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
|
|---|---|---|---|---|
mirceapricop | 2025-02-08T18:40:24 | Same issue as in the original post:
```
File "/mnt/e/LLM_RL/open-r1/src/open_r1/grpo.py", line 196, in <module>
main(script_args, training_args, model_args)
File "/mnt/e/LLM_RL/open-r1/openr1/lib/python3.11/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
return f(*args, **kwargs)
^^^^^^^^^^^^^^^^^^
File "/mnt/e/LLM_RL/open-r1/src/open_r1/grpo.py", line 150, in main
train_result = trainer.train(resume_from_checkpoint=checkpoint)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/e/LLM_RL/open-r1/openr1/lib/python3.11/site-packages/transformers/trainer.py", line 2184, in train
return inner_training_loop(
^^^^^^^^^^^^^^^^^^^^
File "/mnt/e/LLM_RL/open-r1/openr1/lib/python3.11/site-packages/transformers/trainer.py", line 2490, in _inner_training_loop
tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/e/LLM_RL/open-r1/openr1/lib/python3.11/site-packages/transformers/trainer.py", line 3592, in training_step
inputs = self._prepare_inputs(inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/e/LLM_RL/open-r1/openr1/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py", line 422, in _prepare_inputs
llm_model.load_weights(state_dict.items())
File "/mnt/e/LLM_RL/open-r1/openr1/lib/python3.11/site-packages/vllm/model_executor/models/llama.py", line 594, in load_weights
return loader.load_weights(
^^^^^^^^^^^^^^^^^^^^
File "/mnt/e/LLM_RL/open-r1/openr1/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 237, in load_weights
autoloaded_weights = set(self._load_module("", self.module, weights))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/e/LLM_RL/open-r1/openr1/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 226, in _load_module
raise ValueError(msg)
ValueError: There is no module or parameter named 'base_model' in LlamaForCausalLM
``` | 2,698 | 1,412 |
tchang1997 | 2025-02-08T18:47:53 | I've been working off of a custom fork where I merged in #2730. It dynamically saves/reloads LoRA adapters for generation.
The PR is promising but I ran into an issue w/ `trl`'s automatic vLLM device assignment: the `LLM()` initialization in `GRPOTrainer.__init__` works when device=`cuda:0`, but not when device=`cuda:1` (or other non CUDA:0 devices).
I've filed an [issue with vLLM](https://github.com/vllm-project/vllm/issues/12967) describing this. In the meantime, perhaps we need to force everything to be on 1 GPU (`cuda:0`)? This is probably not feasible w/o H100-sized GPUs, so I'm unable to test this. | 2,698 | 1,413 |
mirceapricop | 2025-02-08T20:44:02 | Note that I encounter the above issue even when running on a single GPU, explicitly setting `vllm_device` to `cuda:0`. In contrast, I seem to be able to run LORA GRPO using Unsloth's patches through their notebook: https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Llama3.1_(8B)-GRPO.ipynb | 2,698 | 1,414 |
Ugadot | 2025-02-09T07:40:02 | Any updates on this issue? | 2,698 | 1,415 |
tahafatih | 2025-02-09T23:26:26 | Same issue here with Phi-4:
AttributeError: 'LlamaForCausalLM' object has no attribute 'load_lora' | 2,698 | 1,416 |
GuodongFan | 2025-02-10T12:51:34 | +1 | 2,698 | 1,417 |
qgallouedec | 2025-02-11T07:30:23 | Hey, just fyi, this is being addressed in #2818 #2725 and #2730. We're working hard on it :blush: | 2,698 | 1,418 |
Maghoumi | 2025-02-13T17:29:23 | With #2818 merged a few minutes ago, I confirm I can run training without running into `ValueError: There is no module or parameter named 'base_model' in LlamaForCausalm` | 2,698 | 1,419 |
qgallouedec | 2025-02-13T17:33:46 | Nice!! | 2,698 | 1,420 |
dodekapod | 2025-02-14T16:22:21 | With [v0.15.0](https://github.com/huggingface/trl/releases/tag/v0.15.0) I still have this error message `ValueError: There is no module or parameter named 'base_model' in LlamaForCausalm`
Here are some more mentions of this: https://github.com/vllm-project/vllm/issues/12961 | 2,698 | 1,421 |
HuggingFaceDocBuilderDev | 2025-01-30T16:05:12 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2697). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,697 | 1,422 |
Superskyyy | 2025-01-31T03:16:06 | You need to downgrade accelerate. Idk somehow it may work on other algorithms but PPO RLOO won't work. | 2,696 | 1,423 |
daehuikim | 2025-01-31T03:19:26 | @Superskyyy
PPOTrainer and RLOOTrainer initiate 'accelerator=Accelerator' in their ```__init__``` method.
(These implementations differ from other trainer class in this lib)
I found ```pip install accelerate==0.34.2``` works for these two trainer classes. :)
| 2,696 | 1,424 |
Superskyyy | 2025-01-31T04:15:11 | > @Superskyyy
>
> PPOTrainer and RLOOTrainer initiate 'accelerator=Accelerator' in their ```__init__``` method.
> (These implementations differ from other trainer class in this lib)
> I found ```pip install accelerate==0.34.2``` works for these two trainer classes. :)
>
Right. It may need a fix up. But in general the hype has really shifted to GRPO recently so you can try GRPO too. That one is much cleaner in implementation and works with HF trainer much easier. Note is resuming from checkpoint also doesn't work for RLOO and PPO since they overrode the training loop. | 2,696 | 1,425 |
Tarak200 | 2025-02-17T13:32:06 | which deepspeed verison are you using?
The accelerate==0.34.2 is incompatible with the deepspeed==0.16.3 because of which I am getting the below error.
AssertionError: no_sync context manager is incompatible with gradient partitioning logic of ZeRO stage 2 | 2,696 | 1,426 |
daehuikim | 2025-02-18T01:01:34 | > which deepspeed verison are you using?
>
> The accelerate==0.34.2 is incompatible with the deepspeed==0.16.3 because of which I am getting the below error. AssertionError: no_sync context manager is incompatible with gradient partitioning logic of ZeRO stage 2
As I know, ZeRO stage 2 does not support gradient accumulate step.
Can you run gradient accumulate step with 1 ? | 2,696 | 1,427 |
Tarak200 | 2025-02-18T07:28:57 | # CUDA Out of Memory Issue with ZeRO Stage 1 and Qwen-3B-Instruct PPO Training
I am encountering a CUDA out-of-memory (OOM) issue while training the `Qwen-3B-Instruct` model using **PPO training** on **8 A100 GPUs** with ZeRO Stage 1. can you help me here ?
### Command Used:
```bash
accelerate launch --config_file examples/accelerate_configs/deepspeed_zero1.yaml \
examples/scripts/ppo/ppo.py \
--dataset_name ./ppo_tokenized_dataset \
--dataset_train_split descriptiveness \
--output_dir models/minimal/ppo \
--num_ppo_epochs 1 \
--num_mini_batches 1 \
--learning_rate 3e-6 \
--per_device_train_batch_size 1 \
--gradient_accumulation_steps 32 \
--total_episodes 10000 \
--model_name_or_path Qwen/Qwen2.5-Coder-3B-Instruct \
--sft_model_path Qwen/Qwen2.5-Coder-3B-Instruct \
--reward_model_path Qwen/Qwen2.5-Coder-3B-Instruct \
--local_rollout_forward_batch_size 1 \
--missing_eos_penalty 1.0
| 2,696 | 1,428 |
daehuikim | 2025-02-18T10:18:10 | > # CUDA Out of Memory Issue with ZeRO Stage 1 and Qwen-3B-Instruct PPO Training
> I am encountering a CUDA out-of-memory (OOM) issue while training the `Qwen-3B-Instruct` model using **PPO training** on **8 A100 GPUs** with ZeRO Stage 1. can you help me here ?
>
> ### Command Used:
> accelerate launch --config_file examples/accelerate_configs/deepspeed_zero1.yaml \
> examples/scripts/ppo/ppo.py \
> --dataset_name ./ppo_tokenized_dataset \
> --dataset_train_split descriptiveness \
> --output_dir models/minimal/ppo \
> --num_ppo_epochs 1 \
> --num_mini_batches 1 \
> --learning_rate 3e-6 \
> --per_device_train_batch_size 1 \
> --gradient_accumulation_steps 32 \
> --total_episodes 10000 \
> --model_name_or_path Qwen/Qwen2.5-Coder-3B-Instruct \
> --sft_model_path Qwen/Qwen2.5-Coder-3B-Instruct \
> --reward_model_path Qwen/Qwen2.5-Coder-3B-Instruct \
> --local_rollout_forward_batch_size 1 \
> --missing_eos_penalty 1.0
Running command does not tell much things.
Are you using Wandb? If you are using that, please see your GPU memory allocated and check distribution of memory.
I guess rank0 got 100% while others remain low.
So many hyperparameters can affect memory utilization (ppo_epoch, max_sequence_length, logit preprocessing...)
You can find paramters from trl docs and you can utilze wandb or can print memory utilization :)
| 2,696 | 1,429 |
HuggingFaceDocBuilderDev | 2025-01-30T15:09:43 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2695). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,695 | 1,430 |
HuggingFaceDocBuilderDev | 2025-01-30T12:15:10 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2694). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,694 | 1,431 |
HuggingFaceDocBuilderDev | 2025-01-30T10:26:49 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2692). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,692 | 1,432 |
HuggingFaceDocBuilderDev | 2025-01-30T08:33:34 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2690). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,690 | 1,433 |
HuggingFaceDocBuilderDev | 2025-01-30T08:15:16 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2689). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,689 | 1,434 |
kpfoley | 2025-01-30T06:10:25 | I'm getting the same error on a 4xL4 (4 x 24 GB CUDA RAM) instance with Llama3.2-1b. VLLM loads this model on one GPU device (24gb) with no problems bur it fails with the same OOM error as above when called with accelerate. | 2,688 | 1,435 |
lewtun | 2025-01-30T08:03:35 | Hey @abacaj ! Can you please share a gist / command that reproduces the error (easier for us to debug then) | 2,688 | 1,436 |
qgallouedec | 2025-01-30T08:34:23 | Try to reduce `vllm_gpu_memory_utilization` (default `0.9`):
```python
from trl import GRPOConfig
GRPOConfig(..., vllm_gpu_memory_utilization=0.7)
``` | 2,688 | 1,437 |
imrankh46 | 2025-01-30T15:17:13 | @qgallouedec
Why you are using reward model here ?
```
from datasets import load_dataset
from peft import LoraConfig
from trl import GRPOConfig, GRPOTrainer
# Load the dataset
dataset = load_dataset("trl-lib/tldr", split="train")
training_args = GRPOConfig(
output_dir="Qwen2-0.5B-GRPO",
learning_rate=1e-5,
logging_steps=10,
gradient_accumulation_steps=16,
max_completion_length=128,
)
trainer = GRPOTrainer(
model="Qwen/Qwen2-0.5B-Instruct",
reward_model="weqweasdas/RM-Gemma-2B",
args=training_args,
train_dataset=dataset,
peft_config=LoraConfig(task_type="CAUSAL_LM"),
)
trainer.train()
```
| 2,688 | 1,438 |
imrankh46 | 2025-01-30T15:18:20 | I mean we can use any reward model or we will fine tune special for specific domain? | 2,688 | 1,439 |
qgallouedec | 2025-01-30T15:29:55 | > Why you are using reward model here ?
It's probably an old code from the time when GRPO was not compatible with reward functions. Is this code still accessible somewhere? | 2,688 | 1,440 |
imrankh46 | 2025-01-31T03:37:43 | @qgallouedec I think you mentioned this code on the first release of grpo with example. | 2,688 | 1,441 |
qgallouedec | 2025-01-29T23:15:19 | I usually refer to https://discuss.pytorch.org/t/about-torch-cuda-empty-cache/34232/2 when I consider emptying the cache. In my understanding, when you leave the function, the internal variables no longer exist so the tensor stop being reference and the underlying memory is freed.
