All pull requests

Pull requests

• perf(lanes): TF_FILL_DRAIN=ROWS starts a burst of short prompts in one turn, +56-60% at 8 streams under TF_FILL_MANY (off by default)

  #51·OsakaTX opened Oct 30, 2026·

  ·

• feat(grammar): compile json_schema and guided_choice in Zig (xgrammar 0.2.8 strict)

  #50·nuk3s opened Oct 26, 2026·6 of 8 tasks·

  ·
  1

• feat(dsv41): support response_format json_object by redrawing masked rows on rank 0

  #49·nuk3s opened Oct 26, 2026·6 of 8 tasks·

  ·

• feat(grammar): add a json_object byte matcher and token byte table (no callers yet)

  #48·nuk3s opened Oct 26, 2026·7 of 8 tasks·

  ·

• feat(dsv41): add TF_DS_VISION=0 to serve a vision checkpoint without its tower

  #47·nuk3s opened Oct 26, 2026·6 of 8 tasks·

  ·

• fix(dsv41): don't charge the drafter's weights on a --no-drafts start

  #46·nuk3s opened Oct 26, 2026·7 of 8 tasks·

  ·

• fix(dsv41): run eagerly when a CUDA graph capture fails

  #45·nuk3s opened Oct 26, 2026·7 of 8 tasks·

  ·

• feat(metrics): count prompt tokens restored from kept prompt states

  #44·eleata opened Oct 23, 2026·

  ·
  1

• feat(dsv41): output logprobs on the serial and --parallel engines

  #40·cshintov opened Oct 22, 2026·5 of 8 tasks·

  ·
  1

• feat(dsv41): structured output (response_format) on the serial and --parallel engines

  #39·cshintov opened Oct 22, 2026·5 of 8 tasks·

  ·
  1

• fix(dsv41): enforce native required and named tool prefixes

  #37·swbiggart opened Oct 21, 2026·

  ·
  2

• fix(server): an added token a client quotes in a chat request stays text

  #36·Blackwellboy opened Oct 21, 2026·6 of 8 tasks·

  ·
  1

• perf(dsv41): add opt-in GPU sampling candidate selection

  #11·sayyidfareed opened Oct 11, 2026·4 of 8 tasks·

  ·
  2