user
stringlengths 3
28
| created_at
timestamp[us] | body
stringlengths 1
173k
| issue_number
int64 1
2.95k
| __index_level_0__
int64 0
8.59k
|
|---|---|---|---|---|
zaddy6 | 2025-02-17T19:10:28 | @Maghoumi not the case for me
<img width="1024" alt="Image" src="https://github.com/user-attachments/assets/6e08e4ce-17e6-44f8-910b-05d4dc125a6d" />
purple is with peft and vllm enabled | 2,856 | 212 |
HuggingFaceDocBuilderDev | 2025-02-13T17:26:27 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2855). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,855 | 213 |
qgallouedec | 2025-02-13T17:28:23 | It should always be the case indeed.
The built-in `set` isn't ordered right?
Is vllm faster when you pass 1 prompt with n=N outputs, than with N times the same prompt for n=1? | 2,855 | 214 |
edbeeching | 2025-02-13T17:58:23 | > It should always be the case indeed. The built-in `set` isn't ordered right? Is vllm faster when you pass 1 prompt with n=N outputs, than with N times the same prompt for n=1?
Unfortunately, `set` is not ordered. Yes vllm can share the prefill for the n generations so it is faster, I profiled around 1.5x faster with the changes in this PR at 2k `max_completion_length`. | 2,855 | 215 |
qgallouedec | 2025-02-13T18:24:10 | Nice!! I am surprised, I expected a smaller speedup given that the prefix should already be reused since https://github.com/huggingface/trl/pull/2757.
We should probably do the same with tranformers generation in a future PR, if it makes sense.
Anyway, can you just add comment somewhere to explain why we do this? | 2,855 | 216 |
winglian | 2025-02-13T22:54:59 | My guess is it's an easier optimization for vllm to understand that single prompt has multiple generations than sending the same prompt multiple times from the https://github.com/huggingface/trl/pull/2776 refactor. | 2,855 | 217 |
edbeeching | 2025-02-14T08:45:46 | @qgallouedec, without diving into the codebase of vllm, I would assume that the prefix cache is only used to compare a new batch of prompts with previously processed prompts. The system prompt is shared across all prompts, so this is cached and reused for all batches, whereas a new batch of prompts would first all need to have their prefill calculated and entered into the cache before vllm could identify that there are `num_generations` of the prompts are exactly the same. Hence you get some improvement when you move from `prompt*num_generations` to `n` generations for each prompt.
Let me know if you would like me to clarify. | 2,855 | 218 |
qgallouedec | 2025-02-14T08:57:18 | Thanks Ed! Actually I meant adding a comment in the code to concisely explain why we merge the prompts. Something like `Since 'prompts' contains 'num_generations' duplicates, we first take unique prompts, and generate num_generations outputs for each one. This is faster than generating outputs for each duplicate prompt individually.`. | 2,855 | 219 |
skandermoalla | 2025-02-13T15:34:49 | It's the estimator used by GRPO (ref eq 2 https://arxiv.org/pdf/2501.12948). For more details, you can check the `k3` estimator in this blogpost (http://joschu.net/blog/kl-approx.html). | 2,854 | 220 |
qiaojiim | 2025-02-14T01:40:15 | > It's the estimator used by GRPO (ref eq 2 https://arxiv.org/pdf/2501.12948). For more details, you can check the `k3` estimator in this blogpost (http://joschu.net/blog/kl-approx.html).
great | 2,854 | 221 |
KareemMusleh | 2025-02-14T08:10:30 | It seems that it was moved to DPOConfig | 2,853 | 222 |
llj1113 | 2025-02-14T08:37:19 | > It seems that it was moved to DPOConfig
thanks！I have solved this problem. | 2,853 | 223 |
edbeeching | 2025-02-13T13:29:02 | Can you test with `vllm==0.7.2`, I had a similar issue which I believe was fixed when I bumped vllm version. | 2,851 | 224 |
YunGe0414 | 2025-02-14T03:09:36 | > Can you test with `vllm==0.7.2`, I had a similar issue which I believe was fixed when I bumped vllm version.
