mstar.model.higgs_audio.components.llm

mstar.model.higgs_audio.components.llm#

Higgs-Audio text LLM: dense Qwen3 (1.7B-class) transformer.

Built from the shared transformer components. Qwen3 = Llama-style decoder with GQA, per-head QK-norm, SwiGLU MLP, and plain RoPE (theta 1e6, no llama-3 scaling).

Checkpoint layout is flat (embed_tokens.*, layers.N.*, norm.*) with the text head at audio_decoder_proj.text_lm_head; the weight iterator in higgs_audio_model.py remaps that to lm_head. QKV and gate/up projections are fused from construction via the parallel-linear classes; loading routes the separate HF shards via LLAMA_STACKED_PARAMS.

Classes

HiggsAudioLLM(config[, comm_group])

Qwen3 causal LM; parameter paths mirror the (flat) HF checkpoint.
class mstar.model.higgs_audio.components.llm.HiggsAudioLLM(config, comm_group=None)[source]#

Bases: Module

Qwen3 causal LM; parameter paths mirror the (flat) HF checkpoint.

Parameters:
forward(input_embeds, *, 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