All pull requests

Pull requests

• docs(refit): move the collective refit env vars to the RL configuration guide

  docssize/M
  #947·nicolasnoble opened Oct 27, 2026Contributor·

  ·
  4

• refactor(rl): clarify NIXL refit ownership and replay

  refactorsize/XXL
  #946·zhengluo-nv opened Oct 27, 2026Contributor·

  ·
  7

  

• fix(client): fall back to per-tensor registration when a tensor spans past its allocation

  fixsize/M
  #945·undergg opened Oct 27, 2026Contributor·

  ·
  5

• fix(client): fail clearly when a per-device listener port is taken

  fixsize/L
  #944·undergg opened Oct 27, 2026Contributor·

  ·
  5

• fix(client): flatten and chunk DEBUG tensor checksums

  fixsize/M
  #943·undergg opened Oct 27, 2026Contributor·

  ·
  3

• fix(vllm): drop streamer extra config in native fallback loader

  fixsize/M
  #942·undergg opened Oct 27, 2026Contributor·

  ·
  5

• feat(sglang): refill released weights over P2P on resume

  featsize/L
  #941·galletas1712 opened Oct 27, 2026·

  ·
  1

• feat(sglang): Support p2p metadata transfer

  featsize/XXL
  #940·hzxuzhonghu opened Oct 27, 2026·

  ·
  1

  10

• perf(s3): add opt-in preallocated ranged downloads

  perfsize/XL
  #938·anson627 opened Oct 26, 2026Contributor·

  ·
  6

• fix(crypto): hash with OpenSSL in the openssl build and keep sha2 out of it

  fixsize/L
  #935·madhugoutham opened Oct 26, 2026Contributor·

  ·
  1

  4

• fix(vllm): preserve instance-local FlashInfer MLA counters during RDMA transfer

  fixsize/L
  #933·Ethan-ES opened Oct 25, 2026Contributor·

  ·
  3

• feat(refit): add Azure publication and startup support

  featsize/XL
  #931·ipochi opened Oct 23, 2026Contributor·

  ·
  3

• fix(refit): apply the classic CUDA pool to the registered buffer's device

  fixsize/XL
  #930·yafshar opened Oct 23, 2026Contributor·

  ·
  3

• fix(client): reuse one gRPC channel across an artifact transfer

  fixsize/L
  #928·JiaweiGithub opened Oct 23, 2026·

  ·
  1

  3

• test(ci): add inference parity, modern models, and SGLang coverage

  size/XXLtest
  #926·zhengluo-nv opened Oct 22, 2026Contributor·

  ·
  12

  

• fix(sglang): preserve MXFP8 kernels and scale buffers after peer loading

  fixsize/XL
  #924·KavinKrishnan opened Oct 22, 2026Contributor·

  ·
  1

  15

• test: separate base and RL Python suites

  size/Mtest
  #923·zhengluo-nv opened Oct 21, 2026Contributor·

  ·
  2

• Exclude distributed communicator state from peer weight transfers

  size/M
  #921·jason136 opened Oct 21, 2026Contributor·

  ·
  5

• fix(client): treat unset source_load as unknown in topology_aware

  fixsize/M
  #913·v0ropaev opened Oct 20, 2026·

  ·
  3

• fix(s3): reject '.' and '..' as the bucket of an s3:// model name

  fixsize/S
  #912·v0ropaev opened Oct 20, 2026·

  ·
  4

• feat(client): load weights from S3-over-RDMA object stores via NIXL OBJ

  featsize/XXL
  #909·dvaldivia opened Oct 19, 2026·

  ·
  16

• feat(client): pre-stage receiver to overlap the level-2 weight reload with process restore

  featsize/XXL
  #908·zanes-ops opened Oct 19, 2026·

  ·
  9

• docs(examples): add Mooncake artifact reuse walkthrough

  docssize/L
  #905·CDalezyb opened Oct 19, 2026Contributor·

  ·
  11

• fix(client): reinitialize the model after an unexpected strategy error

  fixsize/S
  #903·JiaweiGithub opened Oct 19, 2026·

  ·
  1

  3

• fix(hf): retry transient download errors and skip optional files

  fixsize/L
  #897·cyclinder opened Oct 14, 2026·

  ·
  7