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
|
|---|---|---|---|---|
pxyWaterMoon | 2025-03-02T15:20:08 | I met the same problem while trianing Alpaca-7B with GRPO on A100，the trl environments are as follows:
```
- Platform: Linux-3.10.0-957.el7.x86_64-x86_64-with-glibc2.31
- Python version: 3.11.11
- TRL version: 0.16.0.dev0
- PyTorch version: 2.6.0
- CUDA device(s): NVIDIA A100-SXM4-80GB, NVIDIA A100-SXM4-80GB, NVIDIA A100-SXM4-80GB, NVIDIA A100-SXM4-80GB, NVIDIA A100-SXM4-80GB, NVIDIA A100-SXM4-80GB, NVIDIA A100-SXM4-80GB, NVIDIA A100-SXM4-80GB
- Transformers version: 4.49.0
- Accelerate version: 1.3.0
- Accelerate config: not found
- Datasets version: 3.3.0
- HF Hub version: 0.28.1
- bitsandbytes version: not installed
- DeepSpeed version: 0.16.3
- Diffusers version: not installed
- Liger-Kernel version: not installed
- LLM-Blender version: not installed
- OpenAI version: not installed
- PEFT version: not installed
- vLLM version: not installed
```
| 2,996 | 412 |
zsychina | 2025-03-02T18:11:46 | Another report
```bash
0%| | 2/87543 [00:18<223:50:10, 9.20s/it]../aten/src/ATen/native/cuda/TensorCompare.cu:110: _assert_async_cuda_kernel: block: [0,0,0], thread: [0,0,0] Assertion `probability tensor contains either `inf`, `nan` or element < 0` failed.
Traceback (most recent call last):
File "/home/zhusiyuan/test_trl/example.py", line 27, in <module>
trainer.train()
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/transformers/trainer.py", line 2241, in train
return inner_training_loop(
^^^^^^^^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/transformers/trainer.py", line 2548, in _inner_training_loop
tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/transformers/trainer.py", line 3692, in training_step
inputs = self._prepare_inputs(inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/trl/trainer/grpo_trainer.py", line 564, in _prepare_inputs
prompt_completion_ids = unwrapped_model.generate(
^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/transformers/generation/utils.py", line 2223, in generate
result = self._sample(
^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/transformers/generation/utils.py", line 3200, in _sample
while self._has_unfinished_sequences(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/transformers/generation/utils.py", line 2401, in _has_unfinished_sequences
elif this_peer_finished:
^^^^^^^^^^^^^^^^^^
RuntimeError: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.
Traceback (most recent call last):
File "/home/zhusiyuan/test_trl/example.py", line 27, in <module>
trainer.train()
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/transformers/trainer.py", line 2241, in train
return inner_training_loop(
^^^^^^^^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/transformers/trainer.py", line 2548, in _inner_training_loop
tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/transformers/trainer.py", line 3692, in training_step
inputs = self._prepare_inputs(inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/trl/trainer/grpo_trainer.py", line 564, in _prepare_inputs
prompt_completion_ids = unwrapped_model.generate(
^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/transformers/generation/utils.py", line 2223, in generate
result = self._sample(
^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/transformers/generation/utils.py", line 3200, in _sample
while self._has_unfinished_sequences(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zhusiyuan/miniconda3/envs/trl/lib/python3.12/site-packages/transformers/generation/utils.py", line 2401, in _has_unfinished_sequences
elif this_peer_finished:
^^^^^^^^^^^^^^^^^^
RuntimeError: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.
``` | 2,996 | 413 |
zsychina | 2025-03-03T06:56:47 | Guys, I think I found the problem.
For distributed training, it has to be
```
accelerate lauch grpo_example.py
```
while
```
python -u grpo_example.py
```
is ok for single gpu training, but may cause above errors in distributed training.
| 2,996 | 414 |
qgallouedec | 2025-03-01T08:41:02 | Indeed. This comes from https://github.com/huggingface/trl/pull/2881.
Our idea on this is, unless we find that it gives worst results, we should align with classsical loss normalization (global).
