feat(smollm3): add native SmolLM3 model family - #1158

Draft

ruiling-smartbear wants to merge 14 commits into

NVIDIA:mainfrom

ruiling-smartbear:feat/smollm3

Commits on Sep 4, 2026

feat(smollm3): scaffold SmolLM3 family with per-layer NoPE schedule
ruiling-smartbear
committed
3d861ac
fix(smollm3): emit the upstream chat template instead of generic ChatML
ruiling-smartbear
committed
350f7cb
fix(smollm3): match the upstream chat template exactly
ruiling-smartbear
committed
5355ea1
test(smollm3): keep the E2E root to the cases this family declares
ruiling-smartbear
committed
53bc69e
feat(smollm3): support the YaRN scaling the model card documents for 128k
ruiling-smartbear
committed
9f6b08c
refactor(smollm3): meet the repository naming and source-quality contracts
ruiling-smartbear
committed
051b283
fix(smollm3): register the checkpoint in the website model metadata
ruiling-smartbear
committed
0d9fcbd
test(smollm3): declare the repository-wide ready-model coverage
ruiling-smartbear
committed
c1501b8
test(smollm3): register the family in the fixed-capacity KV attention contract
ruiling-smartbear
committed
7cd54d1
fix(smollm3): resolve the NoPE schedule from the config the builders receive

ruiling-smartbear
and
Moviw
committed
56a172c
test(smollm3): cover the NoPE gate the builders apply, not just the schedule
ruiling-smartbear
committed
afd9a7f
fix(smollm3): stop gating the native KV path on pretraining_tp
ruiling-smartbear
committed
c6bdb50
fix(smollm3): keep build routing loadable without package context
ruiling-smartbear
committed
ac74495
fix(smollm3): import the schedule resolver at runtime in the dual-profile builder
ruiling-smartbear
committed
0d776fe