Skip to content

Eliminate out-projection remat. #4012

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

Merged
copybara-service merged 1 commit into from
May 30, 2026
+5 −5
Merged

Eliminate out-projection remat. #4012

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
10 changes: 5 additions & 5 deletions src/maxtext/models/deepseek_batchsplit.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1167,7 +1167,7 @@ def batch_split_schedule_bwd(
):
"""Performs the backward pass for a single layer."""
norm_mla_ws, moe_ws = weights
mla_out, mla_bwds = mla_with_norms_remat(
_, mla_bwds = mla_with_norms_remat(
residuals,
norm_mla_ws,
positions,
Expand All @@ -1187,7 +1187,6 @@ def batch_split_schedule_bwd(
dtype=cfg.dtype,
activation_pspec=activation_pspec,
)
residuals["mla_out"] = mla_out
attn_out_grad, moe_ws_grad = moe_bwd(
residuals,
outputs_grad,
Expand Down Expand Up @@ -1286,7 +1285,8 @@ def fn(args):
# Prevent fusion with MoE ops, especially the RMS norm.
# Unfortunately, this seems to be needed to avoid slight numerical differences
# between the fwd pass and remat.
return jax.lax.optimization_barrier(mla_out + x), mla_res
out = jax.lax.optimization_barrier(mla_out + x)
return out, {"mla_out": out, **mla_res}

return staggered_call(fn, list(zip(inputs, yarn_freqs)))

Expand Down Expand Up @@ -1349,11 +1349,11 @@ def remat_fn(args):
mesh=mesh,
activation_pspec=activation_pspec,
)
out = x + mla_out
# Prevent fusion with MoE ops, especially the RMS norm.
# Unfortunately, this seems to be needed to avoid slight numerical differences
# between the fwd pass and remat.
return jax.lax.optimization_barrier(out), (pre_attn_rms_norm_bwd, mla_bwds)
out = jax.lax.optimization_barrier(x + mla_out)
return out, (pre_attn_rms_norm_bwd, mla_bwds)

bwds = [None] * len(xs)
for i, x in enumerate(zip(xs, yarn_freqs, residuals.pop("attn_out"), residuals.pop("lse"))):
Expand Down
Loading