feat(models): add native GPT-OSS implementation - #3424

Merged

S1ro1 merged 16 commits into

mainfrom

feat/gpt-oss-native

Sep 9, 2026

Commits on Aug 28, 2026

refactor(moe): unify runtime and checkpoint layout
S1ro1
authored
45eae7d
refactor(moe): make runtime backends explicit
S1ro1
authored
53b82d0
refactor(moe): simplify canonical execution path
S1ro1
authored
0f8a354
refactor(moe): make shared experts parent-owned
S1ro1
authored
4fd51be
fix(moe): address runtime synchronization review
S1ro1
authored
7b0e567
refactor(trainer): make activation checkpointing operator-based
S1ro1
authored
cf91f97
refactor(models): remove checkpoint-only attention helpers
S1ro1
authored
3c67257
refactor(trainer): clean up selective checkpoint policy
S1ro1
authored
384e781
fix(trainer): enforce varlen ring attention LSE layout
S1ro1
authored
c184918
feat(models): add native GPT-OSS implementation
S1ro1
authored
858113c
fix(inference): reload GPT-OSS weights through vLLM loaders
S1ro1
authored
9bfdb99
refactor(models): isolate GPT-OSS attention
S1ro1
authored
501d435

Commits on Sep 9, 2026