dela ea7167b3f7 Size vocab by max token id so duplicate-piece vocabs (Yi-6B) don't overflow embedding

2026-08-26 14:42:12 +08:00

..

_fla

Initial K3 snapshot: 0.5B KDA/MLA/MoE train path

2026-08-25 14:43:17 +08:00

layers

Keep LatentMoE routed bmm in activation dtype under bf16 autocast

2026-08-25 20:22:51 +08:00

models

Keep attnres on resume, fix final chunk_index, default 0.5b to Yi-6B

2026-08-26 14:23:29 +08:00

ops

Initial K3 snapshot: 0.5B KDA/MLA/MoE train path

2026-08-25 14:43:17 +08:00

training

Size vocab by max token id so duplicate-piece vocabs (Yi-6B) don't overflow embedding

2026-08-26 14:42:12 +08:00

__init__.py

Initial K3 snapshot: 0.5B KDA/MLA/MoE train path

2026-08-25 14:43:17 +08:00