user
stringlengths 3
28
| created_at
timestamp[us]date 2020-04-01 09:48:12
2025-07-30 20:59:07
| body
stringlengths 1
173k
| issue_number
int64 1
3.81k
| __index_level_0__
int64 0
11.8k
|
|---|---|---|---|---|
HuggingFaceDocBuilderDev
| 2025-02-24T15:21:16 |
The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2947). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
| 2,947 | 8,700 |
HuggingFaceDocBuilderDev
| 2025-02-24T10:26:55 |
The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2946). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
| 2,946 | 8,701 |
HuggingFaceDocBuilderDev
| 2025-02-24T09:19:15 |
The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2945). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
| 2,945 | 8,702 |
qgallouedec
| 2025-02-24T09:33:47 |
> will both packed datasets be cached independently
Yes!
```python
from trl import SFTConfig, SFTTrainer
from datasets import load_dataset
dataset = load_dataset("trl-lib/Capybara", split="train[:10%]")
# Processes the dataset
training_args = SFTConfig(output_dir="Qwen/Qwen2.5-0.5B-SFT", max_length=128, packing=True)
trainer = SFTTrainer(
args=training_args,
model="Qwen/Qwen2.5-0.5B",
train_dataset=dataset,
)
# Processes the dataset as well
training_args = SFTConfig(output_dir="Qwen/Qwen2.5-0.5B-SFT", max_length=256, packing=True)
trainer = SFTTrainer(
args=training_args,
model="Qwen/Qwen2.5-0.5B",
train_dataset=dataset,
)
# Uses the cache!
training_args = SFTConfig(output_dir="Qwen/Qwen2.5-0.5B-SFT", max_length=128, packing=True)
trainer = SFTTrainer(
args=training_args,
model="Qwen/Qwen2.5-0.5B",
train_dataset=dataset,
)
```
| 2,945 | 8,703 |
davidhughhenrymack
| 2025-02-24T18:19:24 |
To perform batch generation, you must have right-aligned (e.g. left padded) token strings. Then, by nature of the generate operation, it will append to those strings with some generations hitting end-of-stream before others, causing right padding.
It is possible to run a sequential operation to re-pack all of this (which flash_Attention2 would like...) but in practice not necessary unless you have wildly different sequence lengths and small memory
| 2,944 | 8,704 |
Alex-Songs
| 2025-03-07T03:39:48 |
hello,
Does GRPO currently not support the scenario where batch size > 1, meaning multiple prompts are involved? In such a case, is left padding required?
@davidhughhenrymack
| 2,944 | 8,705 |
davidhughhenrymack
| 2025-03-07T16:03:22 |
It does support multiple prompts, so left padding is needed.
| 2,944 | 8,706 |
qgallouedec
| 2025-02-24T07:37:54 |
Pi = pi_old at this stage. See the line just above in the algorithm (line 6).
| 2,943 | 8,707 |
helloword12345678
| 2025-02-24T07:51:56 |
In my opinion, self.ref_model serves as the old model because the red-marked mathematical formula involves importance sampling. If you were to use the current model (self.model), importance sampling would not be necessary. Regarding your point, line 6 refers to a periodic update, as mentioned in the issue https://github.com/huggingface/trl/issues/2684，By default, this update is turned off.
| 2,943 | 8,708 |
qgallouedec
| 2025-02-24T07:56:30 |
No pi_old and pi_ref are different. pi_ref is only used for kL
| 2,943 | 8,709 |
helloword12345678
| 2025-02-24T08:25:19 |
thanks, Yes pi_old different with pi_ref！ but is pi equals to pi_old in trl code
| 2,943 | 8,710 |
qgallouedec
| 2025-02-24T08:55:28 |
> thanks, Yes pi_old different with pi_ref！ but is pi equals to pi_old in trl code
indeed, when at generation stage, $\pi_\theta = \pi_\mathrm{old}$, that's why we use $\pi_\theta$.
Then, you start to optimise, and $\pi_\theta$ starts to differ from $\pi_\mathrm{old}$. But we've saved the $\pi_\mathrm{old}(o_i)$ (`old_per_token_logps`), so it's not an issue, we use this for importance sampling.
| 2,943 | 8,711 |
qgallouedec
| 2025-02-24T07:43:51 |
Indeed, GRPO doesn't support IterableDataset. I don't think there is an easy fix. Is it blocking for you?
| 2,942 | 8,712 |
Marsella8
| 2025-02-26T04:39:04 |
> Indeed, GRPO doesn't support IterableDataset. I don't think there is an easy fix. Is it blocking for you?
I am currently working on a refinement pipeline, so I need to dynamically change the training data after each epoch. I've tried to get around it by subclassing Dataset and having a Mutable Dataset that would randomly sample from my desired dataset, though this has not been working. Any advice on how to fix this issue? Thank you
| 2,942 | 8,713 |
jiaweiHu-XDU
| 2025-02-26T10:17:35 |
> 事实上，GRPO 不支持 IterableDataset。我认为没有简单的解决办法。它对你来说是阻塞的吗？
Does TRL currently support multi-modal GRPO training? and DPO
| 2,942 | 8,714 |
nsntiw
| 2025-03-11T03:51:33 |
I encountered this problem with iterable datasets, adding `dispatch_batches=False` to GRPOConfig got rid of it.
If relevant to anyone, the generators of iterabledatasets need additional handling for batch sizes>1, this is my implementation:
```
def my_generator(ds, count=[0], batch_size=4):
while True:
i = count[0] // batch_size
yield {'prompt': ds[i]['prompt'], 'answer': ds[i]['answer']}
count[0] += 1
```
| 2,942 | 8,715 |
qgallouedec
| 2025-02-24T07:49:08 |
This is not supported and you'll need to fork trl to have such feature
| 2,941 | 8,716 |
HuggingFaceDocBuilderDev
| 2025-02-23T19:25:39 |
The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2940). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
| 2,940 | 8,717 |
edbeeching
| 2025-02-24T08:52:49 |
@kashif @qgallouedec @lewtun ,
I think Liger models are now supported nativately in in transformers if the `use_liger_kernel==True` flag is set, perhaps we can drop the support for this in the `SFTTrainer` and use the native transformers implementation?
| 2,940 | 8,718 |
kashif
| 2025-02-24T09:13:40 |
i think so too... we will need to pin the transformer version but yes should be a better soluton
| 2,940 | 8,719 |
qgallouedec
| 2025-02-24T11:41:07 |
Thanks @edbeeching!
