Skip to content

fix(quant): dequantize MXFP4 on GPU under the distributed conversion backend #5523

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
mikan-atomoki wants to merge 2 commits into
base: main
Choose a base branch
from
+107 −18
Open

fix(quant): dequantize MXFP4 on GPU under the distributed conversion backend #5523

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
b2d2908
fix(quant): dequantize MXFP4 on GPU under the distributed conversion …
mikan-atomoki Aug 12, 2026
a42b5be
fix(quant): make MXFP4 dequant device-explicit and memory-bounded
mikan-atomoki Aug 14, 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
13 changes: 13 additions & 0 deletions src/megatron/bridge/models/conversion/quantization_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -325,6 +325,19 @@ def dequantize_mxfp4_e2m1_packed(
``scale`` is expected to be one scale per row and per K tile. ``uint8``
E8M0 tensors use exponent bias 127 and are decoded to powers of two.
"""
# HF state dicts loaded via AutoBridge.from_hf_pretrained are CPU-resident,
# so without this every op below (bit unpack, LUT gather, repeat_interleave,
# multiply) runs on CPU even under scripts/conversion's GPU backend, leaving
# GPUs idle while a single CPU process becomes the bottleneck for the whole
# distributed conversion (observed ~20-30x slower than expected on a
# multi-GPU MoE conversion). Guarded on is_initialized() so the
# single-process CPU backend, which never calls init_process_group, is
# unaffected.
if torch.cuda.is_available() and torch.distributed.is_initialized():
device = torch.device("cuda", torch.cuda.current_device())
weight_packed = weight_packed.to(device)
scale = scale.to(device)

w_u8 = weight_packed.view(torch.uint8)
lo = (w_u8 & 0xF).to(torch.int64)
hi = (w_u8 >> 4).to(torch.int64)
Expand Down
Loading