url": "https://api.github.com/users/ArthurZucker/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Thanks a lot!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27999/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27999", "html_url": "https://github.com/huggingface/transformers/pull/27999", "diff_url": "https://github.com/huggingface/transformers/pull/27999.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27999.patch", "merged_at": 1702471031000 }

{ "login": "zhaotyer", "id": 89376832, "node_id": "MDQ6VXNlcjg5Mzc2ODMy", "avatar_url": "https://avatars.githubusercontent.com/u/89376832?v=4", "gravatar_id": "", "url": "https://api.github.com/users/zhaotyer", "html_url": "https://github.com/zhaotyer", "followers_url": "https://api.github.com/users/zhaotyer/followers", "following_url": "https://api.github.com/users/zhaotyer/following{/other_user}", "gists_url": "https://api.github.com/users/zhaotyer/gists{/gist_id}", "starred_url": "https://api.github.com/users/zhaotyer/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/zhaotyer/subscriptions", "organizations_url": "https://api.github.com/users/zhaotyer/orgs", "repos_url": "https://api.github.com/users/zhaotyer/repos", "events_url": "https://api.github.com/users/zhaotyer/events{/privacy}", "received_events_url": "https://api.github.com/users/zhaotyer/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @younesbelkada as the error seems to be being hit in bitsandbytes ", "Hi @zhaotyer -- can you reproduce the issue with a smaller, publically available model like `codellama/CodeLlama-7b-hf`?\r\n\r\nThe `RuntimeError` you got often shows up in OOM cases. It would be weird to get one given that you're using a 34B 4-bit quantized model on a 80GB GPU, but let's rule that out first.", "Oops sorry, this issue somehow got out of my eyes \r\nI second what @gante said! If you can confirm that, it would great\r\nMoreover, the issue \r\n```\r\nRuntimeError: CUDA error: CUBLAS_STATUS_NOT_INITIALIZED when calling `cublasCreate(handle)`\r\n```\r\nSometimes indicates that there is something off with embedding layers and passed indices, to further debug, can you please:\r\n1- Re-run your script with `CUDA_LAUNCH_BLOCKING=1 yourscript.py` --> this will give a new traceback with the exact position where things are off\r\n2- Try out to run the same script on CPU, without quantization\r\nThanks!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27998/reactions", "total_count": 1, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 1 }

{ "login": "ajati", "id": 41211350, "node_id": "MDQ6VXNlcjQxMjExMzUw", "avatar_url": "https://avatars.githubusercontent.com/u/41211350?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ajati", "html_url": "https://github.com/ajati", "followers_url": "https://api.github.com/users/ajati/followers", "following_url": "https://api.github.com/users/ajati/following{/other_user}", "gists_url": "https://api.github.com/users/ajati/gists{/gist_id}", "starred_url": "https://api.github.com/users/ajati/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ajati/subscriptions", "organizations_url": "https://api.github.com/users/ajati/orgs", "repos_url": "https://api.github.com/users/ajati/repos", "events_url": "https://api.github.com/users/ajati/events{/privacy}", "received_events_url": "https://api.github.com/users/ajati/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@kashif " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27997/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27997", "html_url": "https://github.com/huggingface/transformers/pull/27997", "diff_url": "https://github.com/huggingface/transformers/pull/27997.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27997.patch", "merged_at": 1702470866000 }

{ "login": "jiqing-feng", "id": 107918818, "node_id": "U_kgDOBm614g", "avatar_url": "https://avatars.githubusercontent.com/u/107918818?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jiqing-feng", "html_url": "https://github.com/jiqing-feng", "followers_url": "https://api.github.com/users/jiqing-feng/followers", "following_url": "https://api.github.com/users/jiqing-feng/following{/other_user}", "gists_url": "https://api.github.com/users/jiqing-feng/gists{/gist_id}", "starred_url": "https://api.github.com/users/jiqing-feng/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jiqing-feng/subscriptions", "organizations_url": "https://api.github.com/users/jiqing-feng/orgs", "repos_url": "https://api.github.com/users/jiqing-feng/repos", "events_url": "https://api.github.com/users/jiqing-feng/events{/privacy}", "received_events_url": "https://api.github.com/users/jiqing-feng/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @gante @sywangyi ", "Hi @jiqing-feng, thanks for opening this PR! \r\n\r\nI don't think this is a feature we want to add at the moment. Users can already pass compiled models to the pipeline to use, and there they can control which parts will or won't be compiled. There's many optimization choices one might make when loading a model e.g. whether to quantize the weights. To keep the API of the pipeline as simple as possible - we can leave additional configuration of the model outside of the pipeline.\r\n\r\nI'll let @Narsil's opinion weigh more heavily on whether this should be added to the pipeline over mine. \r\n", "The main idea is that users may not know if the `pipeline` uses `model.forward` or `model.generate`. If we integrate `torch.compile` in the `pipeline`, we can decide which function should be compiled. We only need 2 extra params: `torch_compile` and `torch_compile_config`, or we can just pass `torch_compile_backend` to `kwargs`. It will help users to apply torch compile and avoid compiling useless functions easily. \r\n\r\nFor example, in [ASR pipeline](https://github.com/huggingface/transformers/blob/main/src/transformers/pipelines/automatic_speech_recognition.py#L541-L601), it uses `model.generate` if `model_type in {\"seq2seq\", \"seq2seq_whisper\"}`, and use `model.forward` otherwise. \r\n\r\nI think users won't check the `pipeline` code in such detail, so the best way is we define which model function needs to be compiled in `pipeline`.\r\n\r\nBTW, it is important to clarify that `model=torch.compile(model)` won't work on `model.generate.forward`, the model in generation will still use the original forward function to inference.", "yes. user may not be aware of the calling API in each Task of Pipeline in __call__. we could do it for user in each task if user want to use torch.compile for acceleration. for some task, we see 20%-30% perf boost.", "Hi @Narsil . Do you think that we should add `torch.compile` in the pipeline since users may not know which function of model should be compiled?", "No you can and should use ` pipeline(model=torch.compile(model)`.\r\nIf that doesn't work with generate it's up to the caller to know what to do.\r\n\r\nJust as a reminder `pipeline` is meant to make ML models easy to use, while have *decent* performance (shouldn't be 10x slower than it should). However, complexifying the codebase for a potential 10-20% is not really worth it imo (especially when sending directly a compiled model WILL work, you just need to know when you can and when you cannot, but that' s more of a torch.compile limitation than anything related to pipelines... pipelines also can work with Tensorflow models for instance)", "> No you can and should use ` pipeline(model=torch.compile(model)`. If that doesn't work with generate it's up to the caller to know what to do.\r\n> \r\n> Just as a reminder `pipeline` is meant to make ML models easy to use, while have _decent_ performance (shouldn't be 10x slower than it should). However, complexifying the codebase for a potential 10-20% is not really worth it imo (especially when sending directly a compiled model WILL work, you just need to know when you can and when you cannot, but that' s more of a torch.compile limitation than anything related to pipelines... pipelines also can work with Tensorflow models for instance)\r\n\r\nThanks for the clarification!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27996/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27996", "html_url": "https://github.com/huggingface/transformers/pull/27996", "diff_url": "https://github.com/huggingface/transformers/pull/27996.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27996.patch", "merged_at": null }

{ "login": "jiqing-feng", "id": 107918818, "node_id": "U_kgDOBm614g", "avatar_url": "https://avatars.githubusercontent.com/u/107918818?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jiqing-feng", "html_url": "https://github.com/jiqing-feng", "followers_url": "https://api.github.com/users/jiqing-feng/followers", "following_url": "https://api.github.com/users/jiqing-feng/following{/other_user}", "gists_url": "https://api.github.com/users/jiqing-feng/gists{/gist_id}", "starred_url": "https://api.github.com/users/jiqing-feng/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jiqing-feng/subscriptions", "organizations_url": "https://api.github.com/users/jiqing-feng/orgs", "repos_url": "https://api.github.com/users/jiqing-feng/repos", "events_url": "https://api.github.com/users/jiqing-feng/events{/privacy}", "received_events_url": "https://api.github.com/users/jiqing-feng/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @gante @amyeroberts . Would you please help me to review it? Thx!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27995/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27995", "html_url": "https://github.com/huggingface/transformers/pull/27995", "diff_url": "https://github.com/huggingface/transformers/pull/27995.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27995.patch", "merged_at": 1704968700000 }

