Datasets:

DanielPFlorian

/

Transformers-Github-Issues

like 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 }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28557). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "@amyeroberts Thank you. `logger.warning_once` works, however it does not in case we load several time models in the same python process:\r\n\r\n```python\r\nfrom transformers import AutoModelForCausalLM, AutoConfig, LlamaForCausalLM\r\nimport torch\r\n\r\nmodel = LlamaForCausalLM.from_pretrained(\"fxmarty/tiny-llama-fast-tokenizer\", use_flash_attention_2=True, torch_dtype=torch.float32)\r\n\r\nprint(\"LOAD AGAIN\")\r\nmodel = LlamaForCausalLM.from_pretrained(\"fxmarty/tiny-llama-fast-tokenizer\", use_flash_attention_2=True, torch_dtype=torch.float32)\r\n```\r\n\r\nHere the second call does not log anything at all with `logger.warning_once`.\r\n\r\n---\r\n\r\nThe thing is that there are three ways to load models in transformers: `from_pretrained`, `from_config`, and `__init__`. In order to keep the same behavior for all three w.r.t. attention implementation, we call `_autoset_attn_implementation` in these three methods. This results in `_autoset_attn_implementation` being called twice in case we initialize a model from `PreTrainedModel.from_pretrained` or `PreTrainedModel.from_config` (as they call in turn `PreTrainedModel.__init__`).\r\n\r\nFor example, from the above script without any fix, we may get this log:\r\n```\r\nFlash Attention 2.0 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in LlamaForCausalLM is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained(\"openai/whisper-tiny\", attn_implementation=\"flash_attention_2\", torch_dtype=torch.float16)`\r\nYou are attempting to use Flash Attention 2.0 with a model not initialized on GPU. Make sure to move the model to GPU after initializing it on CPU with `model.to('cuda')`.\r\nFlash Attention 2.0 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in LlamaForCausalLM is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained(\"openai/whisper-tiny\", attn_implementation=\"flash_attention_2\", torch_dtype=torch.float16)`\r\nFlash Attention 2.0 only supports torch.float16 and torch.bfloat16 dtypes, but the current dype in LlamaModel is torch.float32. You should run training or inference using Automatic Mixed-Precision via the `with torch.autocast(device_type='torch_device'):` decorator, or load the model with the `torch_dtype` argument. Example: `model = AutoModel.from_pretrained(\"openai/whisper-tiny\", attn_implementation=\"flash_attention_2\", torch_dtype=torch.float16)`\r\n\r\n```\r\n\r\nwhich is the second point in my original post - we want to avoid such duplicate logs (here duplicate for `LlamaForCausalLM`).\r\n\r\n@amyeroberts Currently I fixed with your suggestion of using `logger.warning_once`. Let me know if you prefer it that way despite the issue I mentioned just above.", "> The thing is that there are three ways to load models in transformers: from_pretrained, from_config, and __init__. In order to keep the same behavior for all three w.r.t. attention implementation, we call _autoset_attn_implementation in these three methods. This results in _autoset_attn_implementation being called twice in case we initialize a model from PreTrainedModel.from_pretrained or PreTrainedModel.from_config (as they call in turn PreTrainedModel.__init__).\r\n\r\nWhy do we need `_autoset_attn_implementation` to be in `_from_config` or `from_pretrained` if it's already in `__init__`? ", "@amyeroberts Because of the following: https://github.com/huggingface/transformers/blob/98dda8ed03ac3f4af5733bdddaa1dab6a81e15c1/src/transformers/modeling_utils.py#L3579-L3586\r\n\r\n`device_map`, `use_flash_attention_2`, `torch_dtype` are arguments that are not passed to `XXXModel.__init__`, `PretrainedModel.__init__`.", "I also think that `torch_dtype` and `device_map` should be passed in here: https://github.com/huggingface/transformers/blob/98dda8ed03ac3f4af5733bdddaa1dab6a81e15c1/src/transformers/modeling_utils.py#L1314, as https://github.com/huggingface/transformers/blob/98dda8ed03ac3f4af5733bdddaa1dab6a81e15c1/src/transformers/modeling_utils.py#L1335 will use these parameters.", "yes, that's in part what this PR is for!", "> is an indication that we're probably not drawing the right boundaries around what is set and when.\r\n\r\nAgreed.", "Hey folks! when will the changes of this PR going to be released? I see it got merged 3 weeks ago and the recent release is 2 weeks ago (v4.37.2) which doesn't include these changes.\r\n\r\nhttps://github.com/huggingface/transformers/blob/345b9b1a6a308a1fa6559251eb33ead2211240ac/src/transformers/modeling_utils.py#L1321", "@rohitgr7 This was not picked in the patch release 4.37.2. Unless there are more critical fixes needed, I guess this will make it in 4.38.0." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28557/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/28557", "html_url": "https://github.com/huggingface/transformers/pull/28557", "diff_url": "https://github.com/huggingface/transformers/pull/28557.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28557.patch", "merged_at": 1706258225000 }

{ "login": "Biswajit2902", "id": 10162006, "node_id": "MDQ6VXNlcjEwMTYyMDA2", "avatar_url": "https://avatars.githubusercontent.com/u/10162006?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Biswajit2902", "html_url": "https://github.com/Biswajit2902", "followers_url": "https://api.github.com/users/Biswajit2902/followers", "following_url": "https://api.github.com/users/Biswajit2902/following{/other_user}", "gists_url": "https://api.github.com/users/Biswajit2902/gists{/gist_id}", "starred_url": "https://api.github.com/users/Biswajit2902/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Biswajit2902/subscriptions", "organizations_url": "https://api.github.com/users/Biswajit2902/orgs", "repos_url": "https://api.github.com/users/Biswajit2902/repos", "events_url": "https://api.github.com/users/Biswajit2902/events{/privacy}", "received_events_url": "https://api.github.com/users/Biswajit2902/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi thank you your code saved my day! I think line 535 needs to modify a bit `prompt_tensor = torch.tensor(generate_kwargs[\"prompt_ids\"], dtype=out[\"tokens\"].dtype).cuda() if is_torch_cuda_available else torch.tensor(generate_kwargs[\"prompt_ids\"], dtype=out[\"tokens\"].dtype)`, and add `is_torch_cuda_available` to line 22. without cuda it'll run on cpu which is a lot slower. ", "@kaminwong , this is just to modify the output sequence to avoid showing `inital_prompt` in transcription.\r\n\r\nActual generation has device handles in below line.\r\n``` python\r\n tokens = self.model.generate(\r\n attention_mask=attention_mask,\r\n **generate_kwargs,\r\n )\r\n```\r\n\r\nApart from this token decoding part is serialised implementation which has no effect, that can be misuse of GPU.", "Thanks for the reply! But if I don't make that changes I get the following error, so I assume `prompt_tensor` needs to be in cuda if device is also in cuda? Or is there any other way to correct the error? Thank you for your time.\r\n\r\n`File \"/.../python3.10/site-packages/transformers/pipelines/automatic_speech_recognition.py\", line 538, in _forward\r\n if (tmp_tokens[0:nprompt_token] == prompt_tensor).sum() == nprompt_token:\r\nRuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!`\r\n\r\nI followed the code you posted:\r\n```from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline\r\n\r\ndevice = \"cuda:0\" if torch.cuda.is_available() else \"cpu\"\r\ntorch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32\r\n\r\nmodel_id = \"openai/whisper-small\"\r\nmodel = AutoModelForSpeechSeq2Seq.from_pretrained(\r\n model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True\r\n)\r\nmodel.to(device)\r\n\r\nprocessor = AutoProcessor.from_pretrained(model_id)\r\n\r\npipe = pipeline(\r\n \"automatic-speech-recognition\",\r\n model=model,\r\n tokenizer=processor.tokenizer,\r\n feature_extractor=processor.feature_extractor,\r\n max_new_tokens=128,\r\n chunk_length_s=15,\r\n batch_size=16,\r\n torch_dtype=torch_dtype,\r\n device=device,\r\n processor=processor\r\n)\r\n\r\n", "@kaminwong , Thank you for addressing. I understood the issue. let me verify and reolved it.", "@kaminwong , you can pull latest commit and install it should work now. its fixed.", "Thank you for the elegant solution. It works now!", "Gentle ping @sanchit-gandhi for review ", "@amyeroberts is there any plan to close this in near future? or will it take time?\r\n", "@Biswajit2902 Once @sanchit-gandhi has reviewed and approved, the PR will need a final review from a maintainer. Once approved, then the PR can be merged in. " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28556/reactions", "total_count": 2, "+1": 0, "-1": 0, "laugh": 0, "hooray": 2, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28556", "html_url": "https://github.com/huggingface/transformers/pull/28556", "diff_url": "https://github.com/huggingface/transformers/pull/28556.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28556.patch", "merged_at": null }

{ "login": "jmamou", "id": 19263306, "node_id": "MDQ6VXNlcjE5MjYzMzA2", "avatar_url": "https://avatars.githubusercontent.com/u/19263306?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jmamou", "html_url": "https://github.com/jmamou", "followers_url": "https://api.github.com/users/jmamou/followers", "following_url": "https://api.github.com/users/jmamou/following{/other_user}", "gists_url": "https://api.github.com/users/jmamou/gists{/gist_id}", "starred_url": "https://api.github.com/users/jmamou/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jmamou/subscriptions", "organizations_url": "https://api.github.com/users/jmamou/orgs", "repos_url": "https://api.github.com/users/jmamou/repos", "events_url": "https://api.github.com/users/jmamou/events{/privacy}", "received_events_url": "https://api.github.com/users/jmamou/received_events", "type": "User", "site_admin": false }

[]

[]

[ "It seems that it has been solved yesterday by @ofirzaf https://github.com/huggingface/transformers/pull/28508#issuecomment-1892870681" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28555/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/28555", "html_url": "https://github.com/huggingface/transformers/pull/28555", "diff_url": "https://github.com/huggingface/transformers/pull/28555.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28555.patch", "merged_at": null }

{ "login": "Biswajit2902", "id": 10162006, "node_id": "MDQ6VXNlcjEwMTYyMDA2", "avatar_url": "https://avatars.githubusercontent.com/u/10162006?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Biswajit2902", "html_url": "https://github.com/Biswajit2902", "followers_url": "https://api.github.com/users/Biswajit2902/followers", "following_url": "https://api.github.com/users/Biswajit2902/following{/other_user}", "gists_url": "https://api.github.com/users/Biswajit2902/gists{/gist_id}", "starred_url": "https://api.github.com/users/Biswajit2902/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Biswajit2902/subscriptions", "organizations_url": "https://api.github.com/users/Biswajit2902/orgs", "repos_url": "https://api.github.com/users/Biswajit2902/repos", "events_url": "https://api.github.com/users/Biswajit2902/events{/privacy}", "received_events_url": "https://api.github.com/users/Biswajit2902/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28554/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/28554", "html_url": "https://github.com/huggingface/transformers/pull/28554", "diff_url": "https://github.com/huggingface/transformers/pull/28554.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28554.patch", "merged_at": null }

{ "login": "liboliba", "id": 51449526, "node_id": "MDQ6VXNlcjUxNDQ5NTI2", "avatar_url": "https://avatars.githubusercontent.com/u/51449526?v=4", "gravatar_id": "", "url": "https://api.github.com/users/liboliba", "html_url": "https://github.com/liboliba", "followers_url": "https://api.github.com/users/liboliba/followers", "following_url": "https://api.github.com/users/liboliba/following{/other_user}", "gists_url": "https://api.github.com/users/liboliba/gists{/gist_id}", "starred_url": "https://api.github.com/users/liboliba/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/liboliba/subscriptions", "organizations_url": "https://api.github.com/users/liboliba/orgs", "repos_url": "https://api.github.com/users/liboliba/repos", "events_url": "https://api.github.com/users/liboliba/events{/privacy}", "received_events_url": "https://api.github.com/users/liboliba/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @liboliba, thanks for raising an issue! \r\n\r\nSo that we can best help you, please make sure to follow the [issue template](https://github.com/huggingface/transformers/blob/main/.github/ISSUE_TEMPLATE/bug-report.yml) including: \r\n* The running environment: run `transformers-cli env` in the terminal and copy-paste the output\r\n* The full error traceback\r\n* A minimal code reproducer. Here we don't have access to the weights. Are there weights you could share which reproduce this error? ", "Thank you for the advise!\r\ntransformers-cli env returns:\r\n- `transformers` version: 4.36.2\r\n- Platform: Linux-3.10.0-1160.36.2.el7.x86_64-x86_64-with-glibc2.17\r\n- Python version: 3.11.5\r\n- Huggingface_hub version: 0.19.4\r\n- Safetensors version: 0.4.1\r\n- Accelerate version: 0.25.0\r\n- Accelerate config: not found\r\n- PyTorch version (GPU?): 2.1.1 (False)\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 & No, I am working under HPC environment, my current compute node has no GPU, but it shared hard disk with a GPU node. A GPU node has no internet connection and can only load model via the shared hard disk with the compute node I work on now. \r\n- Using distributed or parallel set-up in script?: No\r\n\r\nFor the other two bullets, sorry I am less sure how to respond because what I did was to download the official meta llama 2 into a folder, and then I git clone the transformer source code and try to run the conversion code. The error I get now is:\r\n\r\npython /scratch/ll1d19/transformers/src/transformers/models/llama/convert_llama_weights_to_hf.py --input_dir /scratch/ll1d19/llama/llama/llama-2-7b-chat/ --model_size 7B --output_dir /scratch/ll1d19/hf_llama2/Llama-2-7b-chat-hf/\r\nYou are using the default legacy behaviour of the <class 'transformers.models.llama.tokenization_llama.LlamaTokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565\r\nTraceback (most recent call last):\r\n File \"/scratch/ll1d19/transformers/src/transformers/models/llama/convert_llama_weights_to_hf.py\", line 319, in <module>\r\n main()\r\n File \"/scratch/ll1d19/transformers/src/transformers/models/llama/convert_llama_weights_to_hf.py\", line 307, in main\r\n write_model(\r\n File \"/scratch/ll1d19/transformers/src/transformers/models/llama/convert_llama_weights_to_hf.py\", line 109, in write_model\r\n tokenizer = tokenizer_class(tokenizer_path)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home/ll1d19/.conda/envs/myiai/lib/python3.11/site-packages/transformers/models/llama/tokenization_llama_fast.py\", line 124, in __init__\r\n super().__init__(\r\n File \"/home/ll1d19/.conda/envs/myiai/lib/python3.11/site-packages/transformers/tokenization_utils_fast.py\", line 117, in __init__\r\n slow_tokenizer = self.slow_tokenizer_class(*args, **kwargs)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home/ll1d19/.conda/envs/myiai/lib/python3.11/site-packages/transformers/models/llama/tokenization_llama.py\", line 178, in __init__\r\n self.sp_model = self.get_spm_processor(kwargs.pop(\"from_slow\", False))\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home/ll1d19/.conda/envs/myiai/lib/python3.11/site-packages/transformers/models/llama/tokenization_llama.py\", line 203, in get_spm_processor\r\n tokenizer.Load(self.vocab_file)\r\n File \"/home/ll1d19/.conda/envs/myiai/lib/python3.11/site-packages/sentencepiece/__init__.py\", line 905, in Load\r\n return self.LoadFromFile(model_file)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n File \"/home/ll1d19/.conda/envs/myiai/lib/python3.11/site-packages/sentencepiece/__init__.py\", line 310, in LoadFromFile\r\n return _sentencepiece.SentencePieceProcessor_LoadFromFile(self, arg)\r\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\nOSError: Not found: \"/scratch/ll1d19/llama/llama/llama-2-7b-chat/tokenizer.model\": No such file or directory Error #2\r\n\r\nThe weights are too large to share as it is about 13GB, the json file is around 100bytes.\r\nAny advise would be grateful!", "Hi @liboliba, thanks for the update! \r\n\r\nBased on the error, I'd suggest making sure you have the latest versions of `tokenizers` and `sentencepiece` installed in your environment. \r\n\r\nThere's no need to convert the official checkpoints though - there's many already available on the hub e.g. [here](https://huggingface.co./huggyllama/llama-7b) which you can access provided you've filled out the access form; or [meta-llama/Llama-2-70b-hf](https://huggingface.co./meta-llama/Llama-2-70b-hf) for llama 2.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28553/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 }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28552). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "Thank you @amyeroberts " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28552/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/28552", "html_url": "https://github.com/huggingface/transformers/pull/28552", "diff_url": "https://github.com/huggingface/transformers/pull/28552.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28552.patch", "merged_at": 1705508959000 }

{ "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 }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28551/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/28551", "html_url": "https://github.com/huggingface/transformers/pull/28551", "diff_url": "https://github.com/huggingface/transformers/pull/28551.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28551.patch", "merged_at": 1705485580000 }

{ "login": "nogifeet", "id": 72322393, "node_id": "MDQ6VXNlcjcyMzIyMzkz", "avatar_url": "https://avatars.githubusercontent.com/u/72322393?v=4", "gravatar_id": "", "url": "https://api.github.com/users/nogifeet", "html_url": "https://github.com/nogifeet", "followers_url": "https://api.github.com/users/nogifeet/followers", "following_url": "https://api.github.com/users/nogifeet/following{/other_user}", "gists_url": "https://api.github.com/users/nogifeet/gists{/gist_id}", "starred_url": "https://api.github.com/users/nogifeet/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/nogifeet/subscriptions", "organizations_url": "https://api.github.com/users/nogifeet/orgs", "repos_url": "https://api.github.com/users/nogifeet/repos", "events_url": "https://api.github.com/users/nogifeet/events{/privacy}", "received_events_url": "https://api.github.com/users/nogifeet/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @nogifeet, thanks for raising this issue! \r\n\r\nHmmm, I'm not sure what's happening here. However, I don't think there's an issue in the code itself. \r\n\r\nClearly the larger model isn't generating the eos token. However, the processing and generation configs for the models are equivalent, and the only difference in the models' configs is their encoder hidden size (expected). \r\n\r\nFollowing the paper, both models have been trained to predict eos and the only difference should be the size of the vision encoder, which matches with the difference in the hidden sizes in the model configs. \r\n\r\nThe only thing that's slightly unexpected is the models are adding a pooling layer, which is randomly initialized (but this would affect both) cc @NielsRogge \r\n\r\nIn terms of your experiments - does the large model always have this behaviour or only occasionally i.e. do we ever observe it correctly generating the eos token? ", "Hello @amyeroberts. There are examples where the eos is generated successfully and the example mentioned is actually an outlier. I can provide more examples from the IAM dataset if you want.\n\nI would also like to point out that I don't think it's a problem isolated to the large checkpoint, I have observed them in the base checkpoints.\n", "@nogifeet OK, in this case I'd recommend looking into [different decoding strategies](https://huggingface.co./docs/transformers/generation_strategies#decoding-strategies) when generating to try and force more sensible sequences. Some more information can be found here: https://huggingface.co./blog/how-to-generate. ", "@amyeroberts Is the handling of decoder layer nodes with shape limitations, managed internally by Hugging Face to prevent undefined behavior?", "> Is the handling of decoder layer nodes with shape limitations, managed internally by Hugging Face to prevent undefined behavior?\r\n\r\nSorry, I don't understand. What do you mean by \"decoder layer nodes with shape limitations\"? \r\nYou should be able to call `model.generate` and all the handling of autoregressive calls and generation logic should be handled for you. " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28550/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "scruel", "id": 16933298, "node_id": "MDQ6VXNlcjE2OTMzMjk4", "avatar_url": "https://avatars.githubusercontent.com/u/16933298?v=4", "gravatar_id": "", "url": "https://api.github.com/users/scruel", "html_url": "https://github.com/scruel", "followers_url": "https://api.github.com/users/scruel/followers", "following_url": "https://api.github.com/users/scruel/following{/other_user}", "gists_url": "https://api.github.com/users/scruel/gists{/gist_id}", "starred_url": "https://api.github.com/users/scruel/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/scruel/subscriptions", "organizations_url": "https://api.github.com/users/scruel/orgs", "repos_url": "https://api.github.com/users/scruel/repos", "events_url": "https://api.github.com/users/scruel/events{/privacy}", "received_events_url": "https://api.github.com/users/scruel/received_events", "type": "User", "site_admin": false }

[ { "id": 3817266200, "node_id": "MDU6TGFiZWwzODE3MjY2MjAw", "url": "https://api.github.com/repos/huggingface/transformers/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": null } ]

{ "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 }

[ { "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 } ]