> i think so too... we will need to pin the transformer version but yes should be a better soluton
After checking, it seems like `use_liger_kernel` exists for at least 4.46, which is the min version in TRL. So we shouldn't need to bump transformers
https://github.com/huggingface/transformers/blob/052e652d6d53c2b26ffde87e039b723949a53493/src/transformers/training_args.py#L1521
| 2,940 | 8,720 |
qgallouedec
| 2025-02-24T11:43:06 |
it was introduced in 4.45
| 2,940 | 8,721 |
lewtun
| 2025-02-24T15:18:41 |
Thanks for the pointer to `transformers`! Just to confirm, the current PR is fine to merge since the model init is taken care of in this line? https://github.com/huggingface/trl/pull/2940/files#r1967113819
If yes, feel free to merge if I'm offline :)
| 2,940 | 8,722 |
kashif
| 2025-02-23T18:42:51 |
@DanFosing which version of TRL are you using?
| 2,939 | 8,723 |
DanFosing
| 2025-02-23T19:02:31 |
I experienced this issue with both v0.15.1 and with the alpha version downloaded using:
`pip install git+https://github.com/huggingface/trl.git`
| 2,939 | 8,724 |
DanFosing
| 2025-02-23T19:05:41 |
Oh and I forgot to mention, max_seq_length didn't seem to work for me for some reason, the warning says it will be deprecated in v.0.20.0 but are you sure it wasn't deprecated already? (that's why I added a comment there in the code, but it's not related to the main fix)
| 2,939 | 8,725 |
kashif
| 2025-02-24T17:27:26 |
@DanFosing ok so kindly remove the `max_seq_length` from the `sft_config.py` and move the chat template logic inside the already defined `if not is_processed:` where then it makes sense... instead of a new `if not is_processed:` block
| 2,939 | 8,726 |
kashif
| 2025-02-24T17:39:56 |
we can fix the warning and say: removed in version `0.16.0`
| 2,939 | 8,727 |
HuggingFaceDocBuilderDev
| 2025-02-24T17:44:15 |
The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2939). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
| 2,939 | 8,728 |
qgallouedec
| 2025-02-24T18:58:21 |
`maybe_apply_chat_template` applies the chat template if needed. hence "maybe". Are encountering a bug? If so what's the traceback?
| 2,939 | 8,729 |
kashif
| 2025-02-24T19:00:50 |
i don't think there is a bug, but please correct me @DanFosing if I am mistaken, the issue is that it's doing this extra work when it's not needed.
| 2,939 | 8,730 |
qgallouedec
| 2025-02-24T19:01:05 |
> Oh and I forgot to mention, max_seq_length didn't seem to work for me for some reason
WDYM "didn't seem to work"? Same question, is an exception raised? if so what's the traceback?
Have you tried to pull the very last commits? Could be related to #2947
| 2,939 | 8,731 |
qgallouedec
| 2025-02-24T19:03:13 |
> i don't think there is a bug, but please correct me @DanFosing if I am mistaken, the issue is that it's doing this extra work when it's not needed.
for clarification, the only extra work done is iterating through the dataset:
https://github.com/huggingface/trl/blob/5c0591319646c171e8ea213d1692058f4bf68ead/trl/data_utils.py#L218-L221
which is usually very fast
| 2,939 | 8,732 |
qgallouedec
| 2025-02-24T19:05:34 |
That being said, I'm ok to add the `if not is_processed:` to avoid extra logging/iteration
| 2,939 | 8,733 |
qgallouedec
| 2025-02-24T20:34:21 |
@bot /style
| 2,939 | 8,734 |
github-actions[bot]
| 2025-02-24T20:34:47 |
Style fixes have been applied. [View the workflow run here](https://github.com/huggingface/trl/actions/runs/13507301969).
| 2,939 | 8,735 |
HuggingFaceDocBuilderDev
| 2025-02-23T16:37:47 |
The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2938). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
| 2,938 | 8,736 |
qgallouedec
| 2025-02-23T16:14:16 |
Thanks for reporting, please provide a refrence code and your system info
| 2,937 | 8,737 |
yzhdut
| 2025-02-24T02:02:38 |
> Thanks for reporting, please provide a refrence code and your system info
Ok Thanks for your reply
system info：
gpu：
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.40.07 Driver Version: 550.40.07 CUDA Version: 12.4 |
|-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 4090 Off | 00000000:01:00.0 Off | Off |
| 44% 79C P2 434W / 450W | 22731MiB / 24564MiB | 100% Default |
`Package Version
----------------------------- ------------
accelerate 1.4.0
aiofiles 23.2.1
aiohttp 3.9.1
aiosignal 1.3.1
altair 5.2.0
annotated-types 0.6.0
antlr4-python3-runtime 4.9.3
anyio 4.2.0
async-timeout 4.0.3
attrs 23.1.0
auto_gptq 0.7.1
bitsandbytes 0.45.2
blinker 1.7.0
cachetools 5.3.2
certifi 2023.11.17
charset-normalizer 3.3.2
click 8.1.7
cmake 3.27.9
colorama 0.4.6
contourpy 1.2.0
cpm-kernels 1.0.11
cycler 0.12.1
datasets 3.3.1
deepspeed 0.12.4
dill 0.3.7
docstring-parser 0.15
et-xmlfile 1.1.0
exceptiongroup 1.2.0
fastapi 0.108.0
ffmpy 0.3.1
filelock 3.13.1
fonttools 4.47.0
frozenlist 1.4.1
fsspec 2023.10.0
gekko 1.2.1
gitdb 4.0.11
GitPython 3.1.40
gradio 3.38.0
gradio_client 0.8.0
h11 0.14.0
hjson 3.1.0
httpcore 1.0.2
httpx 0.26.0
huggingface-hub 0.29.0
idna 3.6
importlib-metadata 6.11.0
importlib-resources 6.1.1
jieba 0.42.1
Jinja2 3.1.2
joblib 1.3.2
jsonschema 4.20.0
jsonschema-specifications 2023.11.2
kiwisolver 1.4.5
linkify-it-py 2.0.3
lit 17.0.6
markdown-it-py 2.2.0
MarkupSafe 2.1.3
matplotlib 3.8.2
mdit-py-plugins 0.3.3
mdurl 0.1.2
mpmath 1.3.0