{ "login": "jerin-scalers-ai", "id": 125901005, "node_id": "U_kgDOB4EYzQ", "avatar_url": "https://avatars.githubusercontent.com/u/125901005?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jerin-scalers-ai", "html_url": "https://github.com/jerin-scalers-ai", "followers_url": "https://api.github.com/users/jerin-scalers-ai/followers", "following_url": "https://api.github.com/users/jerin-scalers-ai/following{/other_user}", "gists_url": "https://api.github.com/users/jerin-scalers-ai/gists{/gist_id}", "starred_url": "https://api.github.com/users/jerin-scalers-ai/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jerin-scalers-ai/subscriptions", "organizations_url": "https://api.github.com/users/jerin-scalers-ai/orgs", "repos_url": "https://api.github.com/users/jerin-scalers-ai/repos", "events_url": "https://api.github.com/users/jerin-scalers-ai/events{/privacy}", "received_events_url": "https://api.github.com/users/jerin-scalers-ai/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The inference time in bfloat16 depends on the hardware you are using and the pytorch version as well. Recommending you to use float16 for inference. ", "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/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27994/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "qihqi", "id": 1719482, "node_id": "MDQ6VXNlcjE3MTk0ODI=", "avatar_url": "https://avatars.githubusercontent.com/u/1719482?v=4", "gravatar_id": "", "url": "https://api.github.com/users/qihqi", "html_url": "https://github.com/qihqi", "followers_url": "https://api.github.com/users/qihqi/followers", "following_url": "https://api.github.com/users/qihqi/following{/other_user}", "gists_url": "https://api.github.com/users/qihqi/gists{/gist_id}", "starred_url": "https://api.github.com/users/qihqi/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/qihqi/subscriptions", "organizations_url": "https://api.github.com/users/qihqi/orgs", "repos_url": "https://api.github.com/users/qihqi/repos", "events_url": "https://api.github.com/users/qihqi/events{/privacy}", "received_events_url": "https://api.github.com/users/qihqi/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@qihqi this is done inside `_save_tpu` https://github.com/huggingface/transformers/pull/27799/files, or we missed something?", "> @qihqi this is done inside `_save_tpu` https://github.com/huggingface/transformers/pull/27799/files, or we missed something?\r\n\r\nI see, we need to make a copy. Should we do that inside `_save_tpu` or do it before here?", "Thanks for the fix! Quick question: I used to see this error `indices should be either on cpu or on the same device as the indexed tensor (XLA). When using XLA, the indexed tensor must be an XLA tensor.` and it seems some index ops failed so I tried to fix the op instead of the model (and it didn't work out). How did you find out it's the model's problem?", "> Thanks for the fix! Quick question: I used to see this error `indices should be either on cpu or on the same device as the indexed tensor (XLA). When using XLA, the indexed tensor must be an XLA tensor.` and it seems some index ops failed so I tried to fix the op instead of the model (and it didn't work out). How did you find out it's the model's problem?\r\n\r\nThanks! So I also started with assumption that the input is wrong, and traced where the input comes from, and arrived here: https://github.com/huggingface/transformers/blob/main/src/transformers/trainer.py#L2668 and here clearly the input is moving to the device. \r\n\r\nAlso I ran the script the first few steps trains fine. So I start suspecting something happened after training is done. ", "Hi @LysandreJik @ArthurZucker would you guys be able to help with a review? Thanks!", "Pinging our TPU expert @muellerzr ", "> Thanks for adding this!\r\n> \r\n> I'm not very familiar with TPUs. My only concern with this is, if it behaves the same as the model being on GPU, then it requires enough space for 2 models on the device when `copy.deepcopy(self.model)` is called i.e. there is no offloading.\r\n\r\nGood point! I have changed the code to first move it CPU then move it back. PTAL.", "@qihqi Just to confirm that you were able to successfully run the script in the PR description with the most recent changes to the PR: \r\n\r\n```\r\npython3 examples/pytorch/text-classification/run_glue.py \\\r\n --model_name_or_path=distilbert-base-uncased \\\r\n --task_name=MNLI \\\r\n --do_train=true \\\r\n --num_train_epochs=1 \\\r\n --max_seq_length=128 \\\r\n --learning_rate=3e-5 \\\r\n --overwrite_output_dir=true \\\r\n --save_steps=3000 \\\r\n --save_strategy=no --output_dir=/workspace/mnli\r\n```\r\n", "> @qihqi Just to confirm that you were able to successfully run the script in the PR description with the most recent changes to the PR:\r\n> \r\n> ```\r\n> python3 examples/pytorch/text-classification/run_glue.py \\\r\n> --model_name_or_path=distilbert-base-uncased \\\r\n> --task_name=MNLI \\\r\n> --do_train=true \\\r\n> --num_train_epochs=1 \\\r\n> --max_seq_length=128 \\\r\n> --learning_rate=3e-5 \\\r\n> --overwrite_output_dir=true \\\r\n> --save_steps=3000 \\\r\n> --save_strategy=no --output_dir=/workspace/mnli\r\n> ```\r\n\r\nYes. (Well, to be exact it's this script: https://github.com/GoogleCloudPlatform/ml-testing-accelerators/blob/master/tests/pytorch/nightly/hf-glue.libsonnet#L41) ", "Hi everyone,\r\nOne question, was this fix supposed to work with XLA FSDP? \r\n\r\nI'm having problems with the model.to ('cpu') part before unwrapping the model (memory access errors), and if I unwrap it before applying .to('cpu'), just one shard of the model is saved." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27993/reactions", "total_count": 2, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 2, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27993", "html_url": "https://github.com/huggingface/transformers/pull/27993", "diff_url": "https://github.com/huggingface/transformers/pull/27993.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27993.patch", "merged_at": 1702980531000 }