[ "@scruel it is expected that `.no_exist` is not populated when using `huggingface-cli download --resume-download meta-llama/Llama-2-7b-hf`. In fact, the \"no exists\" folder is only populated when `transformers` (or any script/library) tries to download a file that does not exist on the repo. `transformers` uses this information to default to another file when that's the case. When you use the `huggingface-cli` we only download existing files but we don't populate \"missing files\" (which wouldn't really make sense). \r\n\r\nSo I agree this is a bug but rather in `transformers` than `huggingface_hub` itself. I think `transformers` should be able to load a model when it has already been downloaded, even if the `.no_exists` folder is not populated.\r\n\r\nI'm therefore transferring this issue to `transformers` repo. Btw, thanks for providing a reproducible example of your problem. @amyeroberts @ArthurZucker @ydshieh would you have time to take a look at how `transformers` is loading files when already cached? ", "Hi @Wauplin When I seared `no_exist` (case insensitive), there are only 4 results, and they are like `if resolved_file is not _CACHED_NO_EXIST`. \r\n\r\nI am happy to take the question you raised, but is it possible you share a bit more how `.no_exists` is related to `transformers`? I mean the logic of creating this is inside `huggingface_hub` right, and `transformers` only check if such path exist or not and perform different actions.", "@ydshieh the `no_exist` folder is indeed populated by `huggingface_hub` each time a library calls `huggingface_hub.hf_hub_download` to download a file that doesn't exist (but the repo does!). `transformers` leverages this information when it tries to load optional files. To avoid making a HEAD call for each file each time the user do a `from_pretrained`, `transformers` looks at the no_exist folder which saves some time (actually the `no_exists` folder is only used by `transformers` as far as I know).\r\n\r\nWhat `transformers` should do is that in the case of missing `no_exists` entries, if the cache is fully downloaded then the model should be loaded correctly. This is what @scruel's PR is doing if I understand correctly. My only comment is that it shouldn't be specific to gated repos but also private/missing repos as well, as long as the files are cached.", "Hi @Wauplin and @scruel \r\n\r\nI am looking into this, but there is one minor stuff I don't get clearly.\r\n\r\nbefore I do `until we run from_XXX method once to have those missed files created as empty files into .no_exist directory`, I checked with\r\n\r\n> !ls -l /root/.cache/huggingface/hub/models--meta-llama--Llama-2-7b-hf/snapshots/8cca527612d856d7d32bd94f8103728d614eb852\r\n\r\nand it gives \r\n\r\n```\r\ntotal 24\r\nlrwxrwxrwx 1 root root 52 Jan 22 14:47 config.json -> ../../blobs/34f901200fa131819b355bc4bed876c957a77a5a\r\nlrwxrwxrwx 1 root root 52 Jan 22 14:47 generation_config.json -> ../../blobs/aa1b3d3486df56a0699ce90c33283b13556fb5a3\r\nlrwxrwxrwx 1 root root 52 Jan 22 14:47 LICENSE.txt -> ../../blobs/51089e27e6764fb9f72c06a0f3710699fb6c9448\r\nlrwxrwxrwx 1 root root 76 Jan 22 14:49 model-00001-of-00002.safetensors -> ../../blobs/4ec71fd53e99766de38f24753b30c9e8942630e9e576a1ba27b0ec531e87be41\r\nlrwxrwxrwx 1 root root 76 Jan 22 14:48 model-00002-of-00002.safetensors -> ../../blobs/41780b5dac322ac35598737e99208d90bdc632a1ba3389ebedbb46a1d8385a7f\r\nlrwxrwxrwx 1 root root 52 Jan 22 14:47 model.safetensors.index.json -> ../../blobs/8b6245796e966e50960a317e4a54aa7bf73b0186\r\nlrwxrwxrwx 1 root root 76 Jan 22 14:51 pytorch_model-00001-of-00002.bin -> ../../blobs/ee62ed2ad7ded505ae47df50bc6c52916860dfb1c009df4715148cc4bfb50d2f\r\nlrwxrwxrwx 1 root root 76 Jan 22 14:51 pytorch_model-00002-of-00002.bin -> ../../blobs/1fd7762035b3ca4f2d6af6bf10129689a119b7c38058025f9842511532ea02fb\r\nlrwxrwxrwx 1 root root 52 Jan 22 14:47 pytorch_model.bin.index.json -> ../../blobs/db7264b24cac7a39947bb5fc02fe5c2d7ac9eaf4\r\nlrwxrwxrwx 1 root root 52 Jan 22 14:47 README.md -> ../../blobs/c7579dcbcbcd699270238173303a9013135a2a7d\r\nlrwxrwxrwx 1 root root 76 Jan 22 14:47 Responsible-Use-Guide.pdf -> ../../blobs/525dc349d71fe257fce4098c146446df6fef4247174f351381e4c3214af126f0\r\nlrwxrwxrwx 1 root root 52 Jan 22 14:47 special_tokens_map.json -> ../../blobs/451134b2ddc2e78555d1e857518c54b4bdc2e87d\r\nlrwxrwxrwx 1 root root 52 Jan 22 14:47 tokenizer_config.json -> ../../blobs/2ef41cbc275000b29afe157ba487f0530b8c26dc\r\nlrwxrwxrwx 1 root root 52 Jan 22 14:47 tokenizer.json -> ../../blobs/a6e931b92caff4c79c5c56282f1e89569a0ae558\r\nlrwxrwxrwx 1 root root 76 Jan 22 14:47 tokenizer.model -> ../../blobs/9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347\r\nlrwxrwxrwx 1 root root 52 Jan 22 14:47 USE_POLICY.md -> ../../blobs/abbcc199b2d1e4feb5d7e40c0bd67e1b0ce29e97\r\n```\r\nAnd after running \r\n\r\n```\r\nAutoTokenizer.from_pretrained('meta-llama/Llama-2-7b-hf', token=token)\r\n```\r\n(with a real token)\r\n\r\nThe same list of entries is shown, and I can't see any `.no_exist`\r\n\r\nBut then running the following\r\n\r\n```\r\nAutoTokenizer.from_pretrained('meta-llama/Llama-2-7b-hf', token=False)\r\n```\r\nindeed works.\r\n\r\nCould you help me to understand where the mentioned `no_exist` is?\r\n\r\n(I am aware of the PR #28566)", "@ydshieh yes so the `.no_exists` folder is separated from the `snapshot` folder. If you do\r\n```\r\n!ls -l /root/.cache/huggingface/hub/models--meta-llama--Llama-2-7b-hf/\r\n```\r\nyou should see `refs/`, `blobs/`, `snapshots/` and `/no_exists` which should help the investigation :)", "Thanks @Wauplin , so when I run\r\n\r\n> until we run from_XXX method once to have those missed files created as empty files into .no_exist directory,\r\n\r\n(with a real token), and checked\r\n\r\n```\r\n!ls -l /root/.cache/huggingface/hub/models--meta-llama--Llama-2-7b-hf/\r\n```\r\n\r\nit still shows \r\n\r\n```\r\ndrwxr-xr-x 2 root root 4096 Jan 22 15:04 blobs\r\ndrwxr-xr-x 2 root root 4096 Jan 22 15:00 refs\r\ndrwxr-xr-x 3 root root 4096 Jan 22 15:00 snapshots\r\n```\r\n\r\nbut again, the subsequent call with `token=False` works. 🤔 \r\n", "wait, maybe let check the huggingface_hub version\r\n\r\nIt is `0.20.2` with `transformers==4.35.2`.", "@ydshieh I can confirm, I have same as both:\r\n\r\nBoth \r\n\r\n```py\r\nfrom transformers import AutoTokenizer\r\nAutoTokenizer.from_pretrained('meta-llama/Llama-2-7b-hf')\r\n```\r\n\r\nand \r\n\r\n```py\r\nfrom transformers import AutoTokenizer\r\nAutoTokenizer.from_pretrained('meta-llama/Llama-2-7b-hf', token=False)\r\n```\r\n\r\nworks for me. \r\n\r\nCache directory only have tokenizers files:\r\n\r\n```\r\n(.venv310) ➜ huggingface_hub git:(main) ✗ tree -h ~/.cache/huggingface/hub/models--meta-llama--Llama-2-7b-hf \r\n[4.0K] /home/wauplin/.cache/huggingface/hub/models--meta-llama--Llama-2-7b-hf\r\n├── [4.0K] blobs\r\n│ ├── [ 776] 2ef41cbc275000b29afe157ba487f0530b8c26dc\r\n│ ├── [ 609] 34f901200fa131819b355bc4bed876c957a77a5a\r\n│ ├── [ 414] 451134b2ddc2e78555d1e857518c54b4bdc2e87d\r\n│ ├── [ 26K] 8b6245796e966e50960a317e4a54aa7bf73b0186\r\n│ ├── [488K] 9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347\r\n│ └── [1.8M] a6e931b92caff4c79c5c56282f1e89569a0ae558\r\n├── [4.0K] refs\r\n│ └── [ 40] main\r\n└── [4.0K] snapshots\r\n └── [4.0K] 8cca527612d856d7d32bd94f8103728d614eb852\r\n ├── [ 52] config.json -> ../../blobs/34f901200fa131819b355bc4bed876c957a77a5a\r\n ├── [ 52] model.safetensors.index.json -> ../../blobs/8b6245796e966e50960a317e4a54aa7bf73b0186\r\n ├── [ 52] special_tokens_map.json -> ../../blobs/451134b2ddc2e78555d1e857518c54b4bdc2e87d\r\n ├── [ 52] tokenizer_config.json -> ../../blobs/2ef41cbc275000b29afe157ba487f0530b8c26dc\r\n ├── [ 52] tokenizer.json -> ../../blobs/a6e931b92caff4c79c5c56282f1e89569a0ae558\r\n └── [ 76] tokenizer.model -> ../../blobs/9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347\r\n\r\n4 directories, 13 files\r\n```", "@scruel are you sure, the failing command was `AutoTokenizer.from_pretrained('meta-llama/Llama-2-7b-hf', token=False)` and not `AutoModel.from_pretrained('meta-llama/Llama-2-7b-hf', token=False)`?", "Yes, I'm pretty sure.\r\nRather than use `ls -l` and `tree -h`, you should use `ls -la` and `tree -ha`, the directory you created is not `no_exiss` but `.no_exist`, it will be treated as a hidden folder in the modern OS.", "Ah yes of course :facepalm: Thanks @scruel! \r\n\r\nHere it is @ydshieh:\r\n```\r\n(.venv310) ➜ huggingface_hub git:(main) ✗ tree -ah ~/.cache/huggingface/hub/models--meta-llama--Llama-2-7b-hf \r\n[4.0K] /home/wauplin/.cache/huggingface/hub/models--meta-llama--Llama-2-7b-hf\r\n├── [4.0K] blobs\r\n│ ├── [ 776] 2ef41cbc275000b29afe157ba487f0530b8c26dc\r\n│ ├── [ 609] 34f901200fa131819b355bc4bed876c957a77a5a\r\n│ ├── [ 414] 451134b2ddc2e78555d1e857518c54b4bdc2e87d\r\n│ ├── [ 26K] 8b6245796e966e50960a317e4a54aa7bf73b0186\r\n│ ├── [488K] 9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347\r\n│ └── [1.8M] a6e931b92caff4c79c5c56282f1e89569a0ae558\r\n├── [4.0K] .no_exist\r\n│ └── [4.0K] 8cca527612d856d7d32bd94f8103728d614eb852\r\n│ └── [ 0] added_tokens.json\r\n├── [4.0K] refs\r\n│ └── [ 40] main\r\n└── [4.0K] snapshots\r\n └── [4.0K] 8cca527612d856d7d32bd94f8103728d614eb852\r\n ├── [ 52] config.json -> ../../blobs/34f901200fa131819b355bc4bed876c957a77a5a\r\n ├── [ 52] model.safetensors.index.json -> ../../blobs/8b6245796e966e50960a317e4a54aa7bf73b0186\r\n ├── [ 52] special_tokens_map.json -> ../../blobs/451134b2ddc2e78555d1e857518c54b4bdc2e87d\r\n ├── [ 52] tokenizer_config.json -> ../../blobs/2ef41cbc275000b29afe157ba487f0530b8c26dc\r\n ├── [ 52] tokenizer.json -> ../../blobs/a6e931b92caff4c79c5c56282f1e89569a0ae558\r\n └── [ 76] tokenizer.model -> ../../blobs/9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347\r\n\r\n6 directories, 14 files\r\n```", "Thank you both for helping me on this! Will continue this along with a review" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28558/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "kenfus", "id": 47979198, "node_id": "MDQ6VXNlcjQ3OTc5MTk4", "avatar_url": "https://avatars.githubusercontent.com/u/47979198?v=4", "gravatar_id": "", "url": "https://api.github.com/users/kenfus", "html_url": "https://github.com/kenfus", "followers_url": "https://api.github.com/users/kenfus/followers", "following_url": "https://api.github.com/users/kenfus/following{/other_user}", "gists_url": "https://api.github.com/users/kenfus/gists{/gist_id}", "starred_url": "https://api.github.com/users/kenfus/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/kenfus/subscriptions", "organizations_url": "https://api.github.com/users/kenfus/orgs", "repos_url": "https://api.github.com/users/kenfus/repos", "events_url": "https://api.github.com/users/kenfus/events{/privacy}", "received_events_url": "https://api.github.com/users/kenfus/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @kenfus, 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\ncc @sanchit-gandhi @ylacombe ", "Hey, thank you!\r\n\r\nI have added a question to this post: https://discuss.huggingface.co/t/finetuning-whisper-with-prompts/43053/3\r\n\r\nHowever, even when looking at the code, I don't quite understand how to pass the prompt to fine tune the model, so I think it's not implemented yet. With some guidance, I am very happy to implement it myself!", "Hi @kenfus, \r\n\r\nI'll let @sanchit-gandhi and @ylacombe answer about fine-tuning. For reference on the intended API, the PR to add prompts is #22395. There is also another open PR looking to add prompts to the pipeline #28556. Although these are not specific to the training regime, they may provide more context on how to pass prompts to the model. ", "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/28549/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "ryan-caesar-ramos", "id": 65334734, "node_id": "MDQ6VXNlcjY1MzM0NzM0", "avatar_url": "https://avatars.githubusercontent.com/u/65334734?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ryan-caesar-ramos", "html_url": "https://github.com/ryan-caesar-ramos", "followers_url": "https://api.github.com/users/ryan-caesar-ramos/followers", "following_url": "https://api.github.com/users/ryan-caesar-ramos/following{/other_user}", "gists_url": "https://api.github.com/users/ryan-caesar-ramos/gists{/gist_id}", "starred_url": "https://api.github.com/users/ryan-caesar-ramos/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ryan-caesar-ramos/subscriptions", "organizations_url": "https://api.github.com/users/ryan-caesar-ramos/orgs", "repos_url": "https://api.github.com/users/ryan-caesar-ramos/repos", "events_url": "https://api.github.com/users/ryan-caesar-ramos/events{/privacy}", "received_events_url": "https://api.github.com/users/ryan-caesar-ramos/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Batch_size is transparent. The model will run the inference batched, but the pipeline still outputs results one by one, meaning you can modify the batch_size according to the actual hardware you are using without actually needing to unpack results manually (it also allows running next batch, before you even finished unpacking the results manually, leading to better GPU usage.", "Thanks! That clears things up" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28548/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "chenbin11200", "id": 5245644, "node_id": "MDQ6VXNlcjUyNDU2NDQ=", "avatar_url": "https://avatars.githubusercontent.com/u/5245644?v=4", "gravatar_id": "", "url": "https://api.github.com/users/chenbin11200", "html_url": "https://github.com/chenbin11200", "followers_url": "https://api.github.com/users/chenbin11200/followers", "following_url": "https://api.github.com/users/chenbin11200/following{/other_user}", "gists_url": "https://api.github.com/users/chenbin11200/gists{/gist_id}", "starred_url": "https://api.github.com/users/chenbin11200/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/chenbin11200/subscriptions", "organizations_url": "https://api.github.com/users/chenbin11200/orgs", "repos_url": "https://api.github.com/users/chenbin11200/repos", "events_url": "https://api.github.com/users/chenbin11200/events{/privacy}", "received_events_url": "https://api.github.com/users/chenbin11200/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @younesbelkada ", "> Makes sense, thank you ! Can you also run the styling checks? `make fixup`\r\n\r\nHi @younesbelkada , I have made the styling checks. If there is still something missing, please informe me :D", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28547). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "> Thanks a lot for fixing!\r\n\r\nMy pleasure :D \r\nI guess it has to be approved by @amyeroberts before being merged?", "Yes @chenbin11200 - it will get reviewed as soon as possible and we'll merge the fix in main ! ", "@chenbin11200 the test would go here ideally: https://github.com/huggingface/transformers/blob/main/tests/peft_integration/test_peft_integration.py let me know if you need help designing the test!", "> @chenbin11200 the test would go here ideally: https://github.com/huggingface/transformers/blob/main/tests/peft_integration/test_peft_integration.py let me know if you need help designing the test!\r\n\r\nHi @younesbelkada @amyeroberts ,\r\nThanks for guiding me. Does this \"resuming from checkpoint\" test(or anything similar) already exist? Will save me some time if I could change on an exisiting one :)", "Hi @chenbin11200 \r\nAfter thinking a bit, to not complexify things, you can write a new test similar than this one: https://github.com/huggingface/transformers/blob/main/tests/trainer/test_trainer.py#L874 by adding the `require_peft` decorator and make sure training runs fine with resume checkpoint for a named adapter", "> Hi @chenbin11200 After thinking a bit, to not complexify things, you can write a new test similar than this one: https://github.com/huggingface/transformers/blob/main/tests/trainer/test_trainer.py#L874 by adding the `require_peft` decorator and make sure training runs fine with resume checkpoint for a named adapter\r\n\r\nThank you @younesbelkada, I will try that. " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28547/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/28547", "html_url": "https://github.com/huggingface/transformers/pull/28547", "diff_url": "https://github.com/huggingface/transformers/pull/28547.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28547.patch", "merged_at": null }

{ "login": "guoyunqingyue", "id": 77528622, "node_id": "MDQ6VXNlcjc3NTI4NjIy", "avatar_url": "https://avatars.githubusercontent.com/u/77528622?v=4", "gravatar_id": "", "url": "https://api.github.com/users/guoyunqingyue", "html_url": "https://github.com/guoyunqingyue", "followers_url": "https://api.github.com/users/guoyunqingyue/followers", "following_url": "https://api.github.com/users/guoyunqingyue/following{/other_user}", "gists_url": "https://api.github.com/users/guoyunqingyue/gists{/gist_id}", "starred_url": "https://api.github.com/users/guoyunqingyue/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/guoyunqingyue/subscriptions", "organizations_url": "https://api.github.com/users/guoyunqingyue/orgs", "repos_url": "https://api.github.com/users/guoyunqingyue/repos", "events_url": "https://api.github.com/users/guoyunqingyue/events{/privacy}", "received_events_url": "https://api.github.com/users/guoyunqingyue/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @guoyunqingyue, thanks for raising an issue! \r\n\r\nMy first suggest would be to update your version of transformers, as there have been many recent updates to weight loading and quantization in the codebase: `pip install -U transformers`. \r\n\r\nSo that we can be best help you, please make sure to follow the [issue template](https://github.com/huggingface/transformers/blob/main/.github/ISSUE_TEMPLATE/bug-report.yml) and provide: \r\n* The running environment: run `transformers-cli env` in the terminal and copy-paste the output\r\n* A minimal code snippet we can run to reproduce the error. We can't run the current code example as many variables are undefined. \r\n* All relevant details about the error including the full error traceback", "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/28546/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "LOVE-YOURSELF-1", "id": 71559440, "node_id": "MDQ6VXNlcjcxNTU5NDQw", "avatar_url": "https://avatars.githubusercontent.com/u/71559440?v=4", "gravatar_id": "", "url": "https://api.github.com/users/LOVE-YOURSELF-1", "html_url": "https://github.com/LOVE-YOURSELF-1", "followers_url": "https://api.github.com/users/LOVE-YOURSELF-1/followers", "following_url": "https://api.github.com/users/LOVE-YOURSELF-1/following{/other_user}", "gists_url": "https://api.github.com/users/LOVE-YOURSELF-1/gists{/gist_id}", "starred_url": "https://api.github.com/users/LOVE-YOURSELF-1/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/LOVE-YOURSELF-1/subscriptions", "organizations_url": "https://api.github.com/users/LOVE-YOURSELF-1/orgs", "repos_url": "https://api.github.com/users/LOVE-YOURSELF-1/repos", "events_url": "https://api.github.com/users/LOVE-YOURSELF-1/events{/privacy}", "received_events_url": "https://api.github.com/users/LOVE-YOURSELF-1/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @LOVE-YOURSELF-1, thanks for opening this feature request! \r\n\r\nIf you want to download a specific file from the hub, you can use `hf_hub_download` from `huggingface_hub`. \r\n\r\nIf you are using the auto classes, they will only download what's necessary. For example: \r\n\r\n```\r\nfrom transformers import AutoTokenizer\r\ntokenizer = AutoTokenizer.from_pretrained(\"bert-base-uncased\")\r\n```\r\n\r\nwill only download the files for the tokenizer, not the model weights. ", "If we have multiple download sources，do you think it is necessary to bulit a new class named downloader?", "@LOVE-YOURSELF-1 Could you explain what you mean in more detail? What do you mean by 'multiple download sources'? ", "https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_utils.py#L3258\r\nCan we rewrite the code about download（config，tokenizer and model）?\r\nrebuilt a class named downloader, which can accomplish download, (from local and hugging hub)", "@LOVE-YOURSELF-1 There's a plan to harmonise how e.g. loading and downloading shards is handled across libraries which will also account for e.g. local versus downloading from the hub. \r\n\r\nI don't see the reason for having a downloader class which handles downloading these all files - it's bundling together responsibilities making the scope of the class ill-defined. What problem would this be solving? ", "As you said, i want to integrate downloading and loading. It can be rewritten in downloader.py.\r\nCan you give me some suggestion or idea?", "@LOVE-YOURSELF-1 Any contributor is welcome to open a PR to demonstrate and new feature or refactor. As this isn't something we or the community are requesting, it's up to you to define the feature you'd like to add. Once there's a clear diff or feature description, then we'll be able to help.", "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/28545/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "Ubadub", "id": 1286898, "node_id": "MDQ6VXNlcjEyODY4OTg=", "avatar_url": "https://avatars.githubusercontent.com/u/1286898?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Ubadub", "html_url": "https://github.com/Ubadub", "followers_url": "https://api.github.com/users/Ubadub/followers", "following_url": "https://api.github.com/users/Ubadub/following{/other_user}", "gists_url": "https://api.github.com/users/Ubadub/gists{/gist_id}", "starred_url": "https://api.github.com/users/Ubadub/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Ubadub/subscriptions", "organizations_url": "https://api.github.com/users/Ubadub/orgs", "repos_url": "https://api.github.com/users/Ubadub/repos", "events_url": "https://api.github.com/users/Ubadub/events{/privacy}", "received_events_url": "https://api.github.com/users/Ubadub/received_events", "type": "User", "site_admin": false }

[]

[]

