vllm.models.muse_spark.nvidia

NVIDIA-specific MuseSpark model implementations.

Modules:

• mega_moe –

  MuseSpark1x routed experts on DeepGEMM BF16 and NVFP4 MegaMoE kernels.

• model –

  Muse Spark HF text decoder with gated FP32 residuals and compressed MoE.

• multimodal –

  MuseSpark1x Omni for NVIDIA GPUs: the shared Omni towers over the CUDA text stack.

• muse_moe_runner –

  MuseSpark1x MoE: routing, experts and the post-expert-norm finalize.

• ops –