When I used galore on orpo, the learning rate was set to 8e-6, but the training rate was 0.01

[Minami-su]

trainer = ORPOTrainer(
        model=model,
        train_dataset=dataset["train"],
        eval_dataset=dataset["test"],

        #peft_config=peft_config,
        tokenizer=tokenizer,
        args= ORPOConfig(
            max_length=cutoff_len,
            max_prompt_length=cutoff_len//2,
            beta=0.1,
            per_device_train_batch_size=micro_batch_size,
            gradient_accumulation_steps=gradient_accumulation_steps,
            warmup_steps=0,
            num_train_epochs=num_epochs,
            lr_scheduler_type="cosine",
            learning_rate=8e-6,
            bf16=True,
            logging_steps=10,
            optim = "galore_adamw_8bit_layerwise",
            optim_target_modules=[r".*attn.*", r".*mlp.*"],
            optim_args="rank=1024, update_proj_gap=500, scale=0.25",
            evaluation_strategy="steps" if val_set_size > 0 else "no",
            save_strategy="steps",
            eval_steps=100 if val_set_size > 0 else None,
            save_steps=100,
            output_dir=output_dir,
            save_total_limit=2,
            gradient_checkpointing=True, 
            gradient_checkpointing_kwargs={'use_reentrant':True},
            load_best_model_at_end=True if val_set_size > 0 else False,
            ddp_find_unused_parameters=False if ddp else None,
            report_to="wandb" if use_wandb else None,
            run_name=wandb_run_name if use_wandb else None,
            do_train=True,
            remove_unused_columns=False,
        )
    )

Activated GaLoRE fine-tuning, depending on your model size and hardware, the training might take a while before starting. Please be patient !
  0%|                                                                                                                                     | 0/495 [00:00<?, ?it/s]Could not estimate the number of tokens of the input, floating-point operations will not be computed
{'loss': 0.3557, 'grad_norm': 0.0, 'learning_rate': 0.001, 'rewards/chosen': -0.015678538009524345, 'rewards/rejected': -0.012379011139273643, 'rewards/accuracies': 0.19999998807907104, 'rewards/margins': -0.003299527335911989, 'logps/rejected': -0.12379010766744614, 'logps/chosen': -0.15678536891937256, 'logits/rejected': 0.7921055555343628, 'logits/chosen': 0.791210412979126, 'nll_loss': 0.2719877064228058, 'log_odds_ratio': -0.8374900817871094, 'log_odds_chosen': -0.25091928243637085, 'epoch': 0.06}
{'loss': 0.2634, 'grad_norm': 0.0, 'learning_rate': 0.001, 'rewards/chosen': -0.012010233476758003, 'rewards/rejected': -0.009977776557207108, 'rewards/accuracies': 0.29999998211860657, 'rewards/margins': -0.0020324576180428267, 'logps/rejected': -0.09977775812149048, 'logps/chosen': -0.12010233104228973, 'logits/rejected': 0.7489851713180542, 'logits/chosen': 0.7482139468193054, 'nll_loss': 0.1832979917526245, 'log_odds_ratio': -0.8010236620903015, 'log_odds_chosen': -0.16869042813777924, 'epoch': 0.12}
{'loss': 0.2482, 'grad_norm': 0.0, 'learning_rate': 0.001, 'rewards/chosen': -0.011346157640218735, 'rewards/rejected': -0.01022450439631939, 'rewards/accuracies': 0.4833333492279053, 'rewards/margins': -0.0011216530110687017, 'logps/rejected': -0.102245032787323, 'logps/chosen': -0.11346157640218735, 'logits/rejected': 0.7105721831321716, 'logits/chosen': 0.7108334898948669, 'nll_loss': 0.17242279648780823, 'log_odds_ratio': -0.7573043704032898, 'log_odds_chosen': -0.08471358567476273, 'epoch': 0.18}
{'loss': 0.2444, 'grad_norm': 0.0, 'learning_rate': 0.001, 'rewards/chosen': -0.012975988909602165, 'rewards/rejected': -0.013058923184871674, 'rewards/accuracies': 0.550000011920929, 'rewards/margins': 8.293241262435913e-05, 'logps/rejected': -0.13058921694755554, 'logps/chosen': -0.12975989282131195, 'logits/rejected': 0.6808757781982422, 'logits/chosen': 0.6832461953163147, 'nll_loss': 0.1756206750869751, 'log_odds_ratio': -0.687309741973877, 'log_odds_chosen': 0.04155167192220688, 'epoch': 0.24}

[younesbelkada]

Hi @Minami-su I think this has been fixed on the recent transformers, what transformers version are you using?

