DeepSpeed with trl #2490
Labels
🐛 bug
Something isn't working
🚀 deepspeed
Related to deepspeed
🏋 DPO
Related to DPO
⏳ needs more info
Additional information or clarification is required to proceed
Comments
qgallouedec
added
🐛 bug
|
Thanks for reporting, please provide a minimal code/steps to reproduce this. |
|
pipeline.zip (edit by maintainer: remove link) thanks @qgallouedec The attached files constitute a pipeline that using the DPOTrainer with DeepSpeed. |
|
Sorry but we don't use zip files. The easy way to provide a MRE is to go line by line, if the error remains when you remove it, then you can discard the line. When there is no line left to remove, you have your MRE |
|
Sorry @qgallouedec , here a minimal code of my pipeline: import argparse
import json
import math
import pandas as pd
import torch
from accelerate import Accelerator
from datasets import Dataset
from torch import optim, nn
from torch.optim.lr_scheduler import LambdaLR
from tqdm import tqdm
from transformers import AutoTokenizer, AutoModelForMaskedLM, AutoModel, AutoModelForCausalLM
from trl import PPOConfig, PPOTrainer, AutoModelForCausalLMWithValueHead, DPOConfig, DPOTrainer
from huggingface_hub import login
import os
import itertools
model_RLRF = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B-Instruct", torch_dtype=torch.float32)
tokenizer_RLRF = AutoTokenizer.from_pretrained(model_RLRF_name_or_path)
tokenizer_RLRF.add_special_tokens({'pad_token': tokenizer_RLRF.eos_token})
tokenizer_RLRF.padding_side = 'left'
DPO_config = DPOConfig(
report_to='tensorboard',
logging_first_step=True,
"per_device_train_batch_size": 4,
"gradient_accumulation_steps": 1,
"sync_ref_model": true,
"ref_model_mixup_alpha": 0.6,
"ref_model_sync_steps": 256,
"bf16": True,
)
# Create reference model:
parameter_names = [n for n, _ in model_RLRF.named_parameters()]
ref_model = deepcopy(model)
# if no layers are shared, return copy of model
for param_name in parameter_names:
param = ref_model.get_parameter(param_name)
param.requires_grad = False
ref_model.eval()
# Set optimizer for RLRF
optimizer_RLRF = optim.AdamW(filter(lambda param: param.requires_grad, model_RLRF.parameters()),
lr=HP_RLRF.get('learning_rate', 1.41e-5))
train_dataset= pd.read_csv("perfernces_dataset_from_ranker_train_queries_and_baseline_doc.csv")
train_dataset= Dataset.from_pandas(train_dataset_RLRF)
dpo_trainer = DPOTrainer(model=model, args=DPO_config, processing_class=tokenizer_RLRF, ref_model=ref_model,
optimizers=(optimizer, None), train_dataset=train_dataset)
dpo_trainer.train()the loaded data file (train_dataset) is: |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
System Info
Describe the bug
I am trying to train meta-llama/Llama-3.1-8B-Instruct with trl DPOTrainer.
After creating the trainer and starting the training loop, I'm getting the following error (in the forward pass):
I tried to downgrade transformers with no success.
System info:
my accelerate config:
trl env output:
Information
Tasks
examplesfolderReproduction
outputs:
Expected behavior
Train my model using the DPOTrainer and deepspeed
Checklist
The text was updated successfully, but these errors were encountered: