All pull requests

Pull requests

• fix(refit): re-register the generator worker before each version lease

  fixsize/M
  #850·c2w-sea opened Oct 5, 2026Contributor·

  ·
  3

• chore: bump version to 0.8.0

  choresize/M
  #849·pvijayakrish opened Oct 5, 2026Contributor·

  ·
  4

• fix(client): resolve NIXL metadata hostnames before fetching

  fixsize/L
  #848·VanyaGlazunov opened Oct 5, 2026·

  ·
  1

  3

• refactor(rl): move Vime weight transfer backend into ModelExpress

  refactorsize/XXL
  #846·nv-hwoo opened Oct 5, 2026Contributor·

  ·
  3

• fix(sglang): read scoped parallel context and preserve shard identity

  fixsize/L
  #845·renault-rai opened Oct 4, 2026·

  ·
  3

• fix(vllm): release idle allocator cache after successful RDMA receive

  fixsize/M
  #844·Ethan-ES opened Oct 4, 2026·

  ·
  3

• fix(client): re-register artifact sources when the server rejects heartbeats

  fixsize/M
  #843·joeltg opened Oct 3, 2026·

  ·
  2

• feat(collective): equal-geometry stacking for the MILES refit path

  featsize/XXL
  #842·ajcasagrande opened Oct 3, 2026·

  ·

• feat(collective): refit sharded SGLang engines from a gathered MILES source

  featsize/XXL
  #841·ajcasagrande opened Oct 3, 2026·

  ·

• fix(cache): keep listing models when one cache directory is unreadable

  fixsize/L
  #840·nicolasnoble opened Oct 3, 2026Contributor·

  ·
  5

• fix(collective): settle lanes before bootstrap and back off bounded waits

  fixsize/XL
  #839·ajcasagrande opened Oct 2, 2026·

  ·

• fix(ci): update async-trait for Rust 1.99 clippy

  fixsize/XS
  #838·nv-hwoo opened Oct 2, 2026Contributor·

  ·
  1

• feat(refit): add trainer mesh APIs and version linkage

  featsize/XXL
  #835·zhengluo-nv opened Oct 2, 2026Contributor·

  ·
  17

  

• fix(sglang): isolate the speculative draft pass from P2P and load only its shards

  fixsize/XL
  #833·joeltg opened Oct 2, 2026·

  ·
  1

• feat(client): partition source discovery by locality with MX_SOURCE_DOMAIN

  featsize/L
  #832·joeltg opened Oct 2, 2026·

  ·
  3

• fix(p2p): validate what a publisher asserts about itself

  fixsize/L
  #831·nicolasnoble opened Oct 1, 2026Contributor·

  ·
  12

• feat(collective): MILES adapter for the NCCL M2N collective refit path

  featsize/XXL
  #830·ajcasagrande opened Oct 1, 2026·

  ·

• feat: support artifact transfer via k8s-service

  featsize/XL
  #829·AndyDai-nv opened Oct 1, 2026Contributor·

  ·
  2

• feat(refit): support incremental S3 bucket staging for Miles

  featsize/XXL
  #826·nv-hwoo opened Sep 30, 2026Contributor·

  ·
  14

• feat: WIP SGLang Qwen3.5 MTP draft P2P and cold-load reuse

  featsize/XXL
  #824·zhy76 opened Sep 30, 2026·

  ·
  2

• fix(refit): abort stranded transfers on epoch bump, TTL the idempotency key

  fixsize/L
  #822·nicolasnoble opened Sep 29, 2026Contributor·

  ·
  3

• feat: make NIXL metadata fetch timeout configurable

  featsize/S
  #820·jason136 opened Sep 29, 2026·

  ·
  4

• ci: trigger model benchmarks with authorized PR comments

  cisize/XXL
  #819·anson627 opened Sep 29, 2026Contributor·

  ·
  8

• ci: add model benchmark harness for cold load and refit

  cisize/XXL
  #818·anson627 opened Sep 29, 2026Contributor·

  ·
  6

• feat(refit): add collective partial checkpoint installation

  featsize/XXL
  #810·anson627 opened Sep 25, 2026Contributor·

  ·
  3