It worked, thank you so much bro. | 2,851 | 225 |
edbeeching | 2025-02-14T10:00:16 | No probs, closing. | 2,851 | 226 |
HuggingFaceDocBuilderDev | 2025-02-13T10:46:17 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2850). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,850 | 227 |
HuggingFaceDocBuilderDev | 2025-02-13T10:03:18 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2848). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,848 | 228 |
qgallouedec | 2025-02-12T19:47:38 | Do you any dataset example that would contain such data? Your concern is that it could corrupt the training in some sense right?
Is it a documented/recurrent issue? | 2,844 | 229 |
shirinyamani | 2025-02-12T20:06:17 | Exactly! It can cause some issues in the training phase. Technically, any sort of rlhf schema that has SFT step with SFT data that you wanna fine-tune on those specific examples. If those specific examples contain any knowledge cutoff, it might cause issues!
So far I did a very simple `re` based search of the cutoff pattern of
```python
r"as of my last update",
r"as of my last knowledge update",
r"as of \d{4}", # Matches "As of 2024", "As of 2023", etc.
r"i do not have access to real-time information",
```
for the [trl-lib/tldr](https://huggingface.co./datasets/trl-lib/tldr) dataset, and could not find anything. I mean I found ~ 1000 examples that has the term "As of the year" or "as it date back" etc but when I took a closer look these examples were referring to some date or sth in the context as this dataset is from reddit and mostly from the relationship subreddit, so could not find any matching example with respect to my concern of model generating a cutoff knowledge completion!
However, note that this specific dataset by nature is not really a good match for the purpose I mentioned earlier but I believe this can happen in other SFT data. Speaking of this, I also saw a similar thing raised in allenai/open-instruct. Therefore, I thought it might be nice if we add such support, WDYT? | 2,844 | 230 |
shirinyamani | 2025-02-12T20:11:47 | Imagine Im a user that only wanna use the Base model as publicly available models but for the sft training, I wanna do it with my local examples using TRL/sft_trainer,
So is there any way we can flag this before fully training on the data that the Data the user is using for sft contains some knowledge cutoff ?
OR the datasets that TRL presents under trl-lib.
Is my question valid? | 2,844 | 231 |
qgallouedec | 2025-02-12T20:39:12 | This information is typically included in the system prompt. So, even if the model has been trained on this "corrupting" data, it shouldn’t pose an issue during generation—but that’s just my intuition.
In any case, this sounds more like a data preparation concern. Unless it’s a severe and recurring issue (i.e., well-documented and frequently reported), I’d consider it slightly beyond the scope of TRL.
That said, now that this issue has been raised, if you have code that can detect or filter such data in a dataset, this would be the right place to share it. | 2,844 | 232 |
shirinyamani | 2025-02-12T21:40:37 | This was the quick code i used for the [trl-lib/tldr](https://huggingface.co./datasets/trl-lib/tldr) dataset.
```python
import re
dataset = load_dataset("trl-lib/tldr", split="train")
# knowledge cutoff-related phrases
cutoff_patterns = [
r"as of my last update",
r"as of my last knowledge update",
r"as of \d{4}", # Matches "As of 2024", etc.
r"i do not have access to real-time information",
r"i was last updated in \d{4}",
]
def check_knowledge_cutoff(text):
text = text.lower() # Normalize to lowercase
return any(re.search(pattern, text) for pattern in cutoff_patterns)
cutoff_mentions = [
(row["prompt"], row["completion"])
for row in dataset
if check_knowledge_cutoff(row["prompt"]) or check_knowledge_cutoff(row["completion"])
]
# Optionally, print a few examples
print("Sample Matches:")
for i, (prompt, completion) in enumerate(cutoff_mentions):
print(f"{i+1}. Prompt: {prompt}\n Completion: {completion}\n")
```
But I also found this [PR](https://github.com/allenai/open-instruct/pull/555) on Allenai/open-instruct relevant to the topic! | 2,844 | 233 |
HuggingFaceDocBuilderDev | 2025-02-13T09:05:24 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2843). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,843 | 234 |
JoelSeniorLiang | 2025-02-12T13:57:04 | try to set per_device_train_batch_size=1 at the config | 2,842 | 235 |
qgallouedec | 2025-02-12T15:34:16 | What trl version do you use? | 2,842 | 236 |
qgallouedec | 2025-02-12T15:37:30 | what is the per device batch size?
