All pull requests

Pull requests

• feat: nanochat model, optimizer and packed-row pretraining data

  #4431·eliebak opened Nov 8, 2026·

  ·

• feat(optim): native Muon, replacing dion

  #4430·faresobeid opened Nov 8, 2026Contributor·

  ·

• perf(optim): chunked optimizer-state offload for Muon via a subset-step interface (builds on #4427)

  #4429·faresobeid opened Nov 8, 2026Contributor·

  ·

• feat(trainer): async checkpointing with pinned staging (supersedes #4331)

  #4428·faresobeid opened Nov 8, 2026Contributor·

  ·

• perf(optim): overlap the offloaded AdamW step with the next forward (builds on #4008)

  #4427·faresobeid opened Nov 8, 2026Contributor·

  ·

• perf(trainer): NUMA-bind ranks for optimizer-state and activation offload

  #4426·faresobeid opened Nov 8, 2026Contributor·

  ·

• feat(inference): carry MoE routed experts from prefill to decode in kv_transfer_params

  #4425·cabrinha opened Nov 8, 2026Collaborator·

  ·

• refactor(inference): vLLM plugin package and stock-vLLM image (vLLM 0.33)

  #4424·cabrinha opened Nov 8, 2026Collaborator·

  ·
  1

• refactor(inference): prime-rl's vLLM changes as a vLLM plugin, plus a stock-vLLM image

  #4423·cabrinha opened Nov 8, 2026Collaborator·

  ·
  1

• fix(trainer): compile DeepEP MoE layers under activation checkpointing

  #4421·faresobeid opened Nov 7, 2026Contributor·

  ·

• feat: log MoE router agreement with replayed experts

  #4420·faresobeid opened Nov 7, 2026Contributor·

  ·

• feat(weights): engines pull ModelExpress weight versions on their own

  #4418·skpulipaka26 opened Nov 7, 2026Contributor·

  ·

• feat: restore rollout retries from filesystem snapshots

  #4416·mikasenghaas opened Nov 7, 2026Member·

  ·

• fix(inference): default Mooncake KV lease to 30 s; correct its docstring

  #4414·faresobeid opened Nov 6, 2026Contributor·

  ·

• fix(trainer): Forward softmax_scale through FA4 custom ops

  #4412·sean-comerford opened Nov 6, 2026·

  ·

• feat(weights): recover NCCL weight transfers without restarting inference

  #4405·hallerite opened Nov 6, 2026Member·2/5·

  ·
  11

• feat(orchestrator): adaptive concurrency on queued share and marginal throughput

  #4403·faresobeid opened Nov 6, 2026Contributor·

  ·

• feat(sft): validate on held-out datatrove token shards

  #4401·snimu opened Nov 5, 2026Collaborator·

  ·

• fix: dequantize Qwen3.5-family fp8 releases on load

  #4399·vikhyat opened Nov 4, 2026·

  ·
  1

• fix(inference): drain every worker reply of a failed collective RPC

  #4397·hallerite opened Nov 4, 2026Member·

  ·

• fix: exit non-zero when a server is stopped by a signal, and save the supervised trainer's resolved config

  #4396·skpulipaka26 opened Nov 4, 2026Contributor·

  ·
  4

• feat: keep training when an inference server fails

  #4395·hallerite opened Nov 4, 2026Member·5/5·

  ·
  56

• feat(moe): Mega MoE mxfp8_qat precision (MXFP8 forward, bf16 backward)

  #4394·nxphi47 opened Nov 4, 2026Contributor·

  ·

• perf(trainer): all-reduce HSDP gradients once per step

  #4393·faresobeid opened Nov 4, 2026Contributor·

  ·

• feat(models): support Nemotron 3.5 Super text-only training

  #4391·KyleWong288 opened Nov 4, 2026Member·

  ·