AI-Hypercomputer / AI-Hypercomputer/maxtext
DEFAULT_MASK_VALUE causes gradient explosion and nan loss on deep models
- Langage dominant
- Python
- Étoiles
- 2.4k
- Forks
- 607
- Merge moyen
- 2 j 19 h
- PR mergées (30 j)
- 158
Description
I was training a llama model on GPU, with a custom embedding. It worked fine with 12 layers, dim 1024, seq length 256, but loss would become nan after the first step if setting num_layers to more than 17. I debugged the gradients, and found after each layer their magnitude would increase by around 100x, until they hit float32_max at around the 18th layer and became inf, leading to nan loss.
The gradient explosion seemed to be coming from
`local_exps = jnp.exp(attn_weights - local_max)`
in attentions.py.
Changing
`DEFAULT_MASK_VALUE = -0.7 * float(jnp.finfo(jnp.dtype("float32")).max)`
to
`DEFAULT_MASK_VALUE = -jnp.inf`
fixed the issue, and the gradients' magnitude stopped increasing after each level.
Presumably the issue wasn't noticed during TPU training as that uses a separate codepath.
Guide de contribution
Ouvrir le guide de contribution
Évaluation
Cette issue n'a pas encore été évaluée.