Where do you get these "output per device"? | 2,842 | 237 |
MAOJIASONG | 2025-02-13T08:50:57 | > try to set per_device_train_batch_size=1 at the config
Hi, I already set `per_device_train_batch_size=1`, so I assume the problem is not happened from training config | 2,842 | 238 |
MAOJIASONG | 2025-02-13T08:51:08 | > What trl version do you use?
0.14.0 | 2,842 | 239 |
MAOJIASONG | 2025-02-13T08:51:59 | > what is the per device batch size? Where do you get these "output per device"?
`per_device_train_batch_size=1` as it is.
my debugging output obtains this output per device. | 2,842 | 240 |
qgallouedec | 2025-02-13T08:57:52 | Ok but there is no such variable as "output_per_device" so please elaborate what line are you refering to. It's not very clear for me at this point | 2,842 | 241 |
MAOJIASONG | 2025-02-14T04:16:11 | > Ok but there is no such variable as "output_per_device" so please elaborate on what line are you referring to. It's not very clear to me at this point
https://github.com/huggingface/trl/blob/49711efab9e0cc3762d3228c9fd5a8064d489503/trl/trainer/grpo_trainer.py#L469
https://github.com/huggingface/trl/blob/49711efab9e0cc3762d3228c9fd5a8064d489503/trl/trainer/grpo_trainer.py#L464
Sorry, the `output_per_device` is defined by myself. In fact, I printed the number of prompts and completions for the above two lines and found the number is not matched with `num_processes=1`, which should be only 1 prompt per device. But this one gives me `avail_num_gpus*per_device_train_batch_size=4*1=4` that many of the prompts, as indicated in the example. | 2,842 | 242 |
zzn1999 | 2025-02-12T10:18:52 | That's Good. I have the same issue. | 2,841 | 243 |
yuting-shi | 2025-02-12T10:33:17 | Solved my problem! | 2,841 | 244 |
kashif | 2025-02-12T13:40:58 | @yuting-shi what problem does this solve? i believe the `generate()` method runs in inference/eval model | 2,841 | 245 |
yuting-shi | 2025-02-13T01:37:41 | @kashif the content generated by inference model is a mess, even though the model has been SFT before. | 2,841 | 246 |
casper-hansen | 2025-02-12T12:04:25 | Offending PR might be https://github.com/huggingface/trl/pull/2817 | 2,840 | 247 |
AndreiCComan | 2025-02-13T19:43:31 | Same issue here. In my case this happened immediately after the checkpoint has been saved. | 2,840 | 248 |
qgallouedec | 2025-02-13T20:15:02 | Can you try to provide the steps to reproduce? Maybe take only a small part of your dataset could help reproduce without having to wait 24 hours | 2,840 | 249 |
Superskyyy | 2025-02-14T00:31:14 | https://github.com/huggingface/open-r1/issues/299 seems to be the same issue referenced in open-r1 | 2,840 | 250 |
casper-hansen | 2025-02-14T08:46:51 | > Can you try to provide the steps to reproduce? Maybe take only a small part of your dataset could help reproduce without having to wait 24 hours
This was with the following dataset https://huggingface.co./datasets/allenai/RLVR-IFeval | 2,840 | 251 |
hezhefly | 2025-02-17T04:58:52 | > Same issue here. In my case this happened immediately after the checkpoint has been saved.
