All pull requests

Pull requests

• [tinker] Decode vLLM completion bodies straight into numpy with pysimdjson

  #2165·avigyabb opened Sep 5, 2026Collaborator·

  ·
  4

• [megatron] Skip frozen-master backload for adapter-only LoRA weight sync

  #2157·avigyabb opened Sep 4, 2026Collaborator·

  ·

• fix(megatron): stage common state for cloud resume

  #2154·bvolpato opened Sep 3, 2026Contributor·

  ·

• fix(megatron): isolate FP32 adapter optimizer state

  #2153·bvolpato opened Sep 3, 2026Contributor·

  ·

• fix(megatron): refresh masters after weights-only load

  #2149·bvolpato opened Sep 3, 2026Contributor·

  ·

• fix(rollout): preserve sampled EOS probabilities

  #2148·bvolpato opened Sep 3, 2026Contributor·

  ·

• fix(inference): reset requests when discarding KV cache

  #2147·bvolpato opened Sep 3, 2026Contributor·

  ·

• fix(megatron): synchronize complete accumulation windows

  #2146·bvolpato opened Sep 3, 2026Contributor·

  ·

• fix(megatron): finalize checkpoints before pruning

  #2145·bvolpato opened Sep 3, 2026Contributor·

  ·

• perf(fsdp): limit actor LM-head work to the action window

  #2144·NenoL2001 opened Sep 3, 2026·

  ·
  1

  3

• TITO support for SkyRL

  #2143·xinze-zheng opened Sep 3, 2026·

  ·
  12

  

• fix(tinker): align forwarding pool deadline

  #2140·hershg opened Sep 2, 2026Contributor·

  ·
  1

• [fix] Build DeepGEMM against selected torch

  #2139·hershg opened Sep 2, 2026Contributor·

  ·
  9

• test: add GLM 5.3 LoRA logprob parity coverage

  x-trajectory
  #2128·hershg opened Sep 1, 2026Contributor·

  ·

• [fix][inference] Cancel abandoned generation

  #2121·bvolpato opened Sep 1, 2026Contributor·

  ·
  3

• perf(megatron): preserve LoRA export dtype

  #2117·hershg opened Sep 1, 2026Contributor·

  ·
  5

• fix(megatron): release LoRA export before reload

  x-trajectory
  #2113·hershg opened Sep 1, 2026Contributor·

  ·

• fix(megatron): release host model replica after backload

  x-trajectory
  #2112·hershg opened Sep 1, 2026Contributor·

  ·
  5

• [feat]Add Group reward-Decoupled Normalization Policy Optimization

  #2104·Athe-kunal opened Aug 31, 2026·

  ·
  14

• Raise a clear error when a batch is smaller than the DP shard stride

  #2103·yurekami opened Aug 30, 2026·

  ·
  1

  3

• docs(megatron): add GLM 5.2 two-node training proof

  x-trajectory
  #2099·hershg opened Aug 28, 2026Contributor·

  ·
  13

• [fix][sft] Locate the assistant response boundary instead of subtracting prompt length

  #2092·bharatmekala opened Aug 24, 2026·3 tasks done·

  ·
  5

• fix(r3): balance replay padding across expert ranks

  #2091·EazyReal opened Aug 23, 2026Contributor·

  ·
  2

• [fsdp] Trim dense padding per final microbatch

  #2089·EazyReal opened Aug 22, 2026Contributor·

  ·
  6

• [train] FSDP2: wrap only input embeddings, not every nn.Embedding (fixes glm4v)

  #2088·dzorlu opened Aug 22, 2026·

  ·
  1

  2