Autoquant and GPTQ in support in Megatron-Core [OMNIML-3151] - #1562

Merged

jenchen13 merged 55 commits into

mainfrom

jennifchen/mcore_autoquant_gptq

Jul 2, 2026

Commits on May 22, 2026

MSE & mixed precision in mcore
jenchen13
committed
676eac4
lint
jenchen13
committed
9143e02

Commits on May 26, 2026

MSE calibration: universal MseCalibrator fallback for non-NVFP4 quantizers
jenchen13
committed
57e5f26
update mse test for mixed precision
jenchen13
committed
2dea94a
fix gpu test, no duplicate backend registration
jenchen13
committed
918ed6a

Commits on May 27, 2026

Merge branch 'main' into feature/mcore_mse_mixed_precision
jenchen13
authored
05a436f
fix unit tests
jenchen13
committed
ff20eca
lazy init autoquant register
jenchen13
committed
985da85
revert breaking change on MSECalibrator
jenchen13
committed
d88e54a
revert autoquant and gptq changes
jenchen13
committed
5f291a6
fallback to copy HF remote code if no dir
jenchen13
committed
5c9cd43

Commits on May 28, 2026

Merge branch 'main' into feature/mcore_mse_mixed_precision
jenchen13
committed
c5c7a2e
fix if else
jenchen13
committed
d63bf70
autoquant and gptq in mcore
jenchen13
committed
2ba29fd

Commits on May 29, 2026

fix logic again
jenchen13
committed
e14fa62
Merge branch 'main' into feature/mcore_mse_mixed_precision
jenchen13
authored
e985e93

Commits on Jun 2, 2026

Commits on Jun 15, 2026

Fix MCore layerwise discovery mutation
jenchen13
committed
967d5ef

Commits on Jun 16, 2026

Merge branch 'main' into jennifchen/mcore_autoquant_gptq
jenchen13
committed
5b141b5
Fix AutoQuant budget denominator for restored stats
jenchen13
committed
55ed710
Update changelog for MCore quantization support
jenchen13
committed
2acbc03

Commits on Jun 22, 2026

Clean up Megatron parallel state in layerwise test
jenchen13
committed
f40849e

Commits on Jun 24, 2026

Merge branch 'main' into jennifchen/mcore_autoquant_gptq
jenchen13
authored
a59eecd

Commits on Jun 25, 2026

fix: sum AutoQuantize cost/score across EP for MCore MoE experts

jenchen13
and
claude
committed
f68bab8
test: strengthen GPTQ mamba hybrid assertions and increase timeout

jenchen13
and
claude
committed
8f94cf8
test: move update_hessian zero-token noop test to CPU unit tests

jenchen13
and
claude
committed
fbaa4a3
separate megatron calib dataloader out into helper
jenchen13
committed
d4d91c8
increase timeout for onnx test
jenchen13
committed
bbf7f35
test: raise unit test timeout to 120 s on Windows
jenchen13
committed
fe80756
update comment
jenchen13
committed
6c2d446

Commits on Jun 26, 2026

move cpu only gptq tests to cpu unit tests
jenchen13
committed
29475fe
move autoquant registration to plugins/megatron.py; fix circular import introduced to model_calib.py
jenchen13
committed
850799b
remove padding hardcode
jenchen13
committed
470b48f
Merge branch 'main' into jennifchen/mcore_autoquant_gptq
jenchen13
authored
b9a1a6b
Increase timeout for Megatron sharded state dict test
jenchen13
committed
0608daf

Commits on Jun 29, 2026

Increase GPU Megatron CI timeout
jenchen13
committed
289dbdc
Adjust GPU Megatron CI timeout
jenchen13
committed
35bd94f
Speed up Mamba GPU quantization tests
jenchen13
committed
f0d1c55
test: tighten timeouts for mamba-hybrid quant tests

jenchen13
and
claude
committed
351c6c6
test: prebuild quant CUDA extensions in a session fixture

jenchen13
and
claude
committed
14cd2df
test: drop redundant quant CUDA extension prebuild fixture

jenchen13
and
claude
committed
9749a49

Commits on Jun 30, 2026

Fix gpu_megatron quant extension timeout
jenchen13
committed
20a888d

Commits on Jul 1, 2026

test: prebuild quant CUDA extensions via session fixture

jenchen13
and
claude
committed
d6b01ac
Merge branch 'main' into jennifchen/mcore_autoquant_gptq
jenchen13
authored
6d75cbc
Reduce compressed gpu_megatron state dict matrix
jenchen13
committed
51ba351
Address docs and gpu_megatron feedback
jenchen13
committed
f518c1a
Add Nemotron Nano AutoQuantize launcher example
jenchen13
committed
848c317
Restore quantization docs cross references
jenchen13
committed
c976a23
Update launcher Megatron submodule for AutoQuantize
jenchen13
committed
485750b
bring gpu_megatron tests back to 60min
jenchen13
committed
8c307de
respond to comments fix tests
jenchen13
committed
eccabc8
explain conftest change
jenchen13
committed
961abac
rename example
jenchen13
committed
2e5100c
refactor: drop dead AutoQuantize denominator fallback path

jenchen13
and
claude
committed
054e353