[P1] AttributeError: 'CausalLMOutputWithPast' object has no attribute 'mean'

[krishnardt]

Hi,

I am training finetuning "meta-llama/Llama-2-7b-hf" with a dataset.
Without eval strategy it is working fine.
But with eval strategy I am facing the issue mentioned in the title.
Could anyone help me in fixing this issue?

This is the code I am using as reference:
https://github.com/stanfordnlp/pyreft/blob/main/examples/alpaca/train.py

Below is the error trace:

Below are the Training arguments I am passing:

training_args = TrainArguments( # output_dir=f"./checkpoints/rank{rank}", output_dir="./trained_model_wt_eval", learning_rate=3e-5, num_train_epochs=2, # evaluation_strategy="epoch", eval_strategy="steps", # do_eval=False, lr_scheduler_type="linear", # warmup_steps=warmup_steps, per_device_train_batch_size=2, per_device_eval_batch_size=2, # gradient_accumulation_steps=grad_acc, weight_decay=0.01, logging_dir=f"./logs/reft_rank{rank}", logging_strategy="steps", logging_steps = 2, save_strategy="steps", save_steps=500, save_total_limit=3, load_best_model_at_end=False, # fp16=True, bf16=True, remove_unused_columns=False )

Data arguments for training:

`def make_supervised_data_module(tokenizer: transformers.PreTrainedTokenizer, model, layers, training_args, data_args) -> Dict:
train_dataset = ReftSupervisedDataset(
"alpaca", data_args.data_path, tokenizer, data_split="train", seed=training_args.seed,
max_n_example=training_args.max_n_train_example,
input_field="input", instruction_field="instruction", output_field="output",
**{"num_interventions": len(layers), "position": training_args.position,
"share_weights": training_args.share_weights}
)

eval_dataset = ReftSupervisedDataset(
    "alpaca", data_args.eval_path, tokenizer, data_split="test", seed=training_args.seed,
    max_n_example=training_args.max_n_train_example,
    input_field="input", instruction_field="instruction", output_field="output",
    **{"num_interventions": len(layers), "position": training_args.position, 
       "share_weights": training_args.share_weights}
)
print(train_dataset)
print(eval_dataset)
data_collator_fn = transformers.DataCollatorForSeq2Seq(
    tokenizer=tokenizer,
    model=model,
    label_pad_token_id=-100,
    padding="longest"
)
data_collator = ReftDataCollator(data_collator=data_collator_fn)
return dict(train_dataset=train_dataset, eval_dataset=eval_dataset, data_collator=data_collator)

data_module = make_supervised_data_module(
tokenizer=tokenizer, model=model, layers=layers,
training_args=training_args, data_args=data_args)

`

[krishnardt]

@frankaging
or anyone, help me resolving this issue?

[BEW111]

I was able to resolve this locally (for ReftTrainerForCausalLM) by changing the output of the compute_loss function:

from pyreft import ReftTrainerForCausalLM


def patched_compute_loss(self, intervenable, inputs, return_outputs=False, **kwargs):
    # run intervened forward pass
    unit_locations = None
    if "intervention_locations" in inputs:
        if inputs["intervention_locations"].dim() == 3:
            unit_locations={"sources->base": (
                None,
                inputs["intervention_locations"].permute(1, 0, 2).tolist()
            )}
        else:
            # this is dummy for lora only baseline
            unit_locations={"sources->base": (None, 0)}
    base_outputs, cf_outputs = intervenable(
        {
            "input_ids": inputs["input_ids"],
            "attention_mask": inputs["attention_mask"]
        },
        unit_locations=unit_locations,
        labels=inputs["labels"],
        subspaces=inputs["subspaces"].permute(1, 0, 2).tolist() if "subspaces" in inputs else None
    )
    # return
    output = cf_outputs
    if cf_outputs is None:
        output = base_outputs # in case of lora only training
    
    loss = output.loss
    return (loss, output) if return_outputs else output.loss

  
ReftTrainerForCausalLM.compute_loss = patched_compute_loss

Specifically, returning (loss, output) rather than (output, output). I'm not sure why the original returns (output, output) but maybe this can help for now?