user
stringlengths 3
28
| created_at
timestamp[us]date 2020-04-01 09:48:12
2025-03-30 02:12:16
| body
stringlengths 1
173k
| issue_number
int64 1
3.18k
| __index_level_0__
int64 0
8.59k
|
|---|---|---|---|---|
qgallouedec | 2024-10-24T18:27:09 | PPO expect `reward_model` to be a model (torch module), not a function. | 2,273 | 700 |
HuggingFaceDocBuilderDev | 2024-10-24T15:52:48 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2272). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,272 | 701 |
HuggingFaceDocBuilderDev | 2024-10-24T10:06:21 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2270). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,270 | 702 |
qgallouedec | 2024-10-25T16:12:25 | Some tests are failing due to PairRM loading: it is fixed in #2276, you can safely ignore it | 2,270 | 703 |
edbeeching | 2024-10-28T09:30:36 | Hi @cutecharmingkid , unfortunately the answer is not trivial. Does the domain of your task match the tasks used to fine-tune the base vision-instruct model? I would imagine 10k-100k example would be enough, but I have not tested extensively. | 2,269 | 704 |
qgallouedec | 2024-10-25T16:02:36 | Thanks for reporting, please share your system info | 2,268 | 705 |
Isaaclgz | 2024-10-27T05:14:50 | > Thanks for reporting, please share your system info
Thanks for looking into this!
System:
Debian 11
Python 3.10
1xA100-80GB
Nvidia driver 550.90.07, CUDA 12.4
(running this on a GCP CE instance based on the c0-deeplearning-common-cu123-v20240922-debian-11-py310 image)
Env:
torch==2.4.0
transformers==4.44.0
trl==0.11.3
flash-attn==2.6.3
accelerate==1.0.1
| 2,268 | 706 |
chenyang399 | 2024-11-08T04:40:19 | is there any chance that we can run KTO script with 24G GPU
| 2,268 | 707 |
qgallouedec | 2024-10-24T18:10:55 | Thanks @cameronphchen! | 2,266 | 708 |
HuggingFaceDocBuilderDev | 2024-10-24T18:15:16 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2266). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,266 | 709 |
qgallouedec | 2024-10-23T08:12:32 | Thanks for reporting, it should have been fixed with #2261. CAN you confirm? | 2,264 | 710 |
ArcherShirou | 2024-10-24T02:28:19 | Thank you for your response. After updating the code and testing it, everything is running smoothly now. For the 14B and 72B models, quantization is necessary when using the 0.5B reward model. However, if I switch to the 70B or 72B reward model, I still encounter out-of-memory (OOM) issues midway, even with quantization and LoRA applied. Do you have any good solutions for this? | 2,264 | 711 |
qgallouedec | 2024-10-24T18:34:55 | You can try reducing the generation length. Closing the issue as the initial question is answered | 2,264 | 712 |
HuggingFaceDocBuilderDev | 2024-10-24T13:49:27 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2263). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,263 | 713 |
qgallouedec | 2024-11-23T12:50:57 | Looks good overall. Feel free to request a final review from me when you think it's ready to be merged | 2,263 | 714 |
yiyepiaoling0715 | 2024-12-25T02:48:30 | same question, has been solved? | 2,262 | 715 |
HuggingFaceDocBuilderDev | 2024-10-21T16:47:46 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2261). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,261 | 716 |
qgallouedec | 2024-10-21T15:04:46 | Thanks @cameronphchen! | 2,259 | 717 |
HuggingFaceDocBuilderDev | 2024-10-21T15:08:51 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2259). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,259 | 718 |
qgallouedec | 2024-10-24T13:01:30 | Thanks for the PR! However, I was actually considering simply removing this bot. In my opinion, it's fine to leave issues open for extended periods. I generally review all the issues and follow up when more information is needed and there hasn't been any activity for a while. From my experience, this bot tends to close issues that should remain open more often than it helps track active ones. See #1949 #1956.
What's more, the bot doesn't seem to have been working for a while, and nobody here seems to miss it.
