AI-Hypercomputer / AI-Hypercomputer/maxtext

DEFAULT_MASK_VALUE causes gradient explosion and nan loss on deep models

Aberta
#614 2 comentários 2 reações 1 responsável Reivindicada por @anfals Ver no GitHub
bug
Linguagem predominante
Python
Estrelas
2.4k
Forks
607
Merge médio
2d 19h
PRs com merge (30d)
158

Descrição

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.

Guia de contribuição

Abrir o guia de contribuição

Avaliação

Esta issue ainda não foi avaliada.

Receba novas issues na sua caixa de entrada

Um resumo curto de issues do GitHub para quem está começando.