Skip to content

Fix incorrect stats for bf16 params in parameter_overview #380

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
uvikas wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix incorrect stats for bf16 params in parameter_overview #380

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions clu/parameter_overview.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,12 +52,21 @@ class _ParamRowWithStatsAndSharding(_ParamRowWithStats):
sharding: tuple[int | None, ...] | str


def _upcast(x):
"""Upcast low-precision floats to float32 for numerically stable stats."""
if hasattr(x, "dtype") and jnp.issubdtype(x.dtype, jnp.floating) and x.dtype != jnp.float32:
return x.astype(jnp.float32)
return x


@jax.jit
def _mean_std_jit(x):
x = jax.tree_util.tree_map(_upcast, x)
return jax.tree_util.tree_map(jnp.mean, x), jax.tree_util.tree_map(jnp.std, x)


def _mean_std(x):
x = jax.tree_util.tree_map(_upcast, x)
mean = jax.tree_util.tree_map(lambda x: x.mean(), x)
std = jax.tree_util.tree_map(lambda x: x.std(), x)
return mean, std
Expand Down