{ "login": "xyx361100238", "id": 19569322, "node_id": "MDQ6VXNlcjE5NTY5MzIy", "avatar_url": "https://avatars.githubusercontent.com/u/19569322?v=4", "gravatar_id": "", "url": "https://api.github.com/users/xyx361100238", "html_url": "https://github.com/xyx361100238", "followers_url": "https://api.github.com/users/xyx361100238/followers", "following_url": "https://api.github.com/users/xyx361100238/following{/other_user}", "gists_url": "https://api.github.com/users/xyx361100238/gists{/gist_id}", "starred_url": "https://api.github.com/users/xyx361100238/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/xyx361100238/subscriptions", "organizations_url": "https://api.github.com/users/xyx361100238/orgs", "repos_url": "https://api.github.com/users/xyx361100238/repos", "events_url": "https://api.github.com/users/xyx361100238/events{/privacy}", "received_events_url": "https://api.github.com/users/xyx361100238/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey @xyx361100238! Do you have a script I can use to reproduce the error and identify the memory increase? It's maybe worth trying adding the flag for full fp16 eval as well to avoid upcasting the weights to fp32 for inference:\r\n\r\n```\r\n--fp16_full_eval\r\n```\r\n\r\n_c.f._ https://huggingface.co./docs/transformers/v4.36.1/en/main_classes/trainer#transformers.TrainingArguments.fp16_full_eval", "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/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27992/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "jcai04", "id": 62895533, "node_id": "MDQ6VXNlcjYyODk1NTMz", "avatar_url": "https://avatars.githubusercontent.com/u/62895533?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jcai04", "html_url": "https://github.com/jcai04", "followers_url": "https://api.github.com/users/jcai04/followers", "following_url": "https://api.github.com/users/jcai04/following{/other_user}", "gists_url": "https://api.github.com/users/jcai04/gists{/gist_id}", "starred_url": "https://api.github.com/users/jcai04/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jcai04/subscriptions", "organizations_url": "https://api.github.com/users/jcai04/orgs", "repos_url": "https://api.github.com/users/jcai04/repos", "events_url": "https://api.github.com/users/jcai04/events{/privacy}", "received_events_url": "https://api.github.com/users/jcai04/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @jcai04, this does not seems to be an issue with transformers. Please submit this issue in the pytorch repo. ", "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/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27991/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "SangbumChoi", "id": 34004152, "node_id": "MDQ6VXNlcjM0MDA0MTUy", "avatar_url": "https://avatars.githubusercontent.com/u/34004152?v=4", "gravatar_id": "", "url": "https://api.github.com/users/SangbumChoi", "html_url": "https://github.com/SangbumChoi", "followers_url": "https://api.github.com/users/SangbumChoi/followers", "following_url": "https://api.github.com/users/SangbumChoi/following{/other_user}", "gists_url": "https://api.github.com/users/SangbumChoi/gists{/gist_id}", "starred_url": "https://api.github.com/users/SangbumChoi/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/SangbumChoi/subscriptions", "organizations_url": "https://api.github.com/users/SangbumChoi/orgs", "repos_url": "https://api.github.com/users/SangbumChoi/repos", "events_url": "https://api.github.com/users/SangbumChoi/events{/privacy}", "received_events_url": "https://api.github.com/users/SangbumChoi/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Also due to change of major arguments I think we need to rewrite the test function for swin_backbone.", "Hi @SangbumChoi thanks so much for your PR. Could you try to make the CI green by running `make fixup` locally and fixing the errors it gives?\r\n\r\nBtw if you would have the time, would be great look into why DETA evaluation numbers on the [open object detection leaderboard](https://huggingface.co./spaces/hf-vision/object_detection_leaderboard) are not as high as reported in the paper. ", "@NielsRogge I think my PR will also affect to this leaderboard. I'm not sure if it will have positive or negative but I hope it is positive :)\r\n\r\nIn a viewpoint of inference pipeline, [the original implementation\r\n](https://huggingface.co./jozhang97/deta-swin-large/discussions/2) of config.json was not using DETA-specific assigner(which was DETAassigner1 and DETAassigner2 in the code that you have written) even though \"assign_first_stage=true\". You may recheck this to confirm this! \r\n(See the line of 1941 in modeling_deta.py)\r\n\r\nIn addition postprocessing pipeline with class_threshold is default by 0.7 however original code doesn't use classification thresholding (=using threshold 0)\r\n\r\nAnd also I will do some extra modification to make all the CI works!", "@NielsRogge I updated and passed the CI. There are few changes that you should know! please don't hestitate to ask details", "Awesome, I'll take a look in more detail in the next coming days.\r\n\r\nAlso, I was working on a notebook to fine-tune DETA on a custom dataset (by mainly copying [this notebook](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/DETR/Fine_tuning_DetrForObjectDetection_on_custom_dataset_(balloon).ipynb) - you can replace `DetrForObjectDetection` by `DetaForObjectDetection`). However, the results were not as expected. If you have time to look into how to successfully fine-tune DETA, that would be awesome.", "@NielsRogge I also saw this notebook and my custom code is very very similar to this notebook which is using pytorch_lightning and distributed training.\r\nOne important fact is that matching lr_backbone and lr with the same hyperparameter with 2e-4. I also used Swin backbone for default training and gave me 41AP in final (not his balloon but the dataset that i mentioned above)\r\n```\r\n results = self.postprocess.post_process_object_detection(\r\n outputs, target_sizes=orig_target_sizes, threshold=0.0\r\n )\r\n```\r\nAlso remind that thresholding and **setting auxiliary_loss to true** is also important\r\n\r\nCan you also share the final AP?\r\n", "@NielsRogge @amyeroberts Ping for reminder", "Hi @SangbumChoi I tried out fine-tuning DETA from your branch on the balloon dataset with [this notebook](https://colab.research.google.com/drive/1fyZLR97V9RcWSjPN6djLdPylOjIV95r1?usp=sharing). Initially results didn't look good, but now they do! Had to set `auxiliary_loss=True` when loading the model.\r\n\r\n\r\n\r\nThe only thing that is still weird is that bounding boxes are shown with confidence > 0.4, even though I've passed in threshold=0.2. Will need to look into this.\r\n\r\nAlready awesome that training seems to have improved a lot!\r\n\r\nThanks!\r\n", "@NielsRogge Sounds great, it will have huge improvement. For the threshold part, because it has NMSpostprocess already implemented on the code, maybe overlapped bounding box with 0.2~0.39 might disappear at the visualization. (Check this part)", "> Thanks for working on this!\r\n> \r\n> There's a few changes we might not be able to add because of backwards compatibility and compatibility with the rest of the codebase. Otherwise looks good.\r\n\r\n@amyeroberts Hi, I also did some modification and resolved several conversations that is obvious. However, I still opened the conversation that has some misalignment or comments! (I think only one comment is not still fixed which is multi-gpu distribute part!)", "@NielsRogge Is there anything that I can help to make this PR merged? ", "@amyeroberts would need to approve the PR.\r\n\r\nFor reference, I don't think anyone was successfully fine-tuning DETA before this PR, so would be great to get it merged. Also haven't shared about this model yet due to differences in reported COCO eval", "@amyeroberts Thanks for your suggestion! Everything is fixed as your recommendation 👍🏼 ", "@amyeroberts There was one mistake importing sorry about that (see last commit)\r\n\r\nfor the slow test\r\n\r\n```\r\nroot@0a2b4fe54761:/mnt/nas2/users/sbchoi/transformers# RUN_SLOW=1 pytest tests/models/deta/\r\n=========================================================== test session starts ============================================================\r\nplatform linux -- Python 3.10.13, pytest-7.4.4, pluggy-1.0.0\r\nrootdir: /mnt/nas2/users/sbchoi/transformers\r\nconfigfile: pyproject.toml\r\nplugins: hypothesis-6.92.0, hydra-core-1.3.2\r\ncollected 150 items\r\n\r\ntests/models/deta/test_image_processing_deta.py .............. [ 9%]\r\ntests/models/deta/test_modeling_deta.py ........s...............ssssss.ssssssssss.....s..............s......s.............ssssssssss [ 70%]\r\ns.sssssssssssssss.s..s.......s.s............ [100%]\r\n\r\n============================================================= warnings summary =============================================================\r\n../../../../../opt/conda/lib/python3.10/site-packages/_pytest/config/__init__.py:1373\r\n /opt/conda/lib/python3.10/site-packages/_pytest/config/__init__.py:1373: PytestConfigWarning: Unknown config option: doctest_glob\r\n\r\n self._warn_or_fail_if_strict(f\"Unknown config option: {key}\\n\")\r\n\r\nsrc/transformers/deepspeed.py:23\r\n /mnt/nas2/users/sbchoi/transformers/src/transformers/deepspeed.py:23: FutureWarning: transformers.deepspeed module is deprecated and will be removed in a future version. Please import deepspeed modules directly from transformers.integrations\r\n warnings.warn(\r\n\r\n../../../../../opt/conda/lib/python3.10/site-packages/torch/utils/cpp_extension.py:28\r\n /opt/conda/lib/python3.10/site-packages/torch/utils/cpp_extension.py:28: DeprecationWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html\r\n from pkg_resources import packaging # type: ignore[attr-defined]\r\n\r\n\r\n declare_namespace(pkg)\r\n /opt/conda/lib/python3.10/site-packages/_pytest/config/__init__.py:1373: PytestConfigWarning: Unknown config option: doctest_glob\r\n\r\n self._warn_or_fail_if_strict(f\"Unknown config option: {key}\\n\")\r\n\r\nsrc/transformers/deepspeed.py:23\r\n /mnt/nas2/users/sbchoi/transformers/src/transformers/deepspeed.py:23: FutureWarning: transformers.deepspeed module is deprecated and will be removed in a future version. Please import deepspeed modules directly from transformers.integrations\r\n warnings.warn(\r\n\r\n../../../../../opt/conda/lib/python3.10/site-packages/torch/utils/cpp_extension.py:28\r\n /opt/conda/lib/python3.10/site-packages/torch/utils/cpp_extension.py:28: DeprecationWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html\r\n from pkg_resources import packaging # type: ignore[attr-defined]\r\n\r\n../../../../../opt/conda/lib/python3.10/site-packages/pkg_resources/__init__.py:2871\r\n../../../../../opt/conda/lib/python3.10/site-packages/pkg_resources/__init__.py:2871\r\n /opt/conda/lib/python3.10/site-packages/pkg_resources/__init__.py:2871: DeprecationWarning: Deprecated call to `pkg_resources.declare_namespace('ruamel')`.\r\n Implementing implicit namespace packages (as specified in PEP 420) is preferred to `pkg_resources.declare_namespace`. See https://setuptools.pypa.io/en/latest/references/keywords.html#keyword-namespace-packages\r\n declare_namespace(pkg)\r\n\r\ntests/models/deta/test_modeling_deta.py::DetaModelTest::test_disk_offload_bin\r\n /opt/conda/lib/python3.10/site-packages/torch/_utils.py:831: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly. To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()\r\n return self.fget.__get__(instance, owner)()\r\n\r\ntests/models/deta/test_modeling_deta.py::DetaModelTest::test_fast_init_context_manager\r\n /mnt/nas2/users/sbchoi/transformers/tests/test_modeling_common.py:460: FutureWarning: `torch.testing.assert_allclose()` is deprecated since 1.12 and will be removed in a future release. Please use `torch.testing.assert_close()` instead. You can find detailed upgrade instructions in https://github.com/pytorch/pytorch/issues/61844.\r\n torch.testing.assert_allclose(init_instance.linear.bias, expected_bias, rtol=1e-3, atol=1e-4)\r\n\r\ntests/models/deta/test_modeling_deta.py::DetaModelTest::test_fast_init_context_manager\r\n /mnt/nas2/users/sbchoi/transformers/tests/test_modeling_common.py:463: FutureWarning: `torch.testing.assert_allclose()` is deprecated since 1.12 and will be removed in a future release. Please use `torch.testing.assert_close()` instead. You can find detailed upgrade instructions in https://github.com/pytorch/pytorch/issues/61844.\r\n torch.testing.assert_allclose(\r\n\r\ntests/models/deta/test_modeling_deta.py::DetaModelTest::test_model_outputs_equivalence\r\n /mnt/nas2/users/sbchoi/transformers/tests/test_modeling_common.py:1894: UserWarning: Use of index_put_ on expanded tensors is deprecated. Please clone() the tensor before performing this operation. This also applies to advanced indexing e.g. tensor[indices] = tensor (Triggered internally at /opt/conda/conda-bld/pytorch_1702400410390/work/aten/src/ATen/native/TensorAdvancedIndexing.cpp:708.)\r\n t[t != t] = 0\r\n\r\ntests/models/deta/test_modeling_deta.py::DetaModelTest::test_model_outputs_equivalence\r\n /mnt/nas2/users/sbchoi/transformers/tests/test_modeling_common.py:1894: UserWarning: Use of masked_fill_ on expanded tensors is deprecated. Please clone() the tensor before performing this operation. This also applies to advanced indexing e.g. tensor[mask] = scalar (Triggered internally at /opt/conda/conda-bld/pytorch_1702400410390/work/aten/src/ATen/native/cuda/Indexing.cu:1564.)\r\n t[t != t] = 0\r\n\r\n-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html\r\n========================================= 100 passed, 50 skipped, 10 warnings in 206.53s (0:03:26) =========================================\r\n```\r\n\r\nI think it has no problem. Does this answer your question of 'reduce' function? \r\nFrom my training code `accelerate launch --num_processes 4 src/train.py` and `accelerate launch --num_processes 1 src/train.py` both works fine.", "@SangbumChoi Yep! That answers everything. Thanks for running the tests. Once the doc test runner finishes I'll merge in. Thanks again for all of your work on this! ", "Finally! Thanks for the fast review @amyeroberts " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27990/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27990", "html_url": "https://github.com/huggingface/transformers/pull/27990", "diff_url": "https://github.com/huggingface/transformers/pull/27990.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27990.patch", "merged_at": 1704464421000 }

{ "login": "hi-sushanta", "id": 93595990, "node_id": "U_kgDOBZQpVg", "avatar_url": "https://avatars.githubusercontent.com/u/93595990?v=4", "gravatar_id": "", "url": "https://api.github.com/users/hi-sushanta", "html_url": "https://github.com/hi-sushanta", "followers_url": "https://api.github.com/users/hi-sushanta/followers", "following_url": "https://api.github.com/users/hi-sushanta/following{/other_user}", "gists_url": "https://api.github.com/users/hi-sushanta/gists{/gist_id}", "starred_url": "https://api.github.com/users/hi-sushanta/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/hi-sushanta/subscriptions", "organizations_url": "https://api.github.com/users/hi-sushanta/orgs", "repos_url": "https://api.github.com/users/hi-sushanta/repos", "events_url": "https://api.github.com/users/hi-sushanta/events{/privacy}", "received_events_url": "https://api.github.com/users/hi-sushanta/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @hi-sushanta, thanks for opening a PR! \r\n\r\nWhy make these changes? ", "Hello, @amyeroberts.\r\n\r\nHere are the reasons why I made the changes described in the code review comment:\r\n\r\n***1. Improved user-friendliness:***\r\n\r\nMaking the \"encoder\" and \"decoder\" properties read-only prevents accidental modifications and simplifies the API for users. This makes it easier to understand and use the code, especially for beginners.\r\nCreating a dedicated method for changing the \"output embeddings\" provides a controlled and documented way to modify this important parameter. This helps ensure consistency and avoid potential issues.\r\n\r\n***2. Enhanced maintainability:***\r\n\r\nRead-only properties make the code more robust and less prone to errors. By separating the access and modification functionality, the code becomes easier to maintain and understand.\r\nThe dedicated \"output_embeddings\" setter allows for cleaner and more concise code, as it avoids inline logic for handling the update. This improves code readability and maintainability.\r\n\r\n***3. Additional benefits:***\r\n\r\nRead-only properties can improve performance by avoiding unnecessary calculations and memory allocations.\r\nHaving a dedicated setter for \"output embeddings\" enables future enhancements, such as adding validations or logging changes.\r\n", "@hi-sushanta Thank you for taking the time to do this. \r\n\r\nHowever, adding the `_encoder` and `_decoder` attributes is not something we're going to merge in. It goes against the pattern of other model implementations in the code base, potentially breaks things as users still expect to be able to use `model.encoder`, generally makes things harder to read and we're yet to have any complaints from users about `model.encoder` being hard to use. \r\n\r\nIf you wish to enable a setter for the output embeddings, you can add this following the pattern for other models e.g. [here for llava](https://github.com/huggingface/transformers/blob/ec43d6870aa1afb42a6d2b1b0a03743d3f9b3ce6/src/transformers/models/llava/modeling_llava.py#L250C1-L250C1).", "I understand your concerns regarding the _encoder and _decoder attributes. I agree that maintaining consistency with the existing pattern is important. I'm happy to revert this change and ensure my contribution aligns with the code base standards.\r\n" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27989/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27989", "html_url": "https://github.com/huggingface/transformers/pull/27989", "diff_url": "https://github.com/huggingface/transformers/pull/27989.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27989.patch", "merged_at": null }