multidict 6.0.4
multiprocess 0.70.15
networkx 3.2.1
ninja 1.11.1.1
nltk 3.8.1
numpy 1.26.2
omegaconf 2.3.0
openpyxl 3.1.2
optimum 1.24.0
orjson 3.9.10
packaging 23.2
pandas 2.1.4
peft 0.10.0
Pillow 10.1.0
pip 25.0.1
protobuf 4.25.1
psutil 5.9.6
py-cpuinfo 9.0.0
pyarrow 19.0.1
pyarrow-hotfix 0.6
pydantic 2.5.2
pydantic_core 2.14.5
pydeck 0.8.1b0
pydub 0.25.1
Pygments 2.17.2
pynvml 11.5.0
pyparsing 3.1.1
pyproject 1.3.1
python-dateutil 2.8.2
python-multipart 0.0.6
pytz 2023.3.post1
PyYAML 6.0.1
referencing 0.32.0
regex 2023.10.3
requests 2.32.3
rich 13.7.0
rouge 1.0.1
rouge-chinese 1.0.3
rpds-py 0.13.2
safetensors 0.5.2
scipy 1.11.4
semantic-version 2.10.0
sentencepiece 0.1.99
setuptools 75.8.0
shellingham 1.5.4
shtab 1.6.5
six 1.16.0
smmap 5.0.1
sniffio 1.3.0
sse-starlette 1.8.2
starlette 0.32.0.post1
streamlit 1.29.0
sympy 1.12
tenacity 8.2.3
tiktoken 0.5.2
tokenizers 0.20.3
toml 0.10.2
tomlkit 0.12.0
toolz 0.12.0
torch 2.0.0+cu118
torchvision 0.15.0+cu118
tornado 6.4
tqdm 4.67.1
transformers 4.46.0
transformers-stream-generator 0.0.4
triton 3.2.0
trl 0.15.1
typer 0.12.3
typing_extensions 4.9.0
tyro 0.6.3
tzdata 2024.2
tzlocal 5.2
uc-micro-py 1.0.3
urllib3 2.1.0
uvicorn 0.25.0
validators 0.22.0
watchdog 3.0.0
websockets 11.0.3
wheel 0.41.2
xxhash 3.4.1
yarl 1.9.4
zipp 3.17.0
zstandard 0.23.0`
reference code：
`
with open('data/RLFTqa_delete_some_answer_is_null.json', 'r') as f:
data = json.load(f)
random.shuffle(data)
data = data[0:200]
formatted_data = []
for item in data:
formatted_data.append({
"prompt": item["question"],
"reference": "||".join(item.get("answer", []))
})
n_total = len(formatted_data)
n_train = int(0.7 * n_total)
n_eval = int(0.15 * n_total)
n_test = n_total - n_train - n_eval
train_data = formatted_data[:1]
eval_data = formatted_data[n_train:n_train+n_eval]
test_data = formatted_data[n_train+n_eval:]
dataset_features = Features({
"prompt": Value("string"),
"reference": Value("string")
})
train_dataset = Dataset.from_list(train_data, features=dataset_features)
eval_dataset = Dataset.from_list(eval_data, features=dataset_features)
test_dataset = Dataset.from_list(test_data, features=dataset_features)
local_model_path = "./Qwen2.5-7B-Instruct-GPTQ-Int4"
tokenizer = AutoTokenizer.from_pretrained(local_model_path)
tokenizer.pad_token = tokenizer.eos_token
base_model = AutoModelForCausalLM.from_pretrained(
local_model_path,
# quantization_config=quant_config,
device_map="auto",
trust_remote_code=True,
torch_dtype=torch.float16
)
base_model.config.gradient_checkpointing = False
lora_config = LoraConfig(
r=16, # 适当增加秩维度
lora_alpha=32,
target_modules=["q_proj", "k_proj", "v_proj", "o_proj"], # 扩展目标模块
lora_dropout=0.1,
bias="none",
task_type="CAUSAL_LM",
)
model = prepare_model_for_kbit_training(base_model)
model = get_peft_model(model, lora_config)
grpo_config = GRPOConfig(
num_generations=8, # 每组生成数量
max_completion_length=512, # 最大生成长度
gradient_accumulation_steps=2,
learning_rate=3e-5,
logging_steps=5,
logging_first_step = True,
save_steps=50,
output_dir="./grpo_checkpoints",
logging_dir="./grpo_checkpoints/log",
log_level="info",
log_completions=True,
eval_strategy="steps",
eval_steps=40,
do_predict=True,
temperature=0.7,
gradient_checkpointing=False,
use_vllm=False
)
model.bfloat16()
trainer = GRPOTrainer(
model=model,
args=grpo_config,
reward_funcs=[reward_func, format_reward_func], # 直接传递奖励函数
train_dataset=train_dataset,
eval_dataset=eval_dataset,
processing_class=tokenizer,
)
train_output = trainer.predict(train_dataset)
trainer.train()
model.save_pretrained("grpo_lora_adapter")`
As mentioned above, the length of train_det is 1. Before training, I first call trainer. predict (train_dataset). Then I printed the completion of the model in the library function, and the obtained location is as follows:
xxx/lib/python3.10/site-packages/trl/trainer/grpo_trainer.py：
` def _prepare_inputs(self, inputs: dict[str, Union[torch.Tensor, Any]]) -> dict[str, Union[torch.Tensor, Any]]:
device = self.accelerator.device
prompts = [x["prompt"] for x in inputs]
prompts_text = [maybe_apply_chat_template(example, self.processing_class)["prompt"] for example in inputs]
prompt_inputs = self.processing_class(
prompts_text, return_tensors="pt", padding=True, padding_side="left", add_special_tokens=False
)
prompt_inputs = super()._prepare_inputs(prompt_inputs)
prompt_ids, prompt_mask = prompt_inputs["input_ids"], prompt_inputs["attention_mask"]
if self.max_prompt_length is not None:
prompt_ids = prompt_ids[:, -self.max_prompt_length :]
prompt_mask = prompt_mask[:, -self.max_prompt_length :]
# Generate completions using either vLLM or regular generation
if self.args.use_vllm:
print("Using VLM model")
# First, have main process load weights if needed
if self.state.global_step != self._last_loaded_step:
self._move_model_to_vllm()
self._last_loaded_step = self.state.global_step
# Generate completions using vLLM: gather all prompts and use them in a single call in the main process
all_prompts_text = gather_object(prompts_text)
if self.accelerator.is_main_process:
outputs = self.llm.generate(all_prompts_text, sampling_params=self.sampling_params, use_tqdm=False)
completion_ids = [out.token_ids for completions in outputs for out in completions.outputs]
else:
completion_ids = [None] * len(all_prompts_text)