[ "IMO, the problem can be generalised through the TrainingArguments parameter `save_only_model` ([link](https://github.com/huggingface/transformers/blob/v4.36.1/src/transformers/training_args.py#L307)):\r\n\r\nIts definition states: \"...when this is true, you won't be able to resume training from checkpoint.\".\r\nAs the current [TrainerState](https://github.com/huggingface/transformers/blob/v4.36.1/src/transformers/trainer_callback.py#L35) implementation stands, we are not _truly_ able to resume training from a checkpoint even on setting `save_only_model=False`.\r\n\r\nAs pointed by @Ubadub here, Trainer's callbacks' states are not persisted along with models (even on setting `save_only_model=False`). To fix this issue (and the auxiliary issue #10290, pointed above), we need the capability to persist callbacks and load them when using `resume_from_checkpoint`.\r\n\r\nIn short, Trainer's callbacks should be a part of the [TrainerState](https://github.com/huggingface/transformers/blob/v4.36.1/src/transformers/trainer_callback.py#L35) object.\r\n\r\nI can help in this implementation if this analysis seems reasonable.\r\n\r\n\r\nCheers!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28544/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "gventuri", "id": 15671184, "node_id": "MDQ6VXNlcjE1NjcxMTg0", "avatar_url": "https://avatars.githubusercontent.com/u/15671184?v=4", "gravatar_id": "", "url": "https://api.github.com/users/gventuri", "html_url": "https://github.com/gventuri", "followers_url": "https://api.github.com/users/gventuri/followers", "following_url": "https://api.github.com/users/gventuri/following{/other_user}", "gists_url": "https://api.github.com/users/gventuri/gists{/gist_id}", "starred_url": "https://api.github.com/users/gventuri/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/gventuri/subscriptions", "organizations_url": "https://api.github.com/users/gventuri/orgs", "repos_url": "https://api.github.com/users/gventuri/repos", "events_url": "https://api.github.com/users/gventuri/events{/privacy}", "received_events_url": "https://api.github.com/users/gventuri/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey! Make sure you are logged in to the hub on `my-company` . For me this works: \r\n```python \r\n>>> from transformers import AutoModel\r\n>>> model = AutoModel.from_pretrained(\"bert-base-uncase\")\r\n>>> model.push_to_hub(\"ArthurZ/my-llm\")\r\nCommitInfo(commit_url='https://huggingface.co./ArthurZ/my-llm/commit/53cb603f0cdd1d6284f21c541f06a3b8a0ddd2d1', commit_message='Upload model', commit_description='', oid='53cb603f0cdd1d6284f21c541f06a3b8a0ddd2d1', pr_url=None, pr_revision=None, pr_num=None)\r\n```", "@ArthurZucker I tried to login both with\r\n```\r\n!huggingface-cli login --token $hf_token\r\n```\r\nand with\r\n```\r\nhuggingface_hub.login\r\n```\r\n\r\nIn both cases I provided my personal token (I'm admin in the company). If I try to login again, it says that I'm already connected.", "cc @Wauplin you might know better than me what this represents! ", "@gventuri Can it be that you have a local directory called `\"my-company/my-llm\"` and that it conflicts with the push_to_hub? \r\n\r\n(to confirm that you are indeed logged in, you can do `huggingface-cli whoami` but it really doesn't seem to be the problem here)", "@Wauplin thanks a lot, it was conflicting with the local folder, it's now working!", "Great to know your problem's solved! :hugs: " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28543/reactions", "total_count": 0, "+1": 0, "-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 }

[]

[]

[ "First draft wrapped up!\r\n\r\n1. It is easier to find `HfDeepSpeedConfig` now\r\n2. Prebuilding troubleshooting steps are in `debugging.md`\r\n3. Testing the DeepSpeed integration is moved to `testing.md`\r\n4. The main DeepSpeed guide is refactored to flow better. For example, installation > memory requirements > selecting a ZeRO stage. I think it's important to address these first before a user finds out they don't have the required memory or chose the wrong ZeRO stage. I've also condensed and made things more concise where appropriate without losing any context or info." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28542/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/28542", "html_url": "https://github.com/huggingface/transformers/pull/28542", "diff_url": "https://github.com/huggingface/transformers/pull/28542.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28542.patch", "merged_at": 1706113888000 }

{ "login": "vallabh001", "id": 88985147, "node_id": "MDQ6VXNlcjg4OTg1MTQ3", "avatar_url": "https://avatars.githubusercontent.com/u/88985147?v=4", "gravatar_id": "", "url": "https://api.github.com/users/vallabh001", "html_url": "https://github.com/vallabh001", "followers_url": "https://api.github.com/users/vallabh001/followers", "following_url": "https://api.github.com/users/vallabh001/following{/other_user}", "gists_url": "https://api.github.com/users/vallabh001/gists{/gist_id}", "starred_url": "https://api.github.com/users/vallabh001/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/vallabh001/subscriptions", "organizations_url": "https://api.github.com/users/vallabh001/orgs", "repos_url": "https://api.github.com/users/vallabh001/repos", "events_url": "https://api.github.com/users/vallabh001/events{/privacy}", "received_events_url": "https://api.github.com/users/vallabh001/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @vallabh001, thanks for raising an issue! \r\n\r\nSo that we can best help you, please make sure to follow the [issue template](https://github.com/huggingface/transformers/blob/main/.github/ISSUE_TEMPLATE/bug-report.yml) and provide:\r\n* The full error traceback\r\n* A minimal code snippet we can use to reproduce the error. In this case, we can't help without knowing what's in `llm_training.py`\r\n* The running environment: run `transformers-cli env` in the terminal and copy-paste the output", "There's the Alignment Handbook which uses Deepspeed to train any decoder LLM like Llama or Mistral on multiple GPUs: https://github.com/huggingface/alignment-handbook. It includes scripts for both supervised fine-tuning (SFT) and human preference fine-tuning using DPO (direct preference optimization): https://github.com/huggingface/alignment-handbook/tree/main/scripts.\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/28541/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_28540). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "Closing for now as we need some funky logs for the no logs case whilst we still support < 3.10 and the logger issue has been resolved. " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28540/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/28540", "html_url": "https://github.com/huggingface/transformers/pull/28540", "diff_url": "https://github.com/huggingface/transformers/pull/28540.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28540.patch", "merged_at": null }

{ "login": "antoine-lizee", "id": 2957716, "node_id": "MDQ6VXNlcjI5NTc3MTY=", "avatar_url": "https://avatars.githubusercontent.com/u/2957716?v=4", "gravatar_id": "", "url": "https://api.github.com/users/antoine-lizee", "html_url": "https://github.com/antoine-lizee", "followers_url": "https://api.github.com/users/antoine-lizee/followers", "following_url": "https://api.github.com/users/antoine-lizee/following{/other_user}", "gists_url": "https://api.github.com/users/antoine-lizee/gists{/gist_id}", "starred_url": "https://api.github.com/users/antoine-lizee/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/antoine-lizee/subscriptions", "organizations_url": "https://api.github.com/users/antoine-lizee/orgs", "repos_url": "https://api.github.com/users/antoine-lizee/repos", "events_url": "https://api.github.com/users/antoine-lizee/events{/privacy}", "received_events_url": "https://api.github.com/users/antoine-lizee/received_events", "type": "User", "site_admin": false }

[]

[]

[ "### Notes\r\n\r\nI realize that this issue probably doesn't arise if the strategy is `epoch`.\r\n\r\nIt seems that using N + epsilon as the `num_train_epochs` would go around this problem in a very hacky way (and evaluate / save the model that corresponds to the first step after the desired epoch that is a multiple of `eval_steps`). Would that be your recommendation?\r\n\r\nedit: Ok digging a bit more, it seems that the proper way of fixing this problem would be to **add a callback** to the trainer which would enforce saving at the end of training.\r\nI will do this, but the default behaviour is still \"wrong\" I believe. (and would warrant at least some clear disclaimer in the doc?)" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28539/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 }

[]

[]

[ "Why do we use reentrant gc by default? It said the non-reentrant gc can be more advantageous than the reentrant version: https://pytorch.org/docs/2.0/checkpoint.html#torch.utils.checkpoint.checkpoint", "@hiyouga the use_reentrant=True is used by default in PT anyway so if you set it to `None`, `use_reentrant` will be set to `True`", "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.", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28538). 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/28538/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/28538", "html_url": "https://github.com/huggingface/transformers/pull/28538", "diff_url": "https://github.com/huggingface/transformers/pull/28538.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28538.patch", "merged_at": 1708392205000 }

{ "login": "gugarosa", "id": 4120639, "node_id": "MDQ6VXNlcjQxMjA2Mzk=", "avatar_url": "https://avatars.githubusercontent.com/u/4120639?v=4", "gravatar_id": "", "url": "https://api.github.com/users/gugarosa", "html_url": "https://github.com/gugarosa", "followers_url": "https://api.github.com/users/gugarosa/followers", "following_url": "https://api.github.com/users/gugarosa/following{/other_user}", "gists_url": "https://api.github.com/users/gugarosa/gists{/gist_id}", "starred_url": "https://api.github.com/users/gugarosa/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/gugarosa/subscriptions", "organizations_url": "https://api.github.com/users/gugarosa/orgs", "repos_url": "https://api.github.com/users/gugarosa/repos", "events_url": "https://api.github.com/users/gugarosa/events{/privacy}", "received_events_url": "https://api.github.com/users/gugarosa/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The `loss=0.0` error while fine-tuning with FP16 is another issue and I do have an ugly fix, but will look into it with more patience (and use a separate PR).", "No problems! Thanks for the merge!", "Thanks very much @gugarosa for the deep dive and the fix! ", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28537). 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/28537/reactions", "total_count": 1, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 1, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28537", "html_url": "https://github.com/huggingface/transformers/pull/28537", "diff_url": "https://github.com/huggingface/transformers/pull/28537.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28537.patch", "merged_at": 1705497765000 }

{ "login": "rosario-purple", "id": 123594463, "node_id": "U_kgDOB13m3w", "avatar_url": "https://avatars.githubusercontent.com/u/123594463?v=4", "gravatar_id": "", "url": "https://api.github.com/users/rosario-purple", "html_url": "https://github.com/rosario-purple", "followers_url": "https://api.github.com/users/rosario-purple/followers", "following_url": "https://api.github.com/users/rosario-purple/following{/other_user}", "gists_url": "https://api.github.com/users/rosario-purple/gists{/gist_id}", "starred_url": "https://api.github.com/users/rosario-purple/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/rosario-purple/subscriptions", "organizations_url": "https://api.github.com/users/rosario-purple/orgs", "repos_url": "https://api.github.com/users/rosario-purple/repos", "events_url": "https://api.github.com/users/rosario-purple/events{/privacy}", "received_events_url": "https://api.github.com/users/rosario-purple/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Thanks for raising! given that we had #27020, this should be fairly easy to fix! cc @younesbelkada ", "@ArthurZucker is this still outstanding?", "Will merge the PR today" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28536/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_28535). 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/28535/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/28535", "html_url": "https://github.com/huggingface/transformers/pull/28535", "diff_url": "https://github.com/huggingface/transformers/pull/28535.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28535.patch", "merged_at": 1705667526000 }

{ "login": "rosario-purple", "id": 123594463, "node_id": "U_kgDOB13m3w", "avatar_url": "https://avatars.githubusercontent.com/u/123594463?v=4", "gravatar_id": "", "url": "https://api.github.com/users/rosario-purple", "html_url": "https://github.com/rosario-purple", "followers_url": "https://api.github.com/users/rosario-purple/followers", "following_url": "https://api.github.com/users/rosario-purple/following{/other_user}", "gists_url": "https://api.github.com/users/rosario-purple/gists{/gist_id}", "starred_url": "https://api.github.com/users/rosario-purple/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/rosario-purple/subscriptions", "organizations_url": "https://api.github.com/users/rosario-purple/orgs", "repos_url": "https://api.github.com/users/rosario-purple/repos", "events_url": "https://api.github.com/users/rosario-purple/events{/privacy}", "received_events_url": "https://api.github.com/users/rosario-purple/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!" } ]

[]

[ "Adding these lines seems to fix it, not sure if this is the best/most general solution though:\r\n\r\n```\r\n tokenizer = AutoTokenizer.from_pretrained(\r\n args.model_name_or_path, use_fast=not args.use_slow_tokenizer, trust_remote_code=args.trust_remote_code\r\n )\r\n tokenizer.pad_token = tokenizer.eos_token\r\n config.pad_token_id = tokenizer.pad_token_id\r\n model = AutoModelForSequenceClassification.from_pretrained(\r\n args.model_name_or_path,\r\n from_tf=bool(\".ckpt\" in args.model_name_or_path),\r\n config=config,\r\n ignore_mismatched_sizes=args.ignore_mismatched_sizes,\r\n trust_remote_code=args.trust_remote_code,\r\n )\r\n```", "Hi @rosario-purple, thanks for raising this issue! \r\n\r\nThe proposed fix is the recommended way to address this. Would you like to open a PR to add this to the script? This way you get the github contribution", "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/28534/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "michaelbenayoun", "id": 25418079, "node_id": "MDQ6VXNlcjI1NDE4MDc5", "avatar_url": "https://avatars.githubusercontent.com/u/25418079?v=4", "gravatar_id": "", "url": "https://api.github.com/users/michaelbenayoun", "html_url": "https://github.com/michaelbenayoun", "followers_url": "https://api.github.com/users/michaelbenayoun/followers", "following_url": "https://api.github.com/users/michaelbenayoun/following{/other_user}", "gists_url": "https://api.github.com/users/michaelbenayoun/gists{/gist_id}", "starred_url": "https://api.github.com/users/michaelbenayoun/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/michaelbenayoun/subscriptions", "organizations_url": "https://api.github.com/users/michaelbenayoun/orgs", "repos_url": "https://api.github.com/users/michaelbenayoun/repos", "events_url": "https://api.github.com/users/michaelbenayoun/events{/privacy}", "received_events_url": "https://api.github.com/users/michaelbenayoun/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28533). 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/28533/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/28533", "html_url": "https://github.com/huggingface/transformers/pull/28533", "diff_url": "https://github.com/huggingface/transformers/pull/28533.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28533.patch", "merged_at": null }

{ "login": "faph", "id": 8397805, "node_id": "MDQ6VXNlcjgzOTc4MDU=", "avatar_url": "https://avatars.githubusercontent.com/u/8397805?v=4", "gravatar_id": "", "url": "https://api.github.com/users/faph", "html_url": "https://github.com/faph", "followers_url": "https://api.github.com/users/faph/followers", "following_url": "https://api.github.com/users/faph/following{/other_user}", "gists_url": "https://api.github.com/users/faph/gists{/gist_id}", "starred_url": "https://api.github.com/users/faph/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/faph/subscriptions", "organizations_url": "https://api.github.com/users/faph/orgs", "repos_url": "https://api.github.com/users/faph/repos", "events_url": "https://api.github.com/users/faph/events{/privacy}", "received_events_url": "https://api.github.com/users/faph/received_events", "type": "User", "site_admin": false }

[]

[]

[ "I think this is already fixed here: https://github.com/huggingface/transformers/blob/002566f398cd6dbf7053a89c26646ac45be540f4/src/transformers/training_args.py#L1830", "Hi @faph, thanks for raising this issue! \r\n\r\nYes, as you identified, a fix was added with #28171 which will be part of the next release. ", "Thanks @amyeroberts " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28532/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "chenbin11200", "id": 5245644, "node_id": "MDQ6VXNlcjUyNDU2NDQ=", "avatar_url": "https://avatars.githubusercontent.com/u/5245644?v=4", "gravatar_id": "", "url": "https://api.github.com/users/chenbin11200", "html_url": "https://github.com/chenbin11200", "followers_url": "https://api.github.com/users/chenbin11200/followers", "following_url": "https://api.github.com/users/chenbin11200/following{/other_user}", "gists_url": "https://api.github.com/users/chenbin11200/gists{/gist_id}", "starred_url": "https://api.github.com/users/chenbin11200/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/chenbin11200/subscriptions", "organizations_url": "https://api.github.com/users/chenbin11200/orgs", "repos_url": "https://api.github.com/users/chenbin11200/repos", "events_url": "https://api.github.com/users/chenbin11200/events{/privacy}", "received_events_url": "https://api.github.com/users/chenbin11200/received_events", "type": "User", "site_admin": false }

[]

[]

[ "A PR is attached, if it is necessary: https://github.com/huggingface/transformers/pull/28547", "Also cc @younesbelkada for PEFT", "Thanks @amyeroberts @chenbin11200 - the PR makes sense to me ! I left a single comment ", "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/28531/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "ManishChandra12", "id": 17062142, "node_id": "MDQ6VXNlcjE3MDYyMTQy", "avatar_url": "https://avatars.githubusercontent.com/u/17062142?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ManishChandra12", "html_url": "https://github.com/ManishChandra12", "followers_url": "https://api.github.com/users/ManishChandra12/followers", "following_url": "https://api.github.com/users/ManishChandra12/following{/other_user}", "gists_url": "https://api.github.com/users/ManishChandra12/gists{/gist_id}", "starred_url": "https://api.github.com/users/ManishChandra12/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ManishChandra12/subscriptions", "organizations_url": "https://api.github.com/users/ManishChandra12/orgs", "repos_url": "https://api.github.com/users/ManishChandra12/repos", "events_url": "https://api.github.com/users/ManishChandra12/events{/privacy}", "received_events_url": "https://api.github.com/users/ManishChandra12/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @ManishChandra12, thanks for raising an issue! \r\n\r\nIn the provided snippet, you don't appear to be passing `compute_metrics` to the `Trainer` class, as required. ", "Hi @amyeroberts. Thanks for the quick response.\r\n\r\nThe error persists even when I pass ```compute_metrics``` to ```Trainer```.\r\n\r\nPS: I have updated the original issue by adding ```compute_metrics```.", "@ManishChandra12 OK, thanks for reporting back. Is the error exactly the same as before? \r\n\r\nCould you update the script to use a public dataset, or provide a sample of the dataset so we can run the code to reproduce on our end? ", "@amyeroberts yes, the error is exactly the same as before.\r\n\r\nHere is a sample of the dataset:\r\n[val.csv](https://github.com/huggingface/transformers/files/13950970/val.csv)\r\n[test.csv](https://github.com/huggingface/transformers/files/13950971/test.csv)\r\n[train.csv](https://github.com/huggingface/transformers/files/13950972/train.csv)\r\n", "Hi @ManishChandra12, I'm able to run the provided code and data without error on main. Could you try updating your transformers version to the latest release v4.36.2? ", "@ManishChandra12 @amyeroberts could you figure out the issue?\r\n\r\nWhen I run the code with the provided example data, it also says\r\n\"early stopping required metric_for_best_model, but did not find eval_f1 so early stopping is disabled\"\r\nand when I run ```trainer.evaluate()``` it only returns the following keys ```'eval_runtime', 'eval_samples_per_second', 'eval_steps_per_second', 'epoch'``` so neither the loss or the F1 score\r\n\r\nI am on v4.36.2.\r\n\r\nI also have the same issue with my code, to recapitulate:\r\n- TrainingArguments.prediction_loss_only = False\r\n- TrainingArguments.metric_for_best_model = \"name_of_my_metric\"\r\n- TrainingArguments.greater_is_better = True\r\n- Trainer.compute_metrics = fn_eval_metric which is a function that returns {\"name_of_my_metric\": score_to_improve}\r\n\r\nI am going through the source code of the Trainer to try to figure out what happens in detail because it seems the compute_metrics function isn't called at all ... I tried to introduce a bug in the function but there is no error as it should happen if the function was called, the only error comes at the end of the evaluation loop as it doesn't find the \"eval_name_of_my_metric\" key ...\r\n\r\nI am puzzled, any hints would be great on how to fix this, thanks!", "@adrienchaton - could you try running on the latest release v4.37? ", "@amyeroberts thanks, I updated to v4.37 but the problem persists in the same way.\r\n\r\nAnd if I set \r\n- Trainer.compute_metrics = None\r\n- TrainingArguments.metric_for_best_model = \"loss\"\r\n- TrainingArguments.greater_is_better = False\r\n- TrainingArguments.prediction_loss_only = True\r\n\r\nThen the same error comes, stating\r\n```\r\ntransformers/trainer.py\", line 1929, in _inner_training_loop\r\n self._maybe_log_save_evaluate(tr_loss, model, trial, epoch, ignore_keys_for_eval)\r\ntransformers/trainer.py\", line 2300, in _maybe_log_save_evaluate\r\n self._save_checkpoint(model, trial, metrics=metrics)\r\ntransformers/trainer.py\", line 2389, in _save_checkpoint\r\n metric_value = metrics[metric_to_check]\r\nKeyError: 'eval_loss'```", "this issue might be related, but setting or not the label_names in TrainingArguments didnt change my issues\r\n\r\nhttps://discuss.huggingface.co/t/why-do-i-get-no-validation-loss-and-why-are-metrics-not-calculated/32373", "@amyeroberts @ManishChandra12 I think I figured out the problem, TrainingArguments.label_names wasn't configured properly and it would silently be ignored by the trainer which seems to run evaluation but doesnt compute any metric, neither the loss, nor the custom metrics if providing a compute_metrics function\r\n\r\nmaking sure that your data_collator returns a \"labels\" key AND setting label_names=[\"labels\"] should fix the problem\r\n\r\nI am not the only one who faced this problem .. I think it would be better to print a warning e.g. \"compute_metrics provided but no label_names, thus metrics will be ignored\" because it is misleading to see the progress bar with all eval samples being processed while the metrics arent computed because label_names keys arent found in the input", "Thanks for updating with the fix @adrienchaton! \r\n\r\nI'll let @pacman100 and @muellerzr decide on the best way to handle / flag on the trainer side " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28530/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "Muskanb", "id": 35324348, "node_id": "MDQ6VXNlcjM1MzI0MzQ4", "avatar_url": "https://avatars.githubusercontent.com/u/35324348?v=4", "gravatar_id": "", "url": "https://api.github.com/users/Muskanb", "html_url": "https://github.com/Muskanb", "followers_url": "https://api.github.com/users/Muskanb/followers", "following_url": "https://api.github.com/users/Muskanb/following{/other_user}", "gists_url": "https://api.github.com/users/Muskanb/gists{/gist_id}", "starred_url": "https://api.github.com/users/Muskanb/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/Muskanb/subscriptions", "organizations_url": "https://api.github.com/users/Muskanb/orgs", "repos_url": "https://api.github.com/users/Muskanb/repos", "events_url": "https://api.github.com/users/Muskanb/events{/privacy}", "received_events_url": "https://api.github.com/users/Muskanb/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @Muskanb, thanks for raising an issue! \r\n\r\nIs the error occurring when calling `LlamaForCausalLM.from_pretrained` or `pa_extractor.load_adapter`? ", "Hey @amyeroberts, Thanks for responding. It occurs while calling ` pa_extractor.load_adapter`", "hi @Muskanb what are your transformers & peft versions? ", "These are the versions : \r\n ```\r\npeft==0.6.2\r\ntransformers==4.35.2\r\n```\r\n", "`token` argument should be supported in the latest transformers version: https://github.com/huggingface/transformers/blob/002566f398cd6dbf7053a89c26646ac45be540f4/src/transformers/integrations/peft.py#L74 , can you try to update peft & transformers? `pip install -U transformers peft`", "@younesbelkada It worked. Appreciate the prompt help!! :)", "Awesome thanks @Muskanb !", "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/28529/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "hhllxx1121", "id": 96508996, "node_id": "U_kgDOBcCcRA", "avatar_url": "https://avatars.githubusercontent.com/u/96508996?v=4", "gravatar_id": "", "url": "https://api.github.com/users/hhllxx1121", "html_url": "https://github.com/hhllxx1121", "followers_url": "https://api.github.com/users/hhllxx1121/followers", "following_url": "https://api.github.com/users/hhllxx1121/following{/other_user}", "gists_url": "https://api.github.com/users/hhllxx1121/gists{/gist_id}", "starred_url": "https://api.github.com/users/hhllxx1121/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/hhllxx1121/subscriptions", "organizations_url": "https://api.github.com/users/hhllxx1121/orgs", "repos_url": "https://api.github.com/users/hhllxx1121/repos", "events_url": "https://api.github.com/users/hhllxx1121/events{/privacy}", "received_events_url": "https://api.github.com/users/hhllxx1121/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @gante ", "Hi @hhllxx1121 👋 \r\n\r\nSee this issue and its comments: https://github.com/huggingface/transformers/issues/28075", "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/28528/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/28527/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/28527", "html_url": "https://github.com/huggingface/transformers/pull/28527", "diff_url": "https://github.com/huggingface/transformers/pull/28527.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28527.patch", "merged_at": 1705419436000 }