But I might be wrong. | 2,686 | 1,442 |
shirinyamani | 2025-01-29T23:20:14 | right, but we wipe the cache in same scenario in [ppo](https://github.com/huggingface/trl/blob/801582ec240c46ecd487a72fcc8944f268380830/trl/trainer/ppo_trainer.py#L454) ,
I am not sure either! | 2,686 | 1,443 |
qgallouedec | 2025-01-29T23:54:16 | It's probably worth doing some tests/profiling with a toy example to see the diff | 2,686 | 1,444 |
qgallouedec | 2025-01-29T23:02:34 | Reward models are called in inference mode.
https://github.com/huggingface/trl/blob/801582ec240c46ecd487a72fcc8944f268380830/trl/trainer/grpo_trainer.py#L485
I think it's the same. But not 100% sure | 2,685 | 1,445 |
qgallouedec | 2025-01-29T20:51:07 | What about using a callback? That the trainer would internally add. | 2,684 | 1,446 |
shirinyamani | 2025-01-29T22:46:38 | @qgallouedec Do you mean `callbacks: Optional[list[TrainerCallback]] ` ? | 2,684 | 1,447 |
qgallouedec | 2025-01-29T22:58:32 | Not really. Like an arg in the `GRPOConfig`, let's say `sync_ref_steps`. And having in the init of `GRPO.__init__` something like
```python
if args.sync_ref_steps is not None:
sync_ref_callback = SyncRefCallback(args.sync_ref_steps)
self.add_callback(sync_ref_callback)
``` | 2,684 | 1,448 |
qgallouedec | 2025-01-29T23:03:24 | Note that we already have such callback in trl, I think it makes sense to reuse it | 2,684 | 1,449 |
howardzhou | 2025-01-29T23:04:40 | nice! I guess you mean SyncRefModelCallback
https://github.com/huggingface/trl/blob/801582ec240c46ecd487a72fcc8944f268380830/trl/trainer/callbacks.py#L96 | 2,684 | 1,450 |
shirinyamani | 2025-01-29T23:15:38 | with the current callbacks existing in trl, it would be sth like;
```python
if args.sync_ref_model:
self.add_callback(SyncRefModelCallback(ref_model=self.ref_model, accelerator=self.accelerator))
| 2,684 | 1,451 |
qgallouedec | 2025-01-29T23:15:42 | Yep | 2,684 | 1,452 |
shirinyamani | 2025-01-30T01:41:57 | I'll take care of it! | 2,684 | 1,453 |
HuggingFaceDocBuilderDev | 2025-01-29T15:39:41 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2683). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,683 | 1,454 |
qgallouedec | 2025-01-29T15:59:51 | ```python
from datasets import load_dataset
from trl import GRPOConfig, GRPOTrainer
dataset = load_dataset("trl-lib/tldr", split="train")
# Make everything deterministic
import torch
torch.manual_seed(0)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Dummy reward function: the closer the completion is to 20 characters, the higher the reward
def reward_len(completions, **kwargs):
return [-abs(20 - len(completion)) for completion in completions]
training_args = GRPOConfig(
output_dir="Qwen2.5-0.5B-GRPO-main",
logging_steps=2,
gradient_accumulation_steps=1,
per_device_train_batch_size=3,
max_steps=20,
)
trainer = GRPOTrainer(
model="Qwen/Qwen2.5-0.5B-Instruct",
reward_funcs=reward_len,
args=training_args,
train_dataset=dataset,
)
trainer.train()
```
OOM with main, with the other: 83% GPU VRAM
| 2,683 | 1,455 |
HuggingFaceDocBuilderDev | 2025-01-30T08:35:17 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2682). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,682 | 1,456 |
qgallouedec | 2025-01-29T09:06:27 | So it's more a torch question, but I guess you refer to
https://github.com/huggingface/trl/blob/4659ad916fe3e05fd81bf1297db7a6767d2c69bc/trl/trainer/grpo_trainer.py#L412-L413
Yes, the result is zero, but you can't discard it, you need the gradient, see https://github.com/huggingface/trl/pull/2565#issuecomment-2595837761
| 2,681 | 1,457 |
qgallouedec | 2025-01-30T16:58:04 | > should help converge faster
IMHO, it's unlikely, because the log prob under the policy would be very low. But I'd very to be wrong. Can you try to fork trl, hardcode your idea, and make some test to see it indeed helps? | 2,680 | 1,458 |
qgallouedec | 2025-01-30T10:11:28 | Your modification seems to do two things:
- add some kind of support for encoder-decoders, for which it seems that the token id label replaces the token id pass. Is this correct? Can you give more info.
- replace the function structure with the old double loop where you check the start of the key and the end of the key. I'm rather skeptical on this point. I refer to PEP20: _Explicit is better than implicit._ To illustrate:
I dare you in both cases to give me the output keys at a glance. 😉
```python
for k in batch:
if k.startswith("pixel") or k == "image_sizes":
new_concatenated_key = k.replace("prompt", "concatenated")
output[new_concatenated_key] = torch.cat([batch[k], batch[k]], dim=0)
```
```python
output["completion_input_ids"] = torch.cat(
(
pad_to_length(batch["chosen_input_ids"], max_completion_length, pad_value=padding_value),
pad_to_length(batch["rejected_input_ids"], max_completion_length, pad_value=padding_value),
),
)
output["completion_attention_mask"] = torch.cat(
(
pad_to_length(batch["chosen_attention_mask"], max_completion_length, pad_value=0),
pad_to_length(batch["rejected_attention_mask"], max_completion_length, pad_value=0),
),
)
```
| 2,677 | 1,459 |
Superskyyy | 2025-01-28T16:21:37 | What about there are cases where a verifiable reward will be nulled if a primary reward is turned out to be unsatisfactory. Like if you rate a code snippet with execution. That code is not runnable at all, then all other rewards should be nulled right? That would need more than a weight but a primary setting. If that generalizes to more use cases. | 2,676 | 1,460 |
hesamsheikh | 2025-01-28T16:45:30 | > Thanks @hesamsheikh! Do you have any references that show this method can be useful?
the paper actually provides a sneak peek on how their rewards are aggregated:
> Finally, we combine the accuracy of
reasoning tasks and the reward for language consistency by directly summing them to form the
final reward. We then apply RL training on the fine-tuned model until it achieves convergence
on reasoning tasks.
However, they only specify **summing** in the case of using **accuracy reward** and **language consistency reward**. In the current implementation, we could provide multiple rewards in the same scope (e.g. format of the output, or accuracy) so it makes sense that a weighted rewarding system can be beneficial. In the example provided in the test file:
```python
def reward_func1(completions, **kwargs):
"""Reward function that rewards longer completions."""
return [float(len(completion)) for completion in completions]
def reward_func2(completions, **kwargs):
"""Reward function that rewards completions with more unique letters."""
return [float(len(set(completion))) for completion in completions]
```
both rewards are related to the output format, but a simple summation doesn't give us control over which is more important (e.g. to make the completions longer is much more important than having more unique letters).
The implementation of the weighted sum is straightforward:
```python
# Sum the rewards from all reward functions
rewards = rewards_per_func.sum(dim=1)
```
is replaced by
```python
rewards = (rewards_per_func * self.reward_weights.to(device).unsqueeze(0)).sum(dim=1)
```
This allows custom weights rather than no weights (replacing 1 x r1 + 1 x r2 with w1 x r1 + w2 x r2) applied only to the final stage (summation of the rewards). It doesn't break the advantages or the loss function as it is applied to all the rewards. This weighted sum makes the reward aggregation much more flexible in cases of using multiple rewards.
| 2,676 | 1,461 |
qgallouedec | 2025-01-28T18:11:21 | Thanks for this detailed explanation. I understand the underlying motivation. I'm wondering if it really helps to get better results? Or whether a naive sum, as is done in the paper actually is enough to get similar results. | 2,676 | 1,462 |
hesamsheikh | 2025-01-28T22:28:40 | > Thanks for this detailed explanation. I understand the underlying motivation. I'm wondering if it really helps to get better results? Or whether a naive sum, as is done in the paper actually is enough to get similar results.
In cases where multiple rewards with different priorities need to be tuned, the weighted reward must be more handy. I'm down with doing some experiments if you suggest some. | 2,676 | 1,463 |
Benjoyo | 2025-01-31T06:55:41 | I mean, you can can just pass a single aggregate reward function and do arbitrary weighting there, no?
I don’t quite understand the need for a list of functions anyway, except it is slightly more self-documenting. But you can do the same and more with a single function and I don’t think we should add additional parameters to fix the problems with separation reward functions. What do you think? | 2,676 | 1,464 |
Superskyyy | 2025-01-31T14:12:53 | > I mean, you can can just pass a single aggregate reward function and do arbitrary weighting there, no?
>
> I don’t quite understand the need for a list of functions anyway, except it is slightly more self-documenting. But you can do the same and more with a single function and I don’t think we should add additional parameters to fix the problems with separation reward functions. What do you think?
Right, a custom aggregating function to rule them and pass that into trainer sounds like the best way to abstract this need of weighting or whatever many functions. Trainer doesn't really need to know of so many reward functions. | 2,676 | 1,465 |
qgallouedec | 2025-01-31T15:28:52 | > I don’t quite understand the need for a list of functions anyway
It allows
1. to be compatible with reward models (you can mix functions and models)
2. to log each reward separately | 2,676 | 1,466 |
Benjoyo | 2025-02-01T10:45:08 | > > I don’t quite understand the need for a list of functions anyway
>
> It allows
>
> 1. to be compatible with reward models (you can mix functions and models)
> 2. to log each reward separately
Ok, fair points! | 2,676 | 1,467 |
qgallouedec | 2025-02-07T09:29:22 | Let's merge this as experimental feature.
Can you just move `reward_weights` to `GRPOConfig` instead? | 2,676 | 1,468 |
hesamsheikh | 2025-02-07T12:12:49 | I changed reward_weights to float only and moved them to GRPOConfig. @qgallouedec
| 2,676 | 1,469 |
HuggingFaceDocBuilderDev | 2025-02-11T08:48:37 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2676). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,676 | 1,470 |
hesamsheikh | 2025-02-11T09:34:04 | > Thanks for the nice feature @hesamsheikh! I think this will be particularly handy at avoiding reward hacking with e.g. the formatting reward. Logic LGTM
Thanks, glad to be part of this! | 2,676 | 1,471 |
Superskyyy | 2025-01-28T21:04:57 | See #2669 | 2,675 | 1,472 |
qgallouedec | 2025-01-28T12:39:07 | Thanks for reporting, please provide a MRE, system info etc. Refer to the bug report template to help you with it | 2,674 | 1,473 |
Tarak200 | 2025-01-31T10:24:57 | ### **Issue**
Unable to log the metrics and loss while doing training of reward model
###
**Sample Code**
```py
from transformers import (
AutoModelForCausalLM,
AutoTokenizer,
BitsAndBytesConfig,
TrainingArguments,
)
from trl import RewardTrainer, RewardConfig
from peft import PeftModel, LoraConfig, TaskType
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.bfloat16
)
model_name = "Qwen/Qwen2.5-Coder-7B-Instruct"
adapter_path = "/raid/ganesh/nagakalyani/nagakalyani/autograding/huggingface_codellama/nithin_zero-shot_2.0/RLHF/qwen/model/final_checkpoint"
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code = True, torch_dtype=torch.bfloat16, device_map="auto", quantization_config = bnb_config)
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code = True)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
model.config.pad_token_id = model.config.eos_token_id
output_dir = "./reward_model"
training_args = RewardConfig(
center_rewards_coefficient=0.01,
output_dir= output_dir,
per_device_train_batch_size=4,
gradient_accumulation_steps= 8,
eval_strategy="steps",
logging_steps=10,
num_train_epochs = 1,
report_to="tensorboard",
max_length = 512,
save_steps = 0.2,
save_strategy="steps",
gradient_checkpointing = True,
fp16 = True,
metric_for_best_model="eval_loss",
optim="paged_adamw_32bit",
save_safetensors=True,
# optim="adamw_torch",
learning_rate=2e-5,
# report_to = "wandb",
logging_dir="./logs"
)
peft_config = LoraConfig(
task_type=TaskType.SEQ_CLS,
inference_mode=False,
r=8,
lora_alpha=32,
lora_dropout=0.1,
target_modules = ["q_proj", "v_proj", "k_proj", "o_proj"],
)
trainer = RewardTrainer(
model=model,
args=training_args,
tokenizer=tokenizer,
train_dataset=formatted_dataset["train"],
eval_dataset=formatted_dataset["test"],
peft_config=peft_config,
)
trainer.train()
```
### **Package versions**
transformers==4.44.0
trl==0.11.4 | 2,674 | 1,474 |
HuggingFaceDocBuilderDev | 2025-01-30T08:52:43 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2672). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,672 | 1,475 |
benjamin-marie | 2025-01-29T19:14:50 | Same issue in open R1:
```
!git clone https://github.com/huggingface/open-r1.git
%cd open-r1/
!python src/open_r1/grpo.py \
--output_dir DeepSeek-R1-Distill-Qwen-7B-GRPO \
--model_name_or_path Qwen/Qwen2.5-1.5B \
--dataset_name AI-MO/NuminaMath-TIR \
--max_prompt_length 256 \
--per_device_train_batch_size 1 \
--gradient_accumulation_steps 16 \
--gradient_checkpointing \
--logging_steps 10 \
--bf16
```
```
/usr/local/lib/python3.11/dist-packages/torch/utils/checkpoint.py:87: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
``` | 2,671 | 1,476 |
benjamin-marie | 2025-01-29T19:56:36 | This part for gradient checkpointing is in the other TRL trainers, but not in the online DPO and GRPO trainers:
```
elif getattr(args, "gradient_checkpointing", False):
# For backward compatibility with older versions of transformers
if hasattr(model, "enable_input_require_grads"):
model.enable_input_require_grads()
else:
def make_inputs_require_grad(module, input, output):
output.requires_grad_(True)
model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
```
Probably an easy fix? | 2,671 | 1,477 |
qgallouedec | 2025-01-29T20:49:25 | Probably. But I not sure to understand the fix at this point | 2,671 | 1,478 |
benjamin-marie | 2025-01-30T19:51:02 | I saw that Philipp uses gradient checkpointing in the following tutorial:
https://www.philschmid.de/mini-deepseek-r1
I tried but it doesn't work either. Gradient checkpointing in this tutorial doesn't trigger the warning because use_reentrant is set to False instead of True. I might be wrong but I think the non_reentrant variant is not implemented in Qwen (and most LLMs). The consequence is that it consumes as much memory as if gradient checkpointing was set to False.
