Issues

• executor: dormant precision-policy block in dispatch_local references undefined expert_count / None expert_list

  EfficientMoE/MoE-Infinity#254 · Filter by author drunkcoding opened on Sep 30, 2026

• route-ahead: speculative_prefetch may issue twice per dispatch (overlap merge artifact)

  EfficientMoE/MoE-Infinity#253 · Filter by author drunkcoding opened on Sep 30, 2026

• fused_moe_ffn_into GEMM0 'Error Internal' on SM120 when built against torch 2.14 (works on 2.12) — pin torch 2.12 for now

  EfficientMoE/MoE-Infinity#245 · Filter by author drunkcoding opened on Sep 29, 2026

• [BUG] test_dispatcher_destruction_drains_inflight_route_callback segfaults on GPU hosts (was hidden by conftest stubbing)

  EfficientMoE/MoE-Infinity#244 · Filter by author drunkcoding opened on Sep 29, 2026

• [Feature] Runtime expert quantization + runtime expert dropping for offloaded MoE inference (phased umbrella)

  enhancement
  EfficientMoE/MoE-Infinity#234 · Filter by author drunkcoding opened on Sep 23, 2026
  8

• [Release] PR5 wheel publication blockers (v0.2.1 release train)

  enhancement
  EfficientMoE/MoE-Infinity#227 · Filter by author drunkcoding opened on Sep 18, 2026
  1

• [Tracking] moe-store split: param-server/ckpt store repo + pure offloading engine

  enhancement
  EfficientMoE/MoE-Infinity#223 · Filter by author drunkcoding opened on Sep 18, 2026
  5

• [Feature Request] Multimodal MoE support: MiniMax-M3 (vision-language) + activate existing VL weights (Qwen3.5-MoE)

  enhancement
  EfficientMoE/MoE-Infinity#221 · Filter by author drunkcoding opened on Sep 18, 2026
  6