All pull requests

Pull requests

• feat(serving): support DeepSeek V4 chat tool calls

  #260·sjduan opened Sep 22, 2026·

  ·
  4

• feat(engram): host-side engram lookup with TP-sharded partial gather

  #259·supurs opened Sep 22, 2026·

  ·
  7

• feat(v41): add model entry, selective weights and execution planning

  #258·hashiqiqixian opened Sep 22, 2026·

  ·
  1

• feat(skill): add DSpark B16 8K profiler

  #255·high-cloud opened Sep 21, 2026·

  ·
  5

• feat(router): launch replicas on demand, locally or over ssh

  #254·lterrac opened Sep 21, 2026·

  ·
  9

• feat(pd): add external-router disaggregated serving

  #253·sjduan opened Sep 21, 2026·

  ·
  1

• ci: skip test and build gates for skill-only changes

  #252·vegetabledoww opened Sep 21, 2026·

  ·
  1

• Use PyPTO's persistent JIT cache in serving

  #251·ndleslx opened Sep 21, 2026·

  ·
  4

• perf(dspark): batch decode metadata preparation

  #250·high-cloud opened Sep 21, 2026·

  ·
  3

• feat(router): route requests across replicas by conversation

  #243·lterrac opened Sep 18, 2026·

  ·
  21

• Fix: consume complete DeepSeek MTP prefill output

  #238·MaxwellF1 opened Sep 17, 2026·

  ·
  5

• Use PyPTO's persistent JIT cache in serving

  #231·lyfne123 opened Sep 11, 2026·

  ·
  1

• docs: add model integration guide, env-vars reference, and model support matrix

  #228·LyuJiayi654 opened Sep 10, 2026·

  ·
  1

  1

• perf(serving): reduce async scheduling overhead

  #221·high-cloud opened Sep 7, 2026·

  ·
  2

• feat(qwen): consume pypto-lib Contract

  #219·bumble0918 opened Sep 6, 2026·

  ·
  1

• Add: integrate DeepSeek V4 DSpark serving

  #208·hashiqiqixian opened Aug 28, 2026·

  ·
  1

• perf(deepseek): build decode block tables without a temporary per row

  #206·lterrac opened Aug 27, 2026·

  ·
  1

• fix(qwen): dispatch decode with runtime batch

  #202·Little-oil opened Aug 27, 2026·

  ·
  1

  1

• fix(deepseek): fence resident decode-metadata refreshes with a verified readback

  #196·ndleslx opened Aug 25, 2026·

  ·
  2

• support mooncake ssd offload

  #193·superxf opened Aug 25, 2026·

  ·
  3

• Revert "perf(deepseek): reduce fused MTP decode host transfers"

  #186·ndleslx opened Aug 24, 2026·

  ·
  1

• feat(deepseek): add optional MoE expert statistics dump

  #185·high-cloud opened Aug 23, 2026·

  ·
  1

  1

• delete useless code

  #168·superxf opened Aug 19, 2026·

  ·
  1

• feat(deepseek): expand continuous batching across fixed kernels

  #145·hashiqiqixian opened Aug 7, 2026·

  ·
  1

• feat(skills): add DeepSeek V4 online SA_PROFILE perf skill

  #121·sunghajung6688 opened Jul 28, 2026·

  ·
  5