vllm.models.minimax_m3.amd

Modules:

• indexer_context_parallel –

  MiniMax-M3 context-parallel Triton indexer for ROCm.

• indexer_cp –

  Indexer-only context parallelism for MiniMax M3 (AMD MSA indexer).

• indexer_msa –

  MSA (gfx950/CDNA4) indexer impl for MiniMax M3.

• indexer_nvfp4 –

  MiniMax M3 lightning indexer over NVFP4 index-K pages on ROCm (gfx950).

• model –

  Inference-only MiniMax M3 (text backbone) model — AMD ROCm implementation.

• model_state –

  MiniMax-M3 model-runner state and specialized-decode cache lifecycle.

• mono –

  MiniMax-M3 fused per-layer decode of up to config.MAX_TOKENS tokens; entered

• mtp –

  MiniMax M3 MTP (multi-token prediction) draft model -- ROCm/AMD variant.

• ops –

  AMD/ROCm fused Triton ops for MiniMax-M3.

• sparse_attention_msa –

  MSA AITER block-sparse attend for MiniMax M3.

• sparse_attention_nvfp4 –

  MiniMax M3 block-sparse attention over an NVFP4 K/V cache on ROCm (gfx950).

• specialized_decode –