What do you think @lewtun @kashif? | 2,258 | 719 |
Ananya54321 | 2024-10-25T02:02:26 | Ohh that makes sense! Thank you for responding!
| 2,258 | 720 |
lewtun | 2024-10-28T20:07:28 | Yes I agree, let's disable the bot since it's more of a nuisance than a help | 2,258 | 721 |
qgallouedec | 2024-11-11T23:16:04 | Close as a consequence of #2300 | 2,258 | 722 |
SinclairCoder | 2024-10-21T18:07:30 | I solved it with torchrun launch. | 2,257 | 723 |
Qinghao-Hu | 2024-10-22T01:37:47 | same problem
| 2,257 | 724 |
SinclairCoder | 2024-10-22T11:50:10 | @Qinghao-Hu launch it with torchrun if also a multigpu training case. | 2,257 | 725 |
innat | 2024-10-24T07:31:44 | what does it mean? , [src](https://huggingface.co/docs/accelerate/usage_guides/big_modeling).
> Multiple GPUs, or “model parallelism”, can be utilized but only one GPU will be active at any given moment. This forces the GPU to wait for the previous GPU to send it the output. You should launch your script normally with Python instead of other tools like torchrun and accelerate launch.
> You may also be interested in pipeline parallelism which utilizes all available GPUs at once, instead of only having one GPU active at a time. This approach is less flexbile though. For more details, refer to the [Memory-efficient pipeline parallelism](https://huggingface.co/docs/accelerate/usage_guides/distributed_inference#memory-efficient-pipeline-parallelism-experimental) guide.
| 2,256 | 726 |
gaetanlop | 2024-10-22T00:27:31 | Hey @mertege, adding the possibility to store teacher logits in the `GKDTrainer` is only useful when setting the parameter `lmbda` to 0 (which corresponds to standard KD). The all point of GKD is to enable on-policy KD (KD on sequences generated by the student) which means that we cannot store teacher logits offline during a pre-processing step. | 2,255 | 727 |
mertege | 2024-10-22T07:03:50 | Thanks for reply @gaetanlop. | 2,255 | 728 |
qgallouedec | 2024-10-21T16:50:10 | > all latest
can you run `trl env` please? | 2,254 | 729 |
qgallouedec | 2024-10-21T16:50:37 | Also please provide the full traceback | 2,254 | 730 |
saxenarohit | 2024-10-21T17:42:36 | Thanks
```
- Platform: Linux-5.4.0-187-generic-x86_64-with-glibc2.35
- Python version: 3.10.12
- PyTorch version: 2.2.0a0+81ea7a4
- CUDA device(s): NVIDIA A100-SXM4-80GB, NVIDIA A100-SXM4-80GB
- Transformers version: 4.45.2
- Accelerate version: 1.0.1
- Accelerate config: not found
- Datasets version: 3.0.1
- HF Hub version: 0.26.0
- TRL version: 0.12.0.dev0
- bitsandbytes version: 0.43.1
- DeepSpeed version: not installed
- Diffusers version: not installed
- Liger-Kernel version: not installed
- LLM-Blender version: not installed
- OpenAI version: not installed
- PEFT version: 0.13.
```
There is no traceback. It's a request to check for a possible bug.
During evaluation in the collate_fn
`labels = batch["input_ids"].clone()`
this will possibly have the gold answer in the input_ids during the evaluation?
| 2,254 | 731 |
edbeeching | 2024-10-23T08:45:08 | Hi @saxenarohit. This is normal, we are just looking at the eval loss. I think you might be thinking of a generative eval, where given a prompt, `model.generate` is used to autoregressively compute an answer, which can then be compared to the ground truth "gold answer". I will close the issue, but feel free to reopen if needed. | 2,254 | 732 |
qgallouedec | 2024-10-19T17:13:40 | This is because you need to provide a split dataset (containing both a training split and an evaluation split) when you use TRL scripts .
I realize the following limitations:
- when you're not evaluating, you still need to have a split dataset
- you may want the script to split the dataset when necessary.
This could be solved by adding something like :
```python
if training_args.eval_strategy != "none" and script_args.dataset_test_split not in dataset :
dataset = dataset[script_args.dataset_train_split].split(test_size=0.05)
...
trainer = AnyTrainer(
...
train_dataset=dataset[script_args.dataset_train_split],
eval_dataset=dataset[script_args.dataset_test_split] if training_args.eval_strategy != "none" else None,
...