| 2,671 | 1,479 |
qgallouedec | 2025-01-30T21:31:34 | Thanks for the follow-up. Can you submit a PR so that we can make some tests? | 2,671 | 1,480 |
benjamin-marie | 2025-01-31T07:06:28 | I don't think I will have the time for this. But If I find a solution that works, I'll definitely submit a PR.
```
model.enable_input_require_grads()
```
Seems to work, without a peft_config. I also quickly tried to patch GRPO with some code from the DPOTrainer (which works with gradient checkpointing and PEFT), but it still failed. This is the code that I tried to insert to get gradient checkpointing with a peft_config:
```
elif is_peft_available() and peft_config is not None:
# if model is a peft model and we have a peft_config, we merge and unload it first
if isinstance(model, PeftModel):
model = model.merge_and_unload()
if ref_model is not None and not args.force_use_ref_model:
raise ValueError(
"You passed both a ref_model and a peft_config. For training PEFT adapters with DPO there is no need to pass a reference"
" model. Please pass `ref_model=None` in case you want to train PEFT adapters, or pass a ref_model with `force_use_ref_model=True` in DPOTrainer's init."
" if you want to use a different ref_model."
)
if getattr(model, "is_loaded_in_8bit", False) or getattr(model, "is_loaded_in_4bit", False):
_support_gc_kwargs = hasattr(
args, "gradient_checkpointing_kwargs"
) and "gradient_checkpointing_kwargs" in list(
inspect.signature(prepare_model_for_kbit_training).parameters
)
prepare_model_kwargs = {"use_gradient_checkpointing": args.gradient_checkpointing}
if _support_gc_kwargs:
prepare_model_kwargs["gradient_checkpointing_kwargs"] = args.gradient_checkpointing_kwargs
model = prepare_model_for_kbit_training(model, **prepare_model_kwargs)
elif getattr(args, "gradient_checkpointing", False):
# For backward compatibility with older versions of transformers
if hasattr(model, "enable_input_require_grads"):
model.enable_input_require_grads()
else:
def make_inputs_require_grad(module, input, output):
output.requires_grad_(True)
model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
# get peft model with the given config
model = get_peft_model(model, peft_config)
if args.bf16 and getattr(model, "is_loaded_in_4bit", False):
peft_module_casting_to_bf16(model)
# If args.bf16 we need to explicitly call `generate` with torch amp autocast context manager
self._peft_has_been_casted_to_bf16 = True
# For models that use gradient_checkpointing, we need to attach a hook that enables input
# to explicitly have `requires_grad=True`, otherwise training will either silently
# fail or completely fail.
elif getattr(args, "gradient_checkpointing", False):
# For backward compatibility with older versions of transformers
if hasattr(model, "enable_input_require_grads"):
model.enable_input_require_grads()
else:
def make_inputs_require_grad(module, input, output):
output.requires_grad_(True)
model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
```
This doesn't seem to have any effects. Something else must be modified but I couldn't find what so far.
We will need the help of someone I think. Maybe someone who worked on the DPOTrainer or SFTTrainer.
| 2,671 | 1,481 |
HuggingFaceDocBuilderDev | 2025-01-28T13:07:18 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2670). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,670 | 1,482 |
qgallouedec | 2025-01-28T19:10:35 | Thanks @SeungyounShin, actually we've drop this branch in favour of #2600. But we used your solution. Actually we used `unwrap_model_for_generation` context. For some reason it's faster. | 2,670 | 1,483 |
andyl98 | 2025-01-27T20:49:26 | Gently adding @qgallouedec to review | 2,669 | 1,484 |
andyl98 | 2025-01-28T18:19:10 | For some reason I cannot upload images but I let the training job ran one night without any issues. | 2,669 | 1,485 |
qgallouedec | 2025-01-28T18:23:16 | Can you profile so that we can compare? | 2,669 | 1,486 |
andyl98 | 2025-01-28T18:29:48 | > Can you profile so that we can compare?
For sure. What are some profiling tools you guys recommend using?
Would `torch.cuda.memory_allocated()` and `torch.cuda.memory_reserved()` be enough? | 2,669 | 1,487 |
qgallouedec | 2025-01-28T18:32:12 | Whatever profiler you're familiar with. Can be wandb, torch profiler... | 2,669 | 1,488 |
andyl98 | 2025-01-28T18:36:10 | > Whatever profiler you're familiar with. Can be wandb, torch profiler...
Sure let me make 2 side-by-side comparisons and I'll share the results. | 2,669 | 1,489 |
andyl98 | 2025-01-28T20:35:24 | Ok done. The results match my expectations and I still have no clue why I cannot directly upload images, but I put down all comparisons in the following PDF and it should give ppl a pretty good overview.
Cheers!
[GRPO VRAM Diff Investigation.pdf](https://github.com/user-attachments/files/18578980/GRPO.VRAM.Diff.Investigation.pdf)
| 2,669 | 1,490 |
qgallouedec | 2025-01-29T08:25:02 | Seems very promising!
What about the training speed? | 2,669 | 1,491 |
ArthurZucker | 2025-01-29T10:09:16 | If you have a static cache, you can also specifically compile this part of the code as the shape will never change | 2,669 | 1,492 |
fkxie | 2025-01-29T11:04:51 | Hi @andyl98, Great thanks for your contributions!
When I test your code, it aborts:
```
utils.py, line 1698, in compute_logps_with_prompt_cache
prompt_out.past_key_values.batch_repeat_interleave(G)
AttributeError: 'tuple' object has no attribute 'batch_repeat_interleave'
```
seems `past_key_values` is a tuple, did I missing something? | 2,669 | 1,493 |
Rocketknight1 | 2025-01-29T13:55:03 | Hmm, I'm not familiar with the TRL code, but is there a danger here that gradient will not be propagated back through the prompt sequence if the cached hidden states for the prompt are computed without gradient? Remember that even if we are not computing a label or loss on the prompt tokens, we still want gradient to flow back to these tokens via their key and value vectors | 2,669 | 1,494 |
qgallouedec | 2025-01-29T14:03:38 | That's a good point @Rocketknight1. I'm currently looking into this | 2,669 | 1,495 |
Rocketknight1 | 2025-01-29T14:25:12 | I think it should still be possible to share the prompt computation and re-use the hidden states! It's just important that those states are computed with gradient, and the graph isn't deleted until you compute the loss on the completion tokens | 2,669 | 1,496 |
andyl98 | 2025-01-29T19:27:16 | Really appreciate all the feedback! Here're changes I've made
- ~~Removed the `torch.no_grad()` context manager when forwarding on prompt in case it blocks gradient flow (need to verify if needed but it doesn't add significant memory overhead)~~ Ok nevermind removing this will break the training script, since `use_cache=True` only stores minimal inference-time required tensors.
- Use the `num_logits_to_keep` as @ArthurZucker suggested. This saves quite a bit of memory
After a bit more research, it seems to me that in RL approaches such as PPO in GRPO, we treat the `prompt` as a `given` state/environment and we don't need to calculate policy gradient on them. This means we should be safe to use `torch.no_grad()` with `use_cache` to save lots of compute and only process the last logprobs from the prompt tokens.
Another angle to think of this compare to `SFT` where we need gradient to flow back to prompt is that
### SFT
- Input: The full sequence [prompt + completion] is processed in one forward pass.
- Loss: Cross-entropy loss is computed only on the completion tokens.
- Gradients: Backpropagation updates all parameters that contributed to the completion predictions, including those processing the prompt.
For example, if the prompt is "Translate to French: 'Hello'" and the completion is "Bonjour", the model must learn how the prompt tokens ("Translate to French: 'Hello'") influence the prediction of "Bonjour". This requires gradients through the prompt's hidden states.
### Contrast with GRPO/PPO
- The prompt is treated as a fixed "state" (like a game board in RL).
- Only gradients from policy decisions (completions) are relevant for reward optimization.
- Detaching the prompt’s hidden states saves memory without harming training, since the policy focuses on adjusting completion-generation behavior, not prompt interpretation.
### My conclusion (?)
Consulted with O1/R1 (not the best option I know, but their responses are identical as mine above).
Maybe as an extra, I can do another round of comparison (with 1 using this branch and 1 using master) and just see how the loss/reward curve looks like for let's say 2 epochs?
