Add AutoQuantize recipe support [OMNIML-5116] - #1856

Merged

juhi10071998 merged 17 commits into

mainfrom

juhim/autoquant-recipe-v2

Jul 5, 2026

Commits on Jul 2, 2026

autoquant: add effective_bits override to the cost model

juhi10071998
and
claude
committed
5512ac9
recipe: add AutoQuantize recipe schema and loader support

juhi10071998
and
claude
committed
77cc7b1
examples/llm_ptq: add recipe-driven auto_quantize path, example recipes, and equivalence tests

juhi10071998
and
claude
committed
698569b
tests: pin cost_weight x effective_bits composition in get_cost

juhi10071998
and
claude
committed
75901a5
recipe: set NVFP4 effective_bits=4.5 library default (Phase D, nvfp4 only)

juhi10071998
and
claude
committed
ddd2f27
recipe: add Qwen3.6 model-specific autoquant recipe with arch disabled-layers

juhi10071998
and
claude
committed
25ad6dd
recipe: add general AutoQuantize example recipes; drop redundant heuristic variant

juhi10071998
and
claude
committed
afdfb03
recipe: make general AutoQuantize recipes self-contained with base disabled_layers

juhi10071998
and
claude
committed
f3dfdfd
examples/llm_ptq: deprecate AutoQuantize CLI flags; recipe-only (Phase G)

juhi10071998
and
claude
committed
4cd4547
AutoQuantize recipe schema: hoist cost_excluded_layers, share disabled_layers via $import

juhi10071998
and
claude
committed
040a8b5
examples/llm_ptq: rename auto_quantize_recipe->auto_quantize + AutoQuantize recipe docs

juhi10071998
and
claude
committed
b2c8f43
examples/hf_ptq: address review feedback (CodeRabbit)

juhi10071998
and
claude
committed
5a4e405
examples/hf_ptq: address review round 2 (Edwardf0t1 + Asma)

juhi10071998
and
claude
committed
2afa594
examples/hf_ptq: harden export guard against effective_bits bypass (CodeRabbit)

juhi10071998
and
claude
committed
2cf69fa
examples/hf_ptq: re-add AutoQuantize CLI as a deprecated on-the-fly-recipe shim

juhi10071998
and
claude
committed
d108b1c
auto_quantize base: disable *shared_expert_gate* (Qwen3.6 MoE export fusion)

juhi10071998
and
claude
committed
6980725
auto_quantize: allow a single candidate_format (one-format + bf16 search)

juhi10071998
and
claude
committed
c53ae9f

Navigation Menu

Uh oh!

Add AutoQuantize recipe support [OMNIML-5116] - #1856

Add AutoQuantize recipe support [OMNIML-5116]#1856
juhi10071998 merged 17 commits into
mainfrom
juhim/autoquant-recipe-v2

Commits on Jul 2, 2026

autoquant: add effective_bits override to the cost model

recipe: add AutoQuantize recipe schema and loader support

examples/llm_ptq: add recipe-driven auto_quantize path, example recipes, and equivalence tests

tests: pin cost_weight x effective_bits composition in get_cost

recipe: set NVFP4 effective_bits=4.5 library default (Phase D, nvfp4 only)

recipe: add Qwen3.6 model-specific autoquant recipe with arch disabled-layers

recipe: add general AutoQuantize example recipes; drop redundant heuristic variant

recipe: make general AutoQuantize recipes self-contained with base disabled_layers

examples/llm_ptq: deprecate AutoQuantize CLI flags; recipe-only (Phase G)

AutoQuantize recipe schema: hoist cost_excluded_layers, share disabled_layers via $import

examples/llm_ptq: rename auto_quantize_recipe->auto_quantize + AutoQuantize recipe docs

examples/hf_ptq: address review feedback (CodeRabbit)

examples/hf_ptq: address review round 2 (Edwardf0t1 + Asma)

examples/hf_ptq: harden export guard against effective_bits bypass (CodeRabbit)

examples/hf_ptq: re-add AutoQuantize CLI as a deprecated on-the-fly-recipe shim

auto_quantize base: disable shared_expert_gate (Qwen3.6 MoE export fusion)

auto_quantize: allow a single candidate_format (one-format + bf16 search)