# Broadcast the completions from the main process to all processes, ensuring each process receives its
# corresponding slice.
completion_ids = broadcast_object_list(completion_ids, from_process=0)
process_slice = slice(
self.accelerator.process_index * len(prompts),
(self.accelerator.process_index + 1) * len(prompts),
)
completion_ids = completion_ids[process_slice]
# Pad the completions, and concatenate them with the prompts
completion_ids = [torch.tensor(ids, device=device) for ids in completion_ids]
completion_ids = pad(completion_ids, padding_value=self.processing_class.pad_token_id)
prompt_completion_ids = torch.cat([prompt_ids, completion_ids], dim=1)
else:
# Regular generation path
with unwrap_model_for_generation(self.model, self.accelerator) as unwrapped_model:
prompt_completion_ids = unwrapped_model.generate(
prompt_ids, attention_mask=prompt_mask, generation_config=self.generation_config
)
# Compute prompt length and extract completion ids
prompt_length = prompt_ids.size(1)
prompt_ids = prompt_completion_ids[:, :prompt_length]
completion_ids = prompt_completion_ids[:, prompt_length:]
# Mask everything after the first EOS token
is_eos = completion_ids == self.processing_class.eos_token_id
eos_idx = torch.full((is_eos.size(0),), is_eos.size(1), dtype=torch.long, device=device)
eos_idx[is_eos.any(dim=1)] = is_eos.int().argmax(dim=1)[is_eos.any(dim=1)]
sequence_indices = torch.arange(is_eos.size(1), device=device).expand(is_eos.size(0), -1)
completion_mask = (sequence_indices <= eos_idx.unsqueeze(1)).int()
# Concatenate prompt_mask with completion_mask for logit computation
attention_mask = torch.cat([prompt_mask, completion_mask], dim=1) # (B*G, P+C)
logits_to_keep = completion_ids.size(1) # we only need to compute the logits for the completion tokens
with torch.inference_mode():
if self.ref_model is not None:
ref_per_token_logps = self._get_per_token_logps(
self.ref_model, prompt_completion_ids, attention_mask, logits_to_keep
)
else:
with self.accelerator.unwrap_model(self.model).disable_adapter():
ref_per_token_logps = self._get_per_token_logps(
self.model, prompt_completion_ids, attention_mask, logits_to_keep
)
# Decode the generated completions
completions_text = self.processing_class.batch_decode(completion_ids, skip_special_tokens=True)
print("Completions text:", completions_text)`
As I mentioned, when predicting, the model shows basic problem-solving ability. However, for the same input, during training, completion appears to significantly reduce model performance due to issues with irrelevant statements and tasks, as well as repetition.
And I found that the log doesn't seem to be saved together with the checkpoint. Is it possible that it will only be saved after the training is completed
| 2,937 | 8,738 |
Wang-Xiaodong1899
| 2025-03-12T08:49:56 |
any update?
| 2,937 | 8,739 |
Wang-Xiaodong1899
| 2025-03-12T08:57:24 |
I have also found that when the input prompt is too long, such as the length of my prompt input_id is 3964, I find that the complete result produces strange results. The normal inference model, on the other hand, is the normal output. Here I am using the model Qwen2-VL-7B.
strange completion such as `[<output>```python\n```]`, or `'<output>2003:000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000'`
Any suggestions?
| 2,937 | 8,740 |
Wang-Xiaodong1899
| 2025-03-12T08:57:45 |
@qgallouedec
| 2,937 | 8,741 |
HuggingFaceDocBuilderDev
| 2025-02-23T13:17:42 |
The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2936). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
| 2,936 | 8,742 |
qgallouedec
| 2025-02-23T16:55:14 |
Super cool PR @August-murr! I'll check in details asap.
> I faced out-of-memory (OOM) issues and couldn't train an agent for more complex tasks.
Can you share your code?
> In the end, we could write a blog post or report to showcase its effectiveness.
Definitely!
> While the results look good, they don’t represent a practical use case
Do you have another simple while practical use case?
| 2,936 | 8,743 |
qgallouedec
| 2025-02-23T17:17:44 |
Can you add some tests and docs as well?
| 2,936 | 8,744 |
August-murr
| 2025-02-23T19:50:09 |
> Can you share your code?
[Kaggle Notebook](https://www.kaggle.com/code/augustmurr/training-agent-to-generate-code)
the biggest issue was really Kaggles 2xT4 having little VRAM.
I did try PEFT but then couldn't use it properly with vllm then decided to do full model instead.
>
> Do you have another simple while practical use case?
no, not simpler than that.
| 2,936 | 8,745 |
qgallouedec
| 2025-02-23T21:22:24 |
Not sure when you tested it but peft + vllm should be fixed now
| 2,936 | 8,746 |
August-murr
| 2025-02-25T12:34:15 |
@qgallouedec I don't get why the tests failed.
8 tests failed with error:
`module 'torch' has no attribute 'hip'`
| 2,936 | 8,747 |
qgallouedec
| 2025-02-28T12:09:49 |
It's because of liger, merging main should solve this
| 2,936 | 8,748 |
kashif
| 2025-02-28T12:12:02 |
they have a fix in the 0.5.4 version
| 2,936 | 8,749 |
qgallouedec
| 2025-02-28T12:36:15 |
Actually,I think it's 0.5.4 that contains the bug, that's why we pinned to 0.5.3: https://github.com/huggingface/trl/pull/2952
| 2,936 | 8,750 |
qgallouedec
| 2025-02-28T18:35:20 |
don't bother too much with windows. It a test fails, you can skip it
| 2,936 | 8,751 |
August-murr
| 2025-03-02T09:27:04 |
@qgallouedec is there anything else needed??
| 2,936 | 8,752 |
August-murr
| 2025-03-03T18:16:32 |
just added parallel code execution using asyncio in https://github.com/huggingface/trl/pull/2936/commits/21973857d6f09868bfbac9c24c8fe5b77129c0b2 to make code generation more scalable.
Right now, we’re working with E2B, but if we find that alternatives like CodeSandbox.io, Modal, or Daytona are cheaper or faster, we’ll create wrappers for those too.
| 2,936 | 8,753 |
August-murr
| 2025-03-04T14:16:49 |
> Not sure when you tested it but peft + vllm should be fixed now
@qgallouedec
Can you share some code for a training run that worked for you? I've been using the example from [docs](https://huggingface.co/docs/trl/main/en/grpo_trainer#quick-start) plus a lora config and `use_vllm=True` and it has not been working. not for agents btw, just training model with the reward function in the docs.
| 2,936 | 8,754 |
qgallouedec
| 2025-03-04T14:58:41 |
Thanks for the feedback, what's the traceback?
| 2,936 | 8,755 |
August-murr
| 2025-03-04T15:25:34 |
> Thanks for the feedback, what's the traceback?
there is no error, Everything's running, but the reward doesn't really seem to improve —the reward is all over the place.
