nemo_automodel.components.models.hy_mt2.state_dict_adapter#

State dict conversion between the on-disk tencent/Hy-MT2-30B-A3B HF checkpoint and Automodel’s native (grouped-experts) format.

The on-disk key layout is identical to tencent/Hy3-preview because both share model_type: "hy_v3" and architectures: ["HYV3ForCausalLM"]:

model.layers.{L}.mlp.expert_bias # [n_experts] model.layers.{L}.mlp.router.gate.weight # [n_experts, hidden] model.layers.{L}.mlp.experts.{E}.gate_proj.weight # [moe_inter, hidden] model.layers.{L}.mlp.experts.{E}.up_proj.weight # [moe_inter, hidden] model.layers.{L}.mlp.experts.{E}.down_proj.weight # [hidden, moe_inter] model.layers.{L}.mlp.shared_mlp.{gate,up,down}_proj.weight # shared expert

Automodel native:

model.layers.{L}.mlp.gate.e_score_correction_bias # [n_local] model.layers.{L}.mlp.gate.weight # [n_experts, hidden] model.layers.{L}.mlp.experts.gate_and_up_projs # grouped model.layers.{L}.mlp.experts.down_projs # grouped model.layers.{L}.mlp.shared_experts.{gate,up,down}_proj.weight

This adapter handles three on-disk-specific renames plus per-expert split/merge (via MoESplitExpertsStateDictMixin). It is functionally a clone of HYV3StateDictAdapter; kept separate so future Hy-MT2-only key changes (e.g. an MTP / aux-head extension that Hy-MT2 ships but Hy3-preview does not) can be added here without affecting Hy3-preview.

Module Contents#

Classes#

HyMT2StateDictAdapter

Bridges Automodel native (grouped experts) and on-disk Hy-MT2 HF format.

Data#

logger

_NATIVE_TO_HF_RENAMES

_HF_TO_NATIVE_RENAMES

API#

nemo_automodel.components.models.hy_mt2.state_dict_adapter.logger#

‘getLogger(…)’

nemo_automodel.components.models.hy_mt2.state_dict_adapter._NATIVE_TO_HF_RENAMES: tuple[tuple[re.Pattern[str], str], ...]#

((), (), ())

nemo_automodel.components.models.hy_mt2.state_dict_adapter._HF_TO_NATIVE_RENAMES: tuple[tuple[re.Pattern[str], str], ...]#

((), (), ())

class nemo_automodel.components.models.hy_mt2.state_dict_adapter.HyMT2StateDictAdapter(
config: Any,
moe_config: nemo_automodel.components.moe.config.MoEConfig,
backend: nemo_automodel.components.models.common.BackendConfig,
dtype: torch.dtype = torch.bfloat16,
)#

Bases: nemo_automodel.components.moe.state_dict_mixin.MoESplitExpertsStateDictMixin, nemo_automodel.components.checkpoint.state_dict_adapter.StateDictAdapter

Bridges Automodel native (grouped experts) and on-disk Hy-MT2 HF format.

Initialization

to_hf(
state_dict: dict[str, Any],
exclude_key_regex: Optional[str] = None,
**kwargs,
) → dict[str, Any]#

Native -> on-disk Hy-MT2 HF: per-expert split + name renames.

from_hf(
hf_state_dict: dict[str, Any],
device_mesh: Optional[torch.distributed.device_mesh.DeviceMesh] = None,
**kwargs,
) → dict[str, Any]#

On-disk Hy-MT2 HF -> native: filter MTP, rename, then merge experts.

convert_single_tensor_to_hf(
fqn: str,
tensor: Any,
**kwargs,
) → list[tuple[str, Any]]#

Per-tensor variant of to_hf for streaming-save code paths.

_is_mtp_key(key: str) → bool#

Return True if key belongs to an MTP layer (index >= num_hidden_layers).

Hy-MT2-30B-A3B does not appear to ship MTP layers in its public checkpoint, but the filter is kept as a defensive no-op so the adapter remains symmetric with HYV3StateDictAdapter.