Issues

• 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 with worker hostnames on IPv6-only deployments

  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

• Runai streamer timeout causes double CUDA memory allocation

  ai-dynamo/modelexpress#638 · Filter by author gshtrasberg opened on Aug 14, 2026

• Cross RDMA domain source selection causes GPU OOM during fallback

  ai-dynamo/modelexpress#623 · Filter by author Broduker opened on Aug 13, 2026

• Declarative coordinator deployment via operator

  ai-dynamo/modelexpress#606 · Filter by author wseaton opened on Aug 11, 2026

• [Vision] Toward ModelExpress 1.0

  ai-dynamo/modelexpress#594 · Filter by author athreesh opened on Aug 6, 2026
  3

• Torch compile cache artifacts transfer between workers with incompatible compile configurations

  ai-dynamo/modelexpress#560 · Filter by author kangclzjc opened on Jul 28, 2026

• Pulling both model weights and MTP from s3:// leads to a crash

  ai-dynamo/modelexpress#537 · Filter by author gshtrasberg opened on Jul 23, 2026
  2

  

• auth: bound TokenReview traffic from unauthenticated callers

  ai-dynamo/modelexpress#530 · Filter by author wseaton opened on Jul 22, 2026
  2

• [RFC] Delta weight sync requirements

  ai-dynamo/modelexpress#524 · Filter by author athreesh opened on Jul 21, 2026
  3

• Clarify usage of model-streamer with s3 compatible object-storage

  ai-dynamo/modelexpress#500 · Filter by author strigazi opened on Jul 9, 2026
  3

• RDMA transfer to a stale donor aborts the process at libfabric with no fallback

  ai-dynamo/modelexpress#499 · Filter by author emricksini-h opened on Jul 9, 2026
  3

• Validate P2P transfers across different GPU SKUs

  enhancement
  ai-dynamo/modelexpress#476 · Filter by author zhengluo-nv opened on Jul 1, 2026