Declarative MTP-Llama head stack + Gemma4 weight aggregation - #528

Merged

jlamypoirier merged 7 commits into

mainfrom

jlp_declarative_aggregator_converters

May 29, 2026

Commits on May 29, 2026

Make Gemma4 base-model weight aggregation declarative

jlamypoirier
and
claude
committed
ec89668
Make MTP-Llama prediction-head stack declarative via RepeatWeightConverter

jlamypoirier
and
claude
committed
eee018d
Gate optional Gemma4 transformers import in HF round-trip test

jlamypoirier
and
claude
committed
74c5db2
Derive HF metadata allowlist from PretrainedConfig for version robustness

jlamypoirier
and
claude
committed
f663d7a
Skip convert-group tests when transformers lacks the format's HF config class

jlamypoirier
and
claude
committed
07d4f46
Trim static HF allowlist entries now covered by the dynamic PretrainedConfig union

jlamypoirier
and
claude
committed
d253886
Drop consumer reference from RepeatWeightConverter docstring

jlamypoirier
and
claude
committed
4b1f769