Same situation | 2,840 | 252 |
hezhefly | 2025-02-17T10:53:48 | 我根据日志分别查阅了trl和deepspeed的源码，发现是`deepspeed.zero.GatheredParameters`中对参数的断言引发的错误，进一步查阅断言的逻辑，发现`free_param(param)`方法希望在执行之前`ds_active_sub_modules`参数值被清空。我不清楚trl中具体是什么原因造成这个这种`ds_active_sub_modules`参数值未清空的现象。
所以，我大胆的尝试了一下手动清空`ds_active_sub_modules`参数值，我尝试在[grpo_trainer.py#L490](https://github.com/huggingface/trl/blob/main/trl/trainer/grpo_trainer.py#L490)中加入以下清空参数的逻辑：
```python
for param in self.model.parameters():
param.ds_active_sub_modules.clear()
```
测试后发现有效，目前已经完成GRPO的训练任务。 | 2,840 | 253 |
wuyifan18 | 2025-02-18T04:20:21 | Same issue | 2,840 | 254 |
Superskyyy | 2025-02-18T04:51:50 | Just cross reference from OpenRLHF issue, seems like related to batch size.
https://github.com/OpenRLHF/OpenRLHF/issues/630 | 2,840 | 255 |
tsrigo | 2025-02-19T02:59:32 | > Same issue here. In my case this happened immediately after the checkpoint has been saved.
@qgallouedec Me too! Have you fix this problem? | 2,840 | 256 |
tsrigo | 2025-02-20T06:58:48 | > > Same issue here. In my case this happened immediately after the checkpoint has been saved.
>
> [@qgallouedec](https://github.com/qgallouedec) Me too! Have you fix this problem?
I fix it by satisfying `save_interval % grad_accum == 0`. | 2,840 | 257 |
loxs123 | 2025-02-21T16:20:46 | > 我根据日志分别查阅了trl和deepspeed的源码，发现是`deepspeed.zero.GatheredParameters`中对参数的断言引发的错误，进一步查阅断言的逻辑，发现`free_param(param)`方法希望在执行之前`ds_active_sub_modules`参数值被清空。我不清楚trl中具体是什么原因造成这个这种`ds_active_sub_modules`参数值未清空的现象。
>
> 所以，我大胆的尝试了一下手动清空`ds_active_sub_modules`参数值，我尝试在[grpo_trainer.py#L490](https://github.com/huggingface/trl/blob/main/trl/trainer/grpo_trainer.py#L490)中加入以下清空参数的逻辑：
>
> for param in self.model.parameters():
> param.ds_active_sub_modules.clear()
> 测试后发现有效，目前已经完成GRPO的训练任务。
我运行代码报了这个错误，`AttributeError: 'Parameter' object has no attribute 'ds_active_sub_modules`，请问你知道该如何解决吗？或许是某个库的版本不太一致？ | 2,840 | 258 |
nikhilchandak | 2025-02-22T12:24:44 | +1, I am also facing the same issue. @tsrigo in your fix, does `save_interval` correspond to `save_steps` which should be set as a multiple of `gradient_accumulation_steps`? I tried that still my runs crash.
