Issues

• [Bug] maru_vllm: with vLLM prefix caching, tokens a turn generates are not stored when the next turn hits the prefix cache, so later lookups stop at the first prompt

  bug
  xcena-dev/maru#148 · Filter by author youngrok-XCENA opened on Nov 6, 2026

• [Feature] maru_vllm: with maru_async_store, a stored chunk is invisible to lookups until its write-behind copy lands

  enhancement
  xcena-dev/maru#143 · Filter by author seohui-XCENA opened on Nov 4, 2026

• [Bug] maru_vllm: a page-locked CPU pool still takes the per-layer load fallback, so returning-turn loads run at about 29 GB/s instead of 52 GB/s

  bug
  xcena-dev/maru#140 · Filter by author youngrok-XCENA opened on Nov 3, 2026

• [Bug] maru_vllm: Mamba-type states are not stored when vLLM writes them inside the forward

  bug
  xcena-dev/maru#139 · Filter by author kihwan-XCENA opened on Nov 3, 2026

• [Bug] maru_vllm: maru_auto_expand is ignored, so maru_pool_size cannot cap an engine's pool

  bug
  xcena-dev/maru#136 · Filter by author youngrok-XCENA opened on Oct 29, 2026

• [Feature] maru_vllm: a lookup asks for every chunk's Mamba state, once per TP rank

  enhancement
  xcena-dev/maru#133 · Filter by author kihwan-XCENA opened on Oct 29, 2026

• [Bug] maru_sglang: every L3 store fails, so SGLang never caches KV in Maru

  bug
  xcena-dev/maru#131 · Filter by author kihwan-XCENA opened on Oct 28, 2026

• [Tracking] maru_vllm: vLLM changes that affect the hybrid (HMA) KV connector

  xcena-dev/maru#127 · Filter by author seohui-XCENA opened on Oct 26, 2026

• [Feature] maru_vllm: key multimodal, LoRA, cache-salt and prompt-embedding requests so they can use Maru

  enhancement
  xcena-dev/maru#116 · Filter by author seohui-XCENA opened on Oct 22, 2026

• [Bug] resource manager: a reclaimed region's range is handed out while cxl Handlers still map it

  bug
  xcena-dev/maru#115 · Filter by author youngrok-XCENA opened on Oct 21, 2026

• [Bug] resource manager: closing one connection reaps a live process's allocations while it keeps another connection open

  bug
  xcena-dev/maru#114 · Filter by author hyunyul-XCENA opened on Oct 21, 2026

• [Feature] maru_vllm: support list for models and parallelism (TP, MLA, DSA, GLM-5.3-Flash)

  enhancement
  xcena-dev/maru#112 · Filter by author jooho-XCENA opened on Oct 20, 2026

• [Bug] maru_vllm: failed loads and deleted objects are mishandled

  bug
  xcena-dev/maru#110 · Filter by author seohui-XCENA opened on Oct 19, 2026
  

• [Feature] maru_vllm: honor kv_role to support prefill/decode disaggregation

  enhancement
  xcena-dev/maru#106 · Filter by author youngrok-XCENA opened on Oct 16, 2026

• [RFC] Extend the native vLLM connector with CPU/CXL L1 and SSD L2 caching

  enhancement
  xcena-dev/maru#87 · Filter by author jooho-XCENA opened on Sep 23, 2026
  1

• feat(maru_vllm): place the layer-overlap load with single_layer_kv_transfer instead of PyTorch indexing

  enhancement
  xcena-dev/maru#86 · Filter by author youngrok-XCENA opened on Sep 22, 2026

• Defense-in-depth: bounds-check server-supplied offset/length before device ops in handler plugins

  xcena-dev/maru#59 · Filter by author hyunyul-XCENA opened on Jul 6, 2026

• prepare pypi organization

  xcena-dev/maru#52 · Filter by author darjeeling opened on May 28, 2026
  1

• feat: event-driven CXL device hotplug monitoring

  xcena-dev/maru#49 · Filter by author seohui-XCENA opened on Apr 17, 2026

• [Feature] Implement eviction and capacity reclaim on CXL pool exhaustion

  high-priority
  Feature

  xcena-dev/maru#4 · Filter by author seohui-XCENA opened on Mar 6, 2026