nemo_automodel.components.models.hy_mt2.dispatch#
Config-shape fingerprint that distinguishes Hy-MT2-30B-A3B from Hy3-preview.
Tencent ships both checkpoints with architectures: ["HYV3ForCausalLM"] and
model_type: "hy_v3" even though the two models differ substantially
(48 vs 80 layers, 128 vs 192 experts, hidden 2048 vs 4096, etc.). The
auto-resolver in _transformers/model_init.py looks up the fingerprint here
so all Hy-MT2-specific knowledge stays inside this module.
Module Contents#
Functions#
Return whether config describes Tencent’s Hy-MT2-30B-A3B checkpoint. |
API#
- nemo_automodel.components.models.hy_mt2.dispatch.is_hy_mt2_config(config: Any) bool#
Return whether config describes Tencent’s Hy-MT2-30B-A3B checkpoint.