vllm.models.step5.amd

Modules:

• checks –

  Startup checks for the ROCm Step5 sparse attention layers.

• csa_attention –

  Step-5 CSA as one graph-splitting custom op: region summaries, scoring,

• csa_cache –

  Step-5 CSA region summaries as their own KV cache group, and the per-step

• model –

  Step5 on ROCm: compressed sparse attention (CSA) and an FP32 residual.

• mtp –

  Step5 MTP on ROCm.

• ops –
• vl –

  Inference-only Step5 vision-language model on ROCm.

• warmup –