{ "login": "DamienAllonsius", "id": 11852475, "node_id": "MDQ6VXNlcjExODUyNDc1", "avatar_url": "https://avatars.githubusercontent.com/u/11852475?v=4", "gravatar_id": "", "url": "https://api.github.com/users/DamienAllonsius", "html_url": "https://github.com/DamienAllonsius", "followers_url": "https://api.github.com/users/DamienAllonsius/followers", "following_url": "https://api.github.com/users/DamienAllonsius/following{/other_user}", "gists_url": "https://api.github.com/users/DamienAllonsius/gists{/gist_id}", "starred_url": "https://api.github.com/users/DamienAllonsius/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/DamienAllonsius/subscriptions", "organizations_url": "https://api.github.com/users/DamienAllonsius/orgs", "repos_url": "https://api.github.com/users/DamienAllonsius/repos", "events_url": "https://api.github.com/users/DamienAllonsius/events{/privacy}", "received_events_url": "https://api.github.com/users/DamienAllonsius/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@amyeroberts thanks a lot for your answer, I'll close this PR then. \r\n\r\nFor the record, I have actually tried to pass one hot labels: \r\n```\r\ndef one_hot_label(batch):\r\n batch[\"label\"] = [torch.functional.F.one_hot(torch.tensor(label), num_classes=18).double() for label in batch[\"label\"]]\r\n return batch\r\nds = ds.map(one_hot_label, batched=True)\r\n```\r\nBut the default collate function complained (data_collator.py line 117) because `first[\"label\"]` is now a vector\r\n```\r\nlabel = first[\"label\"].item() if isinstance(first[\"label\"], torch.Tensor) else first[\"label\"]\r\n```\r\nSo I had to create a new custom collate function and I run into other problems... I hoped to make it less painful with this update.\r\n\r\n(If you have any other idea to use one hot labels in a more straightforward way, that would be great :))" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28526/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/28526", "html_url": "https://github.com/huggingface/transformers/pull/28526", "diff_url": "https://github.com/huggingface/transformers/pull/28526.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28526.patch", "merged_at": null }

{ "login": "monowaranjum", "id": 19803082, "node_id": "MDQ6VXNlcjE5ODAzMDgy", "avatar_url": "https://avatars.githubusercontent.com/u/19803082?v=4", "gravatar_id": "", "url": "https://api.github.com/users/monowaranjum", "html_url": "https://github.com/monowaranjum", "followers_url": "https://api.github.com/users/monowaranjum/followers", "following_url": "https://api.github.com/users/monowaranjum/following{/other_user}", "gists_url": "https://api.github.com/users/monowaranjum/gists{/gist_id}", "starred_url": "https://api.github.com/users/monowaranjum/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/monowaranjum/subscriptions", "organizations_url": "https://api.github.com/users/monowaranjum/orgs", "repos_url": "https://api.github.com/users/monowaranjum/repos", "events_url": "https://api.github.com/users/monowaranjum/events{/privacy}", "received_events_url": "https://api.github.com/users/monowaranjum/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @monowaranjum, thanks for raising an issue! \r\n\r\nThe supported way to save and load models is using the `model.save_pretrained(checkpoint)` and `ModelClass.from_pretrained(checkpoint)` methods. Is there a reason you're using the keras methods instead? ", "Hi @monowaranjum, just to expand on @amyeroberts's comment - all `transformers` TF models are Keras models, but not all Keras models are `transformers` models. When you use Keras save methods, what is saved is a pure Keras model, not a `transformers` model. This is okay when you just want to export a model for inference, but you will lose access to methods like `generate()`, which are part of `transformers`, not Keras.\r\n\r\nIf you want to save and load a `transformers` model, you should use `save_pretrained` and `TFWhisperForConditionalGeneration.from_pretrained()`. We recommend saving pure TF/Keras models only when you want to export the model for inference with e.g. TFLite or something like that.", "HI, Thank you so much for the responses. It clears up a few confusions I had. Before I proceed any farther: I would like to give some context as to why I am trying to do so. \r\n\r\nI am trying to compile the whisper model for a specific GPU. For the compilation process to complete, I am using tensorflow and OpenXLA (IREE) compilation toolchain. In order to lower the model in MLIR and then lowering it all the way to LLVM bitcode that can be processed for the specific GPU backend, I need the input model to be in tensorflow saved_model format. So, the process look like this: \r\n\r\nTF Saved model -> Import into MLIR -> Convert to LLVM -> GPU specific Kernel Generation -> GPU executable for inference. \r\n\r\nNow in order to do that, I tried to save the model using keras's save function. \r\n\r\nQ1: I see when I use keras's save function and load it there are still ```call``` methods in it. Does that perform inference on the model?\r\nQ2: If Q1 is answered no, is there a way I can run inference on the keras saved model ? Which function to call for that purpose and where can I find a documentation/worked out example for that. \r\nQ3: If Q1 is yes, can you please point me to a direction where I can find a worked out example/ documentation on how to use the call function? I am specifically looking for how to set the following parameters: ```decoder_attention_mask, decoder_input_ids```. I tried calling the call function, but could not understand the following error message. \r\n\r\n\r\n```\r\nPositional arguments (16 total):\r\n * <tf.Tensor 'input_features:0' shape=(1, 80, 3000) dtype=float64>\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * False\r\n Keyword arguments: {}\r\n\r\n Expected these arguments to match one of the following 2 option(s):\r\n\r\nOption 1:\r\n Positional arguments (16 total):\r\n * {'decoder_attention_mask': TensorSpec(shape=(None, None), dtype=tf.int32, name='decoder_attention_mask'),\r\n 'decoder_input_ids': TensorSpec(shape=(None, None), dtype=tf.int32, name='decoder_input_ids'),\r\n 'input_features': TensorSpec(shape=(None, 80, None), dtype=tf.float32, name='input_features_input_features')}\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * True\r\n Keyword arguments: {}\r\n\r\nOption 2:\r\n Positional arguments (16 total):\r\n * {'decoder_attention_mask': TensorSpec(shape=(None, None), dtype=tf.int32, name='decoder_attention_mask'),\r\n 'decoder_input_ids': TensorSpec(shape=(None, None), dtype=tf.int32, name='decoder_input_ids'),\r\n 'input_features': TensorSpec(shape=(None, 80, None), dtype=tf.float32, name='input_features_input_features')}\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * None\r\n * False\r\n Keyword arguments: {}\r\n```\r\n\r\nThank you so much again for the wonderful work you are doing for the community. " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28525/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "andersonm-ibm", "id": 63074550, "node_id": "MDQ6VXNlcjYzMDc0NTUw", "avatar_url": "https://avatars.githubusercontent.com/u/63074550?v=4", "gravatar_id": "", "url": "https://api.github.com/users/andersonm-ibm", "html_url": "https://github.com/andersonm-ibm", "followers_url": "https://api.github.com/users/andersonm-ibm/followers", "following_url": "https://api.github.com/users/andersonm-ibm/following{/other_user}", "gists_url": "https://api.github.com/users/andersonm-ibm/gists{/gist_id}", "starred_url": "https://api.github.com/users/andersonm-ibm/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/andersonm-ibm/subscriptions", "organizations_url": "https://api.github.com/users/andersonm-ibm/orgs", "repos_url": "https://api.github.com/users/andersonm-ibm/repos", "events_url": "https://api.github.com/users/andersonm-ibm/events{/privacy}", "received_events_url": "https://api.github.com/users/andersonm-ibm/received_events", "type": "User", "site_admin": false }

[]

[]

[ "@amyeroberts For reference.\r\n\r\nThis is totally OK, pipelines are not supposed to support every option models do. Use the example code if you want more control over the generation.\r\n\r\nClosing for now since this is normal behavior." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28524/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "dashapetr", "id": 54349415, "node_id": "MDQ6VXNlcjU0MzQ5NDE1", "avatar_url": "https://avatars.githubusercontent.com/u/54349415?v=4", "gravatar_id": "", "url": "https://api.github.com/users/dashapetr", "html_url": "https://github.com/dashapetr", "followers_url": "https://api.github.com/users/dashapetr/followers", "following_url": "https://api.github.com/users/dashapetr/following{/other_user}", "gists_url": "https://api.github.com/users/dashapetr/gists{/gist_id}", "starred_url": "https://api.github.com/users/dashapetr/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/dashapetr/subscriptions", "organizations_url": "https://api.github.com/users/dashapetr/orgs", "repos_url": "https://api.github.com/users/dashapetr/repos", "events_url": "https://api.github.com/users/dashapetr/events{/privacy}", "received_events_url": "https://api.github.com/users/dashapetr/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey! Thanks, you are probably right. Would you like to open a PR to change this and make it more friendly? " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28523/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 }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28522). 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/28522/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/28522", "html_url": "https://github.com/huggingface/transformers/pull/28522", "diff_url": "https://github.com/huggingface/transformers/pull/28522.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28522.patch", "merged_at": 1705420202000 }

{ "login": "inisis", "id": 46103969, "node_id": "MDQ6VXNlcjQ2MTAzOTY5", "avatar_url": "https://avatars.githubusercontent.com/u/46103969?v=4", "gravatar_id": "", "url": "https://api.github.com/users/inisis", "html_url": "https://github.com/inisis", "followers_url": "https://api.github.com/users/inisis/followers", "following_url": "https://api.github.com/users/inisis/following{/other_user}", "gists_url": "https://api.github.com/users/inisis/gists{/gist_id}", "starred_url": "https://api.github.com/users/inisis/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/inisis/subscriptions", "organizations_url": "https://api.github.com/users/inisis/orgs", "repos_url": "https://api.github.com/users/inisis/repos", "events_url": "https://api.github.com/users/inisis/events{/privacy}", "received_events_url": "https://api.github.com/users/inisis/received_events", "type": "User", "site_admin": false }

[]

[]

[ "> Hi @inisis, thanks for opening this PR!\r\n> \r\n> A word of caution - functions like these (not importable from the top-level of transformers, not part of the public documentation) are subject to change and not guaranteed to never be renamed, have modified behaviour or deleted. In fact - this PR does just that!\r\n> \r\n> I don't think we should make this change. Semantically the function is performing a check. The equivalent, modified function would be `is_model_supported`, and users can implement that themselves easily.\r\n> \r\n> Moreover, we don't know who might already be using this function, and relying on an exception being raised.\r\n> \r\n> I do however agree it's better to have a single function return the bool and then the user can decide what to do with the output (raise etc.). If you want, we can add `is_model_supported` and have `check_if_model_is_supported` use that.\r\n\r\nHi @amyeroberts, thanks for your comment, I do agree that adding `is_model_supported` is far better than changing existing one, and I do feel this pr is needed, because I'm writing an intergrated tracer, it will choose different tracer based on the input model, this pr would be very helpful and elegant." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28521/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/28521", "html_url": "https://github.com/huggingface/transformers/pull/28521", "diff_url": "https://github.com/huggingface/transformers/pull/28521.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28521.patch", "merged_at": 1705427564000 }

{ "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/28520/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/28520", "html_url": "https://github.com/huggingface/transformers/pull/28520", "diff_url": "https://github.com/huggingface/transformers/pull/28520.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28520.patch", "merged_at": 1705419389000 }

{ "login": "pydaxing", "id": 129026999, "node_id": "U_kgDOB7DLtw", "avatar_url": "https://avatars.githubusercontent.com/u/129026999?v=4", "gravatar_id": "", "url": "https://api.github.com/users/pydaxing", "html_url": "https://github.com/pydaxing", "followers_url": "https://api.github.com/users/pydaxing/followers", "following_url": "https://api.github.com/users/pydaxing/following{/other_user}", "gists_url": "https://api.github.com/users/pydaxing/gists{/gist_id}", "starred_url": "https://api.github.com/users/pydaxing/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/pydaxing/subscriptions", "organizations_url": "https://api.github.com/users/pydaxing/orgs", "repos_url": "https://api.github.com/users/pydaxing/repos", "events_url": "https://api.github.com/users/pydaxing/events{/privacy}", "received_events_url": "https://api.github.com/users/pydaxing/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @pydaxing, thanks for raising an issue! \r\n\r\nSo that we can best help you, could you make sure to follow the issue template and provide: \r\n* The running environment: run `transformers-cli env` in the terminal and copy-paste the output\r\n* A minimal code reproducer\r\n* The full error traceback", "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/28519/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "andysingal", "id": 20493493, "node_id": "MDQ6VXNlcjIwNDkzNDkz", "avatar_url": "https://avatars.githubusercontent.com/u/20493493?v=4", "gravatar_id": "", "url": "https://api.github.com/users/andysingal", "html_url": "https://github.com/andysingal", "followers_url": "https://api.github.com/users/andysingal/followers", "following_url": "https://api.github.com/users/andysingal/following{/other_user}", "gists_url": "https://api.github.com/users/andysingal/gists{/gist_id}", "starred_url": "https://api.github.com/users/andysingal/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/andysingal/subscriptions", "organizations_url": "https://api.github.com/users/andysingal/orgs", "repos_url": "https://api.github.com/users/andysingal/repos", "events_url": "https://api.github.com/users/andysingal/events{/privacy}", "received_events_url": "https://api.github.com/users/andysingal/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @andysingal. To use Kosmos-2 for image grounding, you have to add a special `<grounding>` token before the prompt, as they do in the [paper](https://arxiv.org/abs/2306.14824). Also you can use `<phrase>` token to get bboxes of specific phrases in the format `prev_prompt_text <phrase>the_object</phrase>`\r\n\r\n```from transformers import AutoProcessor, AutoModelForVision2Seq\r\nimport requests\r\nfrom PIL import Image\r\n\r\nprocessor = AutoProcessor.from_pretrained(\"microsoft/kosmos-2-patch14-224\")\r\nmodel = AutoModelForVision2Seq.from_pretrained(\"microsoft/kosmos-2-patch14-224\", device_map=\"cpu\")\r\n\r\nprompt_grounded = \"<grounding>An image of\"\r\nprompt_refer = \"An image of <phrase>a snowman</phrase>\"\r\n\r\nurl = \"https://huggingface.co./microsoft/kosmos-2-patch14-224/resolve/main/snowman.png\"\r\nimage = Image.open(requests.get(url, stream=True).raw)\r\nimage\r\n\r\ninputs = processor(text=prompt_refer, images=image, return_tensors=\"pt\").to(model.device)\r\n\r\n# autoregressively generate completion\r\ngenerated_ids = model.generate(**inputs, max_new_tokens=100)\r\ninput_len = inputs['input_ids'].shape[-1]\r\n\r\n# convert generated token IDs back to strings\r\ngenerated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]\r\n\r\n# By default, the generated text is cleaned up and the entities are extracted.\r\nprocessed_text, entities = processor.post_process_generation(generated_text)\r\n\r\nprint(processed_text)\r\nprint(entities)\r\n>>> An image of a snowman warming himself by a campfire\r\n>>> [('a snowman', (12, 21), [(0.390625, 0.046875, 0.984375, 0.828125)]), ('a campfire', (41, 51), [(0.109375, 0.640625, 0.546875, 0.984375)])]\r\n```", "Thank you very much\r\n\r\nOn Wed, Jan 17, 2024 at 3:08 PM Raushan Turganbay ***@***.***>\r\nwrote:\r\n\r\n> Hi @andysingal <https://github.com/andysingal>. To use Kosmos-2 for image\r\n> grounding, you have to add a special <grounding> token before the prompt,\r\n> as they do in the paper <https://arxiv.org/abs/2306.14824>. Also you can\r\n> use <phrase> token to get bboxes of specific phrases in the format prev_prompt_text\r\n> <phrase>the_object</phrase>\r\n>\r\n> import requests\r\n> from PIL import Image\r\n>\r\n> processor = AutoProcessor.from_pretrained(\"microsoft/kosmos-2-patch14-224\")\r\n> model = AutoModelForVision2Seq.from_pretrained(\"microsoft/kosmos-2-patch14-224\", device_map=\"cpu\")\r\n>\r\n> prompt_grounded = \"<grounding>An image of\"\r\n> prompt_refer = \"An image of <phrase>a snowman</phrase>\"\r\n>\r\n> url = \"https://huggingface.co./microsoft/kosmos-2-patch14-224/resolve/main/snowman.png\"\r\n> image = Image.open(requests.get(url, stream=True).raw)\r\n> image\r\n>\r\n> inputs = processor(text=prompt_refer, images=image, return_tensors=\"pt\").to(model.device)\r\n>\r\n> # autoregressively generate completion\r\n> generated_ids = model.generate(**inputs, max_new_tokens=100)\r\n> input_len = inputs['input_ids'].shape[-1]\r\n>\r\n> # convert generated token IDs back to strings\r\n> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]\r\n>\r\n> # By default, the generated text is cleaned up and the entities are extracted.\r\n> processed_text, entities = processor.post_process_generation(generated_text)\r\n>\r\n> print(processed_text)\r\n> print(entities)\r\n> >>> An image of a snowman warming himself by a campfire\r\n> >>> [('a snowman', (12, 21), [(0.390625, 0.046875, 0.984375, 0.828125)]), ('a campfire', (41, 51), [(0.109375, 0.640625, 0.546875, 0.984375)])]\r\n>\r\n> —\r\n> Reply to this email directly, view it on GitHub\r\n> <https://github.com/huggingface/transformers/issues/28518#issuecomment-1895434657>,\r\n> or unsubscribe\r\n> <https://github.com/notifications/unsubscribe-auth/AE4LJNNJZCFNQ6UKHCMAE3DYO6LYHAVCNFSM6AAAAABB4ENBYKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOJVGQZTINRVG4>\r\n> .\r\n> You are receiving this because you were mentioned.Message ID:\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/28518/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "khaimt", "id": 145790391, "node_id": "U_kgDOCLCVtw", "avatar_url": "https://avatars.githubusercontent.com/u/145790391?v=4", "gravatar_id": "", "url": "https://api.github.com/users/khaimt", "html_url": "https://github.com/khaimt", "followers_url": "https://api.github.com/users/khaimt/followers", "following_url": "https://api.github.com/users/khaimt/following{/other_user}", "gists_url": "https://api.github.com/users/khaimt/gists{/gist_id}", "starred_url": "https://api.github.com/users/khaimt/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/khaimt/subscriptions", "organizations_url": "https://api.github.com/users/khaimt/orgs", "repos_url": "https://api.github.com/users/khaimt/repos", "events_url": "https://api.github.com/users/khaimt/events{/privacy}", "received_events_url": "https://api.github.com/users/khaimt/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28517). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "> Thanks for the update. Can you add a good test to make sure this behaves as expected? 🤗\r\n\r\nSure, I will add a test to make sure that the implementation is correct", "> Thanks for the update. Can you add a good test to make sure this behaves as expected? 🤗\r\n\r\nHi @ArthurZucker, I have just added the tests for load balancing loss ", "@ArthurZucker I don't know why The test is failed on github but was passed in my local test.\r\nIn my local test, I only changed the test_mixtral.py so I ran:\r\npytest --picked\r\nAnd test was passed, but was failed on github.\r\nIt is as if the modeling_mixtral.py was not updated on github test ", "> I don't know why The test is failed on github but was passed in my local test. \r\n\r\nFrom details of [check_code_quality](https://circleci.com/gh/huggingface/transformers/1066145):\r\ntests/models/mixtral/test_modeling_mixtral.py:19:8: F401 [*] `random` imported but unused\r\n\r\nUnused import of the random module in the test file.", "> > I don't know why The test is failed on github but was passed in my local test.\r\n> \r\n> From details of [check_code_quality](https://circleci.com/gh/huggingface/transformers/1066145): tests/models/mixtral/test_modeling_mixtral.py:19:8: F401 [*] `random` imported but unused\r\n> \r\n> Unused import of the random module in the test file.\r\n\r\nI mean this assert: https://app.circleci.com/pipelines/github/huggingface/transformers/82724/workflows/61975ffc-0bdb-4aa7-935e-70f69f11e748/jobs/1066148/parallel-runs/0/steps/0-117", "Some thoughts:\r\n- Regarding `∣actual−expected∣>atol+rtol⋅∣expected∣`.. You aren't getting the fabs() of result.aux_loss before multiplying. \r\n- There's also an [assertNotAlmostEqual](https://github.com/huggingface/transformers/blob/3f69f415adcbdaedec154ba8eac220ef3276975d/tests/models/encoder_decoder/test_modeling_encoder_decoder.py#L817) function for comparing if they're not close.\r\n- Specifying the device to use might give different results on different systems\r\n\r\nThank you for doing this btw!\r\n\r\nEdit for clarity", "> Some thoughts:\r\n> \r\n> * Regarding `∣actual−expected∣>atol+rtol⋅∣expected∣`.. You aren't getting the fabs() of result.aux_loss before multiplying.\r\n> \r\n> * There's also an [assertNotAlmostEqual](https://github.com/huggingface/transformers/blob/3f69f415adcbdaedec154ba8eac220ef3276975d/tests/models/encoder_decoder/test_modeling_encoder_decoder.py#L817) function for comparing if they're not close.\r\n> \r\n> * Specifying the device to use might give different results on different systems\r\n> \r\n> \r\n> Thank you for doing this btw!\r\n> \r\n> Edit for clarity\r\n\r\nHi @congruency Thank you for your comments\r\nActually we no need to get the fabs() of result.aux_loss, because it is already positive, and actually the value is around 2. You can see the value from previous assert: \r\nhttps://github.com/khaimt/transformers/blob/exclude_padding_tokens_in_aux_loss_Mixtral_Moe/tests/models/mixtral/test_modeling_mixtral.py#L478 \r\nOh, thankyou for reminding me of: assertNotAlmostEqual, I will fix my code with this suggestion", "@ArthurZucker The tests have been fully added and passed, can you review ?", "Am a business ", "UC eye care center ", "Sorry for the delay will review! ", "@ArthurZucker I have just fixed my PR:\r\n+ move attention_mask to the last position\r\n+ pad_leng --> pad_length\r\n+ reduce the tol for assert_close" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28517/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/28517", "html_url": "https://github.com/huggingface/transformers/pull/28517", "diff_url": "https://github.com/huggingface/transformers/pull/28517.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28517.patch", "merged_at": 1706087534000 }

