Skip to content

Potential fix for vocab tiling dtype error #2610

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

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

Potential fix for vocab tiling dtype error #2610

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
8 changes: 3 additions & 5 deletions src/MaxText/vocabulary_tiling.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,15 +198,13 @@ def _bwd_scan_body(grad_params_acc, chunk_data):
_bwd_scan_body, initial_grad_params_acc, (reshaped_hidden_states, reshaped_labels, reshaped_segmentation)
)
grad_reshaped_hidden_states = _maybe_shard_with_name(grad_reshaped_hidden_states, reshaped_hidden_spec)
# TODO (chengnuojin): we may want to convert grad_params to bf16 to save memory
# grad_params = jax.tree_util.tree_map(lambda x, y: y.astype(x.dtype), gathered_params, grad_params)
# Chain-rule to accumulate gradients
grad_params = jax.tree_util.tree_map(lambda g: g * loss_cotangent, grad_params)
grad_params = jax.tree_util.tree_map(lambda g, p: (g * loss_cotangent).astype(p.dtype), grad_params, gathered_params)
# Give back sharding constraint
grad_reshaped_hidden_states = _reshape(grad_reshaped_hidden_states, (batch_size, seq_len, emb_dim), hidden_spec)
grad_hidden_states = _reshape(grad_reshaped_hidden_states, (batch_size, seq_len, emb_dim), hidden_spec)
return (
grad_params, # grad for params
grad_reshaped_hidden_states.astype(reshaped_hidden_states.dtype),
grad_hidden_states.astype(config.weight_dtype), # Enforce activation gradients dtype as
None, # grad for reshaped_labels
None, # grad for reshaped_segmentation
)
Expand Down
Loading