python
HuggingFace Training NaN Loss
FloatingPointError: Overflow.*loss became NaN
Fixes
- 1.Enable fp16=True with loss scaling in TrainingArguments
- 2.Reduce learning rate by 10x
- 3.Add max_grad_norm for gradient clipping
huggingfacetrainingnan
Related Errors
python3 fixes
Asyncio event loop already running
RuntimeError: This event loop is already running
- •Use nest_asyncio.apply() to allow nested event loops
- •Use asyncio.run_coroutine_threadsafe() instead of asyncio.run()
python3 fixes
Coroutine never awaited
RuntimeWarning: coroutine '.*' was never awaited
- •Add 'await' before the coroutine call
- •Use asyncio.create_task() to schedule the coroutine
python3 fixes
Asyncio task was cancelled
asyncio\.CancelledError
- •Handle CancelledError in try/except within the task
- •Use asyncio.shield() to protect critical sections from cancellation