)
```
WDYT @kashif @lewtun ? Is this situation common enough to justify this addition?
| 2,253 | 733 |
lewtun | 2024-10-24T09:34:00 | I don't think we should automatically generate a test split for the user (it's a bit too much magic), but I would be in favour of having the logic to set `eval_dataset` to `None` if no eval strategy is provided
| 2,253 | 734 |
qgallouedec | 2024-10-24T09:36:01 | > I don't think we should automatically generate a test split for the user (it's a bit too much magic), but I would be in favour of having the logic to set `eval_dataset` to `None` if no eval strategy is provided
Sounds reasonable. | 2,253 | 735 |
HuggingFaceDocBuilderDev | 2024-10-18T22:38:28 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2252). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,252 | 736 |
qgallouedec | 2024-10-20T13:52:17 | Thanks for the PR! Can you just run `make precommit` | 2,252 | 737 |
ngxson | 2024-10-20T22:25:27 | @qgallouedec Thanks! Should be good now | 2,252 | 738 |
qgallouedec | 2024-10-21T07:35:04 | It seems like this case occurs twice in our tests:
```
FAILED tests/test_dataset_formatting.py::SetupChatFormatTestCase::test_example_with_setup_model - ValueError: Chat template is already added to the tokenizer. If you want to overwrite it, please set it to None
FAILED tests/test_dataset_formatting.py::SetupChatFormatTestCase::test_setup_chat_format - ValueError: Chat template is already added to the tokenizer. If you want to overwrite it, please set it to None
```
Can you update the example so that they use this function correctly? | 2,252 | 739 |
qgallouedec | 2024-10-22T10:39:33 | Lgtm, thanks @ngxson | 2,252 | 740 |
ngxson | 2024-10-22T10:47:07 | Thanks! I don't have merge permission, so please merge when you want 🤗 | 2,252 | 741 |
kashif | 2024-10-21T11:04:55 | @gaetanlop can we use the `pad` helpers?
```py
# Use pad helper to handle padding
padded_query_responses = pad(query_responses, padding_value=pad_token_id, padding_side="right")
padded_logitss = pad(logitss, padding_value=0, padding_side="right")
```
| 2,251 | 742 |
gaetanlop | 2024-10-21T15:05:37 | @kashif, ~~the `pad` function expects the tensor to have no leading dimension corresponding to the batch size.~~
Here is an example `query_responses`:
```python
query_responses = [
torch.randint(vocab_size, (bs, seq_length1)),
torch.randint(vocab_size, (bs, seq_length2)),
torch.randint(vocab_size, (remaining_samples, seq_length3))
]
```
~~Using the `pad` function as it is would require the following change before passing the `query_responses` to the `pad` function:~~
```python
query_responses=[query_reps[i] for query_reps in query_responses for i in range(query_reps.size(0))]
```
~~We can also change the pad function? What do you prefer?~~
After looking more closely to the pad function, you are rigth, we can use the pad function as it is, it just requires reshaping the tensor afterwards.
I am gonna make the update, thanks for pointing it | 2,251 | 743 |
HuggingFaceDocBuilderDev | 2024-10-21T16:26:53 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2251). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,251 | 744 |
gaetanlop | 2024-10-21T16:34:19 | This won't work @kashif, it still requires reshaping the tensors
| 2,251 | 745 |
kashif | 2024-10-21T16:35:13 | ah damn! my bad sorry! | 2,251 | 746 |
gaetanlop | 2024-10-21T16:49:21 | No problem, this should be fixed now
| 2,251 | 747 |
JiahuiSun | 2024-10-27T01:37:26 | I also met the same issue. I use the official example script, dpo_online.py, to train a 75b LLM with a 75b reward model. Even with 60x8 H100 GPUs, the problem still happens. Any help please? | 2,250 | 748 |
lewtun | 2024-10-29T05:53:16 | Hello @hlnchen would you mind sharing a reproducible example that uses the `unwrap_model_for_generation()` method in a simple training loop that simulates your application? | 2,250 | 749 |
KAKSIS | 2024-11-08T06:46:37 | I encountered a similar issue while training a 72B model on an 8x H100 (80G) setup. I’m using the Hugging Face online DPO trainer scripts from [this link](https://huggingface.co/docs/trl/main/en/online_dpo_trainer). To reduce GPU memory usage, I've substituted the reward model with a random judger, so no reward model is loaded in GPU memory.
However, when running the code in zero3-offload mode, I encounter a CUDA out-of-memory (OOM) error at the unwrap_model_for_generation step, specifically in trl.trainer.online_dpo_trainer on line 395.
It seems that when executing this command, each process/graphics card collects parameters distributed across other processes, resulting in OOM. In the debug model, I can observe that the memory usage of each graphics card increases directly from 20GB to 80GB at that point.
Does anyone know the actual function of the command 'unwarp_madel_for_generation' in zero3 mode
here are my scripts.
