Repository navigation
Public preset quant_cfg changed from dict to list in 0.44 (breaking, undocumented?) #2001
Description
Summary
Between nvidia-modelopt 0.43 and 0.44, the quant_cfg field in the public quantization preset configs (e.g. FP8_DEFAULT_CFG, INT8_DEFAULT_CFG, NVFP4_DEFAULT_CFG) changed from a dict to a list. The internal modelopt.torch.quantization.config._default_disabled_quantizer_cfg changed the same way.
Downstream code that composes a quant_cfg by dict-merging these presets/fragments (a natural pattern) now breaks with TypeError: 'list' object is not a mapping. I couldn't find this called out in the changelog / migration notes — is it intentional, and is there a recommended migration path?
Environment
nvidia-modelopt: 0.45.0 (also reproduces on 0.44.0; works on 0.43.0)torch: 2.14.0.dev20260704+cu130 (the objects involved are plain Python, so this is torch-independent)- python: 3.12
Minimal reproducer (modelopt only)
import modelopt.torch.quantization as mtq
# `quant_cfg` in the public preset configs is a dict on <=0.43 and a list on >=0.44:
print(type(mtq.FP8_DEFAULT_CFG["quant_cfg"])) # 0.43: <class 'dict'> | 0.44+: <class 'list'>
print(type(mtq.config._default_disabled_quantizer_cfg)) # 0.43: <class 'dict'> | 0.44+: <class 'list'>
# Code that composes a quant_cfg by dict-merging these now raises:
base = {"*weight_quantizer": {"fake_quant": False}}
merged = {**base, **mtq.config._default_disabled_quantizer_cfg}
# TypeError: 'list' object is not a mappingNote
mtq.quantize(model, quant_cfg=<dict>, ...) still accepts the old dict-style quant_cfg on 0.45, so the documented quantize path itself is intact — the break is specifically for code that reads/merges the preset quant_cfg objects (or the _default_disabled_quantizer_cfg fragment) as mappings.
Questions
- Is the
dict→listchange to presetquant_cfgintentional and stable going forward? - Is there a supported/public way to obtain the "default disabled quantizers" fragment (i.e. a replacement for
config._default_disabled_quantizer_cfg)? - Any migration guidance for merging user overrides into a preset
quant_cfgunder the new list format?
Context: this surfaced in 🤗 diffusers, whose ModelOpt integration composes quant_cfg from these fragments.
Activity
jenchen13 commentedon Jul 21, 2026
Hi @sayakpaul , the change of
quant_cfgfrom from dict to list was in the 0.44 CHANGELOT.rstWe also have documentation on the new list-based
quant_cfgin the docs. The change was motivated to improve granularity and ordering in quant configs -- for example now the last config entry wins: "If two entries both match *weight_quantizer and both carry a cfg, the second entry does not inherit the first entry’s settings — it replaces them entirely."shengliangxu commentedon Jul 21, 2026
Thanks @sayakpaul for reporting. Thanks @jenchen13 for following up.
The change was more than an improvement, it was to fix the semantic. The old dict format is ambiguous when multiple config wildcard items config the same parameter because dict keys are unordered. The list defines a clear ordering and make the semantic deterministic.
However we still accept dict configs for backward compatibility as much as possible by normalizing to list. But old style dict manipulation code will break because of the change of type.
sayakpaul commentedon Jul 22, 2026
Would it be possible to open a PR to Diffusers to fix this?
shengliangxu commentedon Jul 23, 2026
Thanks @sayakpaul , absolutely, I'll submit a PR to fix it soon.
ChenhanYu commentedon Aug 2, 2026
Software Release Triage
release: ModelOpt v0.46.0
fingerprint: f8f01121d056bb252d7ea9fae13de5f778f4259c9a7714c6919e3160b84eb65e
This open issue is in the ModelOpt release sweep. Owner: confirm release impact, linked fix/validation, or that it is non-blocking for this release.
ChenhanYu commentedon Aug 2, 2026
Software Release Triage
release: ModelOpt v0.46.0
fingerprint: 50577401eef36c5af61dda029f69a821312a24c2912316f9dc5596a398115b44
This open issue is in the ModelOpt release sweep. Owner: confirm release impact, linked fix/validation, or that it is non-blocking for this release.
ChenhanYu commentedon Aug 2, 2026
Software Release Triage
release: ModelOpt v0.46.0
fingerprint: 4642a6dbc968c3329a33f75907f59ae879a3faa220dc830429a3a927c67ad94c
Release follow-up: this open ModelOpt issue needs release relevance confirmed. Link its planned fix/validation, or confirm it is not a v0.46.0 blocker.
Fix ModelOpt integration for the >=0.44 list quant_cfg format
kvnloo commentedon Sep 28, 2026
The main downstream problem seems less about accepting lists and more about the lack of a stable public composition mechanism.
Code should not need to import:
_default_disabled_quantizer_cfgor assume that a preset's internal
quant_cfgis a mapping.Could ModelOpt expose a supported helper for something like:
preset + user overrideswhere the helper normalizes either the old dict representation or current ordered-list representation into the canonical form?
The migration documentation could then show:
That would let Diffusers and other downstreams stop depending on private container shape entirely.