performance-distributed-evaluati-runtimeerror-tensors-must-be-contiguous-occurs-dur-4fc7a947

THEN Add `.contiguous()` calls to tensor operations before gathering logits across GPUs. This ensures tensors are stored contiguously in memory, preventing the RuntimeError in `distributed_concat` or similar functions. Modify the model's forward method or the relevant utility to check and enforce contiguity, e.g., `if not logits.is_contiguous(): logits = logits.contiguous()`.

# In trainer_pt_utils.py or model forward:
def distributed_concat(tensor, num_total_examples):
    if not tensor.is_contiguous():
        tensor = tensor.contiguous()
    # ... rest of gathering logic