All pull requests

Pull requests

• Preserve completed routing records across async lifecycle cleanup

  #103·lmcafee-nvidia opened Sep 21, 2026Owner·

  ·

• Keep lazy MLA replacement layers on the source device

  determinism-exempt
  #102·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Preserve tensor pointer types for empty NVLS peers

  #100·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Align dense DeepGEMM physical strides

  #101·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Refresh disaggregated transfer registrations after resume

  #99·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Preserve detached Mamba slots until handoff release

  #98·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Add disaggregated prefill pairwise test foundation

  #97·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Initialize newly private value-cache pages before use

  #94·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Reject default NVFP4 scaling in batch-invariant mode

  #96·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Keep batch-invariant MLA on consistent expanded arithmetic

  #95·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Request FA4 log-sum-exp for attention sink correction

  #93·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Deliver public failure replies for rejected native KV imports

  #92·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Reject batch-dependent tensorwise FP8 in batch-invariant mode

  #91·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Preserve native MoE router width for empty expert peers

  #90·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Cover retained stop-score boundaries in batch-invariant inference

  #88·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Align padded prefill logits with active sampling rows

  #89·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Keep DP coordinator running when no worker accepts a request

  #87·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Fail and retire requests when their DP worker disconnects

  #86·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·
  2

• Correct non-interleaved RoPE for small attention heads

  #85·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Protect pinned disaggregated source state from destructive suspend

  #84·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·
  2

• Keep streamed top-N scores aligned with emitted tokens

  #83·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Keep DP coordinator serving after a client disconnects

  #82·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·
  2

• Add DP coordinator pairwise test foundation

  #81·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Cancel disaggregated handoffs before decode admission

  #80·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·

• Add batch-invariant pairwise test foundation

  #79·lmcafee-nvidia opened Sep 14, 2026Owner·

  ·