[OMNIML-5899] Export IQ checkpoints from HF and Megatron - #2447

Merged

cjluo-nv merged 13 commits into

mainfrom

hungyuehc/omniml-5899-export-v2

Sep 19, 2026

Commits on Sep 17, 2026

Export IQ checkpoints from HF and Megatron
hychiang-git
committed
d789b45
Test IQ checkpoint export
hychiang-git
committed
9f9519b
Remove unused IQ search config from export tests
hychiang-git
committed
13cdcec
Fix IQ checkpoint export review findings
hychiang-git
committed
542b601
Gate unsupported fused-MoE IQ export
hychiang-git
committed
11cd58d
Merge branch 'main' into hungyuehc/omniml-5899-export-v2
hychiang-git
authored
e14a1e9

Commits on Sep 18, 2026

Merge branch 'main' into hungyuehc/omniml-5899-export-v2
cjluo-nv
committed
b8a966b
Close the four open IQ export review findings

cjluo-nv
and
claude
committed
106686d
Fix uses_iq_quantization on SequentialQuantizer and make its guard rank-uniform

cjluo-nv
and
claude
committed
070baa5
Reject fused-MoE IQ export where every rank agrees

cjluo-nv
and
claude
committed
a232c1a
Revert the architecture-keyed fused-MoE IQ rejection

cjluo-nv
and
claude
committed
0ba91aa
Require PP=1 for Megatron IQ export so the fused rejection is rank-uniform

cjluo-nv
and
claude
committed
5d0bbab
Add the PP>1 export test and record the TEGroupedLinear detection gap

cjluo-nv
and
claude
committed
68b7230