[Notebook](https://www.kaggle.com/code/augustmurr/training-model-with-vllm-peft-for-shorter-length)
This exact notebook, without the PEFTConfig, works perfectly, pretty much nailing it with no response (getting the max reward) after just a few steps. So, I’m pretty convinced it has something to do with the PEFT and not the reward function or the data or other hyper-parameters.
| 2,936 | 8,756 |
qgallouedec
| 2025-03-04T15:59:58 |
Possibly related? https://github.com/huggingface/trl/pull/2873
| 2,936 | 8,757 |
August-murr
| 2025-03-04T17:37:32 |
> Possibly related? #2873
Can you share the code you used in https://github.com/huggingface/trl/pull/2873#issuecomment-2663793035
| 2,936 | 8,758 |
qgallouedec
| 2025-02-23T12:11:35 |
@bot /style
| 2,935 | 8,759 |
HuggingFaceDocBuilderDev
| 2025-02-23T12:15:44 |
The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2935). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
| 2,935 | 8,760 |
HuggingFaceDocBuilderDev
| 2025-02-23T12:04:37 |
The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2934). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
| 2,934 | 8,761 |
qgallouedec
| 2025-02-23T13:29:50 |
@bot /style
| 2,934 | 8,762 |
qgallouedec
| 2025-02-23T13:32:58 |
@bot /style
| 2,934 | 8,763 |
qgallouedec
| 2025-02-23T13:35:00 |
@bot /style
| 2,934 | 8,764 |
qgallouedec
| 2025-02-23T16:35:17 |
@bot /style
| 2,934 | 8,765 |
qgallouedec
| 2025-02-23T16:36:09 |
@bot /style
| 2,934 | 8,766 |
github-actions[bot]
| 2025-02-23T16:36:27 |
Style fixes have been applied. [View the workflow run here](https://github.com/huggingface/trl/actions/runs/13484908517).
| 2,934 | 8,767 |
sdpkjc
| 2025-02-25T01:46:31 |
I’ve encountered the same issue too—the KL divergence value is abnormal. But I haven’t been able to pinpoint the problem yet, so I’m glad to see this. I have a question: why do specify tokens exist? Doesn’t VLLM already skip special tokens when generating token sequences? 🤔
| 2,933 | 8,768 |
glowwormX
| 2025-03-05T02:34:47 |
I also encountered a sudden increase in KL. Is there any solution now? @kalomaze
| 2,933 | 8,769 |
qgallouedec
| 2025-02-23T17:21:27 |
Thanks, just a minor suggestion :)
| 2,932 | 8,770 |
qgallouedec
| 2025-02-23T17:21:35 |
@bot /style
| 2,932 | 8,771 |
cuiyuhao1996
| 2025-02-24T07:39:04 |
How is this solved? I also encountered the same problem.
| 2,931 | 8,772 |
qgallouedec
| 2025-02-24T08:00:12 |
Try to upgrade trl, it should solve the issue
| 2,931 | 8,773 |
qgallouedec
| 2025-02-24T08:00:13 |
Try to upgrade trl, it should solve the issue
| 2,931 | 8,774 |
qgallouedec
| 2025-02-24T22:24:11 |
Any reason you closed #2857? (duplicate)
For ref, I suggested using histogram: https://github.com/huggingface/trl/pull/2857#issuecomment-2666618701
| 2,930 | 8,775 |
qgallouedec
| 2025-02-27T10:38:44 |
@bot /style
| 2,930 | 8,776 |
qgallouedec
| 2025-02-27T10:39:29 |
@bot /style
| 2,930 | 8,777 |
github-actions[bot]
| 2025-02-27T10:39:52 |
Style fixes have been applied. [View the workflow run here](https://github.com/huggingface/trl/actions/runs/13564466917).
| 2,930 | 8,778 |
HuggingFaceDocBuilderDev
| 2025-02-22T12:13:19 |
The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2929). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
| 2,929 | 8,779 |
qgallouedec
| 2025-02-24T22:25:52 |
Thanks a lot @ghrua! Huge work! I'll test it asap. It definitely makes sense :)
| 2,929 | 8,780 |
qgallouedec
| 2025-02-25T11:16:05 |
Have you tried with vLLM 0.7.3 btw?
| 2,929 | 8,781 |
ghrua
| 2025-02-25T11:21:49 |
Hey @qgallouedec! Thanks for reviewing my commit.
~~I am sorry that I haven't tried vLLM 0.7.3. Let me do a quick check. Please wait a moment.~~
I checked that the env for this commit is based on `0.7.3`. Sorry for the misremember.
```
▶ pip list | grep vllm
vllm 0.7.3
```
| 2,929 | 8,782 |
qgallouedec
| 2025-02-25T12:46:17 |
So it works with this version?
I'm asking because currently, main branch hangs at some point when you use vLLM 0.7.3, and I'm curious to know if your PR solves it
| 2,929 | 8,783 |
ghrua
| 2025-02-25T13:45:31 |
> So it works with this version? I'm asking because currently, main branch hangs at some point when you use vLLM 0.7.3, and I'm curious to know if your PR solves it
Yes, my commit works well with vLLM 0.7.3. I re-run the code to double check it.
Two parts may be helpful:
1. I use a context manager to control the device that each vLLM can access during inference, because the vLLM may mis-use the devices of distributed training processes: https://github.com/huggingface/trl/blob/79e9af0c7e179f9be5ee1a0f0e88b3e4ac82cb70/trl/trainer/grpo_trainer.py#L670
2. Three additional patches are used for a smoother initialization of vLLM (though not very elegant 😅): https://github.com/huggingface/trl/blob/79e9af0c7e179f9be5ee1a0f0e88b3e4ac82cb70/trl/trainer/grpo_trainer.py#L474
| 2,929 | 8,784 |
ghrua
| 2025-02-27T14:32:39 |
> Ok now I understand better what you are trying to do.
> Why do you need $K$ processes to control $K$ instances of vLLM? Couldn't you do everything in the main process?
That was my initial plan... I tried to initialise a vllm_list for K vllm instances in the main process and use the ThreadExecutor to call the inference in parallel. However, it will meet lots of errors (I forget the details, but seems to be the inconsistency between q and kv cache.) I didn't check the detailed code of vLLM, but those errors seems come from the conflict of some shared static variables. And I find this issue: https://github.com/vllm-project/vllm/issues/1676. Please let me know if I misunderstood anything.
