feat(megatron-bridge): adopt TEGroupedMLP per-expert quantizers (#1550) for W4A16 NVFP4 four_over_six - #2072

Open

yueshen2016 wants to merge 18 commits into

mainfrom

qad/te-per-expert-1550

Commits on Jul 29, 2026

feat(quant): per-expert weight quantizer for TEGroupedMLP

jenchen13
and
claude
committed
87bab30
fix(quant): shard TE grouped per-expert amax with global expert identity

jenchen13
and
claude
committed
b6659bc
feat(export): per-expert HF export for TEGroupedMLP

jenchen13
and
claude
committed
d2b46a4
test(quant): TEGrouped per-expert quantization tests

jenchen13
and
claude
committed
e2742fc
refactor(quant): alias the any-quantizer isinstance tuple in model_calib

jenchen13
and
claude
committed
2883da0
refactor(quant): centralize quantizer type checks on AnyQuantizer

jenchen13
and
claude
committed
5b8f9d7
make tests more robust
jenchen13
committed
4708d85
docs(changelog): document TEGroupedMLP per-expert quantization + breaking change

jenchen13
and
claude
committed
8902e04
feat(quant): make TEGroupedMLP per-expert weight quantizers opt-in + review fixes

jenchen13
and
claude
committed
8bd4685
feat(quant): make TEGroupedMLP per-expert weight quantizers the default

jenchen13
and
claude
committed
39074f2
test(quant): cap real torch.compile TEGrouped test at 90s

jenchen13
and
claude
committed
761d1c1

Commits on Jul 30, 2026

fix(quant/export): restore output_layer static amax + robust TE per-expert export

jenchen13
and
claude
committed
57fe963

Commits on Jul 31, 2026

fix(quant): emit empty extra_state for unquantized modules

jenchen13
and
claude
committed
790a871
Merge remote-tracking branch 'origin/main' into jennifchen/te_per_expert
jenchen13
committed
2b0d5ea

Commits on Aug 3, 2026

feat(megatron-bridge): Nemotron-Nano-3 W4A16 NVFP4 four_over_six PTQ/QAD support
yueshen2016
committed
72d3cb5

Commits on Aug 4, 2026

remove deepcopy
jenchen13
committed
972a57a
Merge branch 'pr1550' into qad/te-per-expert-1550
yueshen2016
committed
c53ee30

Commits on Aug 5, 2026

fix(quantization): resolve default_quant_desc_weight in TEGroupedLinear setup
yueshen2016
committed
5f2f3de