{ "login": "superleesa", "id": 88019950, "node_id": "MDQ6VXNlcjg4MDE5OTUw", "avatar_url": "https://avatars.githubusercontent.com/u/88019950?v=4", "gravatar_id": "", "url": "https://api.github.com/users/superleesa", "html_url": "https://github.com/superleesa", "followers_url": "https://api.github.com/users/superleesa/followers", "following_url": "https://api.github.com/users/superleesa/following{/other_user}", "gists_url": "https://api.github.com/users/superleesa/gists{/gist_id}", "starred_url": "https://api.github.com/users/superleesa/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/superleesa/subscriptions", "organizations_url": "https://api.github.com/users/superleesa/orgs", "repos_url": "https://api.github.com/users/superleesa/repos", "events_url": "https://api.github.com/users/superleesa/events{/privacy}", "received_events_url": "https://api.github.com/users/superleesa/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Actually, the problem occurred even after removing the EarlyStoppingCallback..... ", "cc @muellerzr @pacman100 ", "The weirdest part is that the error always occurs at a training step between 229~235. I was thinking that this is due to a deadlock caused by evaluation / saving but even if I set the the evaluation & saving step to 300 (i.e. the first evaluation & saving step happens at 300 which is after the error occurs), the error still occurs, implying that they are not likely the cause. \r\n\r\nAlso, I tried this with both DDP and Deepspeed but the same problem happens for the both." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28516/reactions", "total_count": 1, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "tshrjn", "id": 8372098, "node_id": "MDQ6VXNlcjgzNzIwOTg=", "avatar_url": "https://avatars.githubusercontent.com/u/8372098?v=4", "gravatar_id": "", "url": "https://api.github.com/users/tshrjn", "html_url": "https://github.com/tshrjn", "followers_url": "https://api.github.com/users/tshrjn/followers", "following_url": "https://api.github.com/users/tshrjn/following{/other_user}", "gists_url": "https://api.github.com/users/tshrjn/gists{/gist_id}", "starred_url": "https://api.github.com/users/tshrjn/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/tshrjn/subscriptions", "organizations_url": "https://api.github.com/users/tshrjn/orgs", "repos_url": "https://api.github.com/users/tshrjn/repos", "events_url": "https://api.github.com/users/tshrjn/events{/privacy}", "received_events_url": "https://api.github.com/users/tshrjn/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @tshrjn, thanks for raising an issue! \r\n\r\nIt looks like this is an issue on how the transformers model is wrapped on unsloth's side. Our model's don't natively have a `merge_and_upload` method. ", "@amyeroberts Thanks for taking a look. \n\nThis might be a PEFT library issue. The docs for which states the `merge_and_unload` function [here](https://huggingface.co./docs/peft/v0.7.1/en/package_reference/lora#peft.LoraModel.merge_and_unload).\n", "@tshrjn Thanks for the pointer! cc @younesbelkada ", "Hi @tshrjn \r\nThanks a lot for the issue! Have you correctly called a `FastLanguageModel.get_peft_model` before running the training? I think that unsloth properly takes care of merging lora weights in the base model if you call `merge_and_unload()` - Can you also print `model` before that error? \r\ncc @danielhanchen as well", "Oh I will check this! I'll run the notebook again to investigate :)", "I just updated the Tiny Llama notebook: https://colab.research.google.com/drive/1AZghoNBQaMDgWJpi4RbffGM1h6raLUj9?usp=sharing\r\n\r\nWill announce tomorrow, but we now support merging to GGUF / float16 directly. The saving modules are on the very bottom.\r\n\r\nBut `model.merge_and_unload()` only works if calling `FastLanguageModel.get_peft_model` as Younes described.\r\n\r\nI'll investigate the HF notebooks I uploaded as well + will update them with the latest Unsloth release!", "Thanks very much @danielhanchen ! \r\n@tshrjn as stated by @danielhanchen can you confirm running the cell with `get_peft_model` fixes your issue ? 🙏 ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md) are likely to be ignored." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28515/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 }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28514). 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@amyeroberts @ArthurZucker What do you think would be a sensible deprecation time frame? I do agree it can become a sensible exception: raising an exception at `save_pretrained` corresponds to breaking the training runs (the model artifact is only stored after these lines) ⚠️ Perhaps 6 minor versions, instead of the standard 2?\r\n\r\n(I wouldn't want to leave it to `v5`, as there would be little incentive for the users to do something about it)", "@gante Definitely not until v5 💀 I'd say 4? I don't feel strongly, so happy to go with what you or @ArthurZucker think is best. ", "Bumped to 4 minor versions 👍 (i.e. set to deprecate in v4.41)", "Sorry saw this a bit late, LGTM" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28514/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/28514", "html_url": "https://github.com/huggingface/transformers/pull/28514", "diff_url": "https://github.com/huggingface/transformers/pull/28514.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28514.patch", "merged_at": 1705429862000 }

{ "login": "khaimt", "id": 145790391, "node_id": "U_kgDOCLCVtw", "avatar_url": "https://avatars.githubusercontent.com/u/145790391?v=4", "gravatar_id": "", "url": "https://api.github.com/users/khaimt", "html_url": "https://github.com/khaimt", "followers_url": "https://api.github.com/users/khaimt/followers", "following_url": "https://api.github.com/users/khaimt/following{/other_user}", "gists_url": "https://api.github.com/users/khaimt/gists{/gist_id}", "starred_url": "https://api.github.com/users/khaimt/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/khaimt/subscriptions", "organizations_url": "https://api.github.com/users/khaimt/orgs", "repos_url": "https://api.github.com/users/khaimt/repos", "events_url": "https://api.github.com/users/khaimt/events{/privacy}", "received_events_url": "https://api.github.com/users/khaimt/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28513/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/28513", "html_url": "https://github.com/huggingface/transformers/pull/28513", "diff_url": "https://github.com/huggingface/transformers/pull/28513.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28513.patch", "merged_at": null }

{ "login": "haixpham", "id": 32718796, "node_id": "MDQ6VXNlcjMyNzE4Nzk2", "avatar_url": "https://avatars.githubusercontent.com/u/32718796?v=4", "gravatar_id": "", "url": "https://api.github.com/users/haixpham", "html_url": "https://github.com/haixpham", "followers_url": "https://api.github.com/users/haixpham/followers", "following_url": "https://api.github.com/users/haixpham/following{/other_user}", "gists_url": "https://api.github.com/users/haixpham/gists{/gist_id}", "starred_url": "https://api.github.com/users/haixpham/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/haixpham/subscriptions", "organizations_url": "https://api.github.com/users/haixpham/orgs", "repos_url": "https://api.github.com/users/haixpham/repos", "events_url": "https://api.github.com/users/haixpham/events{/privacy}", "received_events_url": "https://api.github.com/users/haixpham/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @pacman100 @muellerzr " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28512/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_28511). 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/28511/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/28511", "html_url": "https://github.com/huggingface/transformers/pull/28511", "diff_url": "https://github.com/huggingface/transformers/pull/28511.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28511.patch", "merged_at": 1705338055000 }

{ "login": "haixpham", "id": 32718796, "node_id": "MDQ6VXNlcjMyNzE4Nzk2", "avatar_url": "https://avatars.githubusercontent.com/u/32718796?v=4", "gravatar_id": "", "url": "https://api.github.com/users/haixpham", "html_url": "https://github.com/haixpham", "followers_url": "https://api.github.com/users/haixpham/followers", "following_url": "https://api.github.com/users/haixpham/following{/other_user}", "gists_url": "https://api.github.com/users/haixpham/gists{/gist_id}", "starred_url": "https://api.github.com/users/haixpham/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/haixpham/subscriptions", "organizations_url": "https://api.github.com/users/haixpham/orgs", "repos_url": "https://api.github.com/users/haixpham/repos", "events_url": "https://api.github.com/users/haixpham/events{/privacy}", "received_events_url": "https://api.github.com/users/haixpham/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Not sure if this helps, but I'm also having trouble reproducing accuracy of CodeLlama 13B-I trained with Zero2 using Zero3. ", "Meet the same problem when using [this snippet](https://github.com/huggingface/accelerate/blob/cea6aaa1161d45f7f23ef33fcc3b0a5999ebb5a1/examples/by_feature/deepspeed_with_config_support.py#L712-L723) to save a zero-3 model." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28510/reactions", "total_count": 3, "+1": 3, "-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 }

[]

[]

[ "@ydshieh The same capturelogger errors arose again. A cheeky re-run (with ssh) made them pass, so merging as they're not related to this PR" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28509/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/28509", "html_url": "https://github.com/huggingface/transformers/pull/28509", "diff_url": "https://github.com/huggingface/transformers/pull/28509.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28509.patch", "merged_at": 1705347420000 }

{ "login": "ofirzaf", "id": 18296312, "node_id": "MDQ6VXNlcjE4Mjk2MzEy", "avatar_url": "https://avatars.githubusercontent.com/u/18296312?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ofirzaf", "html_url": "https://github.com/ofirzaf", "followers_url": "https://api.github.com/users/ofirzaf/followers", "following_url": "https://api.github.com/users/ofirzaf/following{/other_user}", "gists_url": "https://api.github.com/users/ofirzaf/gists{/gist_id}", "starred_url": "https://api.github.com/users/ofirzaf/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ofirzaf/subscriptions", "organizations_url": "https://api.github.com/users/ofirzaf/orgs", "repos_url": "https://api.github.com/users/ofirzaf/repos", "events_url": "https://api.github.com/users/ofirzaf/events{/privacy}", "received_events_url": "https://api.github.com/users/ofirzaf/received_events", "type": "User", "site_admin": false }

[]

[]

[ "I was thinking about changing `max_new_tokens` inside the `get_candidates` method so we won't generate more tokens than `max_length - 1`:\r\n```python\r\nmax_new_tokens = min(int(self.num_assistant_tokens), self.generation_config.max_length - new_cur_len - 1)\r\n```\r\nThis will help avoid checking in several places if `n_matches` is too big, it can also cause a problem in `_speculative_sampling`, and prevent unnecessary compute. \r\n\r\nWhat do you think?\r\n", "@ofirzaf good suggestion, the candidate model should indeed be set up such that the resulting candidate sequence has at most a length of `max length - 1`, as the main model will always generate one additional token 👍 ", "> Thanks for fixing this!\r\n> \r\n> +1 on Joao's comment reworking to make more understandable and adding epsilon.\r\n> \r\n> Could you also add at least one test which fails on main but fixes with this fix? It's fine if it's an integration test checks generated outputs\r\n\r\nI've been trying to verify the correctness of the Speculative Decoding (SD) implementation. To do so, I'm planning to add a test which verifies that the token distributions we get using SD are indeed similar (or very close) to the token distributions when using standard sample based decoding. If I understand correctly, this is what the SD paper guarantees. \r\n\r\nSo far, I found that when comparing the token scores of the above methods on a sample input, they are very similar but not identical, perhaps only due to hardware numerics errors.\r\nSince tokens with very low probability do not affect the sampling outcome, I think the test should verify that the top `k` token scores of both methods are close enough to each other (for some chose `k` like 5-10). Do you think there's a better way?\r\n\r\nWhen I benchmark Assisted Generation vs SD, I get very similar latencies, even though SD should theoretically be faster, due to a more relaxed acceptance criteria. In consequence, I think it's important to verify the correctness of the SD implementation e.g. by comparing outputs like above. \r\n\r\n@gante @ofirzaf \r\nWDYT?\r\n\r\nMy results on A100:\r\n```bash\r\nTarget: bigcode/starcoder\r\nAssistant: bigcode/tiny_starcoder_py\r\n\r\nData: openai_humaneval, 20 random examples (args.seed=42)\r\n\r\nSetup: \r\nargs.seed=42\r\nmax_new_tokens=128\r\noutput_scores=True\r\n\r\n======================================================================================================================================================\r\nMethod Token Latency Acceptance Rate Gen. Args\r\n======================================================================================================================================================\r\n\r\nsample 45.19ms ------ {'do_sample': True, 'temperature': 0.2, 'assistant_model': False}\r\nsd 25.53ms 67.81% {'do_sample': True, 'temperature': 0.2, 'assistant_model': True}\r\nag 24.04ms 70.78% {'assistant_model': True}\r\n======================================================================================================================================================\r\n```", "@amyeroberts I am not sure what kind of test do you expect to see. I wrote the following test which tests `_speculative_sampling` directly with dummy input data. Is this the kind of testing you are looking for?\r\n```python\r\nimport unittest \r\nfrom transformers.generation.utils import _speculative_sampling\r\nimport torch\r\n\r\nclass Test(unittest.TestCase):\r\n def test_speculative_sampling(self):\r\n # assume vocab size 10, input length 5 + 3 generated candidates\r\n candidate_input_ids = torch.tensor([[8, 0, 3, 9, 8, 1, 4, 5]]) # input tokens\r\n candidate_logits = torch.tensor([[\r\n [-10., 10., -10., -10., -10., -10., -10., -10., -10., -10.], # generated 1\r\n [-10., -10., -10., -10., 10., -10., -10., -10., -10., -10.], # generated 4\r\n [-10., -10., -10., -10., -10., 10., -10., -10., -10., -10.], # generated 5\r\n ]])\r\n candidate_length = 3\r\n inf = float('inf')\r\n new_logits = torch.tensor([[\r\n [-10., 10., -10., -10., -10., -10., -10., -10., -10., -10.], # accepts 1\r\n [-10., -10., -10., -10., 10., -10., -10., -10., -10., -10.], # accepts 4\r\n [-inf, -inf, -inf, -inf, -inf, -inf, -inf, -inf, 10., -inf], # rejects 5, accepts 8\r\n [-10., -10., -10., -10., -10., -10., -10., -10., -10., -10.], # N/A\r\n ]])\r\n last_assistant_token_is_eos=False\r\n max_matches = 5\r\n validated_tokens, n_matches = _speculative_sampling(\r\n candidate_input_ids,\r\n candidate_logits,\r\n candidate_length,\r\n new_logits,\r\n last_assistant_token_is_eos,\r\n max_matches,\r\n )\r\n self.assertTrue(n_matches.item() == 2)\r\n self.assertTrue(validated_tokens.tolist()[0] == [1, 4, 8])\r\n```", "@danielkorat we want to test that the probability to sample token $t$ from the target model is the same as the probability to sample $t$ using speculative. Maybe the following test will work, given a prompt and a target model, calculate the distribution for the next token by simply running the model on the prompt and applying `softmax` on the resulted logits. Then, we can generate $n$ guesses for the next token using speculative. If speculative is implemented correctly and $n$ is large enough you should get a distribution that would match the calculated distribution from before. \r\n\r\nDoes that make sense? WDYT?", "@ofirzaf Yes we can either do that or compare the logits using `np.allclose()`.\r\nI wonder though if the guarantee holds for the next token only and not for all subsequent tokens.\r\nIf it holds only for the next token, we can have the test feed the SD algo with standard sampled (STD) input_ids one at a time (max_new_tokens=1) and compare outputs (one at a time) with STD in one of the suggested methods above.\r\n________________________________\r\nFrom: Ofir Zafrir ***@***.***>\r\nSent: Thursday, January 18, 2024 3:16:46 AM\r\nTo: huggingface/transformers ***@***.***>\r\nCc: Korat, Daniel ***@***.***>; Mention ***@***.***>\r\nSubject: Re: [huggingface/transformers] Fix `_speculative_sampling` implementation (PR #28508)\r\n\r\n\r\n@danielkorat<https://github.com/danielkorat> we want to test that the probability to sample token $t$ from the target model is the same as the probability to sample $t$ using speculative. Maybe the following test will work, given a prompt and a target model, calculate the distribution for the next token by simply running the model on the prompt and applying softmax on the resulted logits. Then, we can generate $n$ guesses for the next token using speculative. If speculative is implemented correctly and $n$ is large enough you should get a distribution that would match the calculated distribution from before.\r\n\r\nDoes that make sense? WDYT?\r\n\r\n—\r\nReply to this email directly, view it on GitHub<https://github.com/huggingface/transformers/pull/28508#issuecomment-1897597109>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AH26TASLOFOHUY467DYRXPTYPBZX5AVCNFSM6AAAAABB3HU4LOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOJXGU4TOMJQHE>.\r\nYou are receiving this because you were mentioned.Message ID: ***@***.***>\r\n", "@ofirzaf your test suggestion would be a great addition 👍 ", "@danielkorat If the claim in the paper is correct, then @ofirzaf suggestion should work regardless of the distance of the target token to the end of the prompt. In other words, we should see the same distribution in the 1st generated token, in the 2nd generated token given the 1st generated token, and so on.\r\n\r\nMy suggestion would be to generate ~10 tokens for a given prompt ~1M times using super small models (e.g. pythia), to confirm it empirically.", "@gante @amyeroberts I have pushed all the changes we discussed. Can you please take a look?", "@gante @ofirzaf \r\nI wrote this code to check our assumptions about the distributions:\r\n\r\n```python\r\nimport torch\r\nfrom transformers import AutoTokenizer, AutoModelForCausalLM, set_seed\r\n\r\ndef top_k(a, k):\r\n sorted, indices = a.sort(descending=True)\r\n return sorted[:k], indices[:k]\r\ndef get_dist(outputs, num_samples=10**6, norm_factor=1e7):\r\n probs = outputs.scores[0][0].softmax(dim=-1)\r\n print(\"probs:\", top_k(probs, k=5))\r\n next_std_token = torch.multinomial(probs, num_samples=num_samples, replacement=True)\r\n counts = torch.bincount(next_std_token, minlength=probs.shape[0]).float()\r\n print(\"counts:\", top_k(counts, k=5))\r\n norm_counts = counts.float() / norm_factor\r\n print(\"softmaxed_counts:\", top_k(counts.softmax(dim=-1), k=5)[0])\r\n print(\"norm_counts:\", top_k(norm_counts, k=5))\r\n softmaxed_norm_counts = top_k(norm_counts.softmax(dim=-1), k=5)[0]\r\n print(\"softmaxed_norm_counts:\", softmaxed_norm_counts)\r\n return softmaxed_norm_counts\r\n\r\ncheckpoint = \"EleutherAI/pythia-1.4b-deduped\"\r\ntokenizer = AutoTokenizer.from_pretrained(checkpoint)\r\nmodel = AutoModelForCausalLM.from_pretrained(checkpoint)\r\nassistant_model = AutoModelForCausalLM.from_pretrained(\"EleutherAI/pythia-160m-deduped\")\r\ngen_kwargs = dict(**tokenizer(\"In other words, we\", return_tensors=\"pt\"),\r\n pad_token_id=tokenizer.eos_token_id,\r\n max_new_tokens=1,\r\n do_sample=True,\r\n temperature=0.1,\r\n return_dict_in_generate=True,\r\n output_scores=True)\r\nprint(\"Standard sampling:\")\r\nset_seed(0)\r\nstd_dist = get_dist(model.generate(**gen_kwargs))\r\nprint(\"\\nSpeculative decoding:\")\r\nset_seed(0)\r\nsd_dist = get_dist(model.generate(**gen_kwargs, assistant_model=assistant_model))\r\nprint(std_dist == sd_dist)\r\n```\r\n\r\nIn short, I get the token probs for some input for both decoding methods, then I sample 1M times from this prob, and get the token counts (I always look at top 5 scores in all the following computations). As you can see below, the `counts` are very close to each other but not equal of course. If we apply `softmax` to the counts, they produce an insignificant result of `[1., 0., 0., 0., 0.]` (tested other inputs too). So I normalize the counts and then apply `softmax`, and then the output distributions look similar, however, they are still not equal to each other (final output). \r\n\r\nWDYT? Is this the way to go?\r\nIf so, I can extend this test to the next 10 generated tokens as well.\n\r\nOutput::\r\n```bash\r\nStandard sampling:\r\nprobs: (tensor([5.9927e-01, 3.8301e-01, 1.7703e-02, 8.4421e-06, 1.5724e-06]), tensor([452, 403, 476, 878, 513]))\r\ncounts: (tensor([5.9925e+05, 3.8286e+05, 1.7879e+04, 1.7000e+01, 1.0000e+00]), tensor([452, 403, 476, 878, 513]))\r\nsoftmaxed_counts: tensor([1., 0., 0., 0., 0.])\r\nnorm_counts: (tensor([5.9925e-02, 3.8286e-02, 1.7879e-03, 1.7000e-06, 1.0000e-07]), tensor([452, 403, 476, 878, 513]))\r\nsoftmaxed_norm_counts: tensor([2.1106e-05, 2.0654e-05, 1.9914e-05, 1.9878e-05, 1.9878e-05])\r\n\r\nSpeculative decoding:\r\nprobs: (tensor([5.9925e-01, 3.8304e-01, 1.7706e-02, 8.4404e-06, 1.5720e-06]), tensor([452, 403, 476, 878, 513]))\r\ncounts: (tensor([5.9922e+05, 3.8288e+05, 1.7883e+04, 1.7000e+01, 1.0000e+00]), tensor([452, 403, 476, 878, 457]))\r\nsoftmaxed_counts: tensor([1., 0., 0., 0., 0.])\r\nnorm_counts: (tensor([5.9922e-02, 3.8288e-02, 1.7883e-03, 1.7000e-06, 1.0000e-07]), tensor([452, 403, 476, 878, 457]))\r\nsoftmaxed_norm_counts: tensor([2.1106e-05, 2.0654e-05, 1.9914e-05, 1.9878e-05, 1.9878e-05])\r\ntensor([False, False, False, False, False])\r\n```", "@danielkorat there are tiny fluctuations to be expected: doing the forward pass with different shapes (e.g. 1 token at a time vs all at once, 1 row at a time vs batched) will result in slightly different outputs. I've written about it in more detail [here](https://github.com/huggingface/transformers/issues/25420#issuecomment-1775317535).\r\n\r\nFactoring in this source of numerical differences, the results do look similar! Thank you for running the experiment 🤗 ", "> @danielkorat there are tiny fluctuations to be expected: doing the forward pass with different shapes (e.g. 1 token at a time vs all at once, 1 row at a time vs batched) will result in slightly different outputs. I've written about it in more detail [here](https://github.com/huggingface/transformers/issues/25420#issuecomment-1775317535).\n> \n> \n> \n> Factoring in this source of numerical differences, the results do look similar! Thank you for running the experiment 🤗 \n\n@gante My intention is to integrate this code into the repo tests. WDYT?\nShould I do this in a separate PR?", "@danielkorat This test requires a lot of compute, so it's not fit for our CI 😅 " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28508/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/28508", "html_url": "https://github.com/huggingface/transformers/pull/28508", "diff_url": "https://github.com/huggingface/transformers/pull/28508.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28508.patch", "merged_at": 1705673252000 }