[Minami-su]

Hi @Minami-su I think this has been fixed on the recent transformers, what transformers version are you using?

I've tried transformers 4.41.1,transformers 4.42.0.dev0 and they didn't work.

[Minami-su]

I guess because it didn't use trl's orpo.

[Minami-su]

SimPo:https://github.com/princeton-nlp/SimPO/blob/main/scripts/simpo_trainer.py command： CUDA_VISIBLE_DEVICES=1 python ft-SimPo-galore.py \ --base_model Meta-Llama-3-8B-Instruct-hard-iter2 \ --data_path gpt4_1k_hard_SFT_trl_iter3.json \ --output_dir gpt4_1k_hard_SFT_split_3_galore \ --batch_size 6 \ --micro_batch_size 6 \ --num_epochs 3 \ --learning_rate 8e-6 \ --cutoff_len 1400 \ --val_set_size 0 \ --train_on_inputs \ --group_by_length

ft-SimPo-galore.py：

import fire
import torch
import transformers
from datasets import load_dataset
import os
from simpo_trainer import SimPOTrainer
os.environ["NCCL_P2P_DISABLE"] = "1"
os.environ["NCCL_IB_DISABLE"] = "1"

os.environ['CUDA_LAUNCH_BLOCKING'] = '1'

"""
Unused imports:
import torch.nn as nn
import bitsandbytes as bnb
"""

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, DataCollatorForLanguageModeling
from peft import (
    LoraConfig,
    get_peft_model,
    prepare_model_for_kbit_training,
    set_peft_model_state_dict,
)
import signal
import sys
import os
import bitsandbytes as bnb

# torch.autograd.set_detect_anomaly(True)
def find_all_linear_names(model):
    # cls = bnb.nn.Linear8bitLt
    cls = bnb.nn.Linear4bit
    lora_module_names = set()
    for name, module in model.named_modules():
        if isinstance(module, cls):
            names = name.split('.')
            lora_module_names.add(names[0] if len(names) == 1 else names[-1])

    if 'lm_head' in lora_module_names:  # needed for 16-bit
        lora_module_names.remove('lm_head')
    return list(lora_module_names)
os.environ["TOKENIZERS_PARALLELISM"] = "false"
os.environ["WANDB_DISABLED"] = "true"
def train(
    # model/data params
    base_model: str = "",  # the only required argument
    data_path: str = "yahma/alpaca-cleaned",
    output_dir: str = "./lora-alpaca",
    # training hyperparams
    batch_size: int = 128,
    micro_batch_size: int = 4,
    num_epochs: int = 3,
    learning_rate: float = 3e-4,
    cutoff_len: int = 256,
    max_prompt_length: int = 512,
    val_set_size: int = 2000,
    # lora hyperparams
    lora_r: int = 8,
    lora_alpha: int = 16,
    lora_dropout: float = 0.05,
    beta: float = 2.5,
    gamma: float = 1.4,
    # llm hyperparams
    train_on_inputs: bool = True,  # if False, masks out inputs in loss
    add_eos_token: bool = False,
    group_by_length: bool = False,  # faster, but produces an odd training loss curve
    # wandb params
    wandb_project: str = "",
    wandb_run_name: str = "",
    wandb_watch: str = "",  # options: false | gradients | all
    wandb_log_model: str = "",  # options: false | true
    resume_from_checkpoint: str = None,  # either training checkpoint or final adapter
    prompt_template_name: str = "alpaca2",  # The prompt template to use, will default to alpaca.
):
    if int(os.environ.get("LOCAL_RANK", 0)) == 0:
        print(
            f"Training Alpaca-LoRA model with params:\n"
            f"base_model: {base_model}\n"
            f"data_path: {data_path}\n"
            f"output_dir: {output_dir}\n"
            f"batch_size: {batch_size}\n"
            f"micro_batch_size: {micro_batch_size}\n"
            f"num_epochs: {num_epochs}\n"
            f"learning_rate: {learning_rate}\n"
            f"cutoff_len: {cutoff_len}\n"
            f"val_set_size: {val_set_size}\n"
            f"lora_r: {lora_r}\n"
            f"lora_alpha: {lora_alpha}\n"
            f"lora_dropout: {lora_dropout}\n"
            f"train_on_inputs: {train_on_inputs}\n"
            f"add_eos_token: {add_eos_token}\n"
            f"group_by_length: {group_by_length}\n"
            f"wandb_project: {wandb_project}\n"
            f"wandb_run_name: {wandb_run_name}\n"
            f"wandb_watch: {wandb_watch}\n"
            f"wandb_log_model: {wandb_log_model}\n"
            f"resume_from_checkpoint: {resume_from_checkpoint or False}\n"
            f"prompt template: {prompt_template_name}\n"
        )
    assert (
        base_model
    ), "Please specify a --base_model, e.g. --base_model='huggyllama/llama-7b'"
    gradient_accumulation_steps = batch_size // micro_batch_size

    device_map = "auto"
    world_size = int(os.environ.get("WORLD_SIZE", 1))
    ddp = world_size != 1
    if ddp:
        device_map = {"": int(os.environ.get("LOCAL_RANK") or 0)}
        gradient_accumulation_steps = gradient_accumulation_steps // world_size

