MTP support in per-layer fused export - #2259

Draft

Fridah-nv wants to merge 26 commits into

fridah/layerwise-fused-exportfrom

fridah/layerwise-mtp-support

Commits on Aug 19, 2026

fix(hf_ptq): detect layerwise recipes whose algorithm parses as a dict

Fridah-nv
and
claude
committed
e1a06b3
feat(export): export each decoder layer as layerwise calibration finishes it

Fridah-nv
and
claude
committed
b7e1622
feat(export): support accelerate offload in per-layer export

Fridah-nv
and
claude
committed
053b1df
fix(export): close review findings in per-layer export

Fridah-nv
and
claude
committed
f33a399
refactor(export): make ExportContext the single state owner in layerwise export

Fridah-nv
and
claude
committed
88ae24e
docs(export): compact the comments in the layerwise export path

Fridah-nv
and
claude
committed
d3b97dc
fix(hf_ptq): keep the layerwise resume manifest alongside its shards

Fridah-nv
and
claude
committed
21ecccd
fix(export): adopt the name-based tied-weight map in per-layer export

Fridah-nv
and
claude
committed
e887c5a
fix(export): refuse to overwrite layer shards when the resume record is gone

Fridah-nv
and
claude
committed
8c1673f

Commits on Aug 20, 2026

Commits on Aug 21, 2026

Commits on Aug 26, 2026

feat(export): support MTP models in per-layer export [prototype]
Fridah-nv
committed
53b3865