Have you compared the two options? If so, the results would be very useful. | 2,995 | 415 |
tchang1997 | 2025-03-03T21:30:49 | On a related note, is there a reason why the per token loss is globally normalized (L950 of [`grpo_trainer.py`](https://github.com/huggingface/trl/blob/7442d42c21697fd6c0998a75e7478ed4b40490be/trl/trainer/grpo_trainer.py)), but the KL term continues to use per-sequence normalization (L956 of [`grpo_trainer.py`](https://github.com/huggingface/trl/blob/7442d42c21697fd6c0998a75e7478ed4b40490be/trl/trainer/grpo_trainer.py))?
Looks like the [GRPO paper (Eq. 3)](https://arxiv.org/pdf/2402.03300) sequence-normalizes both (expanding the KL divergence term), so I wonder these should be consistent (i.e., both global normalization, or both sequence-norm, not a mix). | 2,995 | 416 |
qgallouedec | 2025-03-03T22:17:16 | Actually L956 is just logging. But you're right that it should use a consistent normalization. Would you like to open a PR to fix this line? | 2,995 | 417 |
tchang1997 | 2025-03-03T22:51:04 | Ah, I see that now. Anyway — opened a [PR as discussed](https://github.com/huggingface/trl/pull/3004)! | 2,995 | 418 |
HuggingFaceDocBuilderDev | 2025-02-28T18:45:42 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2993). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,993 | 419 |
HuggingFaceDocBuilderDev | 2025-02-28T14:00:42 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2991). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,991 | 420 |
qgallouedec | 2025-02-28T14:46:29 | This could help to understand:
https://github.com/huggingface/trl/blob/3d94e4e25c5e63ac56a62cdb9dd5f6ec4153e3c0/trl/trainer/grpo_trainer.py#L580-L598
Feel free to get back to me if it's still not clear.
> Additionally, I noticed that the current implementation determines divisibility by computing the factors, which seems a bit redundant. It might be more concise to directly check:
Indeed, but I wanted to provide the user with values that could solve the issue:
https://github.com/huggingface/trl/blob/3d94e4e25c5e63ac56a62cdb9dd5f6ec4153e3c0/trl/trainer/grpo_trainer.py#L424
| 2,990 | 421 |
Facico | 2025-02-28T14:54:55 | Thanks！ | 2,990 | 422 |
qgallouedec | 2025-02-28T12:48:54 | Thanks for the suggestion. In fact it's already been suggested in #2728, and I think this solution should actually be avoided: https://github.com/huggingface/trl/pull/2728#issuecomment-2635166424
| 2,989 | 423 |
nopepper | 2025-02-28T12:52:54 | @qgallouedec Interesting, thanks for the links, I hadn't seen the discussion. I was working on a project that really benefited from custom sampling kwargs (seq2seq task where target significantly overlaps input) and thought it would be useful for others. Would it be possible to add the generation arguments by name then to GRPOConfig?
Specifically: `top_p`, `top_k`, `min_p`, `repetition_penalty` and `length_penalty` are shared by both `SamplingParams` and `GenerationConfig` | 2,989 | 424 |
qgallouedec | 2025-02-28T14:40:33 | Yes I think it makes sense. I'm not sure that they all share the same default in vLLM and transformers though. You'll need to check | 2,989 | 425 |
nopepper | 2025-02-28T15:32:56 | I added them to a new section called `Parameters that control generation` together with `temperature`. We get around the diverging default values problem by just making them optional and not passing any argument that's `None` :D | 2,989 | 426 |
qgallouedec | 2025-03-04T22:19:14 | Defaults values
| | transformers | vLLM |
| ------------------ | ------------------ | -------------- |
| top_p | 1.0 | 1.0 |
| top_k | 50 (supports None) | -1 (means all) |
| min_p | None | 0.0 |
| repetition_penalty | 1.0 | 1.0 |
| 2,989 | 427 |
HuggingFaceDocBuilderDev | 2025-03-04T23:13:21 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2989). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,989 | 428 |
HuggingFaceDocBuilderDev | 2025-02-28T11:19:49 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2987). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,987 | 429 |
HuggingFaceDocBuilderDev | 2025-02-28T10:46:11 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2986). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,986 | 430 |
qgallouedec | 2025-02-28T08:54:52 | Which trainer? | 2,985 | 431 |
cyr0930 | 2025-02-28T08:59:55 | ah sorry it's DPOTrainer | 2,985 | 432 |
kevinlu1248 | 2025-03-08T02:36:10 | It also seems to not work with Deepspeed stage 1/2, getting:
```
[rank0]: RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)
``` | 2,985 | 433 |
kevinlu1248 | 2025-03-08T03:11:04 | Found a fix for stage 1 & 2 by explicitly initializing the reference model, it looks like it gets garbage collected off of VRAM as well after the initial log prob computation is complete.
| 2,985 | 434 |
jamesbraza | 2025-02-28T07:05:09 | Is there a standard solution for DeepSpeed tests in CI? I think this is the first integration test for DeepSpeed added to the repo.