{ "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 }

[]

[]

[ "Thanks for opening this PR @fxmarty! \r\n\r\nThe reason this happens at the moment is that `model_type` is intended to be a class attribute, not an instance attribute i.e. to be consistent and shared for all instances of the config object and not modified instance to instance. What's the motivation for modifying it like this? ", "Understood!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28507/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/28507", "html_url": "https://github.com/huggingface/transformers/pull/28507", "diff_url": "https://github.com/huggingface/transformers/pull/28507.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28507.patch", "merged_at": null }

{ "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 }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28506). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "failing tests are known flaky" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28506/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/28506", "html_url": "https://github.com/huggingface/transformers/pull/28506", "diff_url": "https://github.com/huggingface/transformers/pull/28506.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28506.patch", "merged_at": 1705575330000 }

{ "login": "khaimt", "id": 145790391, "node_id": "U_kgDOCLCVtw", "avatar_url": "https://avatars.githubusercontent.com/u/145790391?v=4", "gravatar_id": "", "url": "https://api.github.com/users/khaimt", "html_url": "https://github.com/khaimt", "followers_url": "https://api.github.com/users/khaimt/followers", "following_url": "https://api.github.com/users/khaimt/following{/other_user}", "gists_url": "https://api.github.com/users/khaimt/gists{/gist_id}", "starred_url": "https://api.github.com/users/khaimt/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/khaimt/subscriptions", "organizations_url": "https://api.github.com/users/khaimt/orgs", "repos_url": "https://api.github.com/users/khaimt/repos", "events_url": "https://api.github.com/users/khaimt/events{/privacy}", "received_events_url": "https://api.github.com/users/khaimt/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" } ]

[]

[ "cc @ArthurZucker ", "feel free to open a PR for this! Otherwise will mark it as a good second issue 🤗 ", "I would like to work on this issue, i will go through the linked file today and ask any questions i have.", "I was looking at the code.\r\nBelow is what the model outputs\r\n`return MoeModelOutputWithPast(\r\n last_hidden_state=hidden_states,\r\n past_key_values=next_cache,\r\n hidden_states=all_hidden_states,\r\n attentions=all_self_attns,\r\n router_logits=all_router_logits,\r\n )`\r\n \r\nThe attention from the model output can be passed during load_balancing_loss_func, and the function can be changed appropriately to handle the pad tokens.\r\nAm I right in my understanding? @ArthurZucker ", "> feel free to open a PR for this! Otherwise will mark it as a good second issue 🤗\r\n\r\nHi @ArthurZucker, \r\nI have just raised a PR for this issue, can you help me review it?\r\nhttps://github.com/huggingface/transformers/pull/28517\r\n\r\n\r\nPS, I close the old PR: `https://github.com/huggingface/transformers/pull/28513` because it didn't follow the Contributor Guideline" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28505/reactions", "total_count": 1, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 1, "rocket": 0, "eyes": 0 }

{ "login": "StarCycle", "id": 33491471, "node_id": "MDQ6VXNlcjMzNDkxNDcx", "avatar_url": "https://avatars.githubusercontent.com/u/33491471?v=4", "gravatar_id": "", "url": "https://api.github.com/users/StarCycle", "html_url": "https://github.com/StarCycle", "followers_url": "https://api.github.com/users/StarCycle/followers", "following_url": "https://api.github.com/users/StarCycle/following{/other_user}", "gists_url": "https://api.github.com/users/StarCycle/gists{/gist_id}", "starred_url": "https://api.github.com/users/StarCycle/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/StarCycle/subscriptions", "organizations_url": "https://api.github.com/users/StarCycle/orgs", "repos_url": "https://api.github.com/users/StarCycle/repos", "events_url": "https://api.github.com/users/StarCycle/events{/privacy}", "received_events_url": "https://api.github.com/users/StarCycle/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @StarCycle, thanks for this contribution! \r\n\r\nCould you share a small code snippet to show how you're training the model in the PR description and the error? This helps us for anyone coming back to this PR in the future.", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28504). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "I am using another training framework (xtuner), which is a little complicated. But you can easily reproduce the problem with the following code:\r\n\r\n```\r\nimport torch\r\nfrom transformers import AutoImageProcessor, AutoModel\r\nfrom PIL import Image\r\nimport requests\r\n\r\n# URL of the image you want to process\r\nurl = 'http://images.cocodataset.org/val2017/000000039769.jpg'\r\n\r\n# Open the image\r\nimage = Image.open(requests.get(url, stream=True).raw)\r\n\r\n# Load the image processor and the model\r\nprocessor = AutoImageProcessor.from_pretrained('facebook/dinov2-base')\r\nmodel = AutoModel.from_pretrained('facebook/dinov2-base', torch_dtype=torch.bfloat16)\r\n\r\n# Prepare the inputs\r\ninputs = processor(images=image, return_tensors=\"pt\")\r\n\r\n# Get the model outputs\r\noutputs = model(**inputs)\r\n```\r\nNow there will be an error:\r\n```\r\n[/usr/local/lib/python3.10/dist-packages/transformers/models/dinov2/modeling_dinov2.py](https://localhost:8080/#) in forward(self, pixel_values)\r\n 165 f\" Expected {self.num_channels} but got {num_channels}.\"\r\n 166 )\r\n--> 167 embeddings = self.projection(pixel_values).flatten(2).transpose(1, 2)\r\n 168 return embeddings\r\n 169 \r\n\r\n[/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py](https://localhost:8080/#) in _wrapped_call_impl(self, *args, **kwargs)\r\n 1516 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]\r\n 1517 else:\r\n-> 1518 return self._call_impl(*args, **kwargs)\r\n 1519 \r\n 1520 def _call_impl(self, *args, **kwargs):\r\n\r\n[/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py](https://localhost:8080/#) in _call_impl(self, *args, **kwargs)\r\n 1525 or _global_backward_pre_hooks or _global_backward_hooks\r\n 1526 or _global_forward_hooks or _global_forward_pre_hooks):\r\n-> 1527 return forward_call(*args, **kwargs)\r\n 1528 \r\n 1529 try:\r\n\r\n[/usr/local/lib/python3.10/dist-packages/torch/nn/modules/conv.py](https://localhost:8080/#) in forward(self, input)\r\n 458 \r\n 459 def forward(self, input: Tensor) -> Tensor:\r\n--> 460 return self._conv_forward(input, self.weight, self.bias)\r\n 461 \r\n 462 class Conv3d(_ConvNd):\r\n\r\n[/usr/local/lib/python3.10/dist-packages/torch/nn/modules/conv.py](https://localhost:8080/#) in _conv_forward(self, input, weight, bias)\r\n 454 weight, bias, self.stride,\r\n 455 _pair(0), self.dilation, self.groups)\r\n--> 456 return F.conv2d(input, weight, bias, self.stride,\r\n 457 self.padding, self.dilation, self.groups)\r\n 458 \r\n\r\nRuntimeError: Input type (float) and bias type (c10::BFloat16) should be the same\r\n```", "@StarCycle Thanks for sharing and again for this contribution! " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28504/reactions", "total_count": 6, "+1": 3, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 3, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28504", "html_url": "https://github.com/huggingface/transformers/pull/28504", "diff_url": "https://github.com/huggingface/transformers/pull/28504.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28504.patch", "merged_at": 1705518188000 }

{ "login": "hiroshi-matsuda-rit", "id": 40782025, "node_id": "MDQ6VXNlcjQwNzgyMDI1", "avatar_url": "https://avatars.githubusercontent.com/u/40782025?v=4", "gravatar_id": "", "url": "https://api.github.com/users/hiroshi-matsuda-rit", "html_url": "https://github.com/hiroshi-matsuda-rit", "followers_url": "https://api.github.com/users/hiroshi-matsuda-rit/followers", "following_url": "https://api.github.com/users/hiroshi-matsuda-rit/following{/other_user}", "gists_url": "https://api.github.com/users/hiroshi-matsuda-rit/gists{/gist_id}", "starred_url": "https://api.github.com/users/hiroshi-matsuda-rit/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/hiroshi-matsuda-rit/subscriptions", "organizations_url": "https://api.github.com/users/hiroshi-matsuda-rit/orgs", "repos_url": "https://api.github.com/users/hiroshi-matsuda-rit/repos", "events_url": "https://api.github.com/users/hiroshi-matsuda-rit/events{/privacy}", "received_events_url": "https://api.github.com/users/hiroshi-matsuda-rit/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Example model for sudachi_projection option:\r\nhttps://huggingface.co./hiroshi-matsuda-rit/bert-base-sudachitra-v11/blob/main/tokenizer_config.json#L16", "Nice 🔥 Feel free to ping me for a review! ", "The errors which were raised from the test cases for jumanpp and mecab-unidic in local environment were not reproduced in circle-ci test environment, so I removed following section from the top description of this PR.\r\n\r\n> As a side note, some errors have been reported from the test cases for mecab-unidic and jumanpp when turning on `RUN_CUSTOM_TOKENIZERS`.\r\n> I would also like to update these outdated test cases in this PR if required.\r\n```console\r\n$ RUN_CUSTOM_TOKENIZERS=1 pytest -rfs tests/models/bert_japanese/\r\n...\r\n======================================================================================= short test summary info =======================================================================================\r\nFAILED tests/models/bert_japanese/test_tokenization_bert_japanese.py::BertJapaneseTokenizationTest::test_jumanpp_tokenizer - AssertionError: Lists differ: ['アップ[28 chars]', '\\\\ ', 'が', '\\\\ \\\\ ', '\\\\ ', '発売', 'さ', 'れた', '\\\\ ', '。'] != ['アップ[28 chars]', '\\u3000', 'が', '\\u3000', '\\u3000', '\\u3000'[28 chars]...\r\nFAILED tests/models/bert_japanese/test_tokenization_bert_japanese.py::BertJapaneseTokenizationTest::test_jumanpp_tokenizer_ext - AssertionError: Lists differ: ['ありがとう', 'ございます', 'm', '(', '_', '\\\\ _', ')', 'm', '見つける', 'の[15 chars] '。'] != ['ありがとう', 'ございます', 'm(_ _)m', '見つける', 'の', 'が', '大...\r\nFAILED tests/models/bert_japanese/test_tokenization_bert_japanese.py::BertJapaneseTokenizationTest::test_jumanpp_tokenizer_lower - AssertionError: Lists differ: ['アップ[28 chars]', '\\\\ ', 'が', '\\\\ \\\\ ', '\\\\ ', '発売', 'さ', 'れた', '\\\\ ', '。'] != ['アップ[28 chars]', '\\u3000', 'が', '\\u3000', '\\u3000', '\\u3000'[28 chars]...\r\nFAILED tests/models/bert_japanese/test_tokenization_bert_japanese.py::BertJapaneseTokenizationTest::test_jumanpp_tokenizer_no_normalize - AssertionError: Lists differ: ['ｱ',[45 chars]', '\\\\ ', 'が', '\\\\ \\\\ ', '\\\\ ', '発売', 'さ', 'れた', '\\u3000', '。'] != ['ｱ',[45 chars]', '\\u3000', 'が', '\\u3000', '\\u3000', '\\u3000'[28 chars] '。']\r\nFAILED tests/models/bert_japanese/test_tokenization_bert_japanese.py::BertJapaneseTokenizationTest::test_jumanpp_tokenizer_trim_whitespace - AssertionError: Lists differ: ['アップ[21 chars]ne', '8', '\\\\', 'が', '\\\\ \\\\', '\\\\', '発売', 'さ', 'れた', '\\\\', '。'] != ['アップ[21 chars]ne', '8', 'が', '発売', 'さ', 'れた', '。']\r\nFAILED tests/models/bert_japanese/test_tokenization_bert_japanese.py::BertJapaneseTokenizationTest::test_mecab_tokenizer_unidic - RuntimeError: The unidic dictionary itself is not found. See https://github.com/polm/unidic-py for installation.\r\nSKIPPED [2] tests/test_tokenization_common.py:2394: test is PT+TF test\r\nSKIPPED [2] tests/test_tokenization_common.py:2534: test is slow\r\nSKIPPED [2] tests/test_tokenization_common.py:2499: test requires TensorFlow\r\nSKIPPED [2] tests/test_tokenization_common.py:2448: test is slow\r\n======================================================================== 6 failed, 193 passed, 8 skipped, 7 warnings in 4.52s =========================================================================\r\n```", "@ArthurZucker All tests passed! Please review and merge this PR.", "> Let's add a small test where you pass sudachi_kwargs to a BertJapaneseTokenizer to make sure this is usable\r\n\r\nI added some test cases which pass kwargs to word tokenizers.", "> I think we need a sudachi version check for this feature then!\r\n\r\nThanks for your suggestion. Let me explain the background.\r\n\r\nI've been developing a Japanese Dependency parser GiNZA utilizing spacy-transformers and sudachipy, and its [ELECTRA model](https://huggingface.co./megagonlabs/transformers-ud-japanese-electra-base-ginza-520/tree/main) has become widely used for Japanese text analysis in recent years.\r\nI'm wondering that certain users of GiNZA ELECTRA model may hope to keep sudachipy as is but they need to upgrade transformers to use newer models together with GiNZA.\r\n\r\nThis is the reason why I wrote the if else clause.", "@ArthurZucker Finally, I decided to pass projection arg to Dictionary.create() directly.\r\nI would like to inform GiNZA users to update transformers and sudachipy at the same time.", "@ArthurZucker What should I do for this type of errors? It seems I do not have a permission to rerun the failed actions.\r\n\r\n> FAILED tests/test_pipeline_mixin.py::ImageSegmentationPipelineTests::test_small_model_pt - requests.exceptions.ConnectionError: ('Connection aborted.', ConnectionResetError(104, 'Connection reset by peer'))", "Yes if else is alright don't worry! Just wanted to make sure someone that tries to use the new feature with an older version would get an error saying he needs to update the sudachi version to use it! ", "I implemented `is_sudachi_projection_available()` and `@require_sudachi_projection` to check the sudachipy version.\r\nDoes this additional implementation meet your requirements? @ArthurZucker ", "Reviewing! ", "@ArthurZucker Thanks! I just revised the changes as per your instructions.", "@ArthurZucker Thank you. I have completed the fix.", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28503). 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/28503/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/28503", "html_url": "https://github.com/huggingface/transformers/pull/28503", "diff_url": "https://github.com/huggingface/transformers/pull/28503.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28503.patch", "merged_at": 1707796040000 }

{ "login": "hk6an6", "id": 2327624, "node_id": "MDQ6VXNlcjIzMjc2MjQ=", "avatar_url": "https://avatars.githubusercontent.com/u/2327624?v=4", "gravatar_id": "", "url": "https://api.github.com/users/hk6an6", "html_url": "https://github.com/hk6an6", "followers_url": "https://api.github.com/users/hk6an6/followers", "following_url": "https://api.github.com/users/hk6an6/following{/other_user}", "gists_url": "https://api.github.com/users/hk6an6/gists{/gist_id}", "starred_url": "https://api.github.com/users/hk6an6/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/hk6an6/subscriptions", "organizations_url": "https://api.github.com/users/hk6an6/orgs", "repos_url": "https://api.github.com/users/hk6an6/repos", "events_url": "https://api.github.com/users/hk6an6/events{/privacy}", "received_events_url": "https://api.github.com/users/hk6an6/received_events", "type": "User", "site_admin": false }

[]

[]

[ "It seems the problem goes away if I change from `paraphrase-distilroberta-base-v1` to `paraphrase-distilroberta-base-v2`. Could it be that the tokenizer_config for v1 is incompatible with json.dumps?", "Hey! This seems to be related to `sentence-transfromers`, as tokenizers are serializable. \r\nCould you provide the stactrace and open an issue on the [sentence-transformers](https://github.com/UKPLab/sentence-transformers) repo? ", "Hi Arthur!\r\n\r\nYes, I can provide a stack trace.\r\n\r\nI’ll close this issue and open a new issue on the sentence-transformers\r\nrepo.\r\n\r\nOn Mon, Jan 15, 2024 at 02:32 Arthur ***@***.***> wrote:\r\n\r\n> Hey! This seems to be related to sentence-transfromers, as tokenizers are\r\n> serializable.\r\n> Could you provide the stactrace and open an issue on the\r\n> sentence-transformers <https://github.com/UKPLab/sentence-transformers>\r\n> repo?\r\n>\r\n> —\r\n> Reply to this email directly, view it on GitHub\r\n> <https://github.com/huggingface/transformers/issues/28502#issuecomment-1891827117>,\r\n> or unsubscribe\r\n> <https://github.com/notifications/unsubscribe-auth/AARYISEUDT4D3TRGGZYR7GLYOUAUDAVCNFSM6AAAAABB2R4YT6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOJRHAZDOMJRG4>\r\n> .\r\n> You are receiving this because you authored the thread.Message ID:\r\n> ***@***.***>\r\n>\r\n", "[stacktrace.txt](https://github.com/huggingface/transformers/files/13959334/stacktrace.txt)\r\n\r\nKeeping my word, this is the stack trace. I'll cose this ticket and open one in the sentence-transformers repo.", "https://github.com/UKPLab/sentence-transformers/issues/2418" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28502/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "mzbac", "id": 7523197, "node_id": "MDQ6VXNlcjc1MjMxOTc=", "avatar_url": "https://avatars.githubusercontent.com/u/7523197?v=4", "gravatar_id": "", "url": "https://api.github.com/users/mzbac", "html_url": "https://github.com/mzbac", "followers_url": "https://api.github.com/users/mzbac/followers", "following_url": "https://api.github.com/users/mzbac/following{/other_user}", "gists_url": "https://api.github.com/users/mzbac/gists{/gist_id}", "starred_url": "https://api.github.com/users/mzbac/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/mzbac/subscriptions", "organizations_url": "https://api.github.com/users/mzbac/orgs", "repos_url": "https://api.github.com/users/mzbac/repos", "events_url": "https://api.github.com/users/mzbac/events{/privacy}", "received_events_url": "https://api.github.com/users/mzbac/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hey, `AutoTokenizer.from_pretrained('Qwen/Qwen-1_8B', trust_remote_code = True)` should be used as this model is not natively supported", "Otherwise, yes we should prompt the user for that cc @Rocketknight1 if you can have a look! ", "> Hey, `AutoTokenizer.from_pretrained('Qwen/Qwen-1_8B', trust_remote_code = True)` should be used as this model is not natively supported\r\n\r\n@ArthurZucker Thanks for the reply. Yeah, explicitly passing trust_remote_code works. However, it should prompt the user in terminal to ask whether they want to trust the remote code if they didn't explicitly pass trust_remote_code. As I understand it, it should behave similarly to AutoModelForCausalLM.from_pretrained.", "Yes agree with you! ", "Hi @mzbac, I made a PR last week that specifically deals with this issue here: #28419.\r\n\r\nCan you install `transformers` from `main` with `pip install --upgrade git+https://github.com/huggingface/transformers.git` and then try just running `AutoTokenizer.from_pretrained('Qwen/Qwen-1_8B')` again, without needing to edit any repo files? It should work properly now!", "Thanks, @Rocketknight1. I can confirm that using the main branch's transformers resolves the issue.", "Cool! That fix will be included in our next release. I'm going to close this issue for now, but feel free to reopen it if you encounter any other issues with loading custom code tokenizers." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28501/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "akwako", "id": 31602350, "node_id": "MDQ6VXNlcjMxNjAyMzUw", "avatar_url": "https://avatars.githubusercontent.com/u/31602350?v=4", "gravatar_id": "", "url": "https://api.github.com/users/akwako", "html_url": "https://github.com/akwako", "followers_url": "https://api.github.com/users/akwako/followers", "following_url": "https://api.github.com/users/akwako/following{/other_user}", "gists_url": "https://api.github.com/users/akwako/gists{/gist_id}", "starred_url": "https://api.github.com/users/akwako/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/akwako/subscriptions", "organizations_url": "https://api.github.com/users/akwako/orgs", "repos_url": "https://api.github.com/users/akwako/repos", "events_url": "https://api.github.com/users/akwako/events{/privacy}", "received_events_url": "https://api.github.com/users/akwako/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @amyeroberts, \r\n\r\nThank you for your reply and feedback!\r\n\r\nAs far as I can tell, the main issue is when `max_steps` < `args.save_steps`, and only when `save_strategy=\"steps\"`. Since `max_steps` is computed in the `_inner_training_loop`, I have added a check there to make sure training fails fast, as you suggested. However, in case this is happening in other ways, it might be good to keep the warning in the log, too, as per the original PR. What do you think? ", "Possibly related to #27332 ", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28500). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "Hi @amyeroberts, \r\n\r\nThanks for your suggestions! \r\n\r\nI'm looking into [#27332](https://github.com/huggingface/transformers/issues/27332). You may be right that these are related issues. When I run `trainer` with `EarlyStoppingCallback`, I do get the (newly added) message: `ValueError: args.save_steps must be less than max_steps...` I will look further into `EarlyStoppingCallback` to confirm.\r\n\r\nUpdate: In order to determine if the issue [#27332](https://github.com/huggingface/transformers/issues/27332) is due to max_steps < args.save_steps, I think we need to know what the other training arguments are, and how large the training sample is. ", "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/28500/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/28500", "html_url": "https://github.com/huggingface/transformers/pull/28500", "diff_url": "https://github.com/huggingface/transformers/pull/28500.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28500.patch", "merged_at": null }

{ "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 }

[]

[]

