ucm layerwise support MiniMax-m3

[snoopyri]

Purpose

Add layerwise KV-cache support for MiniMax-M3 on Ascend, whose dense and sparse layers use heterogeneous K/V cache layouts. Ensure transfer paths can safely handle padding entries introduced by the unified layout.

Modifications

Modifications

• Detect MiniMax-M3 sparse-attention models and select a dedicated AscendMiniMaxM3Layout when running on Ascend with layerwise transfer enabled.

• Add AscendMiniMaxM3Layout to normalize per-layer KV cache layouts into three logical components:

  • Dense layers: [K, V, null]
  • Sparse layers: [K first half, K second half, V]

• Validate KV tensor types, shapes, and strides before constructing the layout.

• Preserve per-component base addresses, buffer sizes, copy sizes, and physical block strides for correct layerwise addressing.

• Update generic stream and Ascend SDMA-direct transfer paths to skip null/zero-sized device buffers.

• Zero-fill host-side padding regions during device-to-host transfers, ensuring deterministic serialized cache blocks for dense-layer padding.

Test