Remove graph breaks for torch.compile() in padding free branch in DataCollatorForCompletionOnlyLM

[Abhishek-TAMU]

What does this PR do?

This PR adds cu_seq_lens_q, cu_seq_lens_k, max_length_k, max_length_q to the batch in DataCollatorForCompletionOnlyLM. This, together with a PR in transformers (link to be added), removes graph breaks in padding-free tuning, allowing for maximum performance to be obtained.
Specifically, these parameters should be generated here (this PR change), outside of the transformers loop, as they incur a cpu-gpu sync that is unavoidable. Otherwise, this cpu-gpu sync happens here, inside the attention call which causes graph breaks and hence the transformers PR removes this call to remove all graph breaks when torch_compile flag is turned on in Training arguments to use in SFTTrainer.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a GitHub issue? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[Abhishek-TAMU]

CC: @kashif @qgallouedec

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[qgallouedec]

Hi, thanks for the PR.
Can you provide the link of the PR in transformers? Is it huggingface/transformers#33932?

[qgallouedec]

Could you provide a simple test to:

1. Confirm that it is a case of non-functioning.
2. Verify that this addition resolves it.

It might also be helpful to add a few comments, as these lines are unclear without context.

[Abhishek-TAMU]

Thank you @qgallouedec for the review. This is the related transformers PR which is approved and merged.

I added 2 test cases. One where Tuning fails with Padding and another where doesn't fail without padding.

[Abhishek-TAMU]

@kashif @qgallouedec Could you possibly review this PR ? Thank you!

[Abhishek-TAMU]

Hi @kashif @qgallouedec, could you please take another look at this PR when you get the chance? The changes in this PR are urgent for making torch_compile flag in SFTTrainer work for Llama models (LlamaForCausalLM). This is important for users who need to compile the Llama model using SFTTrainer (in padding_free mode) without any graph breaks. Thank you!

[qgallouedec]

is dataset formatting required here, or can we drop it?

[Abhishek-TAMU]

dataset formatting is required because the SFTTrainer and DataCollatorForCompletionOnlyLM expect the dataset to have a specific format—a single text field that combines both the prompt and the completion in a way the model can understand. This function includes both the prompt and completion, ensuring the data collator can correctly identify where the completion starts using the response_template.

[qgallouedec]

Does the issue only occur with cuda device? In other words can we reproduce on cpu?

[Abhishek-TAMU]

Due to usage of flash_attention_2 it would work only on GPU.

[qgallouedec]

Hey @Abhishek-TAMU, to keep you posted with the current status of the PR, I am struggling reproducing the initial error. Do you have a MRE by any chance? The code from the unittest gives

...
  File "/fsx/qgallouedec/miniconda3/envs/trl/lib/python3.11/site-packages/torch/_dynamo/variables/functions.py", line 134, in __init__
    assert isinstance(
AssertionError: expected FunctionType found _lru_cache_wrapper <functools._lru_cache_wrapper object at 0x7f000e67fb60>

from user code:
   File "/fsx/qgallouedec/transformers/src/transformers/models/llama/modeling_llama.py", line 1224, in torch_dynamo_resume_in_forward_at_1199
    loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)

Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information


You can suppress this exception and fall back to eager by setting:
    import torch._dynamo
    torch._dynamo.config.suppress_errors = True

  0%|          | 0/2 [00:07<?, ?it/s]  

and it doesn't seem related

[Abhishek-TAMU]

Thank you @qgallouedec for looking into this. Sharing you the code which would produce graph break.
Using latest release version of transformers (which doesn't have huggingface/transformers#33932 changes) and latest trl including changes from this PR.

If this change huggingface/transformers#33932 is used in transformers then Graph break could be avoided.

import os, tempfile, torch
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer)
from trl import SFTConfig, SFTTrainer
from trl.trainer import DataCollatorForCompletionOnlyLM
from datasets import load_dataset

standard_prompt_completion_dataset = load_dataset(
    "trl-internal-testing/zen", "standard_prompt_completion"
)

os.environ["CUDA_VISIBLE_DEVICES"]="0"
os.environ["CUDA_HOME"]="/home/tuning/.local/cuda-12.1"
model_id = "trl-internal-testing/tiny-random-LlamaForCausalLM"
torch_dtype = getattr(torch, "bfloat16", None)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch_dtype, attn_implementation="flash_attention_2")
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)

formatted_dataset = lambda example: {
"output": f"### prompt:\n{example['prompt'].strip()}\n\n### completion:\n{example['completion'].strip()}{tokenizer.eos_token}"
}

train_dataset = standard_prompt_completion_dataset["train"].map(formatted_dataset)

response_template = "### completion:\n"
data_collator = DataCollatorForCompletionOnlyLM(response_template, tokenizer=tokenizer, padding_free=True)

with tempfile.TemporaryDirectory() as tmp_dir:
    training_args = SFTConfig(
        output_dir=tmp_dir,
        dataloader_drop_last=True,
        max_steps=2,
        per_device_train_batch_size=2,
        gradient_accumulation_steps=1,
        save_steps=2,
        learning_rate=1e-5,
        dataset_text_field="output",
        torch_compile=True,
        torch_compile_backend="inductor",
        torch_compile_mode="default"
    )

    trainer = SFTTrainer(
        model=model,
        tokenizer=tokenizer,
        train_dataset=train_dataset,
        data_collator=data_collator,
        args=training_args,
    )

    # with assertRaises(Exception):
    trainer.train()
    del os.environ["CUDA_VISIBLE_DEVICES"]

[Abhishek-TAMU]

Hi @qgallouedec, were you able to reproduce the initial error with this MRE ?