{ "login": "ccdv-ai", "id": 94319594, "node_id": "U_kgDOBZ8z6g", "avatar_url": "https://avatars.githubusercontent.com/u/94319594?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ccdv-ai", "html_url": "https://github.com/ccdv-ai", "followers_url": "https://api.github.com/users/ccdv-ai/followers", "following_url": "https://api.github.com/users/ccdv-ai/following{/other_user}", "gists_url": "https://api.github.com/users/ccdv-ai/gists{/gist_id}", "starred_url": "https://api.github.com/users/ccdv-ai/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ccdv-ai/subscriptions", "organizations_url": "https://api.github.com/users/ccdv-ai/orgs", "repos_url": "https://api.github.com/users/ccdv-ai/repos", "events_url": "https://api.github.com/users/ccdv-ai/events{/privacy}", "received_events_url": "https://api.github.com/users/ccdv-ai/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Thank you. It's a tradeoff to have between the \"one-model one-file\" philosophy (https://huggingface.co./blog/transformers-design-philosophy) and offloading code to other files/classes.\r\n\r\nWe already kind of deviate from the \"one-model one-file\" philosophy with the KV cache refactor (https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_attn_mask_utils.py) and attention mask refactor (https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_attn_mask_utils.py), and one can indeed argue that we could do the same for the attention.", "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/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27988/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "Decem-Y", "id": 68498490, "node_id": "MDQ6VXNlcjY4NDk4NDkw", "avatar_url": "https://avatars.githubusercontent.com/u/68498490?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Decem-Y", "html_url": "https://github.com/Decem-Y", "followers_url": "https://api.github.com/users/Decem-Y/followers", "following_url": "https://api.github.com/users/Decem-Y/following{/other_user}", "gists_url": "https://api.github.com/users/Decem-Y/gists{/gist_id}", "starred_url": "https://api.github.com/users/Decem-Y/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Decem-Y/subscriptions", "organizations_url": "https://api.github.com/users/Decem-Y/orgs", "repos_url": "https://api.github.com/users/Decem-Y/repos", "events_url": "https://api.github.com/users/Decem-Y/events{/privacy}", "received_events_url": "https://api.github.com/users/Decem-Y/received_events", "type": "User", "site_admin": false }

[]

[]

[ "I find it https://github.com/huggingface/transformers/issues/27925#issuecomment-1851779674", "Please fix the bug caused by this merged PR https://github.com/huggingface/transformers/pull/27820.", "cc @pacman100 @muellerzr " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27987/reactions", "total_count": 2, "+1": 2, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "stevhliu", "id": 59462357, "node_id": "MDQ6VXNlcjU5NDYyMzU3", "avatar_url": "https://avatars.githubusercontent.com/u/59462357?v=4", "gravatar_id": "", "url": "https://api.github.com/users/stevhliu", "html_url": "https://github.com/stevhliu", "followers_url": "https://api.github.com/users/stevhliu/followers", "following_url": "https://api.github.com/users/stevhliu/following{/other_user}", "gists_url": "https://api.github.com/users/stevhliu/gists{/gist_id}", "starred_url": "https://api.github.com/users/stevhliu/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/stevhliu/subscriptions", "organizations_url": "https://api.github.com/users/stevhliu/orgs", "repos_url": "https://api.github.com/users/stevhliu/repos", "events_url": "https://api.github.com/users/stevhliu/events{/privacy}", "received_events_url": "https://api.github.com/users/stevhliu/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27986/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27986", "html_url": "https://github.com/huggingface/transformers/pull/27986", "diff_url": "https://github.com/huggingface/transformers/pull/27986.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27986.patch", "merged_at": 1702670815000 }