In the future, we can expand it to cover https://github.com/huggingface/trl/pull/2871 and https://github.com/huggingface/trl/pull/2963. | 2,984 | 435 |
qgallouedec | 2025-03-24T16:36:18 | I can't reproduce, this runs on my side:
```python
from datasets import load_dataset
from trl import GRPOConfig, GRPOTrainer
def dummy_reward_func(completions, **kwargs):
return [0.0] * len(completions)
dataset = load_dataset("trl-lib/tldr")
training_args = GRPOConfig(output_dir="2983", num_iterations=3)
trainer = GRPOTrainer(
model="Qwen/Qwen2-0.5B-Instruct",
reward_funcs=dummy_reward_func,
args=training_args,
train_dataset=dataset["train"],
)
trainer.train()
```
Maybe try to upgrade TRL?
If you still get the issue, please provide a MRE. | 2,983 | 436 |
Andcircle | 2025-03-24T20:53:19 | Thanks @qgallouedec, this is long time back
I saw the new release note for 0.16.0, num_iteration is added as a feature. I will try it out again
| 2,983 | 437 |
qgallouedec | 2025-03-24T21:30:56 | Thanks, yes, sorry for the delay, the number of open issues is overwhelming, I'm trying to catch up. | 2,983 | 438 |
HuggingFaceDocBuilderDev | 2025-02-28T10:49:46 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2982). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,982 | 439 |
jhinpan | 2025-02-28T01:36:57 | Current Plan of adding alternative inference backend as SGLang:
0. Add global flag as `use_sglang`
1. Init the offline Engine in `def __init__()`
2. Implement function of `_update_sglang_engine_weights` to function like `_move_model_to_vllm`
3. Receive generation requests in `_prepare_inputs()`
4. Shut down SGLang engine after speeding up the generation
| 2,981 | 440 |
jhinpan | 2025-02-28T01:46:48 | Current Issue Summary:
- Successful Standalone Inference:
The SGLang Server works correctly as an inference engine when run in a separate terminal.
- Distributed Engine Initialization Failure:
When switching to using the SGLang engine within the distributed trainer context, the distributed initialization of the offline SGLang engine fails.
- Logging Insights:
- All processes successfully pass the explicit barriers, and the main process reaches the SGLang engine initialization.
- However, no log messages are observed indicating that the generation call (`engine.generate()`) is executed or that a response is returned.
- Conclusion:
These observations suggest that execution is stalling either inside or before the call to engine.generate(). The issue likely lies within SGLang’s internal behavior when operating under the distributed trainer context. | 2,981 | 441 |
jhinpan | 2025-03-01T02:41:57 | It seems the issue is now narrowing down to: `accelerate launch` has some conflicts with SGLang offline engine intialization. Need to make more checks these two days.
- Firstly, maybe we can check whether Accelerate + SGLang, i.e. without things in TRL, behaves normally
- If so, maybe we can check whether manually creating processes (instead of using Accelerate) + SGLang works; otherwise we may need to either make minimal sample or check what global things does TRL change
- If so, maybe we can check environment variable differences between the "manual+SGLang" and "Accelerate+SGLang", and try to remove the different ones, and see whether it works
- If so, we may further dig that single (or several) env var to know what is happening | 2,981 | 442 |
zhaochenyang20 | 2025-03-11T01:52:36 | Amazing work, I will review it quickly.
| 2,981 | 443 |
qgallouedec | 2025-03-12T02:40:12 | Is this PR ready for review? There seems to be many files/lines used for dev that aren't cleaned.
Also, can you make sure to run the pre-commit? It seems like you use a custom format config and it results to many lines changed but not related to the PR | 2,981 | 444 |
zhaochenyang20 | 2025-03-12T02:56:21 | @jhinpan Rebase with the main? Then I can ask them to review. | 2,981 | 445 |
jhinpan | 2025-03-12T03:11:07 | > Is this PR ready for review? There seems to be many files/lines used for dev that aren't cleaned.
