Train LLMs in 50 lines of code. This is a reference code for YouTube tutorial: https://www.youtube.com/watch?v=JNMVulH7fCo&ab_channel=AbhishekThakur

llm_training_sft.py

import torch
from datasets import load_dataset
from peft import LoraConfig, get_peft_model, prepare_model_for_int8_training
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from trl import SFTTrainer


def train():
    train_dataset = load_dataset("tatsu-lab/alpaca", split="train")
    tokenizer = AutoTokenizer.from_pretrained("Salesforce/xgen-7b-8k-base", trust_remote_code=True)
    tokenizer.pad_token = tokenizer.eos_token
    model = AutoModelForCausalLM.from_pretrained(
        "Salesforce/xgen-7b-8k-base", load_in_4bit=True, torch_dtype=torch.float16, device_map="auto"
    )
    model.resize_token_embeddings(len(tokenizer))
    model = prepare_model_for_int8_training(model)
    peft_config = LoraConfig(r=16, lora_alpha=32, lora_dropout=0.05, bias="none", task_type="CAUSAL_LM")
    model = get_peft_model(model, peft_config)

    training_args = TrainingArguments(
        output_dir="xgen-7b-tuned-alpaca-l1",
        per_device_train_batch_size=4,
        optim="adamw_torch",
        logging_steps=100,
        learning_rate=2e-4,
        fp16=True,
        warmup_ratio=0.1,
        lr_scheduler_type="linear",
        num_train_epochs=1,
        save_strategy="epoch",
        push_to_hub=True,
    )
    trainer = SFTTrainer(
        model=model,
        train_dataset=train_dataset,
        dataset_text_field="text",
        max_seq_length=1024,
        tokenizer=tokenizer,
        args=training_args,
        packing=True,
        peft_config=peft_config,
    )
    trainer.train()
    trainer.push_to_hub()


if __name__ == "__main__":
    train()

Hi Abhishek,

I am running the same code in colab notebook, I am getting error :
│ │
│ /usr/local/lib/python3.10/dist-packages/peft/tuners/lora.py:565 in forward │
│ │
│ 562 │ │ │ │ self.unmerge() │
│ 563 │ │ │ result = F.linear(x, transpose(self.weight, self.fan_in_fan_out), bias=self. │
│ 564 │ │ elif self.r[self.active_adapter] > 0 and not self.merged: │
│ ❱ 565 │ │ │ result = F.linear(x, transpose(self.weight, self.fan_in_fan_out), bias=self. │
│ 566 │ │ │ │
│ 567 │ │ │ x = x.to(self.lora_A[self.active_adapter].weight.dtype) │
│ 568 │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
RuntimeError: mat1 and mat2 shapes cannot be multiplied (4096x4096 and 1x8388608)

The same error is encountered when i run the fine-tuning by autotrain command, please help for the same ?

seems like an issue with bitsandbytes. could you please open an issue on bitsandbytes repo?

I also have this same issue using: transformers: 4.30.2
bitsandbytes: 0.40.0. Have created bitsandbytes-foundation/bitsandbytes#600