Skip to content

[JAX] Fix grouped quant checkpointing #2889

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
jberchtold-nvidia wants to merge 4 commits into
base: main
Choose a base branch
from
+15 −0
Open

[JAX] Fix grouped quant checkpointing #2889

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
3063417
Fix grouped quant checkpointing
jberchtold-nvidia Apr 14, 2026
1d2328a
Cleanup
jberchtold-nvidia Apr 16, 2026
0de178e
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Apr 16, 2026
998fe1b
Merge branch 'main' into jberchtold/grouped-quant-checkpointing
jberchtold-nvidia Apr 16, 2026
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
15 changes: 15 additions & 0 deletions transformer_engine/jax/dense.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -429,10 +429,25 @@ def _grouped_dense_fwd_rule(
# rowwise_casted_x.original_shape == (M, K)
# colwise_casted_kernel.original_shape == (G, N, K)
grouped_gemm_x = casted_x.get_tensor(usage=TensorUsage.LHS)
# Checkpoint the rowwise inputs so that te_grouped_quantize_ffi can be DCE'd in the
# backward-scan remat block. Without this, JAX would re-run the quantize kernel to
# obtain grouped_gemm_x / grouped_gemm_kernel for the forward-GEMM recomputation even
# though the colwise residuals (ctx_x / ctx_kernel) are already saved. With both
# orientations checkpointed, all outputs of the custom-call become dead in the remat trace.
grouped_gemm_x = (
grouped_gemm_x.checkpoint(quantizer_set.x)
if isinstance(grouped_gemm_x, ScaledTensor)
else grouped_gemm_x
)
ctx_x = casted_x.get_tensor(usage=TensorUsage.LHS_TRANS)
ctx_kernel = casted_kernel.get_tensor(usage=TensorUsage.RHS_TRANS)

grouped_gemm_kernel = casted_kernel.get_tensor(usage=TensorUsage.RHS)
grouped_gemm_kernel = (
grouped_gemm_kernel.checkpoint(quantizer_set.kernel)
if isinstance(grouped_gemm_kernel, ScaledTensor)
else grouped_gemm_kernel
)
output = tex.grouped_gemm(
grouped_gemm_x,
grouped_gemm_kernel,
Expand Down
Loading