[Python] RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn...

This answer did not solve my problem.

I was trying to fine tune llama model locally using LoRA, peft. Here is my code:

import torch
from datasets import load_dataset
from transformers import (
AutoModelForCausalLM,
AutoTokenizer,
BitsAndBytesConfig,
TrainingArguments,
pipeline,
logging,
)
from peft import LoraConfig, PeftModel, get_peft_model, prepare_model_for_kbit_training
from trl import SFTTrainer

base_model_id = "/ml/models/Meta-Llama-3.1-8B-Instruct"
new_model = "./models/Meta-llama3-8b-SFT2"
dataset = load_dataset('json', data_files='./datasets/done.jsonl', split='train')

def preprocess_function(examples):
examples['text'] = [f"question: {q}, answer: {a}, url: {u}" for q, a, u in zip(examples['question'], examples['answer'], examples['url'])]
return {'text': examples['text']}

preprocessed_dataset = dataset.map(preprocess_function, batched=True)
preprocessed_dataset = preprocessed_dataset.remove_columns(['chunk_id', 'question', 'answer', 'url'])

bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True,
)

model = AutoModelForCausalLM.from_pretrained(base_model_id, quantization_config=bnb_config, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(base_model_id, add_eos_token=True, add_bos_token=True)
tokenizer.pad_token = tokenizer.eos_token

peft_config = LoraConfig(task_type="CAUSAL_LM", r=2, lora_alpha=16, lora_dropout=0.01)
model = prepare_model_for_kbit_training(model)
model = get_peft_model(model, peft_config)
model.print_trainable_parameters()

training_arguments = TrainingArguments(
output_dir="./results",
num_train_epochs=1,
per_device_train_batch_size=1,
gradient_accumulation_steps=16,
optim="paged_adamw_32bit",
save_steps=100,
logging_steps=5,
learning_rate=2e-4,
weight_decay=0.001,
bf16=True,
max_grad_norm=0.3,
max_steps=500,
warmup_ratio=0.03,
group_by_length=True,
gradient_checkpointing=True,
report_to="tensorboard",
)

trainer = SFTTrainer(
model=model,
train_dataset=preprocessed_dataset,
tokenizer=tokenizer,
args=training_arguments,
)

trainer.train()
trainer.model.save_pretrained(new_model)


However, this code throws an error:

---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
Cell In[2], line 68
60 trainer = SFTTrainer(
61 model=model,
62 train_dataset=dataset,
63 tokenizer=tokenizer,
64 args=training_arguments,
65 )
67 # Train the model
---> 68 trainer.train()
70 # Save the trained model
71 trainer.model.save_pretrained(new_model)

File ~/venvs/fine_tuning_llama/lib/python3.11/site-packages/transformers/trainer.py:2164, in Trainer.train(self, resume_from_checkpoint, trial, ignore_keys_for_eval, **kwargs)
2162 hf_hub_utils.enable_progress_bars()
2163 else:
-> 2164 return inner_training_loop(
2165 args=args,
2166 resume_from_checkpoint=resume_from_checkpoint,
2167 trial=trial,
2168 ignore_keys_for_eval=ignore_keys_for_eval,
2169 )

File ~/venvs/fine_tuning_llama/lib/python3.11/site-packages/transformers/trainer.py:2522, in Trainer._inner_training_loop(self, batch_size, args, resume_from_checkpoint, trial, ignore_keys_for_eval)
2516 context = (
2517 functools.partial(self.accelerator.no_sync, model=model)
2518 if i != len(batch_samples) - 1
2519 else contextlib.nullcontext
2520 )
2521 with context():
-> 2522 tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
2524 if (
2525 args.logging_nan_inf_filter
2526 and not is_torch_xla_available()
2527 and (torch.isnan(tr_loss_step) or torch.isinf(tr_loss_step))
2528 ):
2529 # if loss is nan or inf simply add the average of previous logged losses
2530 tr_loss = tr_loss + tr_loss / (1 + self.state.global_step - self._globalstep_last_logged)

File ~/venvs/fine_tuning_llama/lib/python3.11/site-packages/transformers/trainer.py:3688, in Trainer.training_step(***failed resolving arguments***)
3686 scaled_loss.backward()
3687 else:
-> 3688 self.accelerator.backward(loss, **kwargs)
3689 # Finally we need to normalize the loss for reporting
3690 if num_items_in_batch is None:

File ~/venvs/fine_tuning_llama/lib/python3.11/site-packages/accelerate/accelerator.py:2248, in Accelerator.backward(self, loss, **kwargs)
2246 self.lomo_backward(loss, learning_rate)
2247 else:
-> 2248 loss.backward(**kwargs)

File ~/venvs/fine_tuning_llama/lib/python3.11/site-packages/torch/_tensor.py:581, in Tensor.backward(self, gradient, retain_graph, create_graph, inputs)
571 if has_torch_function_unary(self):
572 return handle_torch_function(
573 Tensor.backward,
574 (self,),
(...)
579 inputs=inputs,
580 )
--> 581 torch.autograd.backward(
582 self, gradient, retain_graph, create_graph, inputs=inputs
583 )

File ~/venvs/fine_tuning_llama/lib/python3.11/site-packages/torch/autograd/__init__.py:347, in backward(tensors, grad_tensors, retain_graph, create_graph, grad_variables, inputs)
342 retain_graph = create_graph
344 # The reason we repeat the same comment below is that
345 # some Python versions print out the first line of a multi-line function
346 # calls in the traceback and some print out the last line
--> 347 _engine_run_backward(
348 tensors,
349 grad_tensors_,
350 retain_graph,
351 create_graph,
352 inputs,
353 allow_unreachable=True,
354 accumulate_grad=True,
355 )

File ~/venvs/fine_tuning_llama/lib/python3.11/site-packages/torch/autograd/graph.py:825, in _engine_run_backward(t_outputs, *args, **kwargs)
823 unregister_hooks = _register_logging_hooks_on_whole_graph(t_outputs)
824 try:
--> 825 return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
826 t_outputs, *args, **kwargs
827 ) # Calls into the C++ engine to run the backward pass
828 finally:
829 if attach_logging_hooks:

RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn


What should I do to resolve it?

Continue reading...