### Performance Testing
Compared with master branch (from yesterday) and @qgallouedec 's [PR #2683](https://github.com/huggingface/trl/pull/2683) today and I did run with PyTorch profiler a bit to get a sense. Here're some results
## Experimentation Setting
- Hardware: 4090 GPU
- Model: Qwen 1.5B
- `B` = 2
- `G` = 4
- `P` = 660
- `C` = 32
Old approach (yesterday)
```
------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg Self CUDA Self CUDA % CUDA total CUDA time avg CPU Mem Self CPU Mem CUDA Mem Self CUDA Mem # of Calls
------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
aten::mm 18.18% 69.542ms 18.18% 69.542ms 716.928us 136.159ms 32.18% 136.159ms 1.404ms 0 b 0 b 4.05 Gb 4.05 Gb 97
aten::mul 4.98% 19.049ms 4.98% 19.049ms 86.586us 36.698ms 8.67% 36.698ms 166.809us 0 b 0 b 2.85 Gb 2.85 Gb 220
aten::_log_softmax 15.92% 60.895ms 15.92% 60.895ms 7.612ms 71.630ms 16.93% 71.630ms 8.954ms 0 b 0 b 1.43 Gb 1.43 Gb 8
aten::empty_strided 6.50% 24.862ms 6.50% 24.862ms 246.158us 1.026ms 0.24% 1.026ms 10.158us 0 b 0 b 1.26 Gb 1.26 Gb 101
aten::silu 3.83% 14.654ms 3.83% 14.654ms 610.583us 29.972ms 7.08% 29.972ms 1.249ms 0 b 0 b 1.10 Gb 1.10 Gb 24
aten::pow 5.31% 20.322ms 5.31% 20.322ms 414.735us 12.339ms 2.92% 12.789ms 261.000us 0 b 0 b 880.56 Mb 880.56 Mb 49
aten::empty 0.06% 223.000us 0.06% 223.000us 1.312us 1.862ms 0.44% 1.862ms 10.953us 288 b 288 b 784.39 Mb 784.39 Mb 170
aten::add 0.70% 2.674ms 0.70% 2.674ms 18.441us 21.883ms 5.17% 21.883ms 150.917us 0 b 0 b 659.22 Mb 659.22 Mb 145
aten::addmm 1.15% 4.405ms 1.46% 5.586ms 77.583us 5.216ms 1.23% 7.184ms 99.778us 0 b 0 b 266.88 Mb 266.88 Mb 72
aten::cat 0.24% 918.000us 0.24% 918.000us 18.360us 53.215ms 12.58% 53.215ms 1.064ms 0 b 0 b 243.68 Mb 243.68 Mb 50
aten::neg 0.22% 853.000us 0.22% 853.000us 17.771us 941.000us 0.22% 941.000us 19.604us 0 b 0 b 118.50 Mb 118.50 Mb 48
aten::_to_copy 1.74% 6.671ms 8.69% 33.243ms 329.139us 1.984ms 0.47% 5.408ms 53.545us 0 b 0 b 1.29 Gb 26.64 Mb 101
aten::_efficient_attention_forward 0.69% 2.625ms 0.71% 2.716ms 113.167us 5.236ms 1.24% 6.683ms 278.458us 384 b 96 b 213.94 Mb 26.47 Mb 24
aten::empty_like 0.17% 666.000us 0.20% 783.000us 10.875us 850.000us 0.20% 1.256ms 17.444us 0 b 0 b 622.84 Mb 25.92 Mb 72
aten::resize_ 0.00% 16.000us 0.00% 16.000us 8.000us 18.000us 0.00% 18.000us 9.000us 0 b 0 b 8.65 Mb 8.65 Mb 2
aten::mean 0.31% 1.171ms 0.31% 1.171ms 23.898us 732.000us 0.17% 732.000us 14.939us 0 b 0 b 980.00 Kb 980.00 Kb 49
aten::rsqrt 0.38% 1.452ms 0.38% 1.452ms 29.633us 508.000us 0.12% 508.000us 10.367us 0 b 0 b 980.00 Kb 980.00 Kb 49
aten::cos 0.00% 11.000us 0.00% 11.000us 11.000us 10.000us 0.00% 10.000us 10.000us 0 b 0 b 158.00 Kb 158.00 Kb 1
aten::sin 0.00% 9.000us 0.00% 9.000us 9.000us 13.000us 0.00% 13.000us 13.000us 0 b 0 b 158.00 Kb 158.00 Kb 1
aten::bmm 0.02% 70.000us 0.02% 70.000us 70.000us 69.000us 0.02% 69.000us 69.000us 0 b 0 b 79.00 Kb 79.00 Kb 1
aten::gather 9.33% 35.675ms 9.33% 35.679ms 4.460ms 994.000us 0.23% 1.021ms 127.625us 0 b 0 b 12.00 Kb 12.00 Kb 8
aten::embedding 0.02% 59.000us 0.05% 184.000us 184.000us 41.000us 0.01% 449.000us 449.000us 0 b 0 b 8.64 Mb 0 b 1
aten::reshape 1.52% 5.832ms 2.19% 8.364ms 34.279us 2.623ms 0.62% 8.985ms 36.824us 0 b 0 b 414.75 Mb 0 b 244
aten::view 0.16% 608.000us 0.16% 608.000us 1.778us 3.602ms 0.85% 3.602ms 10.532us 0 b 0 b 0 b 0 b 342
aten::index_select 0.02% 69.000us 0.02% 87.000us 87.000us 343.000us 0.08% 365.000us 365.000us 0 b 0 b 8.64 Mb 0 b 1
aten::arange 0.01% 49.000us 0.02% 90.000us 45.000us 52.000us 0.01% 103.000us 51.500us 0 b 0 b 10.00 Kb 0 b 2
aten::unsqueeze 0.48% 1.847ms 0.49% 1.858ms 17.204us 1.519ms 0.36% 2.095ms 19.398us 0 b 0 b 0 b 0 b 108
aten::as_strided 0.02% 93.000us 0.02% 93.000us 0.100us 7.878ms 1.86% 7.878ms 8.453us 0 b 0 b 0 b 0 b 932
aten::slice 1.28% 4.905ms 1.29% 4.918ms 16.448us 3.443ms 0.81% 6.368ms 21.298us 0 b 0 b 0 b 0 b 299
aten::to 0.63% 2.396ms 9.32% 35.639ms 232.935us 1.105ms 0.26% 6.513ms 42.569us 0 b 0 b 1.29 Gb 0 b 153
aten::expand 0.44% 1.680ms 0.44% 1.680ms 13.659us 2.265ms 0.54% 2.961ms 24.073us 0 b 0 b 0 b 0 b 123
aten::copy_ 0.57% 2.192ms 0.57% 2.192ms 12.671us 4.328ms 1.02% 4.328ms 25.017us 0 b 0 b 0 b 0 b 173
aten::matmul 6.24% 23.865ms 24.97% 95.504ms 974.531us 1.751ms 0.41% 141.257ms 1.441ms 0 b 0 b 4.05 Gb 0 b 98
aten::_unsafe_view 0.07% 261.000us 0.07% 261.000us 1.788us 459.000us 0.11% 459.000us 3.144us 0 b 0 b 0 b 0 b 146
aten::transpose 1.51% 5.783ms 1.53% 5.836ms 16.122us 3.799ms 0.90% 7.312ms 20.199us 0 b 0 b 0 b 0 b 362
aten::result_type 0.00% 0.000us 0.00% 0.000us 0.000us 155.000us 0.04% 155.000us 3.163us 0 b 0 b 0 b 0 b 49
aten::linear 15.10% 57.753ms 43.81% 167.564ms 991.503us 3.572ms 0.84% 158.242ms 936.343us 0 b 0 b 4.31 Gb 0 b 169
aten::t 0.76% 2.914ms 1.53% 5.837ms 34.538us 1.692ms 0.40% 4.697ms 27.793us 0 b 0 b 0 b 0 b 169
aten::clone 0.42% 1.617ms 0.75% 2.882ms 40.028us 1.600ms 0.38% 4.786ms 66.472us 0 b 0 b 622.84 Mb 0 b 72
aten::contiguous 0.07% 270.000us 0.29% 1.101ms 45.875us 279.000us 0.07% 2.080ms 86.667us 0 b 0 b 208.09 Mb 0 b 24
aten::scaled_dot_product_attention 0.14% 517.000us 1.74% 6.661ms 277.542us 176.000us 0.04% 10.428ms 434.500us 384 b 0 b 213.94 Mb 0 b 24
aten::_scaled_dot_product_efficient_attention 0.56% 2.147ms 1.61% 6.144ms 256.000us 638.000us 0.15% 10.252ms 427.167us 384 b 0 b 213.94 Mb 0 b 24
aten::unbind 0.07% 268.000us 0.15% 564.000us 282.000us 48.000us 0.01% 392.000us 196.000us 0 b 0 b 0 b 0 b 2
aten::select 0.08% 287.000us 0.08% 296.000us 18.500us 216.000us 0.05% 344.000us 21.500us 0 b 0 b 0 b 0 b 16
aten::log_softmax 0.03% 122.000us 15.95% 61.017ms 7.627ms 56.000us 0.01% 71.686ms 8.961ms 0 b 0 b 1.43 Gb 0 b 8
aten::squeeze 0.04% 142.000us 0.04% 145.000us 18.125us 102.000us 0.02% 115.000us 14.375us 0 b 0 b 0 b 0 b 8
aten::stack 0.01% 27.000us 0.02% 85.000us 85.000us 4.000us 0.00% 167.000us 167.000us 0 b 0 b 10.00 Kb 0 b 1
[memory] 0.00% 0.000us 0.00% 0.000us 0.000us 0.000us 0.00% 0.000us 0.000us 0 b 0 b -5.30 Gb -5.30 Gb 724
```
My previous approach (before this commit)
```
------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg Self CUDA Self CUDA % CUDA total CUDA time avg CPU Mem Self CPU Mem CUDA Mem Self CUDA Mem # of Calls
------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
aten::mm 4.15% 4.252ms 4.15% 4.252ms 21.918us 13.953ms 11.15% 13.953ms 71.923us 0 b 0 b 1.22 Gb 1.22 Gb 194
aten::mul 5.58% 5.713ms 5.58% 5.713ms 12.955us 9.110ms 7.28% 9.110ms 20.658us 0 b 0 b 937.33 Mb 937.33 Mb 441
aten::empty 0.55% 568.000us 0.55% 568.000us 1.349us 2.172ms 1.74% 2.172ms 5.159us 600 b 600 b 825.25 Mb 825.25 Mb 421
aten::empty_strided 1.65% 1.689ms 1.65% 1.689ms 8.320us 5.890ms 4.71% 5.890ms 29.015us 0 b 0 b 378.82 Mb 378.82 Mb 203
aten::silu 0.39% 395.000us 0.39% 395.000us 8.229us 492.000us 0.39% 492.000us 10.250us 0 b 0 b 335.11 Mb 335.11 Mb 48
aten::pow 3.85% 3.945ms 3.85% 3.945ms 40.255us 6.909ms 5.52% 7.567ms 77.214us 0 b 0 b 265.29 Mb 265.29 Mb 98
aten::add 2.45% 2.509ms 2.45% 2.509ms 8.622us 2.351ms 1.88% 2.351ms 8.079us 0 b 0 b 237.44 Mb 237.44 Mb 291
aten::cat 1.74% 1.780ms 1.74% 1.780ms 12.109us 2.645ms 2.11% 2.645ms 17.993us 0 b 0 b 148.23 Mb 148.23 Mb 147
aten::addmm 4.68% 4.795ms 5.84% 5.984ms 41.556us 5.329ms 4.26% 7.494ms 52.042us 0 b 0 b 92.81 Mb 92.81 Mb 144
aten::neg 0.94% 963.000us 0.94% 963.000us 10.031us 1.508ms 1.21% 1.508ms 15.708us 0 b 0 b 35.77 Mb 35.77 Mb 96
aten::_to_copy 4.61% 4.724ms 7.75% 7.938ms 39.297us 2.577ms 2.06% 10.484ms 51.901us 0 b 0 b 398.67 Mb 21.22 Mb 202
aten::empty_like 2.03% 2.081ms 2.41% 2.473ms 12.490us 1.600ms 1.28% 2.687ms 13.571us 0 b 0 b 764.89 Mb 20.88 Mb 198
aten::resize_ 0.04% 42.000us 0.04% 42.000us 7.000us 65.000us 0.05% 65.000us 10.833us 0 b 0 b 3.17 Mb 3.17 Mb 6
aten::eq 0.08% 78.000us 0.08% 78.000us 26.000us 64.000us 0.05% 64.000us 21.333us 0 b 0 b 1.38 Mb 1.38 Mb 3
aten::constant_pad_nd 0.83% 853.000us 1.52% 1.562ms 65.083us 574.000us 0.46% 1.700ms 70.833us 0 b 0 b 7.50 Mb 1.25 Mb 24
aten::triu 0.03% 26.000us 0.03% 26.000us 13.000us 29.000us 0.02% 29.000us 14.500us 0 b 0 b 743.50 Kb 743.50 Kb 2
aten::_efficient_attention_forward 2.38% 2.437ms 2.51% 2.567ms 53.479us 2.999ms 2.40% 3.939ms 82.062us 768 b 168 b 76.08 Mb 448.00 Kb 48
aten::gt 0.03% 30.000us 0.03% 30.000us 15.000us 30.000us 0.02% 30.000us 15.000us 0 b 0 b 372.00 Kb 372.00 Kb 2
aten::mean 0.94% 965.000us 0.94% 965.000us 9.847us 711.000us 0.57% 711.000us 7.255us 0 b 0 b 294.00 Kb 294.00 Kb 98
aten::rsqrt 0.79% 813.000us 0.79% 813.000us 8.296us 1.089ms 0.87% 1.089ms 11.112us 0 b 0 b 294.00 Kb 294.00 Kb 98
