Skip to content

MXFP8 training bug fixes for quantized_model_init and Torch FSDP fp8 all gather #587

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
sudhu2k merged 4 commits into from
Jun 9, 2026
+11 −3
Merged

MXFP8 training bug fixes for quantized_model_init and Torch FSDP fp8 all gather #587

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
0deec6e
Ensure keep_fp8_weight_transpose_cache is True even for quantized_mod…
sudhu2k May 15, 2026
d9d7a69
Refactor padding logic in MXFP8Tensor to remove unnecessary condition…
sudhu2k May 19, 2026
bab4468
Merge branch 'dev' into sudhu/mxfp8_bug_fixes
sudhu2k Jun 8, 2026
b7b3f8d
Merge conflict cleanup
sudhu2k Jun 8, 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
9 changes: 7 additions & 2 deletions transformer_engine/pytorch/module/base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1088,6 +1088,13 @@ def init_fp8_metadata(self, num_gemms: int = 1) -> None:
meta["num_gemms"] = num_gemms
self.init_fp8_meta_tensors(meta["recipe"])

# Force the transpose cache to be kept whenever the recipe is MXFP8 / MXFP4,
# regardless of whether we are currently inside an fp8_autocast region or not.
# reset_parameters() would disable columnwise_usage for params constructed inside
# `fp8_model_init` / `quantized_model_init`, leaving `_columnwise_data=None`).
if self.fp8_meta["recipe"].mxfp8() or self.fp8_meta["recipe"].mxfp4():
self.keep_fp8_weight_transpose_cache = True

if fp8_enabled:
# Set FP8 and other FP8 metadata
meta["num_gemms"] = num_gemms
Expand All @@ -1103,8 +1110,6 @@ def init_fp8_metadata(self, num_gemms: int = 1) -> None:
self.fast_setattr("fp8_initialized", True)

meta["recipe"] = FP8GlobalStateManager.get_fp8_recipe()
if meta["recipe"].mxfp8() or meta["recipe"].mxfp4():
self.keep_fp8_weight_transpose_cache = True

_current_recipe = meta["recipe"]
if _original_recipe is not None and not (
Expand Down
5 changes: 4 additions & 1 deletion transformer_engine/pytorch/tensor/mxfp8_tensor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -514,7 +514,10 @@ def __torch_dispatch__(cls, func, types, args, kwargs=None):
scale_invs = [tensor._rowwise_scale_inv, tensor._columnwise_scale_inv]
split_sizes_for_scale = [split_size, split_size // MXFP8_BLOCK_SCALING_SIZE]
# Padding requirements: rowwise dim0 should be divisble by 128, columnwise dim0 should be divisble by 4
padding_multiples = [128, 4]
# NOTE: ROCm/HIP backend uses an unpadded scale-inv layout (see `MXFP8Quantizer.make_empty`),
# so applying the padding here would produce a per-shard scale-inv whose dim-0
# does not match the destination scale-inv allocated for the FSDP2 local shard.
padding_multiples = [128, 4] if not IS_HIP_EXTENSION else [1, 1]
Comment thread
alextmagro marked this conversation as resolved.
for scale_inv, scale_split_size, pad_multiple in zip(
scale_invs, split_sizes_for_scale, padding_multiples
):
Expand Down
Loading