    # Check if parameter passed or if set within environ
    use_wandb = len(wandb_project) > 0 or (
        "WANDB_PROJECT" in os.environ and len(os.environ["WANDB_PROJECT"]) > 0
    )
    # Only overwrite environ if wandb param passed
    if len(wandb_project) > 0:
        os.environ["WANDB_PROJECT"] = wandb_project
    if len(wandb_watch) > 0:
        os.environ["WANDB_WATCH"] = wandb_watch
    if len(wandb_log_model) > 0:
        os.environ["WANDB_LOG_MODEL"] = wandb_log_model

    tokenizer = AutoTokenizer.from_pretrained(base_model,trust_remote_code=True)

    if base_model.find("qwen") != -1 or base_model.find("Qwen") != -1:
        tokenizer.add_special_tokens({"bos_token": "<|im_start|>"})
        tokenizer.add_special_tokens({"eos_token": "<|im_end|>"})
        tokenizer.add_special_tokens({"pad_token": "<|endoftext|>"})

    elif base_model.find("Llama-3") != -1:
        tokenizer.add_special_tokens({"eos_token": "<|eot_id|>"})
        tokenizer.add_special_tokens({"pad_token": "<|eot_id|>"})
    #tokenizer.padding_side = "left"  # Allow batched inference
    def save_model(signal, frame):
        print("\nSaving the model...")
        model.save_pretrained(output_dir)
        sys.exit(0)

    def process(row):
        # row["chosen"] = row["chosen"]
        # row["rejected"] = row["rejected"]
        row["prompt"] = row["prompt"]
        row["chosen"] = tokenizer.apply_chat_template(row["chosen"], tokenize=False)
        row["rejected"] = tokenizer.apply_chat_template(row["rejected"], tokenize=False)
        # row["chosen"] = tokenizer.apply_chat_template(row["chosen"], tokenize=False)
        # row["rejected"] = tokenizer.apply_chat_template(row["rejected"], tokenize=False)
        return row

    print(tokenizer.pad_token_id)
    print(tokenizer.pad_token)
    print(tokenizer.bos_token_id)
    print(tokenizer.bos_token)
    print(tokenizer.eos_token_id)
    print(tokenizer.eos_token)
    if data_path.endswith(".json") or data_path.endswith(".jsonl"):
        data = load_dataset("json", data_files=data_path)
    else:
        data = load_dataset(data_path)
    data = data.map(process,num_proc=1)
    if val_set_size > 0:
        train_val = data["train"].train_test_split(
            test_size=val_set_size, shuffle=True, seed=42
        )
        train_data = train_val["train"].shuffle()
        val_data = train_val["test"].shuffle()
    else:
        train_data = data["train"].shuffle()
        val_data = None
    print(len(train_data))
    model = AutoModelForCausalLM.from_pretrained(base_model,
                                            trust_remote_code=True,
                                             attn_implementation="flash_attention_2",
                                             torch_dtype=torch.bfloat16,
                                             device_map=device_map,
                                             )

    if resume_from_checkpoint:
        # Check the available weights and load them
        checkpoint_name = os.path.join(
            resume_from_checkpoint, "pytorch_model.bin"
        )  # Full checkpoint
        if not os.path.exists(checkpoint_name):
            checkpoint_name = os.path.join(
                resume_from_checkpoint, "adapter_model.bin"
            )  # only LoRA model - LoRA config above has to fit
            resume_from_checkpoint = (
                False  # So the trainer won't try loading its state
            )
        # The two files above have a different name depending on how they were saved, but are actually the same.
        if os.path.exists(checkpoint_name):
            print(f"Restarting from {checkpoint_name}")
            adapters_weights = torch.load(checkpoint_name)
            set_peft_model_state_dict(model, adapters_weights)
        else:
            print(f"Checkpoint {checkpoint_name} not found")