```python
from datasets import load_dataset
from trl import OnlineDPOConfig, OnlineDPOTrainer
from transformers import AutoTokenizer
from typing import List, Optional, Union
class TestJudge():
def judge(self, prompts: List[str], completions: List[List[str]], return_scores=False) -> List[Union[int, float]]:
return [0]*len(prompts)
model_path = "Qwen2.5-72B-Instruct" #path to 72B model
judge = TestJudge()
data_path = "trl-lib/ultrafeedback-prompt"#path to dataset
tokenizer = AutoTokenizer.from_pretrained(model_path, local_files_only=True)
train_dataset = load_dataset(data_path, split="train")
training_args = OnlineDPOConfig(output_dir="online-dpo", logging_steps=2, bf16=True, fp16=False, per_device_train_batch_size=1, max_new_tokens=2048,
num_train_epochs=5, gradient_accumulation_steps=2, save_only_model=True,
save_steps=2000, save_total_limit=2)
trainer = OnlineDPOTrainer(
model=model_path,
ref_model=model_path,
judge=judge,
args=training_args,
processing_class=tokenizer,
train_dataset=train_dataset,
)
trainer.train()
#In OnlineDPOTrainer.__init__
#from transformers import AutoModelForCausalLM
#ref_model = AutoModelForCausalLM.from_pretrained(model, local_files_only=True)
#model = AutoModelForCausalLM.from_pretrained(model, local_files_only=True)
``` | 2,250 | 750 |
Namco0816 | 2024-11-13T07:38:02 | It seems like that I encountered the same issue. I also use a dummy reward model which do not take any GPU memory. And the training goes smoothly at the early stage, however after monitoring it for couples of iterations, the GPU memory usage keeps growing and at a specific iteration (in my case, 15 % total training steps for 8 GPUs, 7% total training steps for 4 GPUs), the GPU OOM when performing unwrap generation. I've tried to del as much variables as possible after each iteration and also empty the caches, however not works at all. | 2,250 | 751 |
yiyepiaoling0715 | 2024-12-30T04:57:16 | same question,watiing for method to resolve it | 2,250 | 752 |
Mefisto04 | 2024-10-21T19:31:15 | hey @qgallouedec , i have made a pr for this issue #2237 , please review all the changes that i have made. | 2,249 | 753 |
HuggingFaceDocBuilderDev | 2024-10-24T13:08:26 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2249). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,249 | 754 |
qgallouedec | 2024-10-24T13:22:57 | Thanks for helping improving this @Mefisto04. Can you make sure to run `make precommit`? A few suggestions, but it all looks good to me. | 2,249 | 755 |
Mefisto04 | 2024-10-24T18:37:47 | hey @qgallouedec i have commits all the changes that you have provided, please review this | 2,249 | 756 |
HuggingFaceDocBuilderDev | 2024-10-18T14:23:02 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2248). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,248 | 757 |
qgallouedec | 2024-10-18T10:18:01 | Thanks for reporting, it's about to be fixed: #2246 | 2,247 | 758 |
ArcherShirou | 2024-10-18T10:54:51 | thanks, its work | 2,247 | 759 |
HuggingFaceDocBuilderDev | 2024-10-18T09:31:22 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2246). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,246 | 760 |
kashif | 2024-10-24T08:31:33 | release is out | 2,245 | 761 |
HuggingFaceDocBuilderDev | 2024-10-24T08:35:33 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2245). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,245 | 762 |
HuggingFaceDocBuilderDev | 2024-10-17T11:44:37 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2244). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,244 | 763 |
HuggingFaceDocBuilderDev | 2024-10-16T15:58:00 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2243). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,243 | 764 |
qgallouedec | 2024-10-17T07:13:24 | @kashif can you also add an example in the online dpo documentation? And a test? | 2,243 | 765 |
