mstar.model.orpheus.components.language_model

mstar.model.orpheus.components.language_model#

Orpheus language model: Llama 3.2 3B-style transformer.

Built from the shared transformer components in mstar.model.components. Orpheus uses standard Llama-style RMSNorm (Llama mode, not Gemma), SwiGLU MLP, GQA self-attention with Llama-3 RoPE scaling, and no QK-norm.

QKV and gate/up projections are fused from construction via the parallel-linear classes (with a trivial single-rank comm group for the non-TP case). Weight loading goes through load_weights with stacked shard routing (no post-load consolidate step).

Classes

OrpheusForCausalLM(config[, comm_group])

OrpheusLanguageModel(config[, comm_group])

class mstar.model.orpheus.components.language_model.OrpheusForCausalLM(config, comm_group=None)[source]#

Bases: Module

Parameters:
forward(query_sequence, *, label, **kwargs)[source]#

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Parameters:
Return type:

Tensor

load_weights(weights)[source]#

Load HF Llama-style weights into the fused parameters.

class mstar.model.orpheus.components.language_model.OrpheusLanguageModel(config, comm_group=None)[source]#

Bases: Module

Parameters:
forward(query_sequence, *, label)[source]#

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Parameters:
Return type:

Tensor