aten::cos 0.02% 21.000us 0.02% 21.000us 10.500us 27.000us 0.02% 27.000us 13.500us 0 b 0 b 158.00 Kb 158.00 Kb 2
aten::sin 0.01% 13.000us 0.01% 13.000us 6.500us 18.000us 0.01% 18.000us 9.000us 0 b 0 b 158.00 Kb 158.00 Kb 2
aten::bmm 0.07% 71.000us 0.07% 71.000us 35.500us 45.000us 0.04% 45.000us 22.500us 0 b 0 b 79.00 Kb 79.00 Kb 2
aten::all 0.05% 48.000us 0.05% 49.000us 24.500us 49.000us 0.04% 57.000us 28.500us 0 b 0 b 2.00 Kb 2.00 Kb 2
aten::bitwise_not 0.01% 11.000us 0.01% 11.000us 11.000us 11.000us 0.01% 11.000us 11.000us 0 b 0 b 1.50 Kb 1.50 Kb 1
aten::gather 0.05% 48.000us 0.05% 48.000us 24.000us 17.000us 0.01% 70.000us 35.000us 0 b 0 b 1.00 Kb 1.00 Kb 2
aten::reshape 5.52% 5.658ms 10.11% 10.353ms 21.129us 4.143ms 3.31% 10.605ms 21.643us 0 b 0 b 545.97 Mb 0 b 490
aten::view 0.58% 593.000us 0.58% 593.000us 0.810us 3.276ms 2.62% 3.276ms 4.475us 0 b 0 b 0 b 0 b 732
aten::index_select 0.11% 108.000us 0.14% 144.000us 72.000us 109.000us 0.09% 160.000us 80.000us 0 b 0 b 3.16 Mb 0 b 2
aten::arange 0.15% 151.000us 0.25% 259.000us 32.375us 148.000us 0.12% 300.000us 37.500us 0 b 0 b 31.00 Kb 0 b 8
aten::unsqueeze 2.69% 2.754ms 2.75% 2.814ms 10.949us 2.391ms 1.91% 3.708ms 14.428us 0 b 0 b 0 b 0 b 257
aten::as_strided 0.13% 134.000us 0.13% 134.000us 0.062us 9.062ms 7.24% 9.062ms 4.191us 0 b 0 b 0 b 0 b 2162
aten::is_nonzero 0.03% 28.000us 0.07% 73.000us 73.000us 13.000us 0.01% 78.000us 78.000us 0 b 0 b 0 b 0 b 1
aten::item 0.01% 14.000us 0.04% 45.000us 45.000us 29.000us 0.02% 65.000us 65.000us 0 b 0 b 0 b 0 b 1
aten::_local_scalar_dense 0.03% 31.000us 0.03% 31.000us 31.000us 36.000us 0.03% 36.000us 36.000us 0 b 0 b 0 b 0 b 1
aten::full 0.12% 118.000us 0.16% 166.000us 83.000us 28.000us 0.02% 82.000us 41.000us 0 b 0 b 743.50 Kb 0 b 2
aten::fill_ 0.14% 148.000us 0.14% 148.000us 5.692us 289.000us 0.23% 289.000us 11.115us 0 b 0 b 0 b 0 b 26
aten::mul_ 0.03% 27.000us 0.03% 27.000us 13.500us 30.000us 0.02% 30.000us 15.000us 0 b 0 b 0 b 0 b 2
aten::slice 8.55% 8.756ms 8.59% 8.797ms 10.612us 8.230ms 6.58% 11.979ms 14.450us 0 b 0 b 0 b 0 b 829
aten::expand 3.40% 3.480ms 3.41% 3.493ms 10.125us 3.231ms 2.58% 4.878ms 14.139us 0 b 0 b 0 b 0 b 345
aten::clone 4.50% 4.612ms 8.60% 8.811ms 44.500us 2.904ms 2.32% 8.417ms 42.510us 0 b 0 b 764.89 Mb 0 b 198
aten::copy_ 3.41% 3.495ms 3.41% 3.495ms 8.224us 5.220ms 4.17% 5.220ms 12.282us 0 b 0 b 0 b 0 b 425
aten::masked_fill 0.03% 32.000us 0.13% 133.000us 133.000us 5.000us 0.00% 55.000us 55.000us 0 b 0 b 1.37 Mb 0 b 1
aten::masked_fill_ 0.01% 15.000us 0.01% 15.000us 15.000us 18.000us 0.01% 18.000us 18.000us 0 b 0 b 0 b 0 b 1
aten::to 1.88% 1.923ms 9.63% 9.861ms 32.225us 1.586ms 1.27% 12.070ms 39.444us 0 b 0 b 398.67 Mb 0 b 306
aten::matmul 5.00% 5.119ms 11.23% 11.499ms 58.668us 3.189ms 2.55% 20.354ms 103.847us 0 b 0 b 1.22 Gb 0 b 196
aten::_unsafe_view 0.22% 222.000us 0.22% 222.000us 0.758us 712.000us 0.57% 712.000us 2.430us 0 b 0 b 0 b 0 b 293
aten::transpose 5.79% 5.934ms 5.81% 5.953ms 8.222us 7.495ms 5.99% 9.773ms 13.499us 0 b 0 b 0 b 0 b 724
aten::result_type 0.00% 0.000us 0.00% 0.000us 0.000us 326.000us 0.26% 326.000us 3.327us 0 b 0 b 0 b 0 b 98
aten::linear 8.85% 9.070ms 33.30% 34.116ms 100.935us 4.964ms 3.97% 42.947ms 127.062us 0 b 0 b 1.31 Gb 0 b 338
aten::t 3.29% 3.371ms 5.83% 5.975ms 17.678us 2.938ms 2.35% 8.435ms 24.956us 0 b 0 b 0 b 0 b 338
aten::contiguous 0.30% 304.000us 2.06% 2.113ms 42.260us 272.000us 0.22% 2.606ms 52.120us 0 b 0 b 148.20 Mb 0 b 50
aten::scaled_dot_product_attention 1.64% 1.681ms 10.54% 10.796ms 224.917us 1.034ms 0.83% 11.357ms 236.604us 432 b -336 b 83.58 Mb 0 b 48
aten::_scaled_dot_product_efficient_attention 2.17% 2.220ms 6.24% 6.392ms 133.167us 1.074ms 0.86% 7.481ms 155.854us 768 b 0 b 76.08 Mb 0 b 48
aten::log_softmax 0.01% 13.000us 0.21% 210.000us 105.000us 6.000us 0.00% 490.000us 245.000us 0 b 0 b 72.45 Mb 0 b 2
aten::_log_softmax 0.03% 33.000us 0.19% 197.000us 98.500us 251.000us 0.20% 484.000us 242.000us 0 b 0 b 72.45 Mb 0 b 2
aten::repeat_interleave 2.17% 2.228ms 6.57% 6.731ms 137.367us 1.041ms 0.83% 4.949ms 101.000us 0 b 0 b 67.97 Mb 0 b 49
aten::flatten 0.65% 669.000us 0.66% 672.000us 13.714us 402.000us 0.32% 531.000us 10.837us 0 b 0 b 0 b 0 b 49
aten::squeeze 0.01% 14.000us 0.01% 14.000us 7.000us 7.000us 0.01% 17.000us 8.500us 0 b 0 b 0 b 0 b 2
aten::pad 0.35% 356.000us 1.87% 1.918ms 79.917us 215.000us 0.17% 1.915ms 79.792us 0 b 0 b 7.50 Mb 0 b 24
aten::narrow 0.14% 146.000us 0.35% 361.000us 15.042us 163.000us 0.13% 436.000us 18.167us 0 b 0 b 0 b 0 b 24
aten::embedding 0.11% 112.000us 0.31% 320.000us 160.000us 42.000us 0.03% 431.000us 215.500us 0 b 0 b 3.16 Mb -2.00 Kb 2
[memory] 0.00% 0.000us 0.00% 0.000us 0.000us 0.000us 0.00% 0.000us 0.000us -48 b -48 b -3.61 Gb -3.61 Gb 1946
``` | 2,669 | 1,497 |
andyl98 | 2025-01-29T19:28:25 | Quentin's new approach #2683
```
------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg Self CUDA Self CUDA % CUDA total CUDA time avg CPU Mem Self CPU Mem CUDA Mem Self CUDA Mem # of Calls
------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
aten::mul 19.31% 141.976ms 19.31% 141.976ms 645.345us 137.645ms 18.11% 137.645ms 625.659us 0 b 0 b 2.85 Gb 2.85 Gb 220
aten::mm 16.74% 123.085ms 16.74% 123.085ms 1.269ms 162.452ms 21.37% 162.452ms 1.675ms 0 b 0 b 2.68 Gb 2.68 Gb 97
aten::empty_strided 18.24% 134.109ms 18.24% 134.109ms 1.328ms 125.889ms 16.56% 125.889ms 1.246ms 0 b 0 b 1.25 Gb 1.25 Gb 101
aten::silu 8.87% 65.203ms 8.87% 65.203ms 2.717ms 79.626ms 10.48% 79.626ms 3.318ms 0 b 0 b 1.10 Gb 1.10 Gb 24
aten::pow 17.47% 128.482ms 17.47% 128.483ms 2.622ms 112.970ms 14.86% 113.246ms 2.311ms 0 b 0 b 877.69 Mb 877.69 Mb 49
aten::empty 8.39% 61.655ms 8.39% 61.655ms 360.556us 69.549ms 9.15% 69.549ms 406.719us 288 b 288 b 848.43 Mb 848.43 Mb 171
aten::add 0.38% 2.792ms 0.38% 2.792ms 19.255us 1.920ms 0.25% 1.920ms 13.241us 0 b 0 b 655.46 Mb 655.46 Mb 145
aten::addmm 0.42% 3.113ms 0.58% 4.281ms 59.458us 3.818ms 0.50% 5.348ms 74.278us 0 b 0 b 281.00 Mb 281.00 Mb 72
aten::cat 0.09% 691.000us 0.09% 691.000us 13.820us 988.000us 0.13% 988.000us 19.760us 0 b 0 b 239.90 Mb 239.90 Mb 50
aten::neg 1.50% 11.055ms 1.50% 11.055ms 230.312us 11.046ms 1.45% 11.046ms 230.125us 0 b 0 b 126.09 Mb 126.09 Mb 48
aten::_log_softmax 2.25% 16.546ms 2.25% 16.546ms 2.068ms 16.281ms 2.14% 16.281ms 2.035ms 0 b 0 b 74.19 Mb 74.19 Mb 8
aten::_to_copy 0.58% 4.258ms 18.99% 139.610ms 1.382ms 794.000us 0.10% 128.044ms 1.268ms 0 b 0 b 1.27 Gb 25.92 Mb 101
aten::empty_like 0.10% 761.000us 0.14% 1.016ms 13.918us 593.000us 0.08% 1.203ms 16.479us 0 b 0 b 652.04 Mb 17.28 Mb 73
aten::resize_ 0.00% 19.000us 0.00% 19.000us 9.500us 194.000us 0.03% 194.000us 97.000us 0 b 0 b 8.65 Mb 8.65 Mb 2
aten::mean 0.18% 1.338ms 0.18% 1.338ms 27.306us 752.000us 0.10% 752.000us 15.347us 0 b 0 b 980.00 Kb 980.00 Kb 49
aten::rsqrt 0.13% 992.000us 0.13% 992.000us 20.245us 462.000us 0.06% 462.000us 9.429us 0 b 0 b 980.00 Kb 980.00 Kb 49
aten::_efficient_attention_forward 0.48% 3.558ms 8.83% 64.950ms 2.706ms 2.381ms 0.31% 71.316ms 2.971ms 384 b 96 b 213.94 Mb 280.00 Kb 24
aten::cos 0.00% 10.000us 0.00% 10.000us 10.000us 11.000us 0.00% 11.000us 11.000us 0 b 0 b 158.00 Kb 158.00 Kb 1
aten::sin 0.00% 8.000us 0.00% 8.000us 8.000us 10.000us 0.00% 10.000us 10.000us 0 b 0 b 158.00 Kb 158.00 Kb 1
aten::bmm 0.00% 36.000us 0.00% 36.000us 36.000us 36.000us 0.00% 36.000us 36.000us 0 b 0 b 79.00 Kb 79.00 Kb 1
aten::gather 0.07% 485.000us 0.07% 499.000us 62.375us 253.000us 0.03% 297.000us 37.125us 0 b 0 b 4.00 Kb 4.00 Kb 8
aten::embedding 0.00% 36.000us 0.02% 117.000us 117.000us 8.000us 0.00% 233.000us 233.000us 0 b 0 b 8.64 Mb 0 b 1
aten::reshape 0.40% 2.947ms 0.81% 5.939ms 24.340us 1.915ms 0.25% 8.532ms 34.967us 0 b 0 b 444.67 Mb 0 b 244
aten::view 0.10% 714.000us 0.10% 714.000us 2.094us 3.772ms 0.50% 3.772ms 11.062us 0 b 0 b 0 b 0 b 341
aten::index_select 0.01% 45.000us 0.01% 68.000us 68.000us 25.000us 0.00% 220.000us 220.000us 0 b 0 b 8.64 Mb 0 b 1
aten::arange 0.00% 24.000us 0.01% 43.000us 21.500us 37.000us 0.00% 73.000us 36.500us 0 b 0 b 10.00 Kb 0 b 2
aten::unsqueeze 0.17% 1.257ms 0.17% 1.278ms 11.833us 1.440ms 0.19% 2.012ms 18.630us 0 b 0 b 0 b 0 b 108
aten::as_strided 0.03% 248.000us 0.03% 248.000us 0.266us 4.108ms 0.54% 4.108ms 4.408us 0 b 0 b 0 b 0 b 932
aten::slice 0.52% 3.855ms 0.53% 3.910ms 13.077us 3.530ms 0.46% 5.105ms 17.074us 0 b 0 b 0 b 0 b 299
aten::to 0.17% 1.216ms 19.15% 140.826ms 920.431us 947.000us 0.12% 128.991ms 843.078us 0 b 0 b 1.27 Gb 0 b 153
aten::expand 0.24% 1.796ms 0.25% 1.820ms 14.797us 1.985ms 0.26% 2.516ms 20.455us 0 b 0 b 0 b 0 b 123