@qgallouedec Any known fix for this? | 2,840 | 259 |
hezhefly | 2025-02-24T02:09:09 | > > 我根据日志分别查阅了trl和deepspeed的源码，发现是`deepspeed.zero.GatheredParameters`中对参数的断言引发的错误，进一步查阅断言的逻辑，发现`free_param(param)`方法希望在执行之前`ds_active_sub_modules`参数值被清空。我不清楚trl中具体是什么原因造成这个这种`ds_active_sub_modules`参数值未清空的现象。
> > 所以，我大胆的尝试了一下手动清空`ds_active_sub_modules`参数值，我尝试在[grpo_trainer.py#L490](https://github.com/huggingface/trl/blob/main/trl/trainer/grpo_trainer.py#L490)中加入以下清空参数的逻辑：
> > for param in self.model.parameters():
> > param.ds_active_sub_modules.clear()
> > 测试后发现有效，目前已经完成GRPO的训练任务。
>
> 我运行代码报了这个错误，`AttributeError: 'Parameter' object has no attribute 'ds_active_sub_modules`，请问你知道该如何解决吗？或许是某个库的版本不太一致？
@loxs123 我使用的版本是
Name: deepspeed
Version: 0.15.3 | 2,840 | 260 |
qgallouedec | 2025-02-24T07:48:02 | No because I'm still waiting for someone to provide the sufficient info to reproduce https://github.com/huggingface/trl/issues/2840#issuecomment-2657619732 . Maybe you can help with this? | 2,840 | 261 |
nomadlx | 2025-02-24T10:04:12 | > 不，因为我还在等有人提供足够的信息来重现[#2840（评论）](https://github.com/huggingface/trl/issues/2840#issuecomment-2657619732)。也许你可以帮忙？
I think this information might be helpful for you to reproduce the issue: https://github.com/huggingface/open-r1/issues/299#issuecomment-2667375592.
Because I've verified that with the same training data, when `trainset_size % batch_size == 0`, this error will no longer occur after the first save. | 2,840 | 262 |
qgallouedec | 2025-02-12T07:47:12 | Thanks for the suggestion but it doesn't align with the paper actually. | 2,837 | 263 |
pointerhacker | 2025-02-12T09:58:15 | > Thanks for the suggestion but it doesn't align with the paper actually.
In my understanding, isn't per_token_logps - per_token_logps.detach() always equal to 0? Could you please explain why this is feasible? Thank you! | 2,837 | 264 |
qgallouedec | 2025-02-12T10:46:07 | That's right, answer here: https://github.com/huggingface/trl/pull/2565#issuecomment-2595837761 :)
| 2,837 | 265 |
pointerhacker | 2025-02-12T13:16:29 | Thank you for your reply. | 2,837 | 266 |
pointerhacker | 2025-02-12T13:21:25 | > That's right, answer here: [#2565 (comment)](https://github.com/huggingface/trl/pull/2565#issuecomment-2595837761) :)
Based on what you said that `As a result, the GRPO objective just minimizes the KL divergence between the policy model and the reference policy,` I have another question: How can this approach achieve preference alignment effects? | 2,837 | 267 |
Superskyyy | 2025-02-12T13:58:59 | Since the vllm device patch is growing larger. It might be wise to move them into a utility module instead. Wdyt. | 2,836 | 268 |
baymax591 | 2025-02-14T10:54:03 | This PR helps a lot, I hope it can speed up the integration | 2,836 | 269 |
ji-huazhong | 2025-02-14T13:30:03 | I think this PR is ready to be merged 🤗 @qgallouedec | 2,836 | 270 |
HuggingFaceDocBuilderDev | 2025-02-14T13:52:15 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2836). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,836 | 271 |
qgallouedec | 2025-02-14T14:23:07 | Can you make sure sure to run `make precommit` to apply the style 🙏 | 2,836 | 272 |
ji-huazhong | 2025-02-18T14:56:20 | `make precommit` is successfully executed locally | 2,836 | 273 |
lynnzhiyun | 2025-02-18T15:28:50 | Hi @ji-huazhong, Thank you for your excellent work! This PR has been incredibly helpful in enabling me to train models using GRPO on the NPU smoothly.
I want to ask if this PR is ready to be merged and I'd be extremely grateful if it could be done promptly.
cc @qgallouedec | 2,836 | 274 |
ji-huazhong | 2025-02-19T07:45:00 | [](https://asciinema.org/a/704242)
I did a test on Ascend NPU using the grpo script provided by open-r1, it works 🤗
> Since training grpo for one step takes a long time, only the output of the first 4 steps is shown here, and then I just press ctrl-c to exit. | 2,836 | 275 |
ji-huazhong | 2025-02-19T08:40:49 |
Hi @kashif, the failing test case seems unrelated to this PR. Could you take a look? Thanks! | 2,836 | 276 |
symoon11 | 2025-02-16T06:21:44 | To the best of my knowledge, the "padding free" option works correctly only when FlashAttention is activated. It seems that FlashAttention is not currently activated. I recommend first creating the model and then passing it to the SFTTrainer. | 2,834 | 277 |
YooSungHyun | 2025-02-17T00:22:26 | @symoon11 thx for reply! i give `--attn_implementation=flash_attention_2`, but this is model_config and not training_args...
