All pull requests

Pull requests

• [CI/Build][Hardware][AMD] Use the CUDA Megatron-LM commit and patch in the ROCm image

  #521·jhinpan opened Nov 7, 2026Contributor·

  ·
  3

• [Core] perf: SampleCodec.prepare 5.3x faster on id lists (160 -> 30 ms)

  #520·dexhunter opened Nov 7, 2026·3 of 6 tasks·

  ·
  4

• [Sync] Update Megatron 0.20 compatibility and packed weight-transfer coverage

  #518·aoshen02 opened Nov 5, 2026Collaborator·

  ·
  8

• [NPU] Sync Ascend with main rollout, training and serving updates

  #517·Meihan-chen opened Nov 5, 2026Collaborator·

  ·
  5

• [Bugfix][Hardware][AMD] Fix TP1 colocate IPC weight sync and the NIXL rollout transport on ROCm

  #516·jhinpan opened Nov 5, 2026Contributor·

  ·
  3

• [Bugfix][Hardware][AMD] Make --offload-train work on ROCm and enable it in the AMD example

  #514·jhinpan opened Nov 5, 2026Contributor·

  ·
  1

• [Hardware][Ascend] Support Qwen3.8-27B GRPO with GDN head TP

  #512·Tjh-UKN opened Nov 3, 2026·

  ·
  3

• feat: add online external EAGLE3 and DSpark draft training for NPU

  #511·is-not opened Nov 2, 2026·

  ·
  17

• [Feature][Training] Give the online DSpark draft its own learning rate

  #510·0z5a opened Oct 31, 2026·

  ·
  7

• [Hardware][Ascend] Retire upstreamed weight-transfer backports

  #507·Annarine opened Oct 29, 2026·

  ·
  5

• [Test][Ascend] Validate Qwen3-0.6B disk delta updates against full checkpoint

  #505·wangx700 opened Oct 28, 2026·

  ·
  14

• Add Qwen3-VL loading and tool-use examples

  #493·Meihan-chen opened Oct 23, 2026Collaborator·

  ·
  6

• [Feature][Ascend] Add sparse HCCL weight synchronization

  #492·wangx700 opened Oct 23, 2026·

  ·
  6

• [Hardware][Ascend] Add Ascend NPU support via accelerator extension points

  #489·Meihan-chen opened Oct 21, 2026Collaborator·

  ·
  13

• [Bugfix][Training] Time out weight-transfer requests and resend a failed weight push

  #486·0z5a opened Oct 19, 2026·

  ·
  4

• [Rollout][Training] Speculative-decoding rollout benchmark, tail metrics and an MTP example

  #485·0z5a opened Oct 19, 2026·

  ·
  3

• [Bugfix][Training] Fix DSpark online draft training: grad-norm group and loss memory

  #484·0z5a opened Oct 19, 2026·

  ·
  5

• [NPU] Adapt GPU test suites for Ascend CI

  #481·Meihan-chen opened Oct 19, 2026Collaborator·

  ·
  9

• [Bugfix][Core] Lazy-import aiohttp in rm_hub to unblock reward CPU tests

  #477·Asthenia0412 opened Oct 14, 2026·

  ·
  1

  4

• [Rollout] Add experimental deferred exact-KV Ouro decoding

  #475·Camellia86 opened Oct 14, 2026·

  ·
  3

• [Training] Avoid copies when gathering contiguous TP row shards

  #469·0z5a opened Oct 11, 2026·

  ·
  3

• [Bugfix][Rollout] Honor startup probe timeouts and optional wait budget

  #468·jasonchen505 opened Oct 7, 2026·

  ·
  1

  3

• [Rollout] Add optional per-position OPD teacher scoring

  #467·LinzeShi opened Oct 6, 2026·

  ·
  5

• [Feature][Training] Add Huginn recurrent RL training

  #464·0z5a opened Oct 3, 2026·

  ·
  8

• [Feature][Training] Add Nanbeige recurrent RL training

  #463·0z5a opened Oct 3, 2026·

  ·
  8