Issues

• SGLang MTP draft P2P (#824) fails on AWS EFA/libfabric: second transfer waits 60s and fails with NIXL_ERR_REMOTE_DISCONNECT

  ai-dynamo/modelexpress#929 · Filter by author nskpro-cmd opened on Oct 23, 2026

• Artifact transfer opens a new TCP connection for every per-chunk RPC

  ai-dynamo/modelexpress#927 · Filter by author JiaweiGithub opened on Oct 23, 2026

• [Question] Metadata-only bootstrap for SGLang/vLLM P2P loading from S3-compatible storage

  ai-dynamo/modelexpress#925 · Filter by author zhy76 opened on Oct 22, 2026

• Weight-source health gating is bypassed when MX_ARTIFACT_READY_URL is unset

  bug
  ai-dynamo/modelexpress#922 · Filter by author zhengluo-nv opened on Oct 21, 2026

• SGLang P2P loading selects FP8 fallback from placeholder scales and omits MXFP8 buffers

  ai-dynamo/modelexpress#920 · Filter by author henrylhtsang opened on Oct 21, 2026

• collective refit: debug events and backend timers for epoch changes and evictions

  ai-dynamo/modelexpress#917 · Filter by author nicolasnoble opened on Oct 20, 2026

• fix(client): fail closed on addressless tensors in receive and align register_additional_tensors

  ai-dynamo/modelexpress#916 · Filter by author hmartin-alt opened on Oct 20, 2026

• Runtime tensors included in transfer cause RDMA loading failures

  ai-dynamo/modelexpress#915 · Filter by author vie-serendipity opened on Oct 20, 2026
  3

• Load strategy chain skips model reinit when a strategy raises a plain exception

  ai-dynamo/modelexpress#902 · Filter by author JiaweiGithub opened on Oct 19, 2026

• Source pod dies mid-transfer: TP ranks split between RDMA and storage, load hangs, and the hang spreads to other pods

  ai-dynamo/modelexpress#901 · Filter by author nskpro-cmd opened on Oct 16, 2026

• ModelStreamer / MX_MODEL_URI has no path for a draft checkpoint stored in a separate location from the main model

  ai-dynamo/modelexpress#880 · Filter by author nskpro-cmd opened on Oct 12, 2026
  1

• RL cold start ignores max_replay_chain_length: object-storage replay is hard-coded to 64 entries

  ai-dynamo/modelexpress#851 · Filter by author fbie-droid opened on Oct 5, 2026

• P2P metadata exchange fails when workers advertise DNS hostnames

  ai-dynamo/modelexpress#847 · Filter by author VanyaGlazunov opened on Oct 5, 2026

• Add InstantTensor support to SglangAdapter — SGLang gets zero benefit from ModelExpress on a cold, no-peer start today

  ai-dynamo/modelexpress#825 · Filter by author nskpro-cmd opened on Sep 30, 2026
  5

• Extend P2P/RDMA to cover speculative/draft weights, not just the main model

  ai-dynamo/modelexpress#823 · Filter by author nskpro-cmd opened on Sep 30, 2026
  1

• Make ModelExpress MTP draft loads safe through P2P RDMA and avoid re-streaming the full checkpoint through sglang/vllm

  ai-dynamo/modelexpress#805 · Filter by author zhy76 opened on Sep 24, 2026
  8

  

• feat(refit): make checkpoint cache quota configurable across cold start and refit

  ai-dynamo/modelexpress#796 · Filter by author c2w-sea opened on Sep 22, 2026

• refit: honor KV-cache retention policy when K/V quantization scales change

  ai-dynamo/modelexpress#784 · Filter by author anson627 opened on Sep 17, 2026

• feat(refit): Allow disabling S3 version metadata checks

  ai-dynamo/modelexpress#780 · Filter by author c2w-sea opened on Sep 17, 2026

• feat(refit): Configurable generator model_name and implicit cached seed

  ai-dynamo/modelexpress#779 · Filter by author c2w-sea opened on Sep 17, 2026

• Kimi INT4 S3 refit exhausts host RAM: layerwise accounting retains completed expert weights

  ai-dynamo/modelexpress#776 · Filter by author anson627 opened on Sep 16, 2026
  1

• fix(refit): synchronize cold-start identity with vLLM reporting and active refit

  ai-dynamo/modelexpress#737 · Filter by author anson627 opened on Sep 9, 2026

• Feature request: add a notification channel to external component

  ai-dynamo/modelexpress#692 · Filter by author hzxuzhonghu opened on Aug 28, 2026
  8

• Upstream a vLLM pre-resolution hook and remove the Hugging Face snapshot patch

  enhancement
  ai-dynamo/modelexpress#640 · Filter by author zhengluo-nv opened on Aug 14, 2026
  4

• Support Megatron QKV resharding when KV heads are fewer than trainer TP ranks

  ai-dynamo/modelexpress#639 · Filter by author zhengluo-nv opened on Aug 14, 2026