[ "Hi @getao, thanks for raising an issue! \r\n\r\nCould you provide a minimal reproducer for this error. Specifically the full CLI command used to launch the training job? \r\n\r\ncc @muellerzr @pacman100 ", "Sure.\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\r\n data_prefix = data_args.data_path\r\n \r\n train_file = f\"{data_prefix}.train.json\" # text data for language modeling (the next word prediction task)\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, \"add_special_tokens\": data_args.add_special_tokens}, load_from_cache_file=True)\r\n eval_dataset = eval_dataset.map(tokenize_function, batched=True, fn_kwargs={\"tokenizer\": tokenizer, \"max_seq_length\": data_args.max_seq_length, \"add_special_tokens\": data_args.add_special_tokens, \"dev\": True}, load_from_cache_file=True) \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) # Llama-2\r\n model_download_flag = True\r\n except Exception as e:\r\n print(e)\r\n\r\n train_model(model, train_dataset, eval_dataset, training_args, data_collator)\r\n\r\nmain()\r\n```\r\n\r\nThe CLI looks as follows -- nothing special except 2 --fsdp related flags:\r\n\r\n```\r\ntorchrun --nproc-per-node=6 train_script.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 --fsdp shard_grad_op --fsdp_config my_fsdp_config_path \\\r\n --learning_rate $lr --warmup_steps $warmup --max_grad_norm 1.0 --seed $seed --data_seed $seed --logging_steps 10 --save_strategy 'no' --evaluation_strategy 'steps' --eval_steps $eval_steps \\\r\n --save_steps $save_steps --bf16 --output_dir $OUT_DIR --logging_dir $OUT_DIR --data_path $INPUT_DATA | tee $OUT_DIR/train.log\r\n```" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28499/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "qmeeus", "id": 25608944, "node_id": "MDQ6VXNlcjI1NjA4OTQ0", "avatar_url": "https://avatars.githubusercontent.com/u/25608944?v=4", "gravatar_id": "", "url": "https://api.github.com/users/qmeeus", "html_url": "https://github.com/qmeeus", "followers_url": "https://api.github.com/users/qmeeus/followers", "following_url": "https://api.github.com/users/qmeeus/following{/other_user}", "gists_url": "https://api.github.com/users/qmeeus/gists{/gist_id}", "starred_url": "https://api.github.com/users/qmeeus/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/qmeeus/subscriptions", "organizations_url": "https://api.github.com/users/qmeeus/orgs", "repos_url": "https://api.github.com/users/qmeeus/repos", "events_url": "https://api.github.com/users/qmeeus/events{/privacy}", "received_events_url": "https://api.github.com/users/qmeeus/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @muellerzr ", "Thanks for the suggestions @amyeroberts ! I included them in the PR and synced the branch to the latest commits\r\n\r\nHave a nice day\r\n" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28498/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/28498", "html_url": "https://github.com/huggingface/transformers/pull/28498", "diff_url": "https://github.com/huggingface/transformers/pull/28498.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28498.patch", "merged_at": 1706022498000 }

{ "login": "HamzaFB", "id": 24733081, "node_id": "MDQ6VXNlcjI0NzMzMDgx", "avatar_url": "https://avatars.githubusercontent.com/u/24733081?v=4", "gravatar_id": "", "url": "https://api.github.com/users/HamzaFB", "html_url": "https://github.com/HamzaFB", "followers_url": "https://api.github.com/users/HamzaFB/followers", "following_url": "https://api.github.com/users/HamzaFB/following{/other_user}", "gists_url": "https://api.github.com/users/HamzaFB/gists{/gist_id}", "starred_url": "https://api.github.com/users/HamzaFB/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/HamzaFB/subscriptions", "organizations_url": "https://api.github.com/users/HamzaFB/orgs", "repos_url": "https://api.github.com/users/HamzaFB/repos", "events_url": "https://api.github.com/users/HamzaFB/events{/privacy}", "received_events_url": "https://api.github.com/users/HamzaFB/received_events", "type": "User", "site_admin": false }

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28497). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "Thanks!" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28497/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/28497", "html_url": "https://github.com/huggingface/transformers/pull/28497", "diff_url": "https://github.com/huggingface/transformers/pull/28497.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28497.patch", "merged_at": 1705401027000 }

{ "login": "qxpBlog", "id": 96739096, "node_id": "U_kgDOBcQfGA", "avatar_url": "https://avatars.githubusercontent.com/u/96739096?v=4", "gravatar_id": "", "url": "https://api.github.com/users/qxpBlog", "html_url": "https://github.com/qxpBlog", "followers_url": "https://api.github.com/users/qxpBlog/followers", "following_url": "https://api.github.com/users/qxpBlog/following{/other_user}", "gists_url": "https://api.github.com/users/qxpBlog/gists{/gist_id}", "starred_url": "https://api.github.com/users/qxpBlog/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/qxpBlog/subscriptions", "organizations_url": "https://api.github.com/users/qxpBlog/orgs", "repos_url": "https://api.github.com/users/qxpBlog/repos", "events_url": "https://api.github.com/users/qxpBlog/events{/privacy}", "received_events_url": "https://api.github.com/users/qxpBlog/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @qxpBlog, thanks for raising this issue! \r\n\r\nCould you share a minimal code reproducer for this error and a full traceback of the error? ", "@amyeroberts this question appear on the eighth line(i think that this version of transformers exclude the module):\r\nimport os\r\nimport sys\r\nimport argparse\r\nimport torch\r\nimport numpy as np\r\nimport transformers\r\nfrom transformers import GenerationConfig, LlamaForCausalLM, LlamaTokenizer\r\n**from transformers.activations import SiLUActivation**\r\n\r\nfrom ptflops import get_model_complexity_info\r\nfrom ptflops.pytorch_ops import bn_flops_counter_hook, pool_flops_counter_hook\r\n\r\nfrom LLMPruner.models.hf_llama.modeling_llama import LlamaForCausalLM, LlamaRMSNorm, LlamaAttention, LlamaMLP\r\nfrom LLMPruner.peft import PeftModel\r\n\r\nif torch.cuda.is_available():\r\n device = \"cuda\"\r\nelse:\r\n device = \"cpu\"\r\ntorch_version = int(torch.__version__.split('.')[1])\r\n\r\ndef LlamaAttention_counter_hook(module, input, output):\r\n # (1) Ignore past-key values\r\n # (2) Assume there is no attention mask\r\n # Input will be empty in some pytorch version. use output here since input.shape == output.shape\r\n flops = 0\r\n q_len = output[0].shape[1]\r\n linear_dim = output[0].shape[-1]\r\n num_heads = module.num_heads\r\n head_dim = module.head_dim\r\n\r\n rotary_flops = 2 * (q_len * num_heads * head_dim) * 2\r\n attention_flops = num_heads * (q_len * q_len * head_dim + q_len * q_len + q_len * q_len * head_dim) #QK^T + softmax + AttentionV\r\n linear_flops = 4 * (q_len * linear_dim * num_heads * head_dim) # 4 for q, k, v, o. \r\n flops += rotary_flops + attention_flops + linear_flops\r\n module.__flops__ += int(flops)\r\n\r\ndef rmsnorm_flops_counter_hook(module, input, output):\r\n input = input[0]\r\n\r\n batch_flops = np.prod(input.shape)\r\n batch_flops *= 2\r\n module.__flops__ += int(batch_flops)\r\n\r\ndef main(args):\r\n if args.model_type == 'pretrain':\r\n tokenizer = LlamaTokenizer.from_pretrained(args.base_model)\r\n model = LlamaForCausalLM.from_pretrained(\r\n args.base_model,\r\n low_cpu_mem_usage=True if torch_version >=9 else False\r\n )\r\n elif args.model_type == 'pruneLLM':\r\n pruned_dict = torch.load(args.ckpt, map_location='cpu')\r\n tokenizer, model = pruned_dict['tokenizer'], pruned_dict['model']\r\n else:\r\n raise NotImplementedError\r\n\r\n def input_constructor(x):\r\n return {'input_ids': torch.ones(x).long().to(device)}\r\n\r\n if device == \"cuda\":\r\n model.half()\r\n model = model.cuda()\r\n \r\n with torch.cuda.device(0):\r\n macs, params = get_model_complexity_info(model, (1, 64,), as_strings=True,\r\n input_constructor = input_constructor,\r\n print_per_layer_stat=True, verbose=True,\r\n custom_modules_hooks={\r\n LlamaAttention: LlamaAttention_counter_hook,\r\n LlamaRMSNorm: rmsnorm_flops_counter_hook,\r\n SiLUActivation: pool_flops_counter_hook,\r\n },)\r\n else:\r\n model.float()\r\n macs, params = get_model_complexity_info(model, (1, 64,), as_strings=True,\r\n input_constructor = input_constructor,\r\n print_per_layer_stat=True, verbose=True,\r\n custom_modules_hooks={\r\n LlamaAttention: LlamaAttention_counter_hook,\r\n LlamaRMSNorm: rmsnorm_flops_counter_hook,\r\n SiLUActivation: pool_flops_counter_hook,\r\n },)\r\n\r\n print('{:<30} {:<8}'.format('Computational complexity: ', macs))\r\n print('{:<30} {:<8}'.format('Number of parameters: ', params))\r\n print(\"GPU Memory Requirement: {} MiB\\n\".format(torch.cuda.memory_allocated()/1024/1024))\r\n\r\n\r\nif __name__ == \"__main__\":\r\n parser = argparse.ArgumentParser(description='Tuning Pruned LLaMA (huggingface version)')\r\n\r\n parser.add_argument('--base_model', type=str, default=\"decapoda-research/llama-7b-hf\", help='base model name')\r\n parser.add_argument('--model_type', type=str, required=True, help = 'choose from [pretrain, pruneLLM]')\r\n parser.add_argument('--ckpt', type=str, default=None)\r\n parser.add_argument('--lora_ckpt', type=str, default=None)\r\n \r\n args = parser.parse_args()\r\n main(args)\r\n\r\n", "@qxpBlog OK. So the custom silu acitvation class was removed and so can't be imported from `transformers.activations`. To fix this script, you can use `nn.SiLU` instead. \r\n\r\nAs the class wasn't part of the public API - not importable from the top level of transformers and not documented - we don't guarantee that it'll never be moved, renamed or deleted. We recognise that this can still cause unexpected errors for our users however! I've open a PR to add a dummy class so this goes through a proper deprecation cycle first. ", "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/28496/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "4imothy", "id": 40186632, "node_id": "MDQ6VXNlcjQwMTg2NjMy", "avatar_url": "https://avatars.githubusercontent.com/u/40186632?v=4", "gravatar_id": "", "url": "https://api.github.com/users/4imothy", "html_url": "https://github.com/4imothy", "followers_url": "https://api.github.com/users/4imothy/followers", "following_url": "https://api.github.com/users/4imothy/following{/other_user}", "gists_url": "https://api.github.com/users/4imothy/gists{/gist_id}", "starred_url": "https://api.github.com/users/4imothy/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/4imothy/subscriptions", "organizations_url": "https://api.github.com/users/4imothy/orgs", "repos_url": "https://api.github.com/users/4imothy/repos", "events_url": "https://api.github.com/users/4imothy/events{/privacy}", "received_events_url": "https://api.github.com/users/4imothy/received_events", "type": "User", "site_admin": false }

[]

[]

[ "For the failing tests, you'll need to update the expected strings. " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28495/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/28495", "html_url": "https://github.com/huggingface/transformers/pull/28495", "diff_url": "https://github.com/huggingface/transformers/pull/28495.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28495.patch", "merged_at": 1705343800000 }

{ "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/28494/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/28494", "html_url": "https://github.com/huggingface/transformers/pull/28494", "diff_url": "https://github.com/huggingface/transformers/pull/28494.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28494.patch", "merged_at": 1705338248000 }

{ "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/28493/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/28493", "html_url": "https://github.com/huggingface/transformers/pull/28493", "diff_url": "https://github.com/huggingface/transformers/pull/28493.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28493.patch", "merged_at": 1705177885000 }

{ "login": "bayllama", "id": 142558246, "node_id": "U_kgDOCH9EJg", "avatar_url": "https://avatars.githubusercontent.com/u/142558246?v=4", "gravatar_id": "", "url": "https://api.github.com/users/bayllama", "html_url": "https://github.com/bayllama", "followers_url": "https://api.github.com/users/bayllama/followers", "following_url": "https://api.github.com/users/bayllama/following{/other_user}", "gists_url": "https://api.github.com/users/bayllama/gists{/gist_id}", "starred_url": "https://api.github.com/users/bayllama/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/bayllama/subscriptions", "organizations_url": "https://api.github.com/users/bayllama/orgs", "repos_url": "https://api.github.com/users/bayllama/repos", "events_url": "https://api.github.com/users/bayllama/events{/privacy}", "received_events_url": "https://api.github.com/users/bayllama/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @ArthurZucker ", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28492). 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/28492/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/28492", "html_url": "https://github.com/huggingface/transformers/pull/28492", "diff_url": "https://github.com/huggingface/transformers/pull/28492.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28492.patch", "merged_at": null }

{ "login": "qlwang25", "id": 38132016, "node_id": "MDQ6VXNlcjM4MTMyMDE2", "avatar_url": "https://avatars.githubusercontent.com/u/38132016?v=4", "gravatar_id": "", "url": "https://api.github.com/users/qlwang25", "html_url": "https://github.com/qlwang25", "followers_url": "https://api.github.com/users/qlwang25/followers", "following_url": "https://api.github.com/users/qlwang25/following{/other_user}", "gists_url": "https://api.github.com/users/qlwang25/gists{/gist_id}", "starred_url": "https://api.github.com/users/qlwang25/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/qlwang25/subscriptions", "organizations_url": "https://api.github.com/users/qlwang25/orgs", "repos_url": "https://api.github.com/users/qlwang25/repos", "events_url": "https://api.github.com/users/qlwang25/events{/privacy}", "received_events_url": "https://api.github.com/users/qlwang25/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @qlwang25 👋 \r\n\r\nHave a look at [this comment](https://github.com/huggingface/transformers/issues/25420#issuecomment-1775317535), which I believe answers your question :) ", "Thank you very much for your reply. \r\nThe reason is ```torch type``` instead of ```padding```. Just changing it to ```torch.bfloat16``` (or ```torch.float32```) works." ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28491/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "DjangoPeng", "id": 16943353, "node_id": "MDQ6VXNlcjE2OTQzMzUz", "avatar_url": "https://avatars.githubusercontent.com/u/16943353?v=4", "gravatar_id": "", "url": "https://api.github.com/users/DjangoPeng", "html_url": "https://github.com/DjangoPeng", "followers_url": "https://api.github.com/users/DjangoPeng/followers", "following_url": "https://api.github.com/users/DjangoPeng/following{/other_user}", "gists_url": "https://api.github.com/users/DjangoPeng/gists{/gist_id}", "starred_url": "https://api.github.com/users/DjangoPeng/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/DjangoPeng/subscriptions", "organizations_url": "https://api.github.com/users/DjangoPeng/orgs", "repos_url": "https://api.github.com/users/DjangoPeng/repos", "events_url": "https://api.github.com/users/DjangoPeng/events{/privacy}", "received_events_url": "https://api.github.com/users/DjangoPeng/received_events", "type": "User", "site_admin": false }

[]

[]

[ "cc @SunMarc - let me know if you want me to have a look at this one", "I had the similar issue too. It would be great if someone can take a look at this.\r\n\r\nThere is a builder error on \"allenai--c4\" when I use the following code snippet:\r\n\r\n```python\r\n\r\nraw_datasets = load_dataset(\r\n \"allenai/c4\",\r\n \"allenai--c4\",\r\n data_files={\r\n \"validation\": \"en/c4-validation.00000-of-00008.json.gz\",\r\n },\r\n )\r\n\r\n```\r\n\r\nI found commenting out `allenai--c4` can solve the issue, but the building time seems become much longer than before.", "Hi @DjangoPeng, this is due to a breaking change in datasets library. See related [thread](https://huggingface.co./datasets/allenai/c4/discussions/7).The issue will be fixed after this [PR](https://github.com/huggingface/optimum/pull/1646) is merged (need to install optimum from source) and I recommend you installing the latest version of datasets. The building time should be faster with it. In the meantime, I am switching the the dataset `c4` -> `wikitext2` in the notebook since I don't want users to be blocked by this. ", "SGTM.\r\n\r\nThanks for your reply. :)", "Thanks a lot @SunMarc ! 🎉 ", "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 hit an issue( `... can not find xxx on aws3`) when I use `wikitext2 `. Here, I'd like to append more options for the datasets: [‘wikitext2’,‘c4’,‘c4-new’,‘ptb’,‘ptb-new’]. All these datasets are used in GPTQ paper. So, it is ok to use it. I change to `c4-new` and it works well. \r\n\r\nSource:\r\nhttps://huggingface.co./docs/transformers/v4.33.0/en/main_classes/quantization#transformers.GPTQConfig.dataset" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28490/reactions", "total_count": 0, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "login": "rishit5", "id": 24509842, "node_id": "MDQ6VXNlcjI0NTA5ODQy", "avatar_url": "https://avatars.githubusercontent.com/u/24509842?v=4", "gravatar_id": "", "url": "https://api.github.com/users/rishit5", "html_url": "https://github.com/rishit5", "followers_url": "https://api.github.com/users/rishit5/followers", "following_url": "https://api.github.com/users/rishit5/following{/other_user}", "gists_url": "https://api.github.com/users/rishit5/gists{/gist_id}", "starred_url": "https://api.github.com/users/rishit5/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/rishit5/subscriptions", "organizations_url": "https://api.github.com/users/rishit5/orgs", "repos_url": "https://api.github.com/users/rishit5/repos", "events_url": "https://api.github.com/users/rishit5/events{/privacy}", "received_events_url": "https://api.github.com/users/rishit5/received_events", "type": "User", "site_admin": false }

[]

[]

[]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28489/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/28489", "html_url": "https://github.com/huggingface/transformers/pull/28489", "diff_url": "https://github.com/huggingface/transformers/pull/28489.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28489.patch", "merged_at": 1705337115000 }

{ "login": "abacaj", "id": 7272343, "node_id": "MDQ6VXNlcjcyNzIzNDM=", "avatar_url": "https://avatars.githubusercontent.com/u/7272343?v=4", "gravatar_id": "", "url": "https://api.github.com/users/abacaj", "html_url": "https://github.com/abacaj", "followers_url": "https://api.github.com/users/abacaj/followers", "following_url": "https://api.github.com/users/abacaj/following{/other_user}", "gists_url": "https://api.github.com/users/abacaj/gists{/gist_id}", "starred_url": "https://api.github.com/users/abacaj/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/abacaj/subscriptions", "organizations_url": "https://api.github.com/users/abacaj/orgs", "repos_url": "https://api.github.com/users/abacaj/repos", "events_url": "https://api.github.com/users/abacaj/events{/privacy}", "received_events_url": "https://api.github.com/users/abacaj/received_events", "type": "User", "site_admin": false }

[]

[]