i will test soon and make some result | 2,834 | 278 |
YooSungHyun | 2025-02-17T00:36:23 | I foolishly forgot to include this part in my code:
```python
quantization_config = get_quantization_config(model_args)
model_kwargs = dict(
revision=model_args.model_revision,
trust_remote_code=model_args.trust_remote_code,
attn_implementation=model_args.attn_implementation,
torch_dtype=model_args.torch_dtype,
use_cache=False if training_args.gradient_checkpointing else True,
device_map=get_kbit_device_map() if quantization_config is not None else None,
quantization_config=quantization_config,
)
training_args.model_init_kwargs = model_kwargs
```
Sorry for causing confusion. My code runs correctly now! | 2,834 | 279 |
zaporter | 2025-02-12T04:26:33 | See https://github.com/huggingface/trl/blob/main/docs/source/grpo_trainer.md#computing-the-advantage
It doesn't matter if you have negative or positive weights -- all that matters is the group relative advantage.
Rewards of {1, 0} will result in advantages of `1` and `-1` respectively. That is the same as rewards of {1,-1} which results in `1, -1`
Or consider rewards of `{1, 1, 2}`, this will result in advantages of `-1/sqrt(2), -1/sqrt(2), sqrt(2)` | 2,832 | 280 |
HuggingFaceDocBuilderDev | 2025-02-11T18:12:52 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2831). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,831 | 281 |
HuggingFaceDocBuilderDev | 2025-02-11T13:34:15 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2829). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,829 | 282 |
HuggingFaceDocBuilderDev | 2025-02-11T10:09:22 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2828). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,828 | 283 |
HuggingFaceDocBuilderDev | 2025-02-11T07:55:50 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2827). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,827 | 284 |
MohamedAliRashad | 2025-02-11T06:31:17 | After some inspection, i think this error happens because of `vllm_gpu_memory_utilization`, if it's smaller that what vllm can use to host your model it will give you the error i recieved. | 2,826 | 285 |
qgallouedec | 2025-02-11T06:46:38 | Ah that's right, it a particular case where the error message is misleading. Actually you should set `vllm_device="cuda:0"`. | 2,826 | 286 |
MohamedAliRashad | 2025-02-11T10:16:24 | @qgallouedec I thought the problem was the vllm couldn't find enough VRAM on my single GPU so automatically seeked to get another one but failed at the end.
Anyhow, i am trying to use an evaluation dataset with `GRPOTrainer` and it is giving me this error:
```
Traceback (most recent call last):
File "/workspace/train_grpo2.py", line 173, in <module>
trainer.train(resume_from_checkpoint=False)
File "/usr/local/lib/python3.11/dist-packages/transformers/trainer.py", line 2171, in train
return inner_training_loop(
^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/dist-packages/transformers/trainer.py", line 2598, in _inner_training_loop
self._maybe_log_save_evaluate(
File "/usr/local/lib/python3.11/dist-packages/transformers/trainer.py", line 3071, in _maybe_log_save_evaluate
metrics = self._evaluate(trial, ignore_keys_for_eval)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/dist-packages/transformers/trainer.py", line 3025, in _evaluate
metrics = self.evaluate(ignore_keys=ignore_keys_for_eval)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/dist-packages/transformers/trainer.py", line 4073, in evaluate
output = eval_loop(
^^^^^^^^^^
File "/usr/local/lib/python3.11/dist-packages/transformers/trainer.py", line 4267, in evaluation_loop
losses, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/dist-packages/transformers/trainer.py", line 4436, in prediction_step
has_labels = False if len(self.label_names) == 0 else all(inputs.get(k) is not None for k in self.label_names)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/dist-packages/transformers/trainer.py", line 4436, in <genexpr>
has_labels = False if len(self.label_names) == 0 else all(inputs.get(k) is not None for k in self.label_names)
^^^^^^^^^^
AttributeError: 'list' object has no attribute 'get'
0%| | 10/57951 [01:38<158:47:29, 9.87s/it]
```
Can you tell me what i am doing wrong ? | 2,826 | 287 |
qgallouedec | 2025-02-11T10:18:11 | Thanks for reporting, can you share the output of `trl env`? | 2,826 | 288 |
MohamedAliRashad | 2025-02-11T10:23:38 | INFO 02-11 10:23:26 __init__.py:190] Automatically detected platform cuda.