aten::copy_ 0.25% 1.852ms 0.25% 1.852ms 10.644us 2.710ms 0.36% 2.710ms 15.575us 0 b 0 b 0 b 0 b 174
aten::matmul 0.47% 3.488ms 17.44% 128.210ms 1.308ms 1.436ms 0.19% 167.914ms 1.713ms 0 b 0 b 2.68 Gb 0 b 98
aten::_unsafe_view 0.06% 461.000us 0.06% 461.000us 3.136us 569.000us 0.07% 569.000us 3.871us 0 b 0 b 0 b 0 b 147
aten::transpose 0.61% 4.478ms 0.63% 4.598ms 12.702us 2.775ms 0.37% 4.030ms 11.133us 0 b 0 b 0 b 0 b 362
aten::result_type 0.00% 0.000us 0.00% 0.000us 0.000us 95.000us 0.01% 95.000us 1.939us 0 b 0 b 0 b 0 b 49
aten::linear 0.74% 5.422ms 19.43% 142.880ms 845.444us 2.910ms 0.38% 179.849ms 1.064ms 0 b 0 b 2.95 Gb 0 b 169
aten::t 0.29% 2.136ms 0.54% 3.976ms 23.527us 1.260ms 0.17% 2.869ms 16.976us 0 b 0 b 0 b 0 b 169
aten::clone 0.25% 1.807ms 0.47% 3.432ms 47.014us 1.586ms 0.21% 4.138ms 56.685us 0 b 0 b 652.04 Mb 0 b 73
aten::contiguous 0.04% 258.000us 0.18% 1.357ms 56.542us 223.000us 0.03% 1.501ms 62.542us 0 b 0 b 207.38 Mb 0 b 24
aten::scaled_dot_product_attention 0.05% 360.000us 9.30% 68.377ms 2.849ms 142.000us 0.02% 73.391ms 3.058ms 384 b 0 b 213.94 Mb 0 b 24
aten::_scaled_dot_product_efficient_attention 0.24% 1.786ms 9.25% 68.017ms 2.834ms 582.000us 0.08% 73.249ms 3.052ms 384 b 0 b 213.94 Mb 0 b 24
aten::unbind 0.04% 324.000us 0.08% 599.000us 299.500us 96.000us 0.01% 369.000us 184.500us 0 b 0 b 0 b 0 b 2
aten::select 0.04% 269.000us 0.04% 275.000us 17.188us 155.000us 0.02% 273.000us 17.062us 0 b 0 b 0 b 0 b 16
aten::log_softmax 0.02% 120.000us 2.27% 16.666ms 2.083ms 46.000us 0.01% 16.327ms 2.041ms 0 b 0 b 74.19 Mb 0 b 8
aten::squeeze 0.02% 182.000us 0.03% 190.000us 23.750us 31.000us 0.00% 44.000us 5.500us 0 b 0 b 0 b 0 b 8
aten::stack 0.01% 38.000us 0.01% 63.000us 63.000us 21.000us 0.00% 44.000us 44.000us 0 b 0 b 512 b 0 b 1
[memory] 0.00% 0.000us 0.00% 0.000us 0.000us 0.000us 0.00% 0.000us 0.000us 0 b 0 b -3.96 Gb -3.96 Gb 725
```
Mine after this commit
```
------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg Self CUDA Self CUDA % CUDA total CUDA time avg CPU Mem Self CPU Mem CUDA Mem Self CUDA Mem # of Calls
------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
aten::mm 16.43% 131.428ms 16.43% 131.428ms 677.464us 177.986ms 20.36% 177.986ms 917.454us 0 b 0 b 883.89 Mb 883.89 Mb 194
aten::mul 15.69% 125.517ms 15.69% 125.517ms 284.619us 146.195ms 16.73% 146.195ms 331.508us 0 b 0 b 876.07 Mb 876.07 Mb 441
aten::empty 9.24% 73.895ms 9.24% 73.895ms 176.360us 117.505ms 13.44% 117.505ms 280.442us 728 b 728 b 743.00 Mb 743.00 Mb 419
aten::empty_strided 12.45% 99.564ms 12.45% 99.564ms 490.463us 99.619ms 11.40% 99.619ms 490.734us 0 b 0 b 380.25 Mb 380.25 Mb 203
aten::silu 7.83% 62.628ms 7.83% 62.628ms 1.305ms 74.492ms 8.52% 74.492ms 1.552ms 0 b 0 b 344.19 Mb 344.19 Mb 48
aten::pow 12.34% 98.706ms 12.34% 98.706ms 1.007ms 80.462ms 9.21% 81.197ms 828.541us 0 b 0 b 246.73 Mb 246.73 Mb 98
aten::add 0.82% 6.562ms 0.82% 6.562ms 22.550us 4.767ms 0.55% 4.767ms 16.381us 0 b 0 b 206.86 Mb 206.86 Mb 291
aten::cat 0.43% 3.469ms 0.43% 3.469ms 23.599us 3.973ms 0.45% 3.973ms 27.027us 0 b 0 b 150.40 Mb 150.40 Mb 147
aten::addmm 1.21% 9.707ms 1.54% 12.341ms 85.701us 8.279ms 0.95% 11.663ms 80.993us 0 b 0 b 85.61 Mb 85.61 Mb 144
aten::_log_softmax 2.74% 21.911ms 2.74% 21.911ms 10.956ms 17.248ms 1.97% 17.248ms 8.624ms 0 b 0 b 74.77 Mb 74.77 Mb 2
aten::neg 0.26% 2.057ms 0.26% 2.057ms 21.427us 2.853ms 0.33% 2.853ms 29.719us 0 b 0 b 35.77 Mb 35.77 Mb 96
aten::resize_ 0.01% 40.000us 0.01% 40.000us 6.667us 75.000us 0.01% 75.000us 12.500us 0 b 0 b 2.54 Mb 2.54 Mb 6
aten::eq 0.01% 80.000us 0.01% 80.000us 26.667us 85.000us 0.01% 85.000us 28.333us 0 b 0 b 1.38 Mb 1.38 Mb 3
aten::triu 0.00% 35.000us 0.00% 35.000us 17.500us 45.000us 0.01% 45.000us 22.500us 0 b 0 b 743.50 Kb 743.50 Kb 2
aten::gt 0.01% 45.000us 0.01% 45.000us 22.500us 54.000us 0.01% 54.000us 27.000us 0 b 0 b 372.00 Kb 372.00 Kb 2
aten::mean 0.30% 2.424ms 0.30% 2.424ms 24.735us 1.304ms 0.15% 1.304ms 13.306us 0 b 0 b 294.00 Kb 294.00 Kb 98
aten::rsqrt 0.26% 2.043ms 0.26% 2.043ms 20.847us 1.626ms 0.19% 1.626ms 16.592us 0 b 0 b 294.00 Kb 294.00 Kb 98
aten::cos 0.00% 35.000us 0.00% 35.000us 17.500us 42.000us 0.00% 42.000us 21.000us 0 b 0 b 158.00 Kb 158.00 Kb 2
aten::sin 0.00% 31.000us 0.00% 31.000us 15.500us 39.000us 0.00% 39.000us 19.500us 0 b 0 b 158.00 Kb 158.00 Kb 2
aten::bmm 0.01% 88.000us 0.01% 88.000us 44.000us 101.000us 0.01% 101.000us 50.500us 0 b 0 b 79.00 Kb 79.00 Kb 2
aten::all 0.01% 56.000us 0.01% 57.000us 28.500us 60.000us 0.01% 67.000us 33.500us 0 b 0 b 2.00 Kb 2.00 Kb 2
aten::bitwise_not 0.00% 21.000us 0.00% 21.000us 21.000us 26.000us 0.00% 26.000us 26.000us 0 b 0 b 1.50 Kb 1.50 Kb 1
aten::gather 0.01% 119.000us 0.02% 123.000us 61.500us 62.000us 0.01% 78.000us 39.000us 0 b 0 b 1.00 Kb 1.00 Kb 2
aten::reshape 1.33% 10.662ms 11.48% 91.829ms 187.406us 6.181ms 0.71% 131.618ms 268.608us 0 b 0 b 534.24 Mb 0 b 490
aten::view 0.40% 3.172ms 0.40% 3.172ms 4.339us 8.341ms 0.95% 8.341ms 11.410us 0 b 0 b 0 b 0 b 731
aten::index_select 0.01% 114.000us 0.02% 145.000us 72.500us 102.000us 0.01% 161.000us 80.500us 0 b 0 b 2.52 Mb 0 b 2
aten::arange 0.03% 203.000us 0.04% 339.000us 42.375us 181.000us 0.02% 372.000us 46.500us 0 b 0 b 31.00 Kb 0 b 8
aten::unsqueeze 0.56% 4.461ms 0.59% 4.708ms 18.319us 4.424ms 0.51% 6.021ms 23.428us 0 b 0 b 0 b 0 b 257
aten::as_strided 0.27% 2.121ms 0.27% 2.121ms 0.984us 22.779ms 2.61% 22.779ms 10.565us 0 b 0 b 0 b 0 b 2156
aten::is_nonzero 0.00% 25.000us 0.01% 70.000us 70.000us 14.000us 0.00% 75.000us 75.000us 0 b 0 b 0 b 0 b 1
aten::item 0.00% 14.000us 0.01% 45.000us 45.000us 26.000us 0.00% 61.000us 61.000us 0 b 0 b 0 b 0 b 1
aten::_local_scalar_dense 0.00% 31.000us 0.00% 31.000us 31.000us 35.000us 0.00% 35.000us 35.000us 0 b 0 b 0 b 0 b 1
aten::full 0.01% 59.000us 0.01% 98.000us 49.000us 48.000us 0.01% 109.000us 54.500us 0 b 0 b 743.50 Kb 0 b 2
aten::fill_ 0.04% 343.000us 0.04% 343.000us 13.192us 284.000us 0.03% 284.000us 10.923us 0 b 0 b 0 b 0 b 26
aten::mul_ 0.00% 34.000us 0.00% 34.000us 17.000us 45.000us 0.01% 45.000us 22.500us 0 b 0 b 0 b 0 b 2
aten::slice 3.24% 25.908ms 3.34% 26.732ms 32.481us 14.685ms 1.68% 29.851ms 36.271us 0 b 0 b 0 b 0 b 823
aten::expand 2.15% 17.162ms 2.19% 17.512ms 50.759us 16.360ms 1.87% 18.636ms 54.017us 0 b 0 b 0 b 0 b 345
aten::clone 0.83% 6.630ms 10.75% 85.953ms 438.536us 4.821ms 0.55% 126.932ms 647.612us 0 b 0 b 670.20 Mb 0 b 196
aten::empty_like 0.40% 3.194ms 9.55% 76.405ms 389.821us 2.775ms 0.32% 118.547ms 604.832us 0 b 0 b 670.20 Mb 0 b 196
aten::copy_ 0.77% 6.124ms 0.77% 6.124ms 14.478us 6.428ms 0.74% 6.428ms 15.196us 0 b 0 b 0 b 0 b 423
aten::masked_fill 0.01% 42.000us 0.02% 135.000us 135.000us 37.000us 0.00% 139.000us 139.000us 0 b 0 b 1.45 Mb 0 b 1
aten::masked_fill_ 0.00% 14.000us 0.00% 14.000us 14.000us 18.000us 0.00% 18.000us 18.000us 0 b 0 b 0 b 0 b 1
aten::to 0.45% 3.566ms 14.23% 113.823ms 371.971us 2.572ms 0.29% 109.228ms 356.954us 0 b 0 b 378.80 Mb 0 b 306
aten::_to_copy 0.98% 7.856ms 13.78% 110.257ms 545.827us 4.555ms 0.52% 106.656ms 528.000us 0 b 0 b 378.80 Mb 0 b 202
aten::matmul 1.34% 10.702ms 18.42% 147.355ms 751.811us 4.109ms 0.47% 188.447ms 961.464us 0 b 0 b 883.97 Mb 0 b 196
aten::_unsafe_view 0.15% 1.231ms 0.15% 1.231ms 4.201us 1.300ms 0.15% 1.300ms 4.437us 0 b 0 b 0 b 0 b 293
aten::transpose 1.59% 12.747ms 1.68% 13.440ms 18.564us 9.698ms 1.11% 13.389ms 18.493us 0 b 0 b 0 b 0 b 724
aten::result_type 0.00% 0.000us 0.00% 0.000us 0.000us 401.000us 0.05% 401.000us 4.092us 0 b 0 b 0 b 0 b 98
aten::linear 2.16% 17.273ms 24.20% 193.612ms 572.817us 9.822ms 1.12% 224.641ms 664.618us 0 b 0 b 969.51 Mb 0 b 338
aten::t 0.88% 7.049ms 1.60% 12.819ms 37.926us 4.884ms 0.56% 10.953ms 32.405us 0 b 0 b 0 b 0 b 338
aten::contiguous 0.10% 783.000us 0.52% 4.134ms 86.125us 595.000us 0.07% 3.765ms 78.438us 0 b 0 b 64.89 Mb 0 b 48
aten::scaled_dot_product_attention 0.38% 3.049ms 2.61% 20.884ms 435.083us 1.664ms 0.19% 16.392ms 341.500us 768 b 0 b 73.53 Mb 0 b 48
aten::_scaled_dot_product_efficient_attention 0.57% 4.534ms 1.55% 12.367ms 257.646us 1.935ms 0.22% 10.446ms 217.625us 768 b 0 b 66.03 Mb 0 b 48
aten::_efficient_attention_forward 0.47% 3.763ms 0.54% 4.310ms 89.792us 3.969ms 0.45% 5.436ms 113.250us 768 b 40 b 66.03 Mb 0 b 48
aten::_reshape_alias 0.00% 5.000us 0.00% 5.000us 5.000us 11.000us 0.00% 11.000us 11.000us 0 b 0 b 0 b 0 b 1
aten::log_softmax 0.00% 34.000us 2.74% 21.945ms 10.973ms 8.000us 0.00% 17.256ms 8.628ms 0 b 0 b 74.77 Mb 0 b 2
aten::repeat_interleave 0.38% 3.007ms 1.16% 9.317ms 190.143us 1.992ms 0.23% 9.505ms 193.980us 0 b 0 b 68.16 Mb 0 b 49