> Also, can you make sure to run the pre-commit? It seems like you use a custom format config and it results to many lines changed but not related to the PR
I just cleaned all those dev files and run the pre-commit. Hope that work. Feel free to lmk whether those testing scripts needed to be removed. @qgallouedec @zhaochenyang20 | 2,981 | 446 |
zhaochenyang20 | 2025-03-12T04:21:06 | > This branch has conflicts that must be resolved
> Changes can be cleanly merged.
@jhinpan | 2,981 | 447 |
nopepper | 2025-02-28T07:57:32 | I was wrong, I had `beta=0.0` in all my experiments. Setting `beta=0.001` was enough to prevent the gradient explosion. Perhaps we shouldn't suggest that option in the docs so prominently?
 | 2,980 | 448 |
qgallouedec | 2025-02-28T15:24:36 | I suspected that this could produce surprising results on a long run. https://github.com/huggingface/trl/pull/2806#issuecomment-2645941307
Would you recommend adding some sort of warning in the documentation? | 2,980 | 449 |
nopepper | 2025-02-28T15:30:41 | Sounds good. Perhaps something like this?
```bash
KL coefficient. If `0.0`, the reference model is not loaded, reducing memory usage and improving training speed, but may be numerically unstable for long training runs.
``` | 2,980 | 450 |
qgallouedec | 2025-02-28T15:56:08 | Looks good! Are you willing to open a PR? | 2,980 | 451 |
kenluozhenyu | 2025-03-03T14:50:08 | Same problem, TRL version: 0.15.1, on Windows 11 | 2,979 | 452 |
Tony-yzj | 2025-03-19T03:51:08 | same problem, TRL version: 0.15.2， CUDA12.1, on Windows 11 | 2,979 | 453 |
qgallouedec | 2025-02-27T22:16:39 | Thanks for the suggestion, do you have any such tutorial in mind? | 2,978 | 454 |
ParagEkbote | 2025-02-28T14:23:52 | Yes, we could use a model like smolm2 and use DPO or ORPO with a custom dataset to display the integration. WDYT?
cc: @qgallouedec | 2,978 | 455 |
ParagEkbote | 2025-03-11T18:26:23 | Gentle ping
cc: @qgallouedec | 2,978 | 456 |
qgallouedec | 2025-03-11T18:31:43 | Hi, sorry if this is unclear. In fact this part of the documentation belongs to the community, to share its notebooks. That's why it's called “community tutorials”. If you have a notebook to add, we can add it.
But I think what you're really looking for is documentation for the various TRL integrations? If yes, then we have an “integration" section in the doc. It's not finished yet, and we're very open to contributions. | 2,978 | 457 |
qgallouedec | 2025-02-27T14:41:06 | Try to downgrade vLLM to 0.7.2. See #2952 | 2,977 | 458 |
zaddy6 | 2025-02-27T14:42:40 | vLLM 0.7.2 doesnt support the new phi4 mini any other workaround apart from downgrading? | 2,977 | 459 |
qgallouedec | 2025-02-27T14:15:14 | Answer here: https://github.com/huggingface/open-r1/issues/239#issuecomment-2646297851 😊
| 2,976 | 460 |
L1n111ya | 2025-02-27T14:25:03 | > Answer here: [huggingface/open-r1#239 (comment)](https://github.com/huggingface/open-r1/issues/239#issuecomment-2646297851) 😊
Thank you for your reply. I know the advantage function is 0, but what puzzles me is that since that's the case, the loss only has the KL divergence term. Does it not update based on the reward function? How does the reward converge? | 2,976 | 461 |
iamansinha | 2025-03-01T14:03:22 | This might help: [huggingface/open-r1/issues/239#issuecomment-2692241946](https://github.com/huggingface/open-r1/issues/239#issuecomment-2692241946) | 2,976 | 462 |
L1n111ya | 2025-03-02T04:08:38 | > This might help: [huggingface/open-r1/issues/239#issuecomment-2692241946](https://github.com/huggingface/open-r1/issues/239#issuecomment-2692241946)
Thank you for your reply, it has resolved my doubts. | 2,976 | 463 |
HuggingFaceDocBuilderDev | 2025-02-27T09:15:48 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2975). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,975 | 464 |