Afterwards, I change my strategy by initialising K instances in K processes.
| 2,929 | 8,785 |
SeiunSky0131
| 2025-03-10T02:31:38 |
> This PR allows users to leverage K GPUs for inference and N − K GPUs for training.
Hi, that's a great work! However, when I test this code on my server equipped with 8 NVIDIA A100 80G, with 6 GPUs(GPU 0-5) for training and 2 GPUs(GPU 6-7) for vLLM engine, I find that GPU 7's utilization is always 0 throughout the training process.(See the blue line of GPU7 in the following figure)
<img width="1457" alt="截屏2025-03-10 10 15 57" src="https://github.com/user-attachments/assets/141c8e3a-6720-4d7e-805e-c12f1a2260eb" />
It seems that no generation tasks are allocated to GPU 7. Could you check whether this situation applies to your code?
| 2,929 | 8,786 |
loki369loki
| 2025-03-11T02:27:47 |
> > This PR allows users to leverage K GPUs for inference and N − K GPUs for training.
>
> Hi, that's a great work! However, when I test this code on my server equipped with 8 NVIDIA A100 80G, with 6 GPUs(GPU 0-5) for training and 2 GPUs(GPU 6-7) for vLLM engine, I find that GPU 7's utilization is always 0 throughout the training process.(See the blue line of GPU7 in the following figure) <img alt="截屏2025-03-10 10 15 57" width="1457" src="https://private-user-images.githubusercontent.com/87608284/420733648-141c8e3a-6720-4d7e-805e-c12f1a2260eb.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.-gHuPaDuCEydqN1ULQNmov69ELI5g5Or3sHloLXYIKY">
>
> It seems that no generation tasks are allocated to GPU 7. Could you check whether this situation applies to your code?
Encountered the same issue, GPU7 is not functioning properly for vllm inference.
| 2,929 | 8,787 |
skepsun
| 2025-03-19T09:07:23 |
It seems not to solve the problem that vllm hangs when we want to use large models (for example, 32b) which could only be placed on multiple cards.
| 2,929 | 8,788 |
qgallouedec
| 2025-03-22T18:41:00 |
Closed via #3094
| 2,929 | 8,789 |
mehdiataei
| 2025-02-22T05:57:35 |
Fixed by downgrading to transformers==4.49.0 from dev.
| 2,928 | 8,790 |
dignfei
| 2025-02-22T12:22:32 |
I can train Qwen2.5-3B on 4090 ( 24gb )
| 2,927 | 8,791 |
dignfei
| 2025-02-22T12:22:58 |
Qwen2.5-7B only need 2x H20(80GB)
| 2,927 | 8,792 |
Tuziking
| 2025-02-22T12:39:57 |
> Qwen2.5-7B only need 2x H20(80GB)
I'm sorry to bother you. But can you share your code with me or help me to find the bug in my code as follow? I reference
willccbb's code to train.
``` python
import re
import torch
from datasets import load_dataset, Dataset
from transformers import AutoTokenizer, AutoModelForCausalLM
from trl import GRPOConfig, GRPOTrainer
from peft import LoraConfig, get_peft_model, TaskType
import wandb
import logging
from scripts.utils.replace_grpo_trainer import trigger
#
logging.basicConfig(
filename="GRPO-Qwen2.5-7B.log", # 保存的日志文件名
level=logging.INFO, # 日志等级
format="%(asctime)s - %(message)s", # 日志格式
datefmt="%Y-%m-%d %H:%M:%S"
)
logger = logging.getLogger("logger")
# Load and prep dataset
SYSTEM_PROMPT = """
A conversation between User and Assistant. The user asks a question, and the Assistant solves it. The assistant first thinks about the think process in the mind and then provides the user with the answer. The think process and answer are enclosed within <think> </think> and <answer> </answer> tags, respectively, i.e., <think> think process here </think> <answer> answer here </answer>
"""
XML_COT_FORMAT = """
<think>
{think}
</think>
<answer>
{answer}
</answer>
"""
def extract_xml_answer(text: str) -> str:
answer = text.split("<answer>")[-1]
answer = answer.split("</answer>")[0]
return answer.strip()
def extract_hash_answer(text: str) -> str | None:
if "####" not in text:
return None
return text.split("####")[1].strip()
# uncomment middle messages for 1-shot prompting
def get_gsm8k_questions(split = "train") -> Dataset:
data = load_dataset('dataset/gsm8k', 'main')[split] # type: ignore
data = data.map(lambda x: { # type: ignore
'prompt': [
{'role': 'system', 'content': SYSTEM_PROMPT},
# {'role': 'user', 'content': 'What is the largest single-digit prime number?'},
# {'role': 'assistant', 'content': XML_COT_FORMAT.format(
# think="9 is divisble by 3 and 8 is divisible by 2, but 7 is prime.",
# answer="7"
# )},
{'role': 'user', 'content': x['question']}
],
'answer': extract_hash_answer(x['answer'])
}) # type: ignore
return data # type: ignore
dataset = get_gsm8k_questions()
# print(dataset[0])
# Reward functions
def correctness_reward_func(prompts, completions, answer, **kwargs) -> list[float]:
responses = [completion[0]['content'] for completion in completions]
q = prompts[0][-1]['content']
extracted_responses = [extract_xml_answer(r) for r in responses]
print(len(responses), len(extracted_responses), len(answer))
# for response, extracted_response, _answer in zip(responses, extracted_responses, answer):
logger.info('-'*20)
logger.info(f"Question:\n{q}")
logger.info(f"Answer:\n{answer[0]}")
logger.info(f"Response:\n{responses[0]}")
logger.info(f"Extracted:\n{extracted_responses[0]}")
logger.info(f"Correctness: {1.0 if extracted_responses[0] == answer[0] else 0.0}")
# wandb.log({"Correctness": 1.0 if extracted_responses[0] == answer[0] else 0.0})
# print('-'*20, f"Question:\n{q}", f"\nAnswer:\n{answer[0]}", f"\nResponse:\n{responses[0]}", f"\nExtracted:\n{extracted_responses[0]}")
return [2.0 if r == a else 0.0 for r, a in zip(extracted_responses, answer)]
def int_reward_func(completions, **kwargs) -> list[float]:
# print("int_reward_func")
responses = [completion[0]['content'] for completion in completions]
extracted_responses = [extract_xml_answer(r) for r in responses]
return [0.5 if r.isdigit() else 0.0 for r in extracted_responses]
# def strict_format_reward_func(completions, **kwargs) -> list[float]:
# """Reward function that checks if the completion has a specific format."""
# # print("strict_format_reward_func")
# pattern = r"^<think>\n.*?\n</think>\n<answer>\n.*?\n</answer>\n$"
# responses = [completion[0]["content"] for completion in completions]
# matches = [re.match(pattern, r) for r in responses]
# return [0.5 if match else 0.0 for match in matches]
def soft_format_reward_func(completions, **kwargs) -> list[float]:
# print("soft_format_reward_func")
"""Reward function that checks if the completion has a specific format."""