kashif | 2024-10-17T07:19:39 | working on test thanks! | 2,243 | 766 |
qgallouedec | 2024-10-21T15:17:08 | I'm just updating the doc and running some tests | 2,243 | 767 |
qgallouedec | 2024-10-22T11:23:13 | ```
# 8 GPUs
accelerate launch examples/scripts/dpo_online.py \
--model_name_or_path trl-lib/pythia-1b-deduped-tldr-sft \
--judge pairrm \
--dataset_name trl-lib/tldr \
--learning_rate 5.0e-7 \
--logging_steps 25 \
--output_dir pythia-1b-tldr-online-dpo-reward \
--warmup_ratio 0.1
```
https://wandb.ai/huggingface/huggingface/runs/usqmcs3e
| 2,243 | 768 |
qgallouedec | 2024-10-23T15:44:00 | https://wandb.ai/huggingface/huggingface/runs/mq66mdbt
```
accelerate launch examples/scripts/dpo_online.py \
--model_name_or_path Qwen/Qwen2.5-0.5B-Instruct \
--judge pair_rm \
--dataset_name trl-lib/ultrafeedback-prompt \
--learning_rate 5.0e-7 \
--logging_steps 25 \
--output_dir Qwen2.5-0.5B-Online-DPO-PairRM \
--warmup_ratio 0.1
``` | 2,243 | 769 |
qgallouedec | 2024-10-18T13:49:17 | You can use it, feel free to report if it causes any issues. | 2,242 | 770 |
zwhe99 | 2024-10-20T05:00:09 | Thanks for the response! | 2,242 | 771 |
coding-famer | 2024-10-17T23:41:52 | I'm interested in working on this! | 2,241 | 772 |
qgallouedec | 2024-10-18T13:49:57 | Nice! Thanks @coding-famer. Feel free to open a PR then and request any help if needed | 2,241 | 773 |
August-murr | 2024-10-25T10:28:42 | @lewtun
After reading the paper, I noticed that the DPO checkpoints were combined with a different model rather than the reference model used in DPO training. So, I added an option in my PR to set an external model for merging instead of the reference model. | 2,241 | 774 |
coding-famer | 2024-10-25T18:01:36 | Hi @August-murr , happy to see that you have already worked it out! However I noticed that your implementation only allows merge models in the disk after training, this could be done by user using mergekit directly after training. I think the thing here is to merge the model during the training steps/epochs? | 2,241 | 775 |
August-murr | 2024-10-25T18:41:13 | @coding-famer The callback has an optional parameter called `merge_at_every_checkpoint`, which merges the saved checkpoint at either every step or at the end of each epoch during training. | 2,241 | 776 |
coding-famer | 2024-10-25T19:21:02 | > @coding-famer The callback has an optional parameter called `merge_at_every_checkpoint`, which merges the saved checkpoint at either every step or at the end of each epoch during training.
Sounds great! | 2,241 | 777 |
HuggingFaceDocBuilderDev | 2024-10-17T08:30:51 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2239). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,239 | 778 |
qgallouedec | 2024-10-17T09:03:46 | Thanks @August-murr! | 2,239 | 779 |
qgallouedec | 2024-10-18T14:21:33 | Thanks for pointing this out, #2248 will fix it | 2,238 | 780 |
reihig-ut | 2024-10-24T05:07:42 | Thank you for your PR!
I retried the reproduction process on branch `kto-conv-data-support`, I got this error:
```
/home/hoge/miniconda3/envs/run_kto/lib/python3.11/site-packages/trl/trainer/kto_trainer.py:479: UserWarning: When using DPODataCollatorWithPadding, you should set `max_length` in the KTOTrainer's init it will be set to `512` by default, but you should do it yourself in the future.
warnings.warn(
/home/hoge/miniconda3/envs/run_kto/lib/python3.11/site-packages/trl/trainer/kto_trainer.py:489: UserWarning: When using DPODataCollatorWithPadding, you should set `max_prompt_length` in the KTOTrainer's init it will be set to `128` by default, but you should do it yourself in the future.
warnings.warn(
/home/hoge/miniconda3/envs/run_kto/lib/python3.11/site-packages/trl/trainer/kto_trainer.py:519: UserWarning: When using DPODataCollatorWithPadding, you should set `remove_unused_columns=False` in your KTOConfig we have set it for you, but you should do it yourself in the future.