nbasyl | 2025-02-27T09:49:52 | Thanks for double-checking this! | 2,974 | 465 |
HuggingFaceDocBuilderDev | 2025-02-27T11:10:24 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2974). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,974 | 466 |
canghongjian | 2025-02-27T08:09:29 | Two h20 is enough to start the training, but the speed is quite slow. | 2,972 | 467 |
Tuziking | 2025-02-27T08:42:26 | > Two h20 is enough to start the training, but the speed is quite slow.两个 h20 足够开始训练，但速度相当慢。
I use GRPOtrainer to train with two H20 , but cuda out of memory(I tried more H20 to train but also failed). What's strange is that during the first step, each GPU only uses 40GB, but at the second step, it suddenly fills up and causes the OOM error. I don't know if this is a bug. My configuration is as follows:
``` python
training_args = GRPOConfig(
output_dir=output_dir,
learning_rate=5e-6,
adam_beta1 = 0.9,
adam_beta2 = 0.99,
weight_decay = 0.1,
warmup_ratio = 0.1,
lr_scheduler_type='cosine',
logging_steps=1,
bf16=True,
per_device_train_batch_size=1,
gradient_accumulation_steps=1,
num_generations=2,
max_prompt_length=256,
max_completion_length=512,
num_train_epochs=1,
save_steps=100,
max_grad_norm=0.1,
log_on_each_node=False,
# use_vllm=False,
report_to="wandb",
vllm_gpu_memory_utilization=0.5,
)
trainer = GRPOTrainer(
model = model,
# reward_funcs = xmlcount_reward_func,
reward_funcs = [
xmlcount_reward_func,
soft_format_reward_func,
# strict_format_reward_func,
int_reward_func,
correctness_reward_func,
],
args = training_args,
train_dataset = dataset,
)
``` | 2,972 | 468 |
Fox237 | 2025-03-14T02:22:35 | 同样的问题，解决了吗？ | 2,972 | 469 |
baibizhe | 2025-03-01T15:58:05 | no.i've stuck on this problem for a while. the fact is that trl only support tensor parallel=1 currently. you could switch to some other framework such as verl and openrlhf. they work smoothly | 2,971 | 470 |
xz259 | 2025-03-15T15:17:45 | I think the prompts over the gradient_accumulation_steps should be batched together. 7 generations is under utilizing the inference GPU. | 2,971 | 471 |
lyh1028 | 2025-03-09T03:11:59 | have the same question | 2,970 | 472 |
qgallouedec | 2025-02-27T13:28:40 | Thanks @logicaltrojan!, I took the opportunity to fix it everywhere! Once the CI is green we can merge :) | 2,969 | 473 |
qgallouedec | 2025-02-27T13:28:58 | @bot /style | 2,969 | 474 |
github-actions[bot] | 2025-02-27T13:29:19 | Style fixes have been applied. [View the workflow run here](https://github.com/huggingface/trl/actions/runs/13567545048). | 2,969 | 475 |
HuggingFaceDocBuilderDev | 2025-02-27T13:32:47 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2969). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,969 | 476 |
qgallouedec | 2025-02-26T20:44:38 | Yes, the solution is just to rename the argument :) | 2,968 | 477 |
ErikKankaTrea | 2025-02-27T03:36:14 | Thanks!!! | 2,968 | 478 |
HuggingFaceDocBuilderDev | 2025-02-26T09:54:36 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2966). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,966 | 479 |
jojo23333 | 2025-02-27T04:06:43 | Hi @qgallouedec ,thanks for your great contribution! in this [line](https://github.com/huggingface/trl/blob/019fc6dbaa03b888f9d5c1845f0f690da8ed310c/trl/trainer/grpo_trainer.py#L752) , I wonder why not get logp when sampling but instead do the inference again? vllm should also support getting the logp value. | 2,966 | 480 |
qgallouedec | 2025-02-27T08:41:32 | Yes, but we need the gradient, and the logp returned by vllm are just the values | 2,966 | 481 |
zetian1025 | 2025-03-03T11:07:39 | Same issue. The problem seems to be related to deepseed stage 3 (use stage 2 will avoid the problem) | 2,965 | 482 |