pattern = r"<think>.*?</think>\s*<answer>.*?</answer>"
responses = [completion[0]["content"] for completion in completions]
matches = [re.match(pattern, r) for r in responses]
return [0.5 if match else 0.0 for match in matches]
def count_xml(text) -> float:
count = 0.0
if text.count("<think>\n") == 1:
count += 0.125
if text.count("\n</think>\n") == 1:
count += 0.125
if text.count("\n<answer>\n") == 1:
count += 0.125
count -= len(text.split("\n</answer>\n")[-1])*0.001
if text.count("\n</answer>") == 1:
count += 0.125
count -= (len(text.split("\n</answer>")[-1]) - 1)*0.001
return count
def xmlcount_reward_func(completions, **kwargs) -> list[float]:
# print("xmlcount_reward_func")
contents = [completion[0]["content"] for completion in completions]
return [count_xml(c) for c in contents]
output_dir = "outputs/Qwen2.5-7B-GRPO"
model_name = "models/Qwen2.5-7B"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16)
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=2,
gradient_accumulation_steps=4,
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"
)
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,
)
trainer.train()
trainer.save_model(output_dir)
```
| 2,927 | 8,793 |
Tuziking
| 2025-02-23T09:25:21 |
I test the problem in different per_device_train_batch_size and num_generations, I find that if I use one H20, `per_device_train_batch_size==4, num_generations==4`to train, it can continue some steps before OOM. But if I use 3 x H20, `per_device_train_batch_size==2, num_generations==6`to train, OOM occur early. I don't know why more H20 to train but OOM is occur early. The error log is as follow:
```
[rank0]:[W223 17:38:57.742650271 reducer.cpp:1400] Warning: find_unused_parameters=True was specified in DDP constructor, but did not find any unused parameters in the forward pass. This flag results in an extra traversal of the autograd graph every iteration, which can adversely affect performance. If your model indeed never has any unused parameters in the forward pass, consider turning this flag off. Note that this warning may be a false positive if your model has flow control causing later iterations to have unused parameters. (function operator())
[rank1]:[W223 17:38:57.742968833 reducer.cpp:1400] Warning: find_unused_parameters=True was specified in DDP constructor, but did not find any unused parameters in the forward pass. This flag results in an extra traversal of the autograd graph every iteration, which can adversely affect performance. If your model indeed never has any unused parameters in the forward pass, consider turning this flag off. Note that this warning may be a false positive if your model has flow control causing later iterations to have unused parameters. (function operator())
[rank2]:[W223 17:38:57.745997855 reducer.cpp:1400] Warning: find_unused_parameters=True was specified in DDP constructor, but did not find any unused parameters in the forward pass. This flag results in an extra traversal of the autograd graph every iteration, which can adversely affect performance. If your model indeed never has any unused parameters in the forward pass, consider turning this flag off. Note that this warning may be a false positive if your model has flow control causing later iterations to have unused parameters. (function operator())
Traceback (most recent call last):
File "/online1/sc100010/sc100010/qb_project/MARL/trl_GRPO_train.py", line 176, in <module>
trainer.train()
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/transformers/trainer.py", line 2241, in train
return inner_training_loop(
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/transformers/trainer.py", line 2599, in _inner_training_loop
self.optimizer.step()
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/accelerate/optimizer.py", line 178, in step
self.optimizer.step(closure)
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/lr_scheduler.py", line 137, in wrapper
return func.__get__(opt, opt.__class__)(*args, **kwargs)
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/optimizer.py", line 487, in wrapper
out = func(*args, **kwargs)
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/optimizer.py", line 91, in _use_grad
ret = func(self, *args, **kwargs)
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/adamw.py", line 220, in step
adamw(
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/optimizer.py", line 154, in maybe_fallback
return func(*args, **kwargs)
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/adamw.py", line 782, in adamw
func(
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/adamw.py", line 606, in _multi_tensor_adamw
exp_avg_sq_sqrt = torch._foreach_sqrt(device_exp_avg_sqs)
torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 130.00 MiB. GPU 0 has a total capacity of 94.99 GiB of which 87.19 MiB is free. Including non-PyTorch memory, this process has 94.90 GiB memory in use. Of the allocated memory 90.98 GiB is allocated by PyTorch, and 2.01 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
[rank0]: Traceback (most recent call last):
[rank0]: File "/online1/sc100010/sc100010/qb_project/MARL/trl_GRPO_train.py", line 176, in <module>
[rank0]: trainer.train()
[rank0]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/transformers/trainer.py", line 2241, in train
[rank0]: return inner_training_loop(
[rank0]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/transformers/trainer.py", line 2599, in _inner_training_loop
[rank0]: self.optimizer.step()
[rank0]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/accelerate/optimizer.py", line 178, in step
[rank0]: self.optimizer.step(closure)
[rank0]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/lr_scheduler.py", line 137, in wrapper
[rank0]: return func.__get__(opt, opt.__class__)(*args, **kwargs)
[rank0]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/optimizer.py", line 487, in wrapper
[rank0]: out = func(*args, **kwargs)
[rank0]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/optimizer.py", line 91, in _use_grad
[rank0]: ret = func(self, *args, **kwargs)
[rank0]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/adamw.py", line 220, in step
[rank0]: adamw(
[rank0]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/optimizer.py", line 154, in maybe_fallback
[rank0]: return func(*args, **kwargs)
[rank0]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/adamw.py", line 782, in adamw
[rank0]: func(
[rank0]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/adamw.py", line 606, in _multi_tensor_adamw
[rank0]: exp_avg_sq_sqrt = torch._foreach_sqrt(device_exp_avg_sqs)
[rank0]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 130.00 MiB. GPU 0 has a total capacity of 94.99 GiB of which 87.19 MiB is free. Including non-PyTorch memory, this process has 94.90 GiB memory in use. Of the allocated memory 90.98 GiB is allocated by PyTorch, and 2.01 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
[rank1]: Traceback (most recent call last):
[rank1]: File "/online1/sc100010/sc100010/qb_project/MARL/trl_GRPO_train.py", line 176, in <module>
[rank1]: trainer.train()
[rank1]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/transformers/trainer.py", line 2241, in train
[rank1]: return inner_training_loop(