aten::flatten 0.09% 701.000us 0.11% 895.000us 18.265us 566.000us 0.06% 959.000us 19.571us 0 b 0 b 0 b 0 b 49
aten::squeeze 0.00% 38.000us 0.01% 40.000us 20.000us 21.000us 0.00% 47.000us 23.500us 0 b 0 b 0 b 0 b 2
aten::pad 0.07% 526.000us 0.48% 3.861ms 160.875us 271.000us 0.03% 2.646ms 110.250us 0 b 0 b 7.50 Mb 0 b 24
aten::constant_pad_nd 0.22% 1.745ms 0.42% 3.335ms 138.958us 764.000us 0.09% 2.375ms 98.958us 0 b 0 b 7.50 Mb 0 b 24
aten::narrow 0.05% 414.000us 0.10% 828.000us 34.500us 412.000us 0.05% 829.000us 34.542us 0 b 0 b 0 b 0 b 24
aten::embedding 0.01% 119.000us 0.05% 378.000us 189.000us 63.000us 0.01% 340.000us 170.000us 0 b 0 b 2.52 Mb -2.00 Kb 2
[memory] 0.00% 0.000us 0.00% 0.000us 0.000us 0.000us 0.00% 0.000us 0.000us 0 b 0 b -1.37 Gb -1.37 Gb 1547
``` | 2,669 | 1,498 |
andyl98 | 2025-01-29T19:29:42 | So I think this PR saves quite a bit of memory. The static cache/dynamic cache suggestion is a good one but currently I don't have bandwidth to look at it :) Will work on it in the future if we think it's a viable path. | 2,669 | 1,499 |
HuggingFaceDocBuilderDev | 2025-01-29T19:43:09 | The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/trl/pr_2669). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. | 2,669 | 1,500 |
qgallouedec | 2025-01-29T22:06:36 | Thank you for this detailed report. The rather large tables are not very readable, could you perhaps give a more synthetic overview? Maybe the final sec/it that you get with a training script? | 2,669 | 1,501 |
andyl98 | 2025-01-29T22:26:18 | Speed-wise (model = 1.5b, B = 1, P = 2k, C = 256, G = 4)
New version:
```
{'loss': 0.0, 'grad_norm': 1.8341439962387085, 'learning_rate': 9.997435897435897e-07, 'completion_length': 146.625, 'rewards/concise_reward': -575.4375, 'rewards/format_reward': 0.0, 'reward': -575.4375, 'reward_std': 236.35992431640625, 'kl': 0.0, 'epoch': 0.03}
{'loss': 0.0, 'grad_norm': 2.196131944656372, 'learning_rate': 9.994871794871794e-07, 'completion_length': 201.4375, 'rewards/concise_reward': -775.375, 'rewards/format_reward': 0.0, 'reward': -775.375, 'reward_std': 332.21124267578125, 'kl': 1.6450881958007812e-05, 'epoch': 0.05}
{'loss': 0.0, 'grad_norm': 1.8303061723709106, 'learning_rate': 9.992307692307693e-07, 'completion_length': 123.5625, 'rewards/concise_reward': -498.0625, 'rewards/format_reward': 0.0, 'reward': -498.0625, 'reward_std': 255.6620635986328, 'kl': 3.647804260253906e-05, 'epoch': 0.08}
{'loss': -0.0, 'grad_norm': 2.763676166534424, 'learning_rate': 9.98974358974359e-07, 'completion_length': 109.9375, 'rewards/concise_reward': -444.1875, 'rewards/format_reward': 0.0, 'reward': -444.1875, 'reward_std': 314.7947082519531, 'kl': -2.9325485229492188e-05, 'epoch': 0.1}
{'loss': -0.0, 'grad_norm': 2.452415704727173, 'learning_rate': 9.987179487179487e-07, 'completion_length': 146.4375, 'rewards/concise_reward': -603.4375, 'rewards/format_reward': 0.0, 'reward': -603.4375, 'reward_std': 300.88116455078125, 'kl': -5.555152893066406e-05, 'epoch': 0.13}
0%|▏ | 5/3900 [01:18<16:06:09, 14.88s/it
```
Original version:
```
{'loss': 0.0, 'grad_norm': 2.3698227405548096, 'learning_rate': 9.997435897435897e-07, 'completion_length': 146.625, 'rewards/concise_reward': -575.4375, 'rewards/format_reward': 0.0, 'reward': -575.4375, 'reward_std': 236.35992431640625, 'kl': 0.0, 'epoch': 0.03}
{'loss': -0.0, 'grad_norm': 3.092586040496826, 'learning_rate': 9.994871794871794e-07, 'completion_length': 199.6875, 'rewards/concise_reward': -748.1875, 'rewards/format_reward': 0.0, 'reward': -748.1875, 'reward_std': 301.6230163574219, 'kl': -3.0279159545898438e-05, 'epoch': 0.05}
{'loss': -0.0, 'grad_norm': 2.243756055831909, 'learning_rate': 9.992307692307693e-07, 'completion_length': 116.25, 'rewards/concise_reward': -460.375, 'rewards/format_reward': 0.0, 'reward': -460.375, 'reward_std': 244.493408203125, 'kl': -2.3692846298217773e-06, 'epoch': 0.08}
{'loss': 0.0, 'grad_norm': 2.633629560470581, 'learning_rate': 9.98974358974359e-07, 'completion_length': 121.5, 'rewards/concise_reward': -498.0, 'rewards/format_reward': 0.0, 'reward': -498.0, 'reward_std': 289.4837646484375, 'kl': 5.9604644775390625e-05, 'epoch': 0.1}
0%|▏ | 4/3900 [01:01<15:48:59, 14.61s/it]
```
Memory usage with the new version is about 33%.
I think the new version will be a little slower (because forward twice instead of once) but I'm sure the bottleneck usually comes from the "generation" part and thus this small difference can be dismissed. | 2,669 | 1,502 |
mchen30 | 2025-02-06T06:17:36 | I ran into an odd behavior with the improved version and was wondering if someone might be able to explain it: when running the forward passes (1: get KV cache, 2: get policy logp, 3: get ref logp) in a standalone script, the memory consumption only increases by about 6GB for a mini-batch in Step 2; however, if the training is started from the trl framework with minimal configuration, the same PEFT model consumes an additional 30GB VRAM for one mini-batch in forward pass Step 2. (For reference, the output logp in that mini-batch is about 320MB in fp16.) And the memory consumption keeps increasing with each mini-batch, so with a moderate-sized G, the GPU quickly runs out of memory.
I tried to keep everything constant between the two setups, i.e. same model/config. I also cached the initial generation results for better comparability.
Is this kind of VRAM consumption normal for trl with PEFT enabled? Any help or insight is greatly appreciated! | 2,669 | 1,503 |
qgallouedec | 2025-02-06T10:09:10 | Thanks @andyl98 for pushing in this direction. Why your approach isn't compatible with gradient checkpointing? | 2,669 | 1,504 |
andyl98 | 2025-02-06T16:46:25 | > Thanks @andyl98 for pushing in this direction. Why your approach isn't compatible with gradient checkpointing?
Ah this is because it uses the prompt token’s past kv cache. If we turn gradient checkpointing on, that field will be empty/not-reusable.
Unless we can do sth like
```python
model.gradient_checkpointing_disable()
model.config.use_cache = True
```
before the prompt forwarding part and turn it off after so it is compatible? Not sure if this is an option. cc @ArthurZucker @qgallouedec if you guys know better solutions in this case. | 2,669 | 1,505 |
andyl98 | 2025-02-06T16:47:24 | > I ran into an odd behavior with the improved version and was wondering if someone might be able to explain it: when running the forward passes (1: get KV cache, 2: get policy logp, 3: get ref logp) in a standalone script, the memory consumption only increases by about 6GB for a mini-batch in Step 2; however, if the training is started from the trl framework with minimal configuration, the same PEFT model consumes an additional 30GB VRAM for one mini-batch in forward pass Step 2. (For reference, the output logp in that mini-batch is about 320MB in fp16.) And the memory consumption keeps increasing with each mini-batch, so with a moderate-sized G, the GPU quickly runs out of memory.
>
> I tried to keep everything constant between the two setups, i.e. same model/config. I also cached the initial generation results for better comparability.
>
> Is this kind of VRAM consumption normal for trl with PEFT enabled? Any help or insight is greatly appreciated!
I haven’t tested PEFT but let me try to understand the situation a bit more. When you launch the training script, are you using accelerate/torch distributed? Maybe I can try to reproduce | 2,669 | 1,506 |
mchen30 | 2025-02-06T23:52:19 | > > I ran into an odd behavior with the improved version and was wondering if someone might be able to explain it: when running the forward passes (1: get KV cache, 2: get policy logp, 3: get ref logp) in a standalone script, the memory consumption only increases by about 6GB for a mini-batch in Step 2; however, if the training is started from the trl framework with minimal configuration, the same PEFT model consumes an additional 30GB VRAM for one mini-batch in forward pass Step 2. (For reference, the output logp in that mini-batch is about 320MB in fp16.) And the memory consumption keeps increasing with each mini-batch, so with a moderate-sized G, the GPU quickly runs out of memory.
> > I tried to keep everything constant between the two setups, i.e. same model/config. I also cached the initial generation results for better comparability.
> > Is this kind of VRAM consumption normal for trl with PEFT enabled? Any help or insight is greatly appreciated!