Sampson1107 | 2025-03-12T11:56:54 | Same issue. Same stage3 | 2,965 | 483 |
yeruoforever | 2025-02-26T14:05:44 | TypeError: '>=' not supported between instances of 'list' and 'tuple' | 2,963 | 484 |
qgallouedec | 2025-02-26T16:21:37 | Thanks a lot @jamesbraza | 2,963 | 485 |
qgallouedec | 2025-02-26T16:24:32 | by the way, is this change needed as well? https://github.com/huggingface/trl/pull/2871 | 2,963 | 486 |
jamesbraza | 2025-02-26T18:55:39 | > by the way, is this change needed as well? #2871
Yes, testing it now. Clearly I didn't test this PR previously, as @yeruoforever reported it had a `TypeError` haha, my bad. | 2,963 | 487 |
jamesbraza | 2025-02-27T07:56:10 | Hi @qgallouedec I have completed my validations, this PR is ready for merge. I had to also pull in:
- https://github.com/huggingface/trl/pull/2871 (thanks for sharing it)
- I hit https://github.com/huggingface/trl/issues/2953, and to fix it, I changed [this line](https://github.com/huggingface/trl/blob/019fc6dbaa03b888f9d5c1845f0f690da8ed310c/trl/trainer/grpo_trainer.py#L752) to `with torch.no_grad()` instead of `with torch.inference_mode()` | 2,963 | 488 |
qgallouedec | 2025-02-27T08:51:54 | Thanks, but I can't see the above mentionned changes. Did you pushed them? | 2,963 | 489 |
jamesbraza | 2025-02-27T19:26:59 | > Thanks, but I can't see the above mentionned changes. Did you pushed them?
The other ones I mentioned weren't related to this PR, so they're not here. They are all here: https://github.com/Future-House/trl/tree/working-grpo-2025-02-27 | 2,963 | 490 |
qgallouedec | 2025-03-04T15:46:55 | Thanks! Just commit the suggestions and we are good to merge.
Usually [allowing maintainer to edit the PR](https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/allowing-changes-to-a-pull-request-branch-created-from-a-fork) makes things easier for us. | 2,963 | 491 |
jamesbraza | 2025-03-04T17:54:28 | > Usually [allowing maintainer to edit the PR](https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/allowing-changes-to-a-pull-request-branch-created-from-a-fork) makes things easier for us.
Yeah I always do, but for some reason that checkbox isn't present in this PR's right panel :/
<img width="340" alt="screenshot of right panel" src="https://github.com/user-attachments/assets/63754c9c-f88e-4117-b6f1-5bdaa4935c4d" />
Regardless, PR is ready for review again | 2,963 | 492 |
qgallouedec | 2025-03-05T13:58:44 | Thanks!
| 2,963 | 493 |
HuggingFaceDocBuilderDev | 2025-03-05T14:03:36 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2963). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,963 | 494 |
gengzijun | 2025-02-26T07:11:19 | I'm having the same issue, do you know how to fix this now | 2,962 | 495 |
nopepper | 2025-02-26T13:50:22 | I've noticed a similar issue. DeepSpeed runs will eventually fail catastrophically and the reward plummets to 0. DDP doesn't work anymore when using `vLLM` and just hangs on the first step. | 2,960 | 496 |
qgallouedec | 2025-02-26T14:14:48 | You should set num_processes to 2 if you have 2 GPUs.
Also, to get comparable results you must ensure that the effective batch size (num gpus*per device batch size*grad accum) is the same.
I'm not sure it will solve your problem, let me know | 2,960 | 497 |
Kfkcome | 2025-02-26T15:29:28 | > You should set num_processes to 2 if you have 2 GPUs. Also, to get comparable results you must ensure that the effective batch size (num gpus_per device batch size_grad accum) is the same. I'm not sure it will solve your problem, let me know
If I set num_processes = 1, the first gpu focus train and the second gpu focus sampling(vllm).
And I use the same args, only change the `export CUDA_VISIBLE_DEVICES=0` to `export CUDA_VISIBLE_DEVICES=0,2`
I got the another test use the same args and one gpu.
