Add support for dLLM encoder-decoder models (DiffusionGemma) [tied-weight PTQ export support ]#1707

Open

juhi10071998 wants to merge 12 commits into

NVIDIA:mainfrom

juhi10071998:northbloom

Commits on Jun 17, 2026

Onboard DiffusionGemma (northbloom) block-diffusion text model to hf_ptq.py

juhi10071998
and
claude
committed
moe export: alias bit-identical per-expert buffers between tied modules

juhi10071998
and
claude
committed
export: alias bit-identical buffers between tied dense Linears

juhi10071998
and
claude
committed
export: opt-in reorder of tied-weight aliases to canonical-side names

juhi10071998
and
claude
committed
export: max-merge tied input_quantizer amaxes; alias input_scale buffers

juhi10071998
and
claude
committed
quantization: exclude self_conditioning from default disabled_quantizers

juhi10071998
and
claude
committed
tests: cover tied-weight dedup, canonical reorder, and input-amax sync

juhi10071998
and
claude
committed
docs: add CHANGELOG entry for tied-weight PTQ support

juhi10071998
and
claude
committed
fix: address PR review feedback on tied-weight dedup

juhi10071998
and
claude
committed
recipes: add DiffusionGemma model-specific PTQ recipe

juhi10071998
and
claude
committed
fix: address PR review feedback round 2 on tied-weight dedup

juhi10071998
and
claude
committed
refactor: relocate tied-weight helpers per Edward's review

juhi10071998
and
claude
committed