warnings.warn(
Traceback (most recent call last):
File "/home/hoge/project/test/trl/examples/scripts/kto.py", line 97, in <module>
trainer = KTOTrainer(
^^^^^^^^^^^
File "/home/hoge/miniconda3/envs/run_kto/lib/python3.11/site-packages/trl/trainer/kto_trainer.py", line 721, in __init__
super().__init__(
TypeError: Trainer.__init__() got an unexpected keyword argument 'processing_class'
``` | 2,238 | 781 |
benchay1999 | 2024-10-24T07:47:50 | Changing `processing_class` to `tokenizer` worked for me. | 2,238 | 782 |
kashif | 2024-10-24T08:44:08 | should be fixed now in main with latest transformer release | 2,238 | 783 |
chenyang399 | 2024-11-08T04:35:47 | How much memory it needs to run the KTO script ? is using the KTO script must have a GPU memory more than 24G? i use the 4090 with 24G memory failed. | 2,238 | 784 |
Mefisto04 | 2024-10-16T19:16:43 | hey @qgallouedec, please review this and assign me this issue | 2,237 | 785 |
qgallouedec | 2024-10-18T17:23:07 | Hi, thanks for reporting @Mefisto04. Feel free to open a PR if you can improve it. | 2,237 | 786 |
Mefisto04 | 2024-10-21T19:28:56 | hey @qgallouedec , i have made a pr #2249 , please review that. | 2,237 | 787 |
qgallouedec | 2024-10-25T16:04:41 | Closed via #2249 | 2,237 | 788 |
HuggingFaceDocBuilderDev | 2024-10-21T09:44:29 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2236). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,236 | 789 |
edbeeching | 2024-10-24T06:39:45 | HI @sergiopaniego , thanks for impementing this. Could you run `make precommit` to format the code so the quality tests pass (you may have to `pip install pre-commit`)
We are discussing internally how feasible it is to hormonize this script with the other VLM training scripts, I will let you know when we have a conclusion. | 2,236 | 790 |
sergiopaniego | 2024-10-30T09:12:56 | Updated!
Any updates on the harmonization discussion? I’m happy to make any modifications needed! 😊
| 2,236 | 791 |
mshuffett | 2024-11-04T01:57:33 | @sergiopaniego so is this working in theory? Also OOM'ing for me needs 50 GB and my A100 only has like 40 GB or something. Is there a level I can pull to decrease the memory? Why does it need so much considering it is doing a LORA?
Is it possible to set this up to train on multiple GPUs? | 2,236 | 792 |
sergiopaniego | 2024-11-17T20:25:35 | > @sergiopaniego so is this working in theory? Also OOM'ing for me needs 50 GB and my A100 only has like 40 GB or something. Is there a level I can pull to decrease the memory? Why does it need so much considering it is doing a LORA?
>
> Is it possible to set this up to train on multiple GPUs?
Sorry for the late response @mshuffett. It still needs some polishing. While testing it, it seems like something is still missing from the artifacts for the model shared. You can see more details about it in the [README](https://github.com/2U1/Molmo-Finetune). For example, since the `grad-checkpoint` is disabled, memory consumption increases a lot.
It's also not yet merged in the official transformers repo https://github.com/huggingface/transformers/pull/33962 | 2,236 | 793 |
qgallouedec | 2024-10-18T17:18:01 | This operation replaces tokens outside the attention mask with token 0. This operation has no influence on model output within the attention mask:
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_id = "Qwen/Qwen2.5-0.5B-Instruct"
model = AutoModelForCausalLM.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
pad_token_id = tokenizer.pad_token_id
input_ids = torch.tensor([[pad_token_id, pad_token_id, 1, 2, 3, 4, 5, pad_token_id]])
attention_mask = input_ids != pad_token_id # [[False, False, True, True, True, True, True, False]]
position_ids = attention_mask.cumsum(1) - attention_mask.long() # [[0, 0, 1, 2, 3, 4, 5, 0]]
output_wo_mask_fill = model(input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids)
input_ids = torch.masked_fill(input_ids, ~attention_mask, 0) # [[0, 0, 0, 1, 2, 3, 4, 0]]
output_w_mask_fill = model(input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids)
print(torch.mean(torch.abs(output_wo_mask_fill.logits - output_w_mask_fill.logits), dim=-1)) # [[0.8371, 0.8371, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 3.6457]]
```
This operation is not absolutely necessary, since invalid logits are then masked:
https://github.com/huggingface/trl/blob/a67f2143c38d6520be8735463ce715ad5c281db8/trl/trainer/rloo_trainer.py#L413-L415 | 2,235 | 794 |
Chios-C | 2024-10-19T05:46:57 | Thanks for your great response. | 2,235 | 795 |
HuggingFaceDocBuilderDev | 2024-10-15T10:04:10 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2233). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,233 | 796 |
qgallouedec | 2024-10-15T08:59:33 | Thanks again @DhruvKadam-git. Can you update your branch? | 2,232 | 797 |
DhruvKadam-git | 2024-10-17T07:36:04 | I have updated my branch | 2,232 | 798 |
HuggingFaceDocBuilderDev | 2024-10-18T17:26:18 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2232). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,232 | 799 |
Subsets and Splits