[feat] Support UCM hybrid KV cache connector for mixed Mamba/Attention block layouts

[wangwenxin0312]

Purpose

Add UCM support for vLLM hybrid/HMA KV cache layouts, including group-aware block hashing, external cache lookup, and load/dump metadata generation for mixed full-attention and sliding-window cache groups.

Modifications

• Add per-group request metadata for hybrid KV cache blocks.
• Introduce KVCacheGroupManager to parse kv_cache_groups, derive per-group hash chains, and validate group block/window alignment.
• Implement two-stage external cache lookup:
  • full-attention prefix lookup via lookup_on_prefix
  • sliding-window tail verification via lookup
• Add HMAKVCacheLayout to build cache memory layout from kv_cache_tensors and per-layer KV cache specs.
• Add UCMHMAConnector for scheduler-side hit detection and worker-side KV cache registration.
• Generate HMA dispatch metadata by flattening per-group load/dump blocks into the existing UCM store API format.
• Route hybrid KV cache manager configurations through the HMA connector.

Test

export ENABLE_UCM_PATCH=1
export MODEL_PATH="/home/models/Qwen3-Next-80B-A3B-Instruct/"
python unified-cache-management/examples/offline_inference.py

[ygwpz]

💡 Note: TODO/FIXME marker found. Consider addressing before merge or creating a tracked issue.