TypeError: 'NoneType' object does not support item assignment when using Evaluate

🤗Transformers
1

I’m working with a multinode environment where I’ve got 3 nodes and each node has multiple GPUs

When trying to compute metrics and return the bleu score for my task I get this error

[rank10]:   File "/ext3/miniconda3/envs/venv/lib/python3.8/site-packages/transformers/trainer.py", line 3641, in evaluate
[rank10]:     output = eval_loop(
[rank10]:   File "/ext3/miniconda3/envs/venv/lib/python3.8/site-packages/transformers/trainer.py", line 3933, in evaluation_loop
[rank10]:     metrics[f"{metric_key_prefix}_loss"] = all_losses.mean().item()
[rank10]: TypeError: 'NoneType' object does not support item assignment

This is my compute_metrics function

def compute_metrics(eval_pred, tokenizer):
    rank = int(os.getenv('RANK', '0'))
    world_size = 12

    # Load the metric with the specified num_processes and process_id
    metric = evaluate.load(
        "sacrebleu",
        num_process=12,
        process_id=rank,
    )

    predictions, labels = eval_pred
    decoded_preds = tokenizer.batch_decode(predictions, skip_special_tokens=True)
    # Replace -100 in the labels as we can't decode them
    labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
    decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)

    # Print sample translations and references for debugging
    print("Sample translations: \n")
    print(decoded_preds[:10])
    print("References: ", decoded_labels[:10])
    references = [[label] for label in decoded_labels]
    # Compute the BLEU score
    return metric.compute(predictions=decoded_preds, references=references)

Can someone help point out what to fix here?