vllm.models.step5.nvidia

Modules:

• config –
• csa_tail_ring –

  Per-request ring of the open CSA region's raw indexer rows.

• flashinfer_sparse –

  Step5 sparse selection and attention using FlashInfer and PyTorch.

• layernorm –

  Step5 normalization layers.

• mega_moe –

  Step5 NVFP4/FP8/BF16 DeepGEMM MegaMoE with shared symmetric storage.

• model –

  Inference-only Step5 model.

• mtp –
• ops –
• sparse_attention –

  Hopper and B300 sparse attention using FlashInfer and Triton.

• sparse_summary_cache –

  CSA region summaries, kept as a compressed KV cache layer.

• vl –

  Inference-only Step5 vision-language model.