>
> I haven’t tested PEFT but let me try to understand the situation a bit more. When you launch the training script, are you using accelerate/torch distributed? Maybe I can try to reproduce
Thanks so much! I ran the script directly with python but I think it uses accelerate backend nonetheless. Using accelerate launch results in the same OOM behavior. When I ran your code in a separate script, I used torch/transformers directly without accelerate backend, which end up being significantly more memory efficient despite the computation appearing to be identical. I debugged both setups and the call stacks seem identical. The memory allocations mainly happen during MLP forward pass in each decoder layer, which used code from AutoAWQ GEMM, as well as some from qkv matrices calculation before F.scaled_dot_product_attention. | 2,669 | 1,507 |
Superskyyy | 2025-02-18T15:37:37 | @andyl98 Hello, just a kind reminder is it possible to merge this? Thanks | 2,669 | 1,508 |
andyl98 | 2025-02-18T19:36:25 | > @andyl98 Hello, just a kind reminder is it possible to merge this? Thanks
Hi @Superskyyy , I can resolve the conflicts, but idk if the trl team wants to merge this. Let me do this later tonight | 2,669 | 1,509 |
shayuewt | 2025-03-11T07:49:15 | Hi @andyl98, appreciate for your excellent working. It does save a lot of memory. But the output is a bit different with origin methods.
The version of **transformers==4.45.2**
Below is my test snippet:
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
from transformers.cache_utils import DynamicCache
def initial_context():
tokenizer = AutoTokenizer.from_pretrained('Qwen/Qwen2.5-0.5B-Instruct')
policy_model = AutoModelForCausalLM.from_pretrained(
'Qwen/Qwen2.5-0.5B-Instruct',
torch_dtype=torch.bfloat16,
# _attn_implementation="flash_attention_2",
).cuda()
raw_texts = ["今天天气看起来不错", "三人行，必有我", "乔治奥威尔"]
prompts = [
tokenizer.apply_chat_template(
[
{
"role": "system",
"content": 'you are good assistant.'
},
{
"role": "user",
"content": raw_text
}
],
tokenize=False,
add_generation_prompt=True
)
for raw_text in raw_texts
]
return policy_model, tokenizer, raw_texts, prompts
def complete_texts(policy_model, tokenizer, prompts):
batch_completions = []
gen_config = policy_model.generation_config.to_dict()
gen_config["max_new_tokens"] = 1024
gen_config["num_return_sequences"] = 4
with torch.no_grad():
batch_input_ids = tokenizer(prompts, padding='longest', padding_side='left')
batch_input_ids = torch.LongTensor(batch_input_ids['input_ids']).to(policy_model.device)
# (Batch * Sequence Num, output_token_len)
batch_out_ids = policy_model.generate(
batch_input_ids,
return_dict_in_generate=False,
pad_token_id=tokenizer.pad_token_id,
generation_config=GenerationConfig(**gen_config),
eos_token_id=tokenizer.pad_token_id,
)
batch_size, input_len = batch_input_ids.shape[:2]
for index in range(batch_size):
group_completions = []
for i in range(gen_config["num_return_sequences"]):
curr_completion = tokenizer.decode(
batch_out_ids[index * gen_config["num_return_sequences"] + i][input_len:],
skip_special_tokens=True
)
group_completions.append(curr_completion)
batch_completions.append(group_completions)
return batch_completions, batch_out_ids
def compute_logps(llm_model, text_ids, prompt_len):
"""text_ids is concatenated with prompt_id and response_id """
# 计算每个 position 的 logits
logits = llm_model(text_ids).logits # (B, L, V)
logits = logits[:, :-1, :] # (B, L-1, V), exclude the last logit: it corresponds to the next token pred
text_ids = text_ids[:, 1:] # (B, L-1), exclude the first input ID since we don't have logits for it
# Compute the log probabilities for the input tokens. Use a loop to reduce memory peak.
per_token_logps = []
for logits_row, text_ids_row in zip(logits, text_ids):
log_probs = logits_row.log_softmax(dim=-1)
token_log_prob = torch.gather(log_probs, dim=1, index=text_ids_row.unsqueeze(1)).squeeze(1)
per_token_logps.append(token_log_prob)
per_token_logps = torch.stack(per_token_logps)
per_token_logps = per_token_logps[:, prompt_len - 1:]
return per_token_logps
# Borrow from https://github.com/huggingface/trl/pull/2669
def compute_logps_with_prompt_cache(
model: torch.nn.Module,
prompt_inputs: dict,
completion_ids: torch.LongTensor,
mini_batch_size: int = 0,
requires_grad_for_completion: bool = True,
) -> torch.FloatTensor:
"""
The method will compute the log probabilities of the completion tokens by using the prompt cache.
1) Forward pass on the prompt with torch.no_grad() to get `past_key_values`.
2) Forward pass (with or without grad) on the completion tokens using that cache.
3) Compute per-token log probabilities for the completion.
Args:
model (`nn.Module`): A causal LM (transformers.AutoModelForCausalLM) or similar.
prompt_inputs (`dict`): The dict of prompt tensors, e.g. {"input_ids", "attention_mask", ...}.
completion_ids (`torch.LongTensor`): Shape [B*G, completion_len].
mini_batch_size (`int`): The number of completion rows to process at once.
requires_grad_for_completion (`bool`): Whether to enable gradient for the completion pass.
Returns:
per_token_logps (`torch.FloatTensor`): shape [B*G, completion_len],
where per_token_logps[i, t] is the logprob of ith completion's t-th completion token,
given all preceding tokens in the prompt + the partial completion up to t-1.
"""
# Get the batch size (B), number of completions (G), and completion length (C)
B = prompt_inputs["input_ids"].size(0)
G = completion_ids.size(0) // B
C = completion_ids.size(1)
# If the user did not specify a mini_batch_size, use the full batch size (B*G)
if mini_batch_size <= 0:
mini_batch_size = completion_ids.size(0)
# Forward pass over prompt tokens to get 2 things with torch.no_grad:
# 1) `past_key_values` (KV cache)
# 2) `prompt_last_logps` (the logprobs of the first completion token prediction)
with torch.no_grad():
prompt_out = model(**prompt_inputs, use_cache=True, num_logits_to_keep=1)
# Only keep the last prompt logit, immediately convert to log probabilities and expand to B*G
prompt_last_logps = prompt_out.logits[:, -1:].log_softmax(dim=-1).repeat_interleave(G, dim=0)
# Gather the these log probs as they relates to the first completion token
first_completion_token_logps = torch.gather(
prompt_last_logps, dim=-1, index=completion_ids[:, :1].unsqueeze(-1)
).squeeze(-1)
# Expand the KV Cache `G` times to match the dimension of completion_ids (B -> B*G) and split into mini-batches
# repeated_kv_cache = prompt_out.past_key_values # a DynamicCache
repeated_kv_cache = DynamicCache.from_legacy_cache(prompt_out.past_key_values)
repeated_kv_cache.batch_repeat_interleave(G)
mini_batch_kv_caches = repeated_kv_cache.batch_split(full_batch_size=B * G, split_size=mini_batch_size)
# Process completion tokens in mini-batches
completion_token_logps = []
for batch_idx, mini_batch_kv_cache in enumerate(mini_batch_kv_caches):
start_idx = batch_idx * mini_batch_size
end_idx = start_idx + mini_batch_size
mini_batch_ids = completion_ids[start_idx:end_idx] # (mini_batch_size, C)
with torch.set_grad_enabled(requires_grad_for_completion):
mini_batch_logits = model(
input_ids=mini_batch_ids,
past_key_values=mini_batch_kv_cache,
num_logits_to_keep=C,
use_cache=False,
).logits[:, -C:-1, :]
# # Original method
# mini_batch_log_probs = mini_batch_logits.log_softmax(dim=-1)
# del mini_batch_logits
# mini_batch_token_log_prob = torch.gather(mini_batch_log_probs, dim=-1, index=mini_batch_index).squeeze(-1)
# del mini_batch_log_probs
# More optimized method (https://github.com/huggingface/trl/pull/2773)
# Get the corresponding completion token ids and gather the logits for completion_ids w/ idx >= 1
mini_batch_index = mini_batch_ids[:, 1:].unsqueeze(-1) # (mini_batch_size, C-1, 1)
# (mini_batch_size, C-1)
mini_batch_token_logits = torch.gather(mini_batch_logits, dim=-1, index=mini_batch_index).squeeze(-1)
# (mini_batch_size, C-1)
mini_batch_logsumexp_values = torch.stack([torch.logsumexp(_logits, dim=-1) for _logits in mini_batch_logits])
del mini_batch_logits
mini_batch_token_log_prob = mini_batch_token_logits - mini_batch_logsumexp_values # (mini_batch_size, C-1)
completion_token_logps.append(mini_batch_token_log_prob)
del mini_batch_token_logits, mini_batch_logsumexp_values, mini_batch_token_log_prob
# Combine results
all_completion_token_logps = torch.cat(completion_token_logps, dim=0) # (B*G, C-1)
return torch.cat([first_completion_token_logps, all_completion_token_logps], dim=1) # (B*G, C)
def compare():
print("initializing...")
policy_model, tokenizer, raw_texts, prompts = initial_context()
print("generating texts...")
completions, prompt_completion_ids = complete_texts(policy_model, tokenizer, prompts)
prompt_completion_ids = prompt_completion_ids.to(policy_model.device)
print("tokenizing...")
prompt_inputs = tokenizer(
prompts,
return_tensors="pt",
padding=True,
padding_side="left",
add_special_tokens=False
)
for key in prompt_inputs.keys():
prompt_inputs[key] = prompt_inputs[key].to(policy_model.device)
prompt_length = prompt_inputs['input_ids'].size(1)
print("comparing...")
logps_0 = compute_logps(policy_model, prompt_completion_ids, prompt_length)
print(logps_0.shape)
logps_1 = compute_logps_with_prompt_cache(
policy_model,
prompt_inputs,
prompt_completion_ids[:, prompt_length:],
prompt_completion_ids.size(0)
)
print(logps_1.shape)
idx = 3
print(logps_0[idx, :10])
print(logps_1[idx, :10])
compare()
```
Here is the output:
Is it reasonable with `use_cache=True`?
Looking forward to your reply.
| 2,669 | 1,510 |
andyl98 | 2025-03-11T18:46:28 | Hi @shayue111 , thanks for letting me know and this does raise an alarm from my side about how we handle different precisions `float32/bfloat16/float16` etc.
TL; DR:
1. The `compute_logps` method in your example is not 100% correct because you have a batch of prompts with different lens but you are not putting the correct `attention_mask` for log_prob calculation, more details in https://github.com/huggingface/trl/pull/2708. Therefore, I changed your version to only test the case where `B=1`. You can change your `raw_text` to `raw_texts = ["今天天气看起来不错"]`
2. You raised a very good concern where I also noticed there's quite some discrepancies if I use `bfloat16` instead of `float32`.
If I load the model in `float32`, I'll get all close.
```python
initializing...
generating texts...
tokenizing...
comparing...
torch.Size([4, 39])
torch.Size([4, 39])
tensor([-1.3976, -0.0308, -0.0051, -2.8322, -1.4581, -0.0333, -0.0924, -1.6681,
-0.0110, -2.8074], device='cuda:0', grad_fn=<SliceBackward0>)
tensor([-1.3976, -0.0308, -0.0051, -2.8322, -1.4581, -0.0333, -0.0924, -1.6681,
-0.0110, -2.8074], device='cuda:0', grad_fn=<SliceBackward0>)
```
but if I load in `bfloat16` it's indeed different
```python
initializing...
generating texts...
tokenizing...
comparing...
torch.Size([4, 57])
torch.Size([4, 57])
tensor([-1.4609, -0.0300, -0.0054, -2.8438, -3.2344, -0.4453, -0.0659, -1.5625,
-0.0057, -2.6094], device='cuda:0', dtype=torch.bfloat16,
grad_fn=<SliceBackward0>)
tensor([-1.4297, 0.0000, 0.0000, -2.8750, -3.1250, -0.3750, 0.0000, -1.6250,
0.0000, -2.6250], device='cuda:0', dtype=torch.bfloat16,
grad_fn=<SliceBackward0>)
```
I then tried a different implementation for the prompt_cache method with upcasting some tensors to `float32` (e.g. `past_key_values`, logits, etc) and I still find some gap between the original method vs mine.
I think it would be tricky to "fix" as I need to read the entire code base and see where upcasting happens when dealing with lower precisions in the model loading process (i.e., maybe certain tensors such as `past_key_values` requiers to be upcasted but others do not). I'll try to investigate further but really appreciate you raising up the issue.
Therefore, at this moment, feel free to use this branch "only if" you use float32 for now. 😢 | 2,669 | 1,511 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.