{ "login": "MohamedAliRashad", "id": 26205298, "node_id": "MDQ6VXNlcjI2MjA1Mjk4", "avatar_url": "https://avatars.githubusercontent.com/u/26205298?v=4", "gravatar_id": "", "url": "https://api.github.com/users/MohamedAliRashad", "html_url": "https://github.com/MohamedAliRashad", "followers_url": "https://api.github.com/users/MohamedAliRashad/followers", "following_url": "https://api.github.com/users/MohamedAliRashad/following{/other_user}", "gists_url": "https://api.github.com/users/MohamedAliRashad/gists{/gist_id}", "starred_url": "https://api.github.com/users/MohamedAliRashad/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/MohamedAliRashad/subscriptions", "organizations_url": "https://api.github.com/users/MohamedAliRashad/orgs", "repos_url": "https://api.github.com/users/MohamedAliRashad/repos", "events_url": "https://api.github.com/users/MohamedAliRashad/events{/privacy}", "received_events_url": "https://api.github.com/users/MohamedAliRashad/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @MohamedAliRashad, thanks for raising this issue! \r\n\r\nCould you share the error encountered including its full traceback as well? ", "@amyeroberts \r\nYou are unable to reproduce ?", "Hi! \r\nI'm Having the same Error: \r\nonly happens when the token length is greater than the sliding window size\r\n(I do not have the error with transformers version 4.34.0, but when I upgrade to 4.36.0 I get the error)\r\n\r\nThanks! :) \r\n\r\n- transformers version: 4.36.0\r\n- Platform: Linux 5.10.0-26-cloud-amd64\r\n- Python version: 3.10.13\r\n- Huggingface_hub version: 0.19.4\r\n- Safetensors version: 0.4.1\r\n- PyTorch version (GPU?): 2.1.1+cu121 (True)\r\n- Tensorflow version (GPU?): not installed (NA)\r\n- Flax version (CPU?/GPU?/TPU?): not installed (NA)\r\n- Jax version: not installed\r\n- JaxLib version: not installed\r\n- Using GPU in script?: Yes\r\n- Using distributed or parallel set-up in script?: Yes\r\n\r\nTraceback (most recent call last):\r\nFile \"/opt/conda/lib/python3.10/site-packages/uvicorn/protocols/http/httptools_impl.py\", line 426, in run_asgi\r\n result = await app( # type: ignore[func-returns-value]\r\n File \"/opt/conda/lib/python3.10/site-packages/uvicorn/middleware/proxy_headers.py\", line 84, in __call__\r\n return await self.app(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/fastapi/applications.py\", line 1106, in __call__\r\n await super().__call__(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/applications.py\", line 122, in __call__\r\n await self.middleware_stack(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/middleware/errors.py\", line 184, in __call__\r\n raise exc\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/middleware/errors.py\", line 162, in __call__\r\n await self.app(scope, receive, _send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/middleware/cors.py\", line 83, in __call__\r\n await self.app(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/middleware/exceptions.py\", line 79, in __call__\r\n raise exc\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/middleware/exceptions.py\", line 68, in __call__\r\n await self.app(scope, receive, sender)\r\n File \"/opt/conda/lib/python3.10/site-packages/fastapi/middleware/asyncexitstack.py\", line 20, in __call__\r\n raise e\r\n File \"/opt/conda/lib/python3.10/site-packages/fastapi/middleware/asyncexitstack.py\", line 17, in __call__\r\n await self.app(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/routing.py\", line 718, in __call__\r\n await route.handle(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/routing.py\", line 276, in handle\r\n await self.app(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/routing.py\", line 66, in app\r\n response = await func(request)\r\n File \"/opt/conda/lib/python3.10/site-packages/fastapi/routing.py\", line 274, in app\r\n raw_response = await run_endpoint_function(\r\n File \"/opt/conda/lib/python3.10/site-packages/fastapi/routing.py\", line 191, in run_endpoint_function\r\n return await dependant.call(**values)\r\n File \"/home/mmockus/dev/chatbot/rGPT/host.py\", line 301, in benchmark_model\r\n response, time, previous_prompt = rgpt(\r\n File \"/home/mmockus/dev/chatbot/rGPT/RGPT.py\", line 384, in __call__\r\n output = self.__generate_text(final_prompt)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/pipelines/base.py\", line 1140, in __call__\r\n return self.run_single(inputs, preprocess_params, forward_params, postprocess_params)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/pipelines/base.py\", line 1147, in run_single\r\n model_outputs = self.forward(model_inputs, **forward_params)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/pipelines/base.py\", line 1046, in forward\r\n model_outputs = self._forward(model_inputs, **forward_params)\r\n File \"/home/mmockus/dev/chatbot/rGPT/instruct_pipeline.py\", line 60, in _forward\r\n generated_sequence = self.model.generate(\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/utils/_contextlib.py\", line 115, in decorate_context\r\n return func(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/generation/utils.py\", line 1764, in generate\r\n return self.sample(\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/generation/utils.py\", line 2861, in sample\r\n outputs = self(\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/accelerate/hooks.py\", line 165, in new_forward\r\n output = module._old_forward(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/models/mixtral/modeling_mixtral.py\", line 1212, in forward\r\n outputs = self.model(\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/models/mixtral/modeling_mixtral.py\", line 1080, in forward\r\n layer_outputs = decoder_layer(\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/accelerate/hooks.py\", line 165, in new_forward\r\n output = module._old_forward(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/models/mixtral/modeling_mixtral.py\", line 796, in forward\r\n hidden_states, self_attn_weights, present_key_value = self.self_attn(\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/accelerate/hooks.py\", line 165, in new_forward\r\n output = module._old_forward(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/models/mixtral/modeling_mixtral.py\", line 441, in forward\r\n past_key = past_key_value[0]\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/cache_utils.py\", line 78, in __getitem__\r\n raise KeyError(f\"Cache only has {len(self)} layers, attempted to access layer with index {layer_idx}\")\r\nKeyError: 'Cache only has 0 layers, attempted to access layer with index 0'", "@manueljmockus Thanks for providing these details! As Mixtral was only part of the most recent release - does this mean this error message was generated loading and using a different model than in the issue example? \r\n\r\n> @amyeroberts You are unable to reproduce ?\r\n\r\n@MohamedAliRashad We ask that anyone posting an issue posts the full traceback. We have many issues opened and PRs to review a day; to help get your issue resolved as soon as possible, you need to help us help you. With partial information we can't know if we're trying to solve the right issue - or we might be able to reply with a solution from just seeing the stack trace. ", "cc @tomaarsen as this seems like a cache issue ", "@amyeroberts \r\nI understand.\r\n\r\nI am asking if you were unable to produce the issue with the code i provided because i think it is an OOM probelm (the cached layers can't be accessed because the input has overwritten them).", "@amyeroberts \r\nHi, sorry if I was not clear. I was Using \"mistralai/Mistral-7B-Instruct-v0.1\" with transformers version 4.34.0 without an issue. \r\nYesterday after upgrading to 4.36.0 to test mixtral I am getting the error mentioned, Both with \"mistralai/Mixtral-8x7B-Instruct-v0.1\" and \"mistralai/Mistral-7B-Instruct-v0.1\" models. Here are the logs from the mistral execution:\r\n\r\nTraceback (most recent call last):\r\n File \"/opt/conda/lib/python3.10/site-packages/uvicorn/protocols/http/httptools_impl.py\", line 426, in run_asgi\r\n result = await app( # type: ignore[func-returns-value]\r\n File \"/opt/conda/lib/python3.10/site-packages/uvicorn/middleware/proxy_headers.py\", line 84, in __call__\r\n return await self.app(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/fastapi/applications.py\", line 1106, in __call__\r\n await super().__call__(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/applications.py\", line 122, in __call__\r\n await self.middleware_stack(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/middleware/errors.py\", line 184, in __call__\r\n raise exc\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/middleware/errors.py\", line 162, in __call__\r\n await self.app(scope, receive, _send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/middleware/cors.py\", line 83, in __call__\r\n await self.app(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/middleware/exceptions.py\", line 79, in __call__\r\n raise exc\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/middleware/exceptions.py\", line 68, in __call__\r\n await self.app(scope, receive, sender)\r\n File \"/opt/conda/lib/python3.10/site-packages/fastapi/middleware/asyncexitstack.py\", line 20, in __call__\r\n raise e\r\n File \"/opt/conda/lib/python3.10/site-packages/fastapi/middleware/asyncexitstack.py\", line 17, in __call__\r\n await self.app(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/routing.py\", line 718, in __call__\r\n await route.handle(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/routing.py\", line 276, in handle\r\n await self.app(scope, receive, send)\r\n File \"/opt/conda/lib/python3.10/site-packages/starlette/routing.py\", line 66, in app\r\n response = await func(request)\r\n File \"/opt/conda/lib/python3.10/site-packages/fastapi/routing.py\", line 274, in app\r\n raw_response = await run_endpoint_function(\r\n File \"/opt/conda/lib/python3.10/site-packages/fastapi/routing.py\", line 191, in run_endpoint_function\r\n return await dependant.call(**values)\r\n File \"/home/mmockus/dev/chatbot/rGPT/host.py\", line 302, in benchmark_model\r\n response, time, previous_prompt = rgpt(\r\n File \"/home/mmockus/dev/chatbot/rGPT/RGPT.py\", line 386, in __call__\r\n output = self.__generate_text(final_prompt)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/pipelines/base.py\", line 1140, in __call__\r\n return self.run_single(inputs, preprocess_params, forward_params, postprocess_params)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/pipelines/base.py\", line 1147, in run_single\r\n model_outputs = self.forward(model_inputs, **forward_params)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/pipelines/base.py\", line 1046, in forward\r\n model_outputs = self._forward(model_inputs, **forward_params)\r\n File \"/home/mmockus/dev/chatbot/rGPT/instruct_pipeline.py\", line 60, in _forward\r\n generated_sequence = self.model.generate(\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/utils/_contextlib.py\", line 115, in decorate_context\r\n return func(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/generation/utils.py\", line 1764, in generate\r\n return self.sample(\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/generation/utils.py\", line 2861, in sample\r\n outputs = self(\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/accelerate/hooks.py\", line 165, in new_forward\r\n output = module._old_forward(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/models/mistral/modeling_mistral.py\", line 1044, in forward\r\n outputs = self.model(\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/models/mistral/modeling_mistral.py\", line 929, in forward\r\n layer_outputs = decoder_layer(\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/accelerate/hooks.py\", line 165, in new_forward\r\n output = module._old_forward(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/models/mistral/modeling_mistral.py\", line 654, in forward\r\n hidden_states, self_attn_weights, present_key_value = self.self_attn(\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/accelerate/hooks.py\", line 165, in new_forward\r\n output = module._old_forward(*args, **kwargs)\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/models/mistral/modeling_mistral.py\", line 391, in forward\r\n past_key = past_key_value[0]\r\n File \"/opt/conda/lib/python3.10/site-packages/transformers/cache_utils.py\", line 78, in __getitem__\r\n raise KeyError(f\"Cache only has {len(self)} layers, attempted to access layer with index {layer_idx}\")\r\nKeyError: 'Cache only has 0 layers, attempted to access layer with index 0'\r\n\r\nThanks for yout help!", "I am facing the same problem while using the official transformers examples (pytorch/run_clm with trainer). \r\n\r\nSpecifically, I only get this exact error when 1. training context window is larger than sliding window (4096), and 2. I am using zero3/fsdp. When context window is less than 4096, I get a different error, and when using zero1/2, training works as expected.", "This indeed seems like a caching issue. cc @gante\r\nIt seems like this snippet was not updated to work with the new Cache class: https://github.com/huggingface/transformers/blob/2788f8d8d5f9cee2fe33a9292b0f3570bd566a6d/src/transformers/models/mistral/modeling_mistral.py#L388-L407\r\n\r\nI suspect that not using Flash Attention 2 may solve the issue in the meantime.\r\n\r\n- Tom Aarsen", "> This indeed seems like a caching issue. cc @gante It seems like this snippet was not updated to work with the new Cache class:\r\n> \r\n> https://github.com/huggingface/transformers/blob/2788f8d8d5f9cee2fe33a9292b0f3570bd566a6d/src/transformers/models/mistral/modeling_mistral.py#L388-L407\r\n> \r\n> I suspect that not using Flash Attention 2 may solve the issue in the meantime.\r\n> \r\n> * Tom Aarsen\r\n\r\nHi, Thanks for the quick response. I can confirm that removing flash attention solves this issue. However removing it means that long contexts take up too much memory and i get a Cuda OOM.", "Opening a PR to fix it :)", "FYI, this seems to be a distributed error: in a single GPU device, the following script runs with no exceptions\r\n\r\n**EDIT**: it doesn't fail because `mistralai/Mistral-7B-Instruct-v0.2` doesn't use a sliding window 👀 \r\n\r\n```py\r\nimport torch\r\nfrom transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer\r\n\r\nlong_text = \"Foo Bar \" * 5000\r\n\r\nmodel_id = \"mistralai/Mistral-7B-Instruct-v0.2\"\r\ntokenizer = AutoTokenizer.from_pretrained(model_id)\r\n\r\nmodel = AutoModelForCausalLM.from_pretrained(model_id, use_flash_attention_2=True, torch_dtype=torch.float16, device_map=\"auto\")\r\n\r\nmessages = [\r\n {\"role\": \"user\", \"content\": f\"Summarize the following:\\n{long_text}\"}\r\n]\r\n\r\ninputs = tokenizer.apply_chat_template(messages, return_tensors=\"pt\").to(model.device)\r\nprint(inputs.shape)\r\n\r\noutputs = model.generate(inputs, max_new_tokens=8192, do_sample=True)\r\nprint(outputs.shape)\r\n```", "I can confirm, this lines up with my experience since zero1/2 works but zero3/fsdp does not.", "The PR linked above fixes it :)", "Thank you so much! I can confirm training works now :)", "Hi @gante , my transformers version is 4.36.2 （having viewed this issue I pip -U immediatly ） but still face the same question above when using standard llama2-7b-hf.\r\nWhen loading model, I get a warning \r\n\r\n> 'Instantiating LlamaAttentionLayerBetterTransformer without passing `layer_idx` is not recommended and will to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` when creating this class.'\r\n\r\nEventually\r\n\r\n> KeyError: 'Cache only has 0 layers, attempted to access layer with index 0'\r\n\r\n`self.model = AutoModelForCausalLM.from_pretrained(model_name,torch_dtype='auto',low_cpu_mem_usage=True).to_bettertransformer()`\r\nI dive into the code about modeling_llama.py \r\n\r\nSo I check my config.json also, it seems nothing wrong with num_hidden_layers.\r\n`print(self.model.config.num_hidden_layers)` ->32\r\nis there any possible reason cause this? (btw since this issue is closed, shall I open a new one? : | Thanks for your help in advance! \r\n", "Hey @rangehow 👋 \r\n\r\nFrom your error, it seems that you are using BetterTransformer + Llama, which is deprecated. See [here](https://github.com/huggingface/optimum/releases/tag/v1.16.1) for more information (and how to update it)", "> > This indeed seems like a caching issue. cc @gante It seems like this snippet was not updated to work with the new Cache class:\r\n> > https://github.com/huggingface/transformers/blob/2788f8d8d5f9cee2fe33a9292b0f3570bd566a6d/src/transformers/models/mistral/modeling_mistral.py#L388-L407\r\n> > \r\n> > I suspect that not using Flash Attention 2 may solve the issue in the meantime.\r\n> > \r\n> > * Tom Aarsen\r\n> \r\n> Hi, Thanks for the quick response. I can confirm that removing flash attention solves this issue. However removing it means that long contexts take up too much memory and i get a Cuda OOM.\r\n\r\n\r\n\r\nDo you know if have same issue? I am using TheBloke/Llama-2-13B-chat-GPTQ with GPU A100 80GB? if yes, please help me which file to change or update?\r\n\r\n\r\n File \"/home//miniconda3/envs/NetAI_B/lib/python3.11/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home//miniconda3/envs/NetAI_B/lib/python3.11/site-packages/transformers/models/llama/modeling_llama.py\", line 1183, in forward\r\n outputs = self.model(\r\n ^^^^^^^^^^^\r\n File \"/home//miniconda3/envs/NetAI_B/lib/python3.11/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home//miniconda3/envs/NetAI_B/lib/python3.11/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home//miniconda3/envs/NetAI_B/lib/python3.11/site-packages/transformers/models/llama/modeling_llama.py\", line 1070, in forward\r\n layer_outputs = decoder_layer(\r\n ^^^^^^^^^^^^^^\r\n File \"/home//miniconda3/envs/NetAI_B/lib/python3.11/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home//miniconda3/envs/NetAI_B/lib/python3.11/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home//miniconda3/envs/NetAI_B/lib/python3.11/site-packages/transformers/models/llama/modeling_llama.py\", line 798, in forward\r\n hidden_states, self_attn_weights, present_key_value = self.self_attn(\r\n ^^^^^^^^^^^^^^^\r\n File \"/home//miniconda3/envs/NetAI_B/lib/python3.11/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home//miniconda3/envs/NetAI_B/lib/python3.11/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home//miniconda3/envs/NetAI_B/lib/python3.11/site-packages/auto_gptq/nn_modules/fused_llama_attn.py\", line 62, in forward\r\n kv_seq_len += past_key_value[0].shape[-2]\r\n ~~~~~~~~~~~~~~^^^\r\n File \"/home//miniconda3/envs/NetAI_B/lib/python3.11/site-packages/transformers/cache_utils.py\", line 78, in __getitem__\r\n raise KeyError(f\"Cache only has {len(self)} layers, attempted to access layer with index {layer_idx}\")\r\nKeyError: 'Cache only has 0 layers, attempted to access layer with index 0'", "@bp020108 make sure you're running the latest `transformers` version. If the issue persists after updating `transformers` to `v4.38`, we'll need a script to reproduce your issue 🤗 ", "looks like I have 4.37.2. \r\n\r\n-vm:~/miniconda3/LLAMA/localchat$ pip3.11 list | grep transf\r\nsentence-transformers 2.2.2\r\ntransformers 4.37.2\r\n\r\nyou want me to upgrade this?\r\npip3.11 install transformers==4.38?\r\n" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27985/reactions", "total_count": 4, "+1": 4, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "dberenbaum", "id": 2308172, "node_id": "MDQ6VXNlcjIzMDgxNzI=", "avatar_url": "https://avatars.githubusercontent.com/u/2308172?v=4", "gravatar_id": "", "url": "https://api.github.com/users/dberenbaum", "html_url": "https://github.com/dberenbaum", "followers_url": "https://api.github.com/users/dberenbaum/followers", "following_url": "https://api.github.com/users/dberenbaum/following{/other_user}", "gists_url": "https://api.github.com/users/dberenbaum/gists{/gist_id}", "starred_url": "https://api.github.com/users/dberenbaum/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/dberenbaum/subscriptions", "organizations_url": "https://api.github.com/users/dberenbaum/orgs", "repos_url": "https://api.github.com/users/dberenbaum/repos", "events_url": "https://api.github.com/users/dberenbaum/events{/privacy}", "received_events_url": "https://api.github.com/users/dberenbaum/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27983/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27983", "html_url": "https://github.com/huggingface/transformers/pull/27983", "diff_url": "https://github.com/huggingface/transformers/pull/27983.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27983.patch", "merged_at": 1702416599000 }