[rank1]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/transformers/trainer.py", line 2599, in _inner_training_loop
[rank1]: self.optimizer.step()
[rank1]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/accelerate/optimizer.py", line 178, in step
[rank1]: self.optimizer.step(closure)
[rank1]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/lr_scheduler.py", line 137, in wrapper
[rank1]: return func.__get__(opt, opt.__class__)(*args, **kwargs)
[rank1]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/optimizer.py", line 487, in wrapper
[rank1]: out = func(*args, **kwargs)
[rank1]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/optimizer.py", line 91, in _use_grad
[rank1]: ret = func(self, *args, **kwargs)
[rank1]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/adamw.py", line 220, in step
[rank1]: adamw(
[rank1]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/optimizer.py", line 154, in maybe_fallback
[rank1]: return func(*args, **kwargs)
[rank1]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/adamw.py", line 782, in adamw
[rank1]: func(
[rank1]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/adamw.py", line 606, in _multi_tensor_adamw
[rank1]: exp_avg_sq_sqrt = torch._foreach_sqrt(device_exp_avg_sqs)
[rank1]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 130.00 MiB. GPU 1 has a total capacity of 94.99 GiB of which 127.19 MiB is free. Including non-PyTorch memory, this process has 94.86 GiB memory in use. Of the allocated memory 92.27 GiB is allocated by PyTorch, and 682.32 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
[rank2]: Traceback (most recent call last):
[rank2]: File "/online1/sc100010/sc100010/qb_project/MARL/trl_GRPO_train.py", line 176, in <module>
[rank2]: trainer.train()
[rank2]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/transformers/trainer.py", line 2241, in train
[rank2]: return inner_training_loop(
[rank2]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/transformers/trainer.py", line 2599, in _inner_training_loop
[rank2]: self.optimizer.step()
[rank2]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/accelerate/optimizer.py", line 178, in step
[rank2]: self.optimizer.step(closure)
[rank2]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/lr_scheduler.py", line 137, in wrapper
[rank2]: return func.__get__(opt, opt.__class__)(*args, **kwargs)
[rank2]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/optimizer.py", line 487, in wrapper
[rank2]: out = func(*args, **kwargs)
[rank2]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/optimizer.py", line 91, in _use_grad
[rank2]: ret = func(self, *args, **kwargs)
[rank2]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/adamw.py", line 220, in step
[rank2]: adamw(
[rank2]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/optimizer.py", line 154, in maybe_fallback
[rank2]: return func(*args, **kwargs)
[rank2]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/adamw.py", line 782, in adamw
[rank2]: func(
[rank2]: File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/optim/adamw.py", line 606, in _multi_tensor_adamw
[rank2]: exp_avg_sq_sqrt = torch._foreach_sqrt(device_exp_avg_sqs)
[rank2]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 130.00 MiB. GPU 2 has a total capacity of 94.99 GiB of which 1.19 MiB is free. Including non-PyTorch memory, this process has 94.98 GiB memory in use. Of the allocated memory 92.43 GiB is allocated by PyTorch, and 703.27 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
wandb:
wandb: 🚀 View run outputs/Qwen2.5-7B-GRPO at: https://wandb.ai/bobo1398861921-nus/huggingface/runs/eorh7fyx
wandb: Find logs at: ../../../../../../../../online1/sc100010/sc100010/qb_project/MARL/wandb/run-20250223_173847-eorh7fyx/logs
W0223 17:39:00.122000 40366 /online1/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/api.py:897] Sending process 40439 closing signal SIGTERM
W0223 17:39:00.125000 40366 /online1/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/api.py:897] Sending process 40440 closing signal SIGTERM
E0223 17:39:00.741000 40366 /online1/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/api.py:869] failed (exitcode: 1) local_rank: 2 (pid: 40441) of binary: /home/export/base/sc100010/sc100010/.conda/envs/torch/bin/python
Traceback (most recent call last):
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/bin/accelerate", line 8, in <module>
sys.exit(main())
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/accelerate/commands/accelerate_cli.py", line 48, in main
args.func(args)
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/accelerate/commands/launch.py", line 1163, in launch_command
multi_gpu_launcher(args)
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/accelerate/commands/launch.py", line 792, in multi_gpu_launcher
distrib_run.run(args)
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/distributed/run.py", line 910, in run
elastic_launch(
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 138, in __call__
return launch_agent(self._config, self._entrypoint, list(args))
File "/home/export/base/sc100010/sc100010/.conda/envs/torch/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 269, in launch_agent
raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
============================================================
trl_GRPO_train.py FAILED
------------------------------------------------------------
Failures:
<NO_OTHER_FAILURES>
------------------------------------------------------------
Root Cause (first observed failure):
[0]:
time : 2025-02-23_17:39:00
host : gpu018
rank : 2 (local_rank: 2)
exitcode : 1 (pid: 40441)
error_file: <N/A>
traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
============================================================
```
| 2,927 | 8,794 |
willccbb
| 2025-02-24T23:58:47 |
try setting `vllm_gpu_memory_utilization=0.7` (default is 0.9)
GRPO needs room on the inference node to load the model weights so they can be used to update the vLLM engine. 3B will take 6GB, so it fits in the ~8GB available by default, but 7B won't fit unless you increase the headroom.
| 2,927 | 8,795 |
Tuziking
| 2025-02-27T03:34:52 |
> try setting `vllm_gpu_memory_utilization=0.7` (default is 0.9)尝试设置 `vllm_gpu_memory_utilization=0.7` （默认为 0.9）
>
> GRPO needs room on the inference node to load the model weights so they can be used to update the vLLM engine. 3B will take 6GB, so it fits in the ~8GB available by default, but 7B won't fit unless you increase the headroom.GRPO 需要在推理节点上留出空间来加载模型权重，以便它们可以用来更新 vLLM 引擎。3B 将占用 6GB，因此它适合默认的~8GB 可用空间，但 7B 将无法适应，除非你增加可用空间。
I'm sorry to bother you, but this doesn't seem to help. It still encounters the OOM error at the second step. 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.
| 2,927 | 8,796 |
Fox237
| 2025-03-14T02:41:19 |
same problem
| 2,927 | 8,797 |
kashif
| 2025-02-21T16:47:42 |
i might need to update the loss on the liger side with respect to the multi-turn PR in TRL's GRPOTrainer
| 2,926 | 8,798 |
SalmanMohammadi
| 2025-02-21T17:10:03 |
> i might need to update the loss on the liger side with respect to the multi-turn PR in TRL's GRPOTrainer
Let me know if I can help test!
| 2,926 | 8,799 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.