[OMNIML-5899] Add IQ post-training quantization recipes - #2449

Merged

cjluo-nv merged 17 commits into

mainfrom

hungyuehc/omniml-5899-recipes-v2

Sep 22, 2026

Commits on Sep 17, 2026

Add IQ post-training quantization recipes
hychiang-git
committed
f3fca65
Document IQ quantization recipes
hychiang-git
committed
b9dfae3
Remove unused IQ search recipe option
hychiang-git
committed
d658c5f

Commits on Sep 18, 2026

Merge branch 'main' into hungyuehc/omniml-5899-recipes-v2
hychiang-git
authored
57b7f88
[OMNIML-5899] Pin IQ recipe packing contracts
hychiang-git
committed
a2849c6

Commits on Sep 20, 2026

Merge branch 'main' into hungyuehc/omniml-5899-recipes-v2
hychiang-git
authored
dfec594

Commits on Sep 21, 2026

Update the IQ recipe docs now that export has landed

cjluo-nv
and
claude
committed
6b5d603
Add end-to-end hf_ptq coverage for IQ1_S and IQ2_XS

cjluo-nv
and
claude
committed
eaee227
Give the IQ end-to-end cases an explicit timeout

cjluo-nv
and
claude
committed
8020935
Reuse the packed IQ weight across forwards instead of re-encoding it

cjluo-nv
and
claude
committed
3390b98
Tie the packed IQ weight cache to the parameter, not the forward's view

cjluo-nv
and
claude
committed
f11eb63
Say what the IQ recipes actually quantize and who reads effective_bits

cjluo-nv
and
claude
committed
871d4ee

Commits on Sep 22, 2026

Size the IQ decode chunk separately from the encode chunk

cjluo-nv
and
claude
committed
e707df7
Merge remote-tracking branch 'origin/main' into local/omniml-5899-recipes-review
hychiang-git
committed
0ec5aa7
Merge remote-tracking branch 'origin/hungyuehc/omniml-5899-recipes-v2' into hungyuehc/omniml-5899-recipes-v2
cjluo-nv
committed
b820e47
Migrate the IQ recipes off the deprecated recipe_type

cjluo-nv
and
claude
committed
bcabea8
Let decode_chunk_size be tuned per quantizer like block_chunk_size

cjluo-nv
and
claude
committed
9a6ce16