{ "login": "dberenbaum", "id": 2308172, "node_id": "MDQ6VXNlcjIzMDgxNzI=", "avatar_url": "https://avatars.githubusercontent.com/u/2308172?v=4", "gravatar_id": "", "url": "https://api.github.com/users/dberenbaum", "html_url": "https://github.com/dberenbaum", "followers_url": "https://api.github.com/users/dberenbaum/followers", "following_url": "https://api.github.com/users/dberenbaum/following{/other_user}", "gists_url": "https://api.github.com/users/dberenbaum/gists{/gist_id}", "starred_url": "https://api.github.com/users/dberenbaum/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/dberenbaum/subscriptions", "organizations_url": "https://api.github.com/users/dberenbaum/orgs", "repos_url": "https://api.github.com/users/dberenbaum/repos", "events_url": "https://api.github.com/users/dberenbaum/events{/privacy}", "received_events_url": "https://api.github.com/users/dberenbaum/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27982/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27982", "html_url": "https://github.com/huggingface/transformers/pull/27982", "diff_url": "https://github.com/huggingface/transformers/pull/27982.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27982.patch", "merged_at": null }

{ "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 }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27981/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27981", "html_url": "https://github.com/huggingface/transformers/pull/27981", "diff_url": "https://github.com/huggingface/transformers/pull/27981.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27981.patch", "merged_at": 1702413162000 }

{ "login": "amyeroberts", "id": 22614925, "node_id": "MDQ6VXNlcjIyNjE0OTI1", "avatar_url": "https://avatars.githubusercontent.com/u/22614925?v=4", "gravatar_id": "", "url": "https://api.github.com/users/amyeroberts", "html_url": "https://github.com/amyeroberts", "followers_url": "https://api.github.com/users/amyeroberts/followers", "following_url": "https://api.github.com/users/amyeroberts/following{/other_user}", "gists_url": "https://api.github.com/users/amyeroberts/gists{/gist_id}", "starred_url": "https://api.github.com/users/amyeroberts/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/amyeroberts/subscriptions", "organizations_url": "https://api.github.com/users/amyeroberts/orgs", "repos_url": "https://api.github.com/users/amyeroberts/repos", "events_url": "https://api.github.com/users/amyeroberts/events{/privacy}", "received_events_url": "https://api.github.com/users/amyeroberts/received_events", "type": "User", "site_admin": false }

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

[]

[ "Hi @amyeroberts , Can I take up the task of porting this model ?\r\n", "@venkateshtata Certainly! Feel free to open a PR and ping me when it's ready for review. Let us know if you have any questions about porting into transformers in the meantime :) " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27980/reactions", "total_count": 2, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 2, "rocket": 0, "eyes": 0 }

{ "login": "gante", "id": 12240844, "node_id": "MDQ6VXNlcjEyMjQwODQ0", "avatar_url": "https://avatars.githubusercontent.com/u/12240844?v=4", "gravatar_id": "", "url": "https://api.github.com/users/gante", "html_url": "https://github.com/gante", "followers_url": "https://api.github.com/users/gante/followers", "following_url": "https://api.github.com/users/gante/following{/other_user}", "gists_url": "https://api.github.com/users/gante/gists{/gist_id}", "starred_url": "https://api.github.com/users/gante/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/gante/subscriptions", "organizations_url": "https://api.github.com/users/gante/orgs", "repos_url": "https://api.github.com/users/gante/repos", "events_url": "https://api.github.com/users/gante/events{/privacy}", "received_events_url": "https://api.github.com/users/gante/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@patrickvonplaten tagging you here for a 2nd set of eyes on the speculative decoding method (changes in `utils.py`), which I'm assuming you're familiar with. Feel free to delegate to someone else who is familiar with the method! 🤗 ", "Thanks for adding this! Can we split this up into two separate PRs: one changing the assisted generation and the other adding speculative decoding?", "@amyeroberts pulled the assisted generation changes into this PR: https://github.com/huggingface/transformers/pull/28030\r\n\r\nAfter it is merged, I will rebase this one and ping you again -- this one will become exclusively about speculative decoding 🤗 ", "@amyeroberts I've rerun the slow tests, and I can confirm they are passing. Ready for a review :)", "@patrickvonplaten the two types of sampling are needed :D \r\n\r\nNew candidate-based methods are popping up (e.g. https://github.com/huggingface/transformers/pull/27775), and they don't necessarily have logits. As such, speculative decoding, which needs the candidates' logits, can't be applied to those methods. ", "> @patrickvonplaten the two types of sampling are needed :D\r\n> \r\n> New candidate-based methods are popping up (e.g. #27775), and they don't necessarily have logits. As such, speculative decoding, which needs the candidates' logits, can't be applied to those methods.\r\n\r\nBut shouldn't they just be the \"own\" method now? I.e. I don't think we should put https://github.com/huggingface/transformers/pull/27775 into the speculative decoding method no? \r\n\r\n", "@patrickvonplaten #27775 does not introduce changes to assisted generation 🤗 In #28030 I've abstracted the candidate generation part of assisted generation. We now load candidate generators the same way as we load the logits processors:\r\n\r\nhttps://github.com/huggingface/transformers/blob/e6dcf8abd6f65bb4b6dfc1831b20d9ba49ce00e2/src/transformers/generation/utils.py#L899-L919\r\n\r\nIn assisted generation, we call the candidate generator to get candidate sequences (which may or may not contain associated logits, depending on the method)\r\n\r\nhttps://github.com/huggingface/transformers/blob/e6dcf8abd6f65bb4b6dfc1831b20d9ba49ce00e2/src/transformers/generation/utils.py#L4588\r\n\r\nThe technique in #27775 can thus be added by adding a new candidate generator in ` _get_candidate_generator`. Other candidate generators may be added the same way, enabling users to experiment with the concept of candidates!\r\n\r\nBecause needing the logits (for speculative decoding) is a very limiting constraint, I'd rather keep the two sampling paths.", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27979). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "@amyeroberts PR comments addressed 🤗 \r\n\r\n@patrickvonplaten Unless you don't strongly oppose, I'd like to keep the two sampling paths, for the reasons I've written [here](https://github.com/huggingface/transformers/pull/27979#issuecomment-1860125398) -- I think it will be beneficial in the long run! :) (otherwise, a whole new generation method has to be written for #27775)", "@amyeroberts -- @patrickvonplaten and I had a chat about whether to keep the two sampling paths or not. For context, here's what we agreed on:\r\n- It's okay to leave it as is, and perhaps abstract the different ways we accept candidates into a `candidate_checker` block. \r\n- Be conservative on adding new candidate generators, so we don't end up with unused methods\r\n- [in a follow-up PR] squash other cases where the decoding method is the same except for the token selection, like `greedy_decoding` + `sample`\r\n- [in a follow-up PR] mode each decoding method into its own file. There are several private functions in `generation/utils.py` that are exclusively used with one generation method.", "@gante \r\nAccording to experiments reported in Leviathan's paper, speculative decoding (SD) has higher speedup with greedy decoding (temp=0). However, in the current implementation, SD works only with do_sample=True. ", "@jmamou speculative decoding with `do_sample=False` (or `temp=0`) was already encoded in `assisted_generation`, before this PR -- try calling `model.generate(input_ids, do_sample=False, assistant_model=assistant_model)` :)", "@gante \r\nSince acceptance criteria are different between speculative decoding and assisted generation, I think that it would be great to be able to run both speculative decoding and assisted generation with no sampling.", "@gante \r\nI implemented it. I can submit a PR.", "@gante \r\nIn previous implementation of assisted generation (4.33) with heuristical update of `num_assistant_tokens` (or `max_assistant_tokens`), the value of `num_assistant_tokens` was preserved between 2 consecutive generate() calls.\r\n\r\nIn current implementation (4.38), `num_assistant_tokens` is updated by the `candidate_generator` during the generation but `assistant_model.generation_config.num_assistant_tokens` is not updated at the end of the generation. Therefore, next call to generate will start with the initial value of `assistant_model.generation_config.num_assistant_tokens` (5).\r\n\r\nIs it intentional? If that's a bug, I can open a PR to fix it. ", "@jmamou \r\n\r\n> Since acceptance criteria are different between speculative decoding and assisted generation, I think that it would be great to be able to run both speculative decoding and assisted generation with no sampling.\r\n\r\nNot sure if this is a good idea\r\n1. if we see greedy decoding as applying `temperature=0`, the model probability will be `1` at the most likely token and `0` everywhere else. In turn, this implies that `p_i/q_i` is `>=1` at all positions, and thus all candidate tokens would be accepted 👉 speculative decoding would be the same as simply using the assistant model \r\n2. If we don't apply `temperature=0`, then it would be sampling -- in other words, it wouldn't be greedy decoding\r\n\r\n> In previous implementation of assisted generation (4.33) with heuristical update of num_assistant_tokens (or max_assistant_tokens), the value of num_assistant_tokens was preserved between 2 consecutive generate() calls.\r\nIn current implementation (4.38), num_assistant_tokens is updated by the candidate_generator during the generation but assistant_model.generation_config.num_assistant_tokens is not updated at the end of the generation. Therefore, next call to generate will start with the initial value of assistant_model.generation_config.num_assistant_tokens (5).\r\nIs it intentional? If that's a bug, I can open a PR to fix it.\r\n\r\nThis is a good point! A PR to revert to the previous behaviour (with a test) would be appreciated 🙏 " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27979/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27979", "html_url": "https://github.com/huggingface/transformers/pull/27979", "diff_url": "https://github.com/huggingface/transformers/pull/27979.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27979.patch", "merged_at": 1702994310000 }