[ "I experienced the same thing! Over 3 epochs same set up just updated code and flash attention, the loss went from 6 to 2. And on the old code without flash attention it was .60 to ~.29 . Very strange.", "cc @younesbelkada @ArthurZucker ", "Hi @abacaj, as per @pacman100 guidelines in https://github.com/huggingface/transformers/pull/28142 / https://github.com/huggingface/transformers/pull/28142#issuecomment-1869513914 you need to make sure to load your model in full-precision and train with autocast (bf16=True). Also can you share more insights on how you train your model? (do you load the model in bf16/fp16, do you use PEFT, packing, etc.) ?", "Hi @younesbelkada, this is a full fine tune using HF trainer. Padding only. Model is loaded in bf16. I try loading in \"fp32\" but get error:\r\n\r\n```\r\nValueError: Flash Attention 2.0 only supports torch.float16 and torch.bfloat16 dtypes. You passed torch.float32, this might lead to unexpected behaviour.\r\n```\r\n\r\n```python\r\n model = AutoModelForCausalLM.from_pretrained(\r\n model_args.model_name_or_path,\r\n trust_remote_code=True,\r\n config=config,\r\n attn_implementation=\"flash_attention_2\",\r\n torch_dtype=torch.float32,\r\n cache_dir=training_args.cache_dir,\r\n )\r\n```\r\n", "Ok thanks @abacaj for getting back ! I think you get that error because the patch #28142 has not been released on pypi - can you try to build transformers from source? \r\n```bash\r\npip install -U git+https://github.com/huggingface/transformers.git\r\n```\r\nThat should hopefully solve it, let me know if you face into more issues!", "Ok so I remove the explicit `torch_dtype` following the comments in your link. The loss is still very high with flash-attn-2 using [phi-2](https://huggingface.co./microsoft/phi-2) model", "@abacaj which padding side are you using for training?", "I use `padding_side=\"left\"`. Here is how the loss goes with and without FA2 (green line has FA2) using phi-2:\r\n\r\n\r\n", "FWIW changing padding side doesn't do anything to the loss, it's the same", "I see, as a sanity check, can you share your `TrainingArguments` ?", "```\r\nadafactor=False,\r\nadam_beta1=0.9,\r\nadam_beta2=0.95,\r\nadam_epsilon=1e-08,\r\nauto_find_batch_size=False,\r\nbf16=True,\r\nbf16_full_eval=False,\r\ncache_dir=None,\r\ndata_seed=None,\r\ndataloader_drop_last=False,\r\ndataloader_num_workers=0,\r\ndataloader_persistent_workers=False,\r\ndataloader_pin_memory=True,\r\nddp_backend=None,\r\nddp_broadcast_buffers=None,\r\nddp_bucket_cap_mb=None,\r\nddp_find_unused_parameters=None,\r\nddp_timeout=1800,\r\ndebug=[],\r\ndeepspeed=src/configs/deepspeed_2_config.json,\r\ndisable_tqdm=False,\r\ndispatch_batches=None,\r\ndo_eval=False,\r\ndo_predict=False,\r\ndo_train=True,\r\neval_accumulation_steps=None,\r\neval_delay=0,\r\neval_steps=0.0,\r\nevaluation_strategy=no,\r\nfp16=False,\r\nfp16_backend=auto,\r\nfp16_full_eval=False,\r\nfp16_opt_level=O1,\r\nfsdp=[],\r\nfsdp_config={'min_num_params': 0, 'xla': False, 'xla_fsdp_grad_ckpt': False},\r\nfsdp_min_num_params=0,\r\nfsdp_transformer_layer_cls_to_wrap=None,\r\nfull_determinism=False,\r\ngradient_accumulation_steps=2,\r\ngradient_checkpointing=True,\r\ngradient_checkpointing_kwargs=None,\r\ngreater_is_better=None,\r\ngroup_by_length=False,\r\nhalf_precision_backend=auto,\r\nhub_always_push=False,\r\nhub_model_id=None,\r\nhub_private_repo=False,\r\nhub_strategy=every_save,\r\nhub_token=<HUB_TOKEN>,\r\nignore_data_skip=False,\r\ninclude_inputs_for_metrics=False,\r\ninclude_num_input_tokens_seen=False,\r\ninclude_tokens_per_second=False,\r\ninference_length=2048,\r\njit_mode_eval=False,\r\nlabel_names=None,\r\nlabel_smoothing_factor=0.0,\r\nlearning_rate=5e-05,\r\nlength_column_name=length,\r\nload_best_model_at_end=False,\r\nlocal_rank=1,\r\nlog_level=passive,\r\nlog_level_replica=warning,\r\nlog_on_each_node=True,\r\nlogging_dir=checkpoints/results/2k-2k-dynamic-5e-5/runs/Jan15_13-42-15_sgpu,\r\nlogging_first_step=False,\r\nlogging_nan_inf_filter=True,\r\nlogging_steps=1.0,\r\nlogging_strategy=steps,\r\nlr_scheduler_kwargs={},\r\nlr_scheduler_type=cosine,\r\nmax_grad_norm=1.0,\r\nmax_steps=-1,\r\nmetric_for_best_model=None,\r\nmodel_max_position_embeddings=2048,\r\nmp_parameters=,\r\nneftune_noise_alpha=None,\r\nno_cuda=False,\r\nnum_train_epochs=3.0,\r\noptim=adamw_torch,\r\noptim_args=None,\r\noutput_dir=checkpoints/results/2k-2k-dynamic-5e-5,\r\noverwrite_output_dir=False,\r\npast_index=-1,\r\nper_device_eval_batch_size=4,\r\nper_device_train_batch_size=4,\r\nprediction_loss_only=False,\r\npush_to_hub=False,\r\npush_to_hub_model_id=None,\r\npush_to_hub_organization=None,\r\npush_to_hub_token=<PUSH_TO_HUB_TOKEN>,\r\nray_scope=last,\r\nremove_unused_columns=True,\r\nreport_to=['tensorboard'],\r\nresume_from_checkpoint=None,\r\nrope_scaling_factor=1.0,\r\nrope_scaling_type=dynamic,\r\nrun_name=checkpoints/results/2k-2k-dynamic-5e-5,\r\nsave_on_each_node=False,\r\nsave_only_model=False,\r\nsave_safetensors=True,\r\nsave_steps=100.0,\r\nsave_strategy=epoch,\r\nsave_total_limit=None,\r\nseed=70,\r\nskip_memory_metrics=True,\r\nsplit_batches=False,\r\ntf32=None,\r\ntorch_compile=False,\r\ntorch_compile_backend=None,\r\ntorch_compile_mode=None,\r\ntorchdynamo=None,\r\ntpu_metrics_debug=False,\r\ntpu_num_cores=None,\r\nuse_cpu=False,\r\nuse_ipex=False,\r\nuse_legacy_prediction_loop=False,\r\nuse_mps_device=False,\r\nwarmup_ratio=0.02,\r\nwarmup_steps=0,\r\nweight_decay=0.1\r\n```", "During my testing, I used bf16, trust remote code, no gradient ckpt, for SFT, with flshattn. The resulting model was terrible, I knew off the of (6 to 2)loss it wasn’t going to preform, but during testing it was worse than expected, very mangled answers. However when I trained the model, same arguments; just using the old phi repo code, and no flshattnt I got a great model. The loss from .6 to .29. Both were full fine tunes. Flash attention is critical for 24gb cards as without it it’s training off shared memory. I can help out more with testing when it’s done training in ~30 hours off shared mem 😭. The script I used is on #28381 . (Keep in mind the script doesn’t reflect me using bf16, however both times I trained the model I did have compute dtype set to bf16.)", "Hello everyone!\r\n\r\nCould you all please test using the latest revision on `microsoft/phi-2` and report the results? We might have found the issue.\r\n\r\nRegards,\r\nGustavo.", "FWIW - the model still comes out significantly worse using FA2. If anyone wants to fine-tune this model, I recommend you use it without FA2 currently. Running code benchmarks with FA2 < 50% on heval. Without FA2 (and all other hparams are identical, including seed) > 60% heval.", "The first graph is a comparison between using and not using flash attention 2. It seems that the loss doesn't change much with fa2 (yellowish curve).\r\n<img width=\"1037\" alt=\"截屏2024-01-18 16 43 57\" src=\"https://github.com/huggingface/transformers/assets/86560128/b2722683-6b1f-4cf0-88a0-f6928ceb3efd\">\r\n", "@abacaj could you please provide a minimal snippet to reproduce your fine-tuning?\r\n\r\nWe want to investigate it further more and attempt to find the root of the problem. We are doing a line-by-line comparison between the new model's code and the previous one.", "> FWIW - the model still comes out significantly worse using FA2. If anyone wants to fine-tune this model, I recommend you use it without FA2 currently. Running code benchmarks with FA2 < 50% on heval. Without FA2 (and all other hparams are identical, including seed) > 60% heval.\r\n\r\nI second this, just woke up and checked training after 3 epochs with FA2 I went from .54 to ~.40, meanwhile, with no FA2 I went from .60 to .30. Both full fine tunes. I’m gonna train the fa2 check point on an additional epoch to see if it gives me the same loss as with out FA2. Or to see if it over fits.\r\n\r\n\r\n\r\nEDIT:\r\nThe loss is off to a terrible start. Went as low as .39 then up to as high as .49.( It’s only at .07 of an epoch. But i’m training a check point that has trained on this exact dataset already for 3 epochs.) Significantly better than before with the soft max scaling issues, but there is still something up.\r\n\r\n\r\n\r\nThe loss is acting quite random, in comparison to no FA2. Where the loss consistently went down.\r\n\r\nSECOND UPDATE: I restarted the training with the same checkpoint and upped the learning rate by a order of 1, so from 2e5 to 2e6 and now the loss is more consistent, confusing why this hyper parameter differs in training when using fa2 and not using fa2.\r\n\r\n\r\nNot perfect but better.\r\n\r\nTHIRD UPDATE: I tried retraining the base model with fa2 and the loss isnt going anywhere. After 1.5 epochs. Its almost as if the weights aren’t being updated at all, and if so very marginally. Just consistently staying between .5 and .4 but random at every logging step.", "sorry to comment on this closed issue but I still have issues with FA2\r\n\r\n1. loss is different with FA2 compared to without\r\n2. loss is different even between two FA2 runs (used `set_seed`. doesn't happen without FA2 - loss always exactly the same)\r\n\r\n\r\n ", "I do another two runs using FA2 and without FA2 (blue line). Testing the models out using vLLM, the model without FA2 (blue line) scores 58% on humaneval, the model with FA2 scores 49%. I basically stopped using FA2 because the model comes out so much worse and I can't pinpoint why (the runs are identical with exception of use_flash_2)\r\n\r\n\r\n", "Hi @ArthurZucker,\r\nCould you reconsider opening this issue again? I think it’s worth opening, as training with flash attention on phi-2 is still not viable. The performance gains are almost essential though. I appreciate it thank you!", "Just wanted to acknowledge I have the same issue with using Fast Attention 2 with phi-2, the training loss hardly decreases with FA2 turned on, and works pretty well with it turned off.", "same question...", "> We want to investigate it further more and attempt to find the root of the problem. We are doing a line-by-line comparison between the new model's code and the previous one.\r\n\r\n@gugarosa is there any update on fixing FA2 for this amazing model? \r\n", "There is a PR that should fix it but is hard to merge #28673", "I am also seeing similar issue where loss is trending downwards but quite unstable and it seems to learn very slowly. I am running full fine-tuning of latest Phi2 model on my dataset.\r\n<img width=\"497\" alt=\"Screenshot 2024-02-12 at 9 22 57 PM\" src=\"https://github.com/huggingface/transformers/assets/5215386/26dec373-a4a3-4180-b86b-3c9a841ef042\">\r\n\r\n@ArthurZucker I just started another run after reinstalling transformers with changes from [#28673](https://github.com/huggingface/transformers/pull/28673) to see if it fixes the issue (still using FA2). will post loss curve in next few hours. \r\n\r\n**[Incorrect] Update-1**: loss curve after reinstalling transformers with changes from [#28673]. Looks like there is no change..\r\n\r\n\r\n**Update-2**: Looks like my new transformer installation didn't include changes from [#28673](https://github.com/huggingface/transformers/pull/28673) so essentially both plot should be same. I tried reinstalling transformers again with PR changes and now training is failing:\r\n\r\n```\r\n File \"/home/minimalist/miniconda3/envs/axolotl_Feb12/lib/python3.9/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/home/minimalist/miniconda3/envs/axolotl_Feb12/lib/python3.9/site-packages/torch/amp/autocast_mode.py\", line 16, in decorate_autocast\r\n return func(*args, **kwargs)\r\n File \"/home/minimalist/miniconda3/envs/axolotl_Feb12/lib/python3.9/site-packages/torch/amp/autocast_mode.py\", line 16, in decorate_autocast\r\n return func(*args, **kwargs)\r\n File \"/home/minimalist/work/projects/transformers/src/transformers/models/phi/modeling_phi.py\", line 318, in forward\r\n query_states = self.q_proj(hidden_states)\r\n File \"/home/minimalist/miniconda3/envs/axolotl_Feb12/lib/python3.9/site-packages/torch/nn/modules/module.py\", line 1518, in _wrapped_call_impl\r\n return self._call_impl(*args, **kwargs)\r\n File \"/home/minimalist/miniconda3/envs/axolotl_Feb12/lib/python3.9/site-packages/torch/nn/modules/module.py\", line 1527, in _call_impl\r\n return forward_call(*args, **kwargs)\r\n File \"/home/minimalist/miniconda3/envs/axolotl_Feb12/lib/python3.9/site-packages/torch/nn/modules/linear.py\", line 114, in forward\r\n return F.linear(input, self.weight, self.bias)\r\nRuntimeError: mat1 and mat2 must have the same dtype, but got Float and BFloat16\r\n```", "Update, it was a torch error, its training now, but the loss is the same as before, I lowered my dataset to 1k examples over 3 epochs with a lr of 2e6 and still the loss is random. Never consistently going down.", "How are you guys testing this? It does seem to matter when doing a full fine tune, and a lora fine tune. Using FA2 I could never get the loss to even be consistent with a full fine tune ( with SFT). Right now I am doing a DPO of phi2 with QLORA, and the loss is not only consistent, it’s consistently going down; from .69 to .27 at just a single epoch.\r\n\r\nI have not tried SFT with a lora, but maybe if we wanna use FA2 its just better to stick with using lora.", "hi there, now that SDPA has been merged #29108 you can use FA-2 through pytorch interface:\r\n\r\n0- Install pytorch 2.2\r\n1- make sure to load SDPA attention by passing `attn_implementation=\"sdpa\"` in `from_pretrained`\r\n2- Force-dispatch the SDPA kernel to use FA2 as follows:\r\n```diff\r\n- trainer.train()\r\n+ with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):\r\n+ trainer.train()\r\n```\r\nPerhaps this will fix all instability issues with respect to FA2 !", "@younesbelkada Hey! Thanks for your response!(before starting my training run) I got pytorch2.2, and I pulled the latest commits from transformers and installed from source. I’m using the DPO.py, from TRL, and I saw the commit, so I tried to pass “—attn_implementation SPDA” but it gave me a SPDA not currently supported error, I wish I still had the error up, I’ll try it out again, once my training run ends in a little less than an hour. However I had only tried and pass it as a flag, not how you are just now telling me.", "Hi @NickWithBotronics ! \r\nYou need to use transformers from source: `pip install -U git+https://github.com/huggingface/transformers`" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28488/reactions", "total_count": 2, "+1": 1, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 0, "eyes": 1 }

{ "login": "not-lain", "id": 70411813, "node_id": "MDQ6VXNlcjcwNDExODEz", "avatar_url": "https://avatars.githubusercontent.com/u/70411813?v=4", "gravatar_id": "", "url": "https://api.github.com/users/not-lain", "html_url": "https://github.com/not-lain", "followers_url": "https://api.github.com/users/not-lain/followers", "following_url": "https://api.github.com/users/not-lain/following{/other_user}", "gists_url": "https://api.github.com/users/not-lain/gists{/gist_id}", "starred_url": "https://api.github.com/users/not-lain/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/not-lain/subscriptions", "organizations_url": "https://api.github.com/users/not-lain/orgs", "repos_url": "https://api.github.com/users/not-lain/repos", "events_url": "https://api.github.com/users/not-lain/events{/privacy}", "received_events_url": "https://api.github.com/users/not-lain/received_events", "type": "User", "site_admin": false }

[ { "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" } ]

[]

[ "also it would be cool to use something like this : \r\n```python\r\n# Use a pipeline as a high-level helper\r\nfrom transformers import pipeline\r\n\r\npipe = pipeline(\"custom\", model=\"model_id\", trust_remote_code=True)\r\n```", "Hi @not-lain, thanks for raising this! \r\n\r\nYou might be interested in [using tools](https://huggingface.co./docs/transformers/transformers_agents#tools), which you can [define and use from the hub](https://huggingface.co./docs/transformers/custom_tools). \r\n\r\nWith regards to things not working, could you open a new issue following the [issue template](https://github.com/huggingface/transformers/blob/main/.github/ISSUE_TEMPLATE/bug-report.yml) including details about the running environment and the errors encountered? ", "@amyeroberts although it is interesting to find about tools but what i'm trying to adress that \r\n* pipeline doesn't work for custom model architectures on the hub [ feature requested] \r\n* the documentation says you can only work with a new pipeline if you add it to the transformers github repo (which will bloat the library on the long run)", "@amyeroberts, i found the mistake \r\ni needed to call the model as : \r\n```python\r\nfrom transformers import pipeline\r\n\r\nclassifier = pipeline(model=\"sgugger/test-dynamic-pipeline\", trust_remote_code=True)\r\n```\r\nnot as:\r\n```python\r\n# Use a pipeline as a high-level helper\r\nfrom transformers import pipeline\r\n\r\npipe = pipeline(\"text-classification\", model=\"sgugger/test-dynamic-pipeline\")\r\n```\r\ni will close this issue " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28487/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 }

[]

[]

[ "The problem with the current solution is that it's a patch over a larger issue: the ASR pipeline is a child of `Pipeline` but doesn't share the `__init__`. So, if there's any new logic added to `Pipeline` which relies on attributes being set or new kwargs this will still fail. \r\n\r\nIt also doesn't make sense for this class to accept an image processor. \r\n\r\nThe simplest solution (but doesn't address the divergence here) would be making certain attributes - `tokenizer`, `feature_extractor`, `image_processor` - to class attributes with default `None`. This way, any new pipeline which is similar to ASR will already have these required objects without having to handle / set them. \r\n\r\nA better solution is to update the `__init__` so that `super().__init__(*args, **kwargs)` is called at some point either modifying the kwargs before, setting self.type where necessary or adding any additional logic after. ", "Thanks for your comments. I agree that calling the parent `__init__` method is a more elegant solution. Resolved in https://github.com/huggingface/transformers/pull/28486/commits/e942642024aff014d5e7cd739ce7e5eb85abbfc7", "@sanchit-gandhi Before merging - could you update the title to reflect the change in fix applied? ", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/transformers/pr_28486). 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/28486/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/28486", "html_url": "https://github.com/huggingface/transformers/pull/28486", "diff_url": "https://github.com/huggingface/transformers/pull/28486.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28486.patch", "merged_at": 1705594310000 }

{ "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/28485/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/28485", "html_url": "https://github.com/huggingface/transformers/pull/28485", "diff_url": "https://github.com/huggingface/transformers/pull/28485.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28485.patch", "merged_at": 1705594335000 }

{ "login": "qmeeus", "id": 25608944, "node_id": "MDQ6VXNlcjI1NjA4OTQ0", "avatar_url": "https://avatars.githubusercontent.com/u/25608944?v=4", "gravatar_id": "", "url": "https://api.github.com/users/qmeeus", "html_url": "https://github.com/qmeeus", "followers_url": "https://api.github.com/users/qmeeus/followers", "following_url": "https://api.github.com/users/qmeeus/following{/other_user}", "gists_url": "https://api.github.com/users/qmeeus/gists{/gist_id}", "starred_url": "https://api.github.com/users/qmeeus/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/qmeeus/subscriptions", "organizations_url": "https://api.github.com/users/qmeeus/orgs", "repos_url": "https://api.github.com/users/qmeeus/repos", "events_url": "https://api.github.com/users/qmeeus/events{/privacy}", "received_events_url": "https://api.github.com/users/qmeeus/received_events", "type": "User", "site_admin": false }

[]

[]

[ "Hi @qmeeus - thanks for opening a PR! \r\n\r\nAt the moment, there's a very large diff including unrelated changes to those in the PR description. Could you make sure to resolve all conflicts and include the latest updates in `main` using `rebase` or `merge` into this branch? ", "> ed changes to those in the PR description. Could you make sure to resolve all conflicts and include the latest updates in `m`\r\n\r\nYes, this is a work in progress, I was not aware that this was available yet for reviewing. I will sort it out, thank you :)", "@amyeroberts I closed this one and created a new one [available here](https://github.com/huggingface/transformers/pull/28498)" ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28484/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/28484", "html_url": "https://github.com/huggingface/transformers/pull/28484", "diff_url": "https://github.com/huggingface/transformers/pull/28484.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28484.patch", "merged_at": null }

{ "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_28483). 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/28483/reactions", "total_count": 1, "+1": 0, "-1": 0, "laugh": 0, "hooray": 1, "confused": 0, "heart": 0, "rocket": 0, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28483", "html_url": "https://github.com/huggingface/transformers/pull/28483", "diff_url": "https://github.com/huggingface/transformers/pull/28483.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28483.patch", "merged_at": 1705078594000 }

{ "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/28482/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/28482", "html_url": "https://github.com/huggingface/transformers/pull/28482", "diff_url": "https://github.com/huggingface/transformers/pull/28482.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28482.patch", "merged_at": 1705315101000 }

{ "login": "NimaYaqmuri", "id": 62163525, "node_id": "MDQ6VXNlcjYyMTYzNTI1", "avatar_url": "https://avatars.githubusercontent.com/u/62163525?v=4", "gravatar_id": "", "url": "https://api.github.com/users/NimaYaqmuri", "html_url": "https://github.com/NimaYaqmuri", "followers_url": "https://api.github.com/users/NimaYaqmuri/followers", "following_url": "https://api.github.com/users/NimaYaqmuri/following{/other_user}", "gists_url": "https://api.github.com/users/NimaYaqmuri/gists{/gist_id}", "starred_url": "https://api.github.com/users/NimaYaqmuri/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/NimaYaqmuri/subscriptions", "organizations_url": "https://api.github.com/users/NimaYaqmuri/orgs", "repos_url": "https://api.github.com/users/NimaYaqmuri/repos", "events_url": "https://api.github.com/users/NimaYaqmuri/events{/privacy}", "received_events_url": "https://api.github.com/users/NimaYaqmuri/received_events", "type": "User", "site_admin": false }

[]

[]

[ "To observe the issue addressed by this PR, visit the [Hugging Face SpeechT5 Fine-Tuning Tutorial](https://huggingface.co./learn/audio-course/chapter6/fine-tuning). During training, you encounter a concatenation error due to incorrect dimensions.", "Thanks for your advice! I'm going to tweak the `generate_speech` for these two cases.", "Hello @ylacombe , @sanchit-gandhi and @ArthurZucker\r\n\r\nI've implemented your suggestions and made updates to the SpeechT5 Text to Speech class. My focus was on improving how we handle speaker embeddings in two key scenarios:\r\n\r\n1. **Matching Batch Size**: When there's one embedding per sample.\r\n2. **One-to-Many**: When a single embedding is used for all samples in the batch.\r\n\r\nKey updates include:\r\n--------------------\r\n\r\n* **Embedding Replication**: Introduced logic to replicate a single speaker embedding across multiple samples when necessary.\r\n* **Error Handling**: Implemented a ValueError to alert for dimension mismatches in speaker embeddings.\r\n* **Testing**: Added comprehensive test cases to ensure robust functionality across both scenarios.\r\n\r\nAdditionally, I've adapted the handling of speaker embedding dimensions outside the main model classes, aligning with the approach used in the original [SpeechT5 implementation by Microsoft](https://github.com/microsoft/SpeechT5). This decision avoids altering the SpeechT5 speech decoder's pre-net forward method, maintaining consistency with the existing model structure.\r\n\r\nPlease let me know if there are other scenarios or considerations I should account for. Your feedback is greatly appreciated.\r\n\r\nThank you for your guidance,\r\n\r\nNima Yaqmuri", "It seems there exists a related issue, [Issue #28189](https://github.com/huggingface/transformers/issues/28189), which my PR's speaker embedding updates could potentially address.", "Alongside the issue highlighted by @ylacombe in the Hugging Face thread \"[Audio Course Unit 6: Unable to Train SpeechT5](https://discuss.huggingface.co/t/audio-course-unit-6-unable-to-train-speech-t5/68888)\", my PR also aims to resolve a similar problem outlined in \"[SpeechT5 Text-to-Speech Fine-Tuning Runtime Error](https://discuss.huggingface.co/t/speecht5-text-to-speech-fine-tuning-runtime-error/67472)\", which seems to be related to the solution.", "Hi @ylacombe and @amyeroberts,\r\n\r\nThanks a lot for your valuable feedback and approvals!\r\n\r\nI've implemented the suggested changes, introducing randomized speaker embeddings with a fixed seed in the `speecht5` tests. \r\n\r\nAdditionally, I've run all the relevant slow tests for `speecht5` on my end, and everything works as expected. It seems we're ready for the merge.\r\n\r\nI appreciate your support and guidance throughout this process!\r\n\r\nBest.", "Thanks again @NimaYaqmuri for this great contribution! " ]

{ "url": "https://api.github.com/repos/huggingface/transformers/issues/28481/reactions", "total_count": 3, "+1": 0, "-1": 0, "laugh": 0, "hooray": 0, "confused": 0, "heart": 0, "rocket": 3, "eyes": 0 }

{ "url": "https://api.github.com/repos/huggingface/transformers/pulls/28481", "html_url": "https://github.com/huggingface/transformers/pull/28481", "diff_url": "https://github.com/huggingface/transformers/pull/28481.diff", "patch_url": "https://github.com/huggingface/transformers/pull/28481.patch", "merged_at": 1705414469000 }

{ "login": "hugo-syn", "id": 61210734, "node_id": "MDQ6VXNlcjYxMjEwNzM0", "avatar_url": "https://avatars.githubusercontent.com/u/61210734?v=4", "gravatar_id": "", "url": "https://api.github.com/users/hugo-syn", "html_url": "https://github.com/hugo-syn", "followers_url": "https://api.github.com/users/hugo-syn/followers", "following_url": "https://api.github.com/users/hugo-syn/following{/other_user}", "gists_url": "https://api.github.com/users/hugo-syn/gists{/gist_id}", "starred_url": "https://api.github.com/users/hugo-syn/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/hugo-syn/subscriptions", "organizations_url": "https://api.github.com/users/hugo-syn/orgs", "repos_url": "https://api.github.com/users/hugo-syn/repos", "events_url": "https://api.github.com/users/hugo-syn/events{/privacy}", "received_events_url": "https://api.github.com/users/hugo-syn/received_events", "type": "User", "site_admin": false }

[]