Copy-paste the following information when reporting an issue:
- Platform: Linux-5.4.0-167-generic-x86_64-with-glibc2.35
- Python version: 3.11.10
- PyTorch version: 2.5.1
- CUDA device(s): NVIDIA L40
- Transformers version: 4.48.3
- Accelerate version: 1.3.0
- Accelerate config: not found
- Datasets version: 3.2.0
- HF Hub version: 0.28.1
- TRL version: 0.14.0
- bitsandbytes version: not installed
- DeepSpeed version: not installed
- Diffusers version: not installed
- Liger-Kernel version: not installed
- LLM-Blender version: not installed
- OpenAI version: 1.61.1
- PEFT version: not installed
| 2,826 | 289 |
qgallouedec | 2025-02-11T10:49:42 | @MohamedAliRashad can you try after install from source:
```
pip install git+https://github.com/huggingface/trl.git@main
``` | 2,826 | 290 |
lidh15 | 2025-02-11T02:13:48 | with very similar environment setup (except for trl 0.15.0.dev0, where is that version? I can only find 0.14.0) I encountered this issue [No inf checks were recorded for this optimizer](https://discuss.pytorch.org/t/no-inf-checks-were-recorded-for-this-optimizer/140505).
when I turn off vllm, the error will not be triggered, I wonder if there is any clue for this error. | 2,825 | 291 |
HuggingFaceDocBuilderDev | 2025-02-10T20:28:56 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2824). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,824 | 292 |
mark-mcl | 2025-02-11T02:07:13 | this seems to also affect sampling, tried patching this and I get a different prompt on each device now | 2,824 | 293 |
kiddj | 2025-02-11T02:46:58 | thanks! does using a different seed for each process affect training? | 2,824 | 294 |
qgallouedec | 2025-02-11T07:18:27 | True @kiddj!! Fixed in 95fcfeb003d8ba0ab339faeabcb41365786cee58 | 2,824 | 295 |
qgallouedec | 2025-02-11T08:58:31 | I've checked it locally, and yes it's now working as expected.
I'll see in a follow up PR if I can add a test for this | 2,824 | 296 |
qgallouedec | 2025-02-10T18:41:19 | Hi and thanks for your contribution! The idea seems quite natural. Do you have any quantitative results?
I'd like to keep the codebase simple, so for the moment I'm in favor of leaving this PR open for the community to reference, and if it's a feature in high demand, then we'll merge it. | 2,823 | 297 |
mandeep511 | 2025-02-10T19:04:10 | > Hi and thanks for your contribution! The idea seems quite natural. Do you have any quantitative results? I'd like to keep the codebase simple, so for the moment I'm in favor of leaving this PR open for the community to reference, and if it's a feature in high demand, then we'll merge it.
Hey, I'm doing a couple of runs to test how well this performs compared to the vanilla implementation. Will post the results here once it is done. | 2,823 | 298 |
willccbb | 2025-02-11T01:08:42 | I'm not sure it makes sense to directly add features like this which are not part of the canonical algorithm, and which add significant complexity to the codebase, making further maintenance + feature compatibility more difficult.