{ "login": "ArthurZucker", "id": 48595927, "node_id": "MDQ6VXNlcjQ4NTk1OTI3", "avatar_url": "https://avatars.githubusercontent.com/u/48595927?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ArthurZucker", "html_url": "https://github.com/ArthurZucker", "followers_url": "https://api.github.com/users/ArthurZucker/followers", "following_url": "https://api.github.com/users/ArthurZucker/following{/other_user}", "gists_url": "https://api.github.com/users/ArthurZucker/gists{/gist_id}", "starred_url": "https://api.github.com/users/ArthurZucker/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ArthurZucker/subscriptions", "organizations_url": "https://api.github.com/users/ArthurZucker/orgs", "repos_url": "https://api.github.com/users/ArthurZucker/repos", "events_url": "https://api.github.com/users/ArthurZucker/events{/privacy}", "received_events_url": "https://api.github.com/users/ArthurZucker/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_27978). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "Glad to know it doesn't cause (many) extra issues 😆 ", "Not at all! It's really nice for reviews, we are certain that only this part is what we need to look", "Picking this back soon 🤗 " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27978/reactions", "total_count": 3, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 3, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27978", "html_url": "https://github.com/huggingface/transformers/pull/27978", "diff_url": "https://github.com/huggingface/transformers/pull/27978.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27978.patch", "merged_at": null }

{ "login": "lombardata", "id": 39915110, "node_id": "MDQ6VXNlcjM5OTE1MTEw", "avatar_url": "https://avatars.githubusercontent.com/u/39915110?v=4", "gravatar_id": "", "url": "https://api.github.com/users/lombardata", "html_url": "https://github.com/lombardata", "followers_url": "https://api.github.com/users/lombardata/followers", "following_url": "https://api.github.com/users/lombardata/following{/other_user}", "gists_url": "https://api.github.com/users/lombardata/gists{/gist_id}", "starred_url": "https://api.github.com/users/lombardata/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/lombardata/subscriptions", "organizations_url": "https://api.github.com/users/lombardata/orgs", "repos_url": "https://api.github.com/users/lombardata/repos", "events_url": "https://api.github.com/users/lombardata/events{/privacy}", "received_events_url": "https://api.github.com/users/lombardata/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @lombardata,\r\n\r\nYou can specify both the size the image is resized to during the `resize` call and the crop size e.g.: \r\n\r\n```\r\nimage_processor = AutoImageProcessor.from_pretrained(checkpoint, crop_size={\"height\": 320, \"width\": 256})\r\n```\r\n\r\nnote: I don't know if in this timm example if the dimensions are in (h,w) or (w,h) order. ", "hi @amyeroberts , thank you very much for your quick reply. In this case what is the \"meaning\" of the **image_size** parameter in the model config file?\r\nLet's say that my images are 1080 x 1920 and I process them with : \r\n`image_processor = AutoImageProcessor(checkpoint, crop_size={\"height\": 1080, \"width\": 1920})\r\n`then if in the model config.json file I keep the default parameter : \r\n` \"image_size\": 518,\r\n`What would be the behaviour of the training?\r\nThank you a lot !", "You're hitting on one of the tricky coupling issues between models and their data!\r\n\r\nFor the image processor, `crop_size` and `size` control the processing logic. This is independent of the model and modifying the behaviour of the image processor won't automatically update the necessary model params. \r\n\r\nFor the DinoV2 model, `image_size` refers to the input image size i.e. the dimensions of the processed images. It controls how the [patches are extracted when creating the embeddings](https://github.com/huggingface/transformers/blob/78172dcdb7cdaf04ec6697d4747c505b2e7b0df0/src/transformers/models/dinov2/modeling_dinov2.py#L147). However, the model employs [interpolation on the embeddings](https://github.com/huggingface/transformers/blob/78172dcdb7cdaf04ec6697d4747c505b2e7b0df0/src/transformers/models/dinov2/modeling_dinov2.py#L83) if the input image is of a different resolution. So for inference, you should be able to pass in different sized images and run things fine. \r\n\r\nIf you want to train a model, then I'd suggest aligning the processor and model configurations. \r\n\r\nTo align the two, you'll need to do something like this: \r\n\r\n```py\r\nfrom transformers import Dinov2Config, Dinov2ForImageClassification, AutoImageProcessor\r\n\r\nimage_height, image_width = 1080, 1920\r\n\r\ncheckpoint = \"facebook/dinov2-base\"\r\n\r\n# Create a new model with randomly initialized weights\r\nmodel_config = Dinov2Config.from_pretrained(checkpoint, image_size=(image_height, image_width))\r\nmodel = Dinov2ForImageClassification(model_config)\r\nimage_processor = AutoImageProcessor.from_pretrained(\r\n checkpoint, image_size={\"height\": image_heaight, \"width\": image_width}\r\n)\r\n```", "> You're hitting on one of the tricky coupling issues between models and their data!\r\n> \r\n> For the image processor, `crop_size` and `size` control the processing logic. This is independent of the model and modifying the behaviour of the image processor won't automatically update the necessary model params.\r\n> \r\n> For the DinoV2 model, `image_size` refers to the input image size i.e. the dimensions of the processed images. It controls how the [patches are extracted when creating the embeddings](https://github.com/huggingface/transformers/blob/78172dcdb7cdaf04ec6697d4747c505b2e7b0df0/src/transformers/models/dinov2/modeling_dinov2.py#L147). However, the model employs [interpolation on the embeddings](https://github.com/huggingface/transformers/blob/78172dcdb7cdaf04ec6697d4747c505b2e7b0df0/src/transformers/models/dinov2/modeling_dinov2.py#L83) if the input image is of a different resolution. So for inference, you should be able to pass in different sized images and run things fine.\r\n> \r\n> If you want to train a model, then I'd suggest aligning the processor and model configurations.\r\n> \r\n> To align the two, you'll need to do something like this:\r\n> \r\n> ```python\r\n> from transformers import Dinov2Config, Dinov2ForImageClassification, AutoImageProcessor\r\n> \r\n> image_height, image_width = 1080, 1920\r\n> \r\n> checkpoint = \"facebook/dinov2-base\"\r\n> \r\n> # Create a new model with randomly initialized weights\r\n> model_config = Dinov2Config.from_pretrained(checkpoint, image_size=(image_height, image_width))\r\n> model = Dinov2ForImageClassification(model_config)\r\n> image_processor = AutoImageProcessor.from_pretrained(\r\n> checkpoint, image_size={\"height\": image_heaight, \"width\": image_width}\r\n> )\r\n> ```\r\n\r\nThank you very much @amyeroberts for your complete reply.\r\nLooking at the source code of `Dinov2Config` I found that the `image_size` parameter must be an int (and not a dict of heigth and width) : \r\n` image_size (`int`, *optional*, defaults to 224):`\r\nso that I don't know if we are allowed to pass rectangular images to this specific model. \r\nMoreover the `AutoImageProcessor` (which in our case is a `BitImageProcessor`) should accept as you said an input size : \r\n`size: Dict[str, int] = None,\r\n`\r\nand\r\n\r\n` if \"shortest_edge\" in size:\r\n size = size[\"shortest_edge\"]\r\n default_to_square = False\r\n elif \"height\" in size and \"width\" in size:\r\n size = (size[\"height\"], size[\"width\"])\r\n else:\r\n raise ValueError(\"Size must contain either 'shortest_edge' or 'height' and 'width'.\")\r\n`\r\nbut when I try to instantiate it : \r\n```\r\nimage_processor2= AutoImageProcessor.from_pretrained(checkpoint_name, \r\n size={\"height\": 720, \"width\": 1080},\r\n #size={\"shortest_edge\": 518},\r\n do_center_crop=True, \r\n do_resize=True, \r\n do_rescale = True, \r\n do_normalize=True)\r\n```\r\nI get the following error : \r\n```\r\nFile .../lib/python3.8/site-packages/transformers/models/bit/image_processing_bit.py:152, in BitImageProcessor.resize(self, image, size, resample, data_format, input_data_format, **kwargs)\r\n size = get_size_dict(size, default_to_square=False)\r\n if \"shortest_edge\" not in size:\r\n--> raise ValueError(f\"The `size` parameter must contain the key `shortest_edge`. Got {size.keys()}\")\r\n output_size = get_resize_output_image_size(\r\n image, size=size[\"shortest_edge\"], default_to_square=False, input_data_format=input_data_format\r\n )\r\n```\r\nwhich is strange since in the corresponding doc : \r\nhttps://github.com/huggingface/transformers/blob/main/src/transformers/models/bit/image_processing_bit.py\r\nthe error is different : \r\n` raise ValueError(\"Size must contain either 'shortest_edge' or 'height' and 'width'.\")\r\n`\r\nDo you know where I'm wrong ?\r\nSorry for bothering you but I'm a little bit lost :) ", "Hi @lombardata, \r\n\r\n`image_size` for the model is an int, but `size` for the image processors should be a dictionary. This is because, when images are passed to the model, their (h, w) dimensions are defined. \r\n\r\nHowever, when processing an image, the output size isn't always fixed, the output height, width it can be calculated based on the input dimensions. For example `size={\"shortest_edge\": s}`, will resize the image so that the shortest edge of the image matches, `\"shortest_edge\"` and it will rescale the other edge to match the input aspect ratio. \r\n\r\nWith regards to the error you're encountering, which version of transformers are you running from? I was able to run the example snippet without error on the most recent version. ", "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/main/CONTRIBUTING.md) are likely to be ignored.", "Hi @amyeroberts , thank you very much for your reply.\r\n\r\n> With regards to the error you're encountering, which version of transformers are you running from? I was able to run the example snippet without error on the most recent version.\r\n\r\nI was running version : 4.34.1, now I upgraded to the latest and it's wirking fine. Thanks !\r\n" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27977/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "Rocketknight1", "id": 12866554, "node_id": "MDQ6VXNlcjEyODY2NTU0", "avatar_url": "https://avatars.githubusercontent.com/u/12866554?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Rocketknight1", "html_url": "https://github.com/Rocketknight1", "followers_url": "https://api.github.com/users/Rocketknight1/followers", "following_url": "https://api.github.com/users/Rocketknight1/following{/other_user}", "gists_url": "https://api.github.com/users/Rocketknight1/gists{/gist_id}", "starred_url": "https://api.github.com/users/Rocketknight1/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Rocketknight1/subscriptions", "organizations_url": "https://api.github.com/users/Rocketknight1/orgs", "repos_url": "https://api.github.com/users/Rocketknight1/repos", "events_url": "https://api.github.com/users/Rocketknight1/events{/privacy}", "received_events_url": "https://api.github.com/users/Rocketknight1/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27976/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/27976", "html_url": "https://github.com/huggingface/transformers/pull/27976", "diff_url": "https://github.com/huggingface/transformers/pull/27976.diff", "patch_url": "https://github.com/huggingface/transformers/pull/27976.patch", "merged_at": 1702392115000 }