The format reward increase very fast just like the one before I test....
 | 2,960 | 498 |
I-l-l-I | 2025-03-12T11:22:54 | @Kfkcome Something similar happened to me. There may be a communication problem between the GPUs. I solved it by replacing
```python
llm_model.load_weights(state_dict.items())
```
in `GRPOTrainer._move_model_to_vllm` with
```python
llm_device = next(llm_model.parameters()).device
for name, param in state_dict.items():
weight = param.to('cpu').to(llm_device)
llm_model.load_weights(weights=[(name, weight)])
del weight
```
| 2,960 | 499 |
HuggingFaceDocBuilderDev | 2025-02-25T13:18:00 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2956). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,956 | 500 |
qgallouedec | 2025-02-25T13:35:22 | Thank for this great work!! It seems very close to GRPO, can you summarize the key differences to make the reviewing a bit easier for me? | 2,955 | 501 |
liziniu | 2025-02-25T15:13:06 | Hi, ReMax differs from GRPO in two key aspects: baseline estimation and advantage calculation.
## Key Conceptual Differences
**Baseline Estimation:**
- GRPO uses the averaged empirical reward as the baseline value
- ReMax simply uses the reward value of a greedy decoding completion as the baseline
**Advantage Calculation:**
- GRPO calculates the grouped mean and standard deviation for normalization
- ReMax does not require this normalization step
## Implementation Details
The implementation of `remax_config.py` is basically same with `grpo_config.py`, with modifications primarily in the trainer code's (`remax_trainer.py`) `_generate_and_score_completions` function (lines 690-880):
### Key Modifications
1. **Lines 690-760:** Modified generation to incorporate greedy decoding for baseline estimation
- Sampling parameters for vllm/HF generation slightly changed to accommodate this
2. **Lines 760-808:** Minimal changes to existing code
3. **Lines 810-860:** Added calculation of rewards for greedy completions
4. **Lines 860-870:** Direct advantage calculation without additional operations like gathering
### Additional Changes
- **`__init__` method:**
- Changed vllm sampling parameters by setting `n = 1`
- This preserves the function of generating multiple completions since prompts are repeated in lines 690-760
- **`compute_loss` method (line 951):**
- Added `dim=1` when calculating averaged loss
- Loss is first normalized across timesteps then across different batches
- This implementation follows the description in the paper
I also provide an introduction to ReMax at the [docs](https://github.com/huggingface/trl/pull/2955/commits/c0fcdd350103d0f22a17645f44de8cf41719cd06?short_path=15a860f#diff-15a860fec3744a3eecf254dff5060ee5d30d3ec245dbc8dfad7417bcc3bc513b).
If you have additional questions, please feel free to let me know. | 2,955 | 502 |
qgallouedec | 2025-02-27T11:07:37 | Thanks! Can you try to integrate the very latest changes in GRPO? | 2,955 | 503 |
liziniu | 2025-03-03T08:27:12 | Hi, I’ve integrated the latest changes from GRPO. Below is a summary of the updates:
- Lines 736-802: Modified the sampling strategy for ReMax to incorporate greedy decoding, which improves baseline estimation for ReMax.
- Lines 857-906: Added reward calculations for greedy completions. These rewards will be used to compute the advantage function.
- Lines 908-915: Implemented a customized advantage calculation specifically tailored for ReMax.
Let me know if you have any questions or need further details! | 2,955 | 504 |
kashif | 2025-03-03T08:31:55 | currently, the remax trainer file is a copy of the remax config file... is that a mistake? | 2,955 | 505 |
liziniu | 2025-03-03T11:44:48 | Hi @kashif, thank you for pointing that out! It was my mistake to copy the wrong content. I’ve now fixed it. | 2,955 | 506 |
liziniu | 2025-03-12T02:36:54 | Hi @qgallouedec
Could you please review the code when you have a moment? If you need any additional information to assist with the review, I’d be happy to provide it.
Thanks in advance! | 2,955 | 507 |
HuggingFaceDocBuilderDev | 2025-02-25T10:14:24 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2954). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,954 | 508 |
qgallouedec | 2025-02-25T10:41:41 | Nice! When it works, can you also add a few lines in https://huggingface.co/docs/trl/en/reducing_memory_usage? 🙏 | 2,954 | 509 |
kashif | 2025-02-25T13:03:41 | sure! | 2,954 | 510 |
casper-hansen | 2025-03-03T16:47:10 | Looks like this PR was parked for now. @kashif did the implementation not work? This is super relevant to me if I am going to use TRL for training long-context reasoners | 2,954 | 511 |
Subsets and Splits