I believe that this approach probably works, and could yield higher performance, but my opinion is that this shouldn't be the primary goal of GRPOTrainer. There are lots of such tricks that could be added, but each one makes the code harder to read and modify, and is no longer "GRPO" in the literal sense.
Mentioning my PR https://github.com/huggingface/trl/pull/2810 here because I think it's directly relevant: if people want to try out non-canonical sampling methods like MCTS, reward thresholds, reward score diversity constraints etc., or add things like tool calls or multi-step interactions they should have a way to do this without needing to modify the core Trainer. The `Environment` abstraction would allow users to have full control over the sampling step, and perhaps implementations of these could live in a unified place in TRL (e.g. `RolloutSamplers`) to support easy hot-swapping. We may also want to allow users to return rewards directly in the rollout stage. On the whole, I think it will be easier for more people to use and adapt TRL trainers if the primary code stays simple while supporting modular customization. | 2,823 | 299 |
winglian | 2025-02-11T22:58:10 | It could be worthwhile to refactor the GRPOTrainer to make it easier to extend the class without having to duplicate whole swaths of code in a method in order to add retries in a subclass. | 2,823 | 300 |
Rocketknight1 | 2025-02-11T15:23:10 | Not all models are expected to support tool use! When they do support tool use, we encourage support for that in their chat template, but I'm not sure if models like Deepseek-R1 are trained to use tools.
cc @aymeric-roucher for agentic workflows, though! | 2,821 | 301 |
qgallouedec | 2025-02-10T14:52:11 | @winglian just to point out a different approach: #2730 | 2,818 | 302 |
winglian | 2025-02-10T15:04:15 | @qgallouedec The downside there is that you're limited to the lora support in vllm, which means no DoRA support. This approach almost any peft adapter type could be used. While LoRA does converge pretty quickly too compared to full parameter training, dora seems to be more performant.
<img width="1327" alt="Screenshot 2025-02-10 at 9 25 06 AM" src="https://github.com/user-attachments/assets/b37e7935-a18b-4ed1-8146-8578041b1d5c" />
| 2,818 | 303 |
qgallouedec | 2025-02-10T15:06:04 | This seems quite reasonable, thank you for the clear explanation.
| 2,818 | 304 |
qgallouedec | 2025-02-10T15:09:47 | Another pointer that could be useful:
> It is possible to call `model.merge_adapter` (optionally with `adapter_names` argument), then `model.state_dict()`, then `model.unmerge_adapter`.
> The `state_dict` may require some clean up though, depending on what you need to do with it (I couldn't infer that from the PR).
> By clean up, I mean: After `merge_and_unload` the model looks like the base model. But `merge_adapter` keeps the LoRA structure, with the wrapped base model, LoRA weights etc. still being present in the `state_dict`.
From @BenjaminBossan | 2,818 | 305 |
winglian | 2025-02-10T16:05:52 | I tried
```
unwrapped_model.merge_and_unload()
state_dict = unwrapped_model.base_model.model.state_dict()
unwrapped_model.unmerge_adapter()
```
but the state dict results still has the prefix of `base_model.model.` | 2,818 | 306 |
qgallouedec | 2025-02-10T18:30:41 | I've added the suggested modification to this branch: https://github.com/huggingface/trl/pull/2725 it seems to work...! EDIT: DORA included | 2,818 | 307 |
BenjaminBossan | 2025-02-11T11:25:26 | > I've added the suggested modification to this branch: #2725 it seems to work...! EDIT: DORA included
Nice, I added a comment there. Hopefully, one of these branches can be merged soon :) | 2,818 | 308 |
winglian | 2025-02-13T00:45:53 | I re-did this PR to account for the other changes, and also updated the test to use lora. | 2,818 | 309 |
qgallouedec | 2025-02-13T13:41:38 | thanks for the followup @BenjaminBossan ! | 2,818 | 310 |
HuggingFaceDocBuilderDev | 2025-02-13T13:52:31 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2818). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,818 | 311 |
Subsets and Splits