{ "login": "elena-soare20", "id": 114069526, "node_id": "U_kgDOBsyQFg", "avatar_url": "https://avatars.githubusercontent.com/u/114069526?v=4", "gravatar_id": "", "url": "https://api.github.com/users/elena-soare20", "html_url": "https://github.com/elena-soare20", "followers_url": "https://api.github.com/users/elena-soare20/followers", "following_url": "https://api.github.com/users/elena-soare20/following{/other_user}", "gists_url": "https://api.github.com/users/elena-soare20/gists{/gist_id}", "starred_url": "https://api.github.com/users/elena-soare20/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/elena-soare20/subscriptions", "organizations_url": "https://api.github.com/users/elena-soare20/orgs", "repos_url": "https://api.github.com/users/elena-soare20/repos", "events_url": "https://api.github.com/users/elena-soare20/events{/privacy}", "received_events_url": "https://api.github.com/users/elena-soare20/received_events", "type": "User", "site_admin": false }

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

[]

[ "Hi @elena-soare20, thanks for raising this issue! \r\n\r\nYes, at the moment InstructBLIP isn't compatible with the pipeline because of the specific processing it does - which is different from many other models. Specifically, it has two tokenizers to create `qformer_input_ids` and `input_ids` to be passed to the model. There's some ongoing work to unify our processors so that hopefully more models like these can be quickly integrated. \r\n\r\nHappy to review any PRs for anyone in the community who would like to enable this. See also: #21110 \r\n", "hey @amyeroberts I would be happy to work on this\r\n", "@nakranivaibhav Awesome! Feel free to ping me for review when you have a PR ready 🤗 ", "@amyeroberts Give me some time on this. The models are very large to reproduce the error. I am figuring out where to reproduce the error to start working on it.", "@nakranivaibhav If all you need is a model to test functionality i.e. a randomly initialized model that outputs nonsense is fine, then the small model used during tests might help here. The config to build the model and test inputs can be [found here](https://github.com/huggingface/transformers/blob/f40b87de0ca234df61f76928956c4a2118c0b548/tests/models/instructblip/test_modeling_instructblip.py#L410).", "@amyeroberts Yes that i what I need. Thank you for pointing it out." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/27975/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "muziyongshixin", "id": 21971718, "node_id": "MDQ6VXNlcjIxOTcxNzE4", "avatar_url": "https://avatars.githubusercontent.com/u/21971718?v=4", "gravatar_id": "", "url": "https://api.github.com/users/muziyongshixin", "html_url": "https://github.com/muziyongshixin", "followers_url": "https://api.github.com/users/muziyongshixin/followers", "following_url": "https://api.github.com/users/muziyongshixin/following{/other_user}", "gists_url": "https://api.github.com/users/muziyongshixin/gists{/gist_id}", "starred_url": "https://api.github.com/users/muziyongshixin/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/muziyongshixin/subscriptions", "organizations_url": "https://api.github.com/users/muziyongshixin/orgs", "repos_url": "https://api.github.com/users/muziyongshixin/repos", "events_url": "https://api.github.com/users/muziyongshixin/events{/privacy}", "received_events_url": "https://api.github.com/users/muziyongshixin/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @muziyongshixin, thanks for raising an issue!\r\n\r\nThis is a question best placed in our [forums](https://discuss.huggingface.co/). We try to reserve the github issues for feature requests and bug reports.\r\n\r\nTo learn about how to how to modify the tokenizers, you can check out the documentation, [1](https://huggingface.co./docs/transformers/v4.36.0/en/main_classes/tokenizer#tokenizer), [2](https://huggingface.co./docs/transformers/v4.36.0/en/main_classes/tokenizer#transformers.PreTrainedTokenizer.add_tokens). For example, you can add tokens to the tokenzers vocabulary by using the [`add_tokens` method](https://huggingface.co./docs/transformers/v4.36.0/en/main_classes/tokenizer#transformers.PreTrainedTokenizer.add_tokens). ", "> Hi @muziyongshixin, thanks for raising an issue!\r\n> \r\n> This is a question best placed in our [forums](https://discuss.huggingface.co/). We try to reserve the github issues for feature requests and bug reports.\r\n> \r\n> To learn about how to how to modify the tokenizers, you can check out the documentation, [1](https://huggingface.co./docs/transformers/v4.36.0/en/main_classes/tokenizer#tokenizer), [2](https://huggingface.co./docs/transformers/v4.36.0/en/main_classes/tokenizer#transformers.PreTrainedTokenizer.add_tokens). For example, you can add tokens to the tokenzers vocabulary by using the [`add_tokens` method](https://huggingface.co./docs/transformers/v4.36.0/en/main_classes/tokenizer#transformers.PreTrainedTokenizer.add_tokens).\r\n\r\nthanks for your reply@amyeroberts \r\n\r\nI know the `add_tokens` function, but I don't want to change the vocabulary size, this will introduce some inconvience in finetuning process(like resize the embedding layer and the lm_head).\r\nSo I am wondering whether there is a method to replace the existing tokens.\r\n\r\nP.S. I am using a sentencepiece tokenizer which is released in [this baichuan2 repo](https://github.com/baichuan-inc/Baichuan2) \r\n", "@ArthurZucker will be best placed to answer this. I believe `add_tokens` or `add_special_tokens` is the safe way to do this, as it correctly keeps track of the modifications vs. from the loaded checkpoint and ensures the correct preprocessing is applied before mapping to the token id. It should be possible to define your own vocab files to load a tokenizer but I don't know about the guarantees of the behaviour vs the original tokenizer. ", "hey! You should modify manually both the `added_tokens_decoder` field (saved in the `tokenizer_config.json` ) and the `added_tokens` field (saved in the `tokenizer.json`). We don't really support this manually, but that is the recommended way to do it! (If the reserved tokens were already part of the vocab, so not AddedTokens, then you have to overwrite the vocab as well, the vocab files, to make sure they are removed) that would be hard than if it's just the content of the added tokens that you are trying to modify 😉 ", "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/main/CONTRIBUTING.md) are likely to be ignored." ]