All pull requests

Pull requests

• [5/5] refactor(attention): integrate model bindings and state lifecycle

  #1721·sufubao opened Nov 12, 2026Collaborator·

  ·

• [4/5] refactor(attention): use varlen operators for vision and audio

  #1720·sufubao opened Nov 12, 2026Collaborator·

  ·

• [3/5] refactor(attention): add provider backends and explicit selection

  #1719·sufubao opened Nov 12, 2026Collaborator·

  ·

• [2/5] refactor(attention): relocate kernels without changing execution

  #1718·sufubao opened Nov 12, 2026Collaborator·

  ·

• [1/5] refactor(attention): define execution contracts and device workspace

  #1717·sufubao opened Nov 12, 2026Collaborator·

  ·

• fix(pd): align handshake deadlines with resource wait budgets

  #1715·shihaobai opened Nov 12, 2026Collaborator·

  ·

• fix(config): support TCPStore startup in daemon workers

  #1714·shihaobai opened Nov 12, 2026Collaborator·

  ·

• perf(pd): retain bounded affinity for long cached prefixes

  #1713·shihaobai opened Nov 12, 2026Collaborator·

  ·

• perf(glm): fuse MTP input normalization and concatenation

  #1712·shihaobai opened Nov 12, 2026Collaborator·

  ·

• perf(glm): fuse dense/shared FFN activation and FP8 quantization

  #1711·shihaobai opened Nov 12, 2026Collaborator·

  ·

• perf(moe): fuse GLM prefill SwiGLU and UE8M0 quantization

  #1709·shihaobai opened Nov 12, 2026Collaborator·

  ·

• feat: support Qwen3.8 Flash Next with ReplaySSM

  #1706·sufubao opened Nov 11, 2026Collaborator·

  ·

• feat(openai): support output logprobs for chat completions

  #1701·WANDY666 opened Nov 11, 2026Contributor·

  ·

• perf: reduce attention cache write and transfer overhead

  #1681·sufubao opened Nov 2, 2026Collaborator·

  ·
  1

• fix: authenticate PD registration and restrict node destinations

  #1673·mvanhorn opened Nov 2, 2026·

  ·
  1

  9

• fix(glm5): align FP8 precision and preserve reasoning tool calls

  #1657·shihaobai opened Oct 22, 2026Collaborator·

  ·
  12

• perf(mtp): project accepted-tail logits in the first EAGLE draft

  #1653·hiworldwzj opened Oct 22, 2026Collaborator·

  ·

• fix: parse GLM tool calls inside reasoning

  #1651·shihaobai opened Oct 21, 2026Collaborator·

  ·

• feat: Intel XPU backend on the hardware platform registry

  #1650·siju-samuel opened Oct 21, 2026·3 of 4 tasks·

  ·

• perf: streamline greedy sampling and skip identity penalties

  #1643·sufubao opened Oct 20, 2026Collaborator·

  ·

• Add tool-call-aware KV retention and bounded continuation admission

  #1639·sufubao opened Oct 17, 2026Collaborator·

  ·

• refactor(logging): clarify startup and bound runtime diagnostics

  #1638·sufubao opened Oct 17, 2026Collaborator·

  ·
  5

• Defer decode KV allocation until the first prefill chunk is ready

  #1637·sufubao opened Oct 17, 2026Collaborator·

  ·

• feat: decouple device backends and add Metal support for Qwen3.5

  #1630·sufubao opened Oct 15, 2026Collaborator·

  ·

• perf: batch KV page allocation and fuse token and KV index gathering

  #1617·blueswhen opened Oct 9, 2026Collaborator·

  ·
  3