Issues

• GLM DMTCP checkpoint stalls draining socket after successful CUDA checkpoint

  TensorRails/warmctl#72 · Filter by author kbespalov opened on Nov 2, 2026
  1

• Validate and bind the complete CUDA runtime across bake and restore; reject compiler/header drift before model loading

  TensorRails/warmctl#71 · Filter by author kbespalov opened on Nov 2, 2026
  3

• Detect and cover lazy initialization and Triton JIT after restore readiness

  TensorRails/warmctl#70 · Filter by author kbespalov opened on Nov 1, 2026
  1

• Preserve a recoverable snapshot when network publication fails

  TensorRails/warmctl#67 · Filter by author kbespalov opened on Nov 1, 2026

• DMTCP: a bake of a spawned vLLM engine hangs in the socket drain (about 1 in 5)

  TensorRails/warmctl#62 · Filter by author kbespalov opened on Nov 1, 2026

• Registry: zstd-compress every snapshot layer, weights included (byte planes for 16-bit), and re-encode published snapshots in place without a rebake

  enhancement
  TensorRails/warmctl#60 · Filter by author kbespalov opened on Oct 31, 2026

• Snapshot vLLM engines that run a KV-transfer connector (NixlConnector prefill/decode, KV offloading)

  enhancement
  TensorRails/warmctl#40 · Filter by author kbespalov opened on Oct 25, 2026

• Warm up and canary the chat and tool-call path: xgrammar initializes on the first structured request after a restore

  enhancement
  TensorRails/warmctl#39 · Filter by author kbespalov opened on Oct 25, 2026

• Weights reload at 3.3 GiB/s per GPU: split the phase timing, then pipeline read, copy and fingerprint

  enhancement
  TensorRails/warmctl#38 · Filter by author kbespalov opened on Oct 25, 2026
  1

• Compress the weight parts on the GPU with nvCOMP ANS (float16 and FP8 E4M3)

  enhancement
  TensorRails/warmctl#37 · Filter by author kbespalov opened on Oct 25, 2026

• Compress non-weight snapshot layers with zstd on push

  enhancement
  TensorRails/warmctl#35 · Filter by author kbespalov opened on Oct 24, 2026

• Drive CRIU through go-criu (swrk RPC) and start the CUDA restore on PostRestore

  enhancement
  TensorRails/warmctl#9 · Filter by author kbespalov opened on Oct 14, 2026