    if not ddp and torch.cuda.device_count() > 1:
        # keeps Trainer from trying its own DataParallelism when more than 1 gpu is available
        model.is_parallelizable = True
        model.model_parallel = True
   # data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
    training_args=transformers.TrainingArguments(
            per_device_train_batch_size=micro_batch_size,
            gradient_accumulation_steps=gradient_accumulation_steps,
            warmup_steps=0,#(num_epochs*len(train_data)/batch_size)//10,
            num_train_epochs=num_epochs,
            learning_rate=learning_rate,
            bf16=True,
            logging_steps=10,
            optim = "galore_adamw_8bit_layerwise",
            optim_target_modules=[r".*attn.*", r".*mlp.*"],
            optim_args="rank=1024, update_proj_gap=500, scale=0.25,proj_type=std",
            evaluation_strategy="steps" if val_set_size > 0 else "no",
            save_strategy="steps",
            eval_steps=100 if val_set_size > 0 else None,
            save_steps=200,
            output_dir=output_dir,
            save_total_limit=2,
            gradient_checkpointing=True, 
            gradient_checkpointing_kwargs={'use_reentrant':True},
            load_best_model_at_end=True if val_set_size > 0 else False,
            #ddp_find_unused_parameters=False if ddp else None,
            report_to="wandb" if use_wandb else None,
            run_name=wandb_run_name if use_wandb else None,
            do_train=True,
            remove_unused_columns=False,

        )
    ref_model=None
    trainer = SimPOTrainer(
        model=model,
        ref_model=ref_model, # pass in to bypass DPO Trainer check for ref model but is not actually used
        #model_init_kwargs=model_kwargs,
        args=training_args,
        beta=beta,
        gamma=gamma,
        train_dataset=train_data,
        eval_dataset=val_data,
        tokenizer=tokenizer,
        max_length=cutoff_len,
        max_prompt_length=cutoff_len//2,
        #loss_type=training_args.loss_type,
    )
    signal.signal(signal.SIGINT, save_model)

    trainer.train()
    model.save_pretrained(output_dir)

    print(
        "\n If there's a warning about missing keys above, please disregard :)"
    )

if __name__ == "__main__":
    fire.Fire(train)

output:{'loss': 1.7515, 'grad_norm': 0.0, 'learning_rate': 0.001, 'rewards/chosen': -0.8618572354316711, 'rewards/rejected': -0.702195405960083, 'rewards/accuracies': 0.0833333358168602, 'rewards/margins': -0.15966185927391052, 'logps/rejected': -0.28087812662124634, 'logps/chosen': -0.34474286437034607, 'logits/rejected': 0.4580184519290924, 'logits/chosen': 0.4565271735191345, 'epoch': 0.06} {'loss': 1.6516, 'grad_norm': 0.0, 'learning_rate': 0.001, 'rewards/chosen': -0.9827778935432434, 'rewards/rejected': -0.9460451006889343, 'rewards/accuracies': 0.4166666865348816, 'rewards/margins': -0.03673281893134117, 'logps/rejected': -0.3784180283546448, 'logps/chosen': -0.3931111693382263, 'logits/rejected': 0.47764071822166443, 'logits/chosen': 0.47824805974960327, 'epoch': 0.12} {'loss': 1.5422, 'grad_norm': 0.0, 'learning_rate': 0.001, 'rewards/chosen': -1.096510648727417, 'rewards/rejected': -1.19998037815094, 'rewards/accuracies': 0.6000000238418579, 'rewards/margins': 0.10346974432468414, 'logps/rejected': -0.479992151260376, 'logps/chosen': -0.4386042654514313, 'logits/rejected': 0.5151220560073853, 'logits/chosen': 0.5120567679405212, 'epoch': 0.18} {'loss': 1.4713, 'grad_norm': 0.0, 'learning_rate': 0.001, 'rewards/chosen': -1.1208069324493408, 'rewards/rejected': -1.3170217275619507, 'rewards/accuracies': 0.7500000596046448, 'rewards/margins': 0.1962147206068039, 'logps/rejected': -0.5268087387084961, 'logps/chosen': -0.4483228325843811, 'logits/rejected': 0.5599152445793152, 'logits/chosen': 0.5529050827026367, 'epoch': 0.24} {'loss': 1.2346, 'grad_norm': 0.0, 'learning_rate': 0.001, 'rewards/chosen': -1.534162163734436, 'rewards/rejected': -2.1145620346069336, 'rewards/accuracies': 0.800000011920929, 'rewards/margins': 0.5803996920585632, 'logps/rejected': -0.8458248376846313, 'logps/chosen': -0.6136649250984192, 'logits/rejected': 0.6048283576965332, 'logits/chosen': 0.5854496359825134, 'epoch': 0.3}

[Minami-su]

trl version 0.8.6,transformers version 4.41.1

[younesbelkada]

Thanks, will have a look !

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.