gate

Workflow runs for gate

• feat(graniteswitch): Granite Switch 4.1 8B with its adapters, systemone decisions and Granite Speech audio

  completed successfully:
  gate#37:Pull request#237openedbyasher
  feat/graniteswitch

  Oct 28, 202612s
  More actions for workflow run 37

• feat(load): nvfp4 weight scales apply inside kq kernels and per-expert scales load

  completed successfully:
  gate#36:Pull request#236openedbyasher
  nvfp4-out-scale

  Oct 27, 202615s
  More actions for workflow run 36

• fix(gdn): --kv-bits decode on gated-delta models drops a per-token host sync

  completed successfully:
  gate#35:Pull request#235openedbyasher
  fix/gdn-decode-mask-sync

  Oct 27, 202615s
  More actions for workflow run 35

• release: 0.5.3

  completed successfully:
  gate#34:Pull request#234openedbyasher
  release-0.5.3

  Oct 26, 202616s
  More actions for workflow run 34

• fix(serve): coarse decode caps, 2048 MoE prefill chunk and fewer per-step host reads on pre-M5 GPUs

  completed successfully:
  gate#33:Pull request#233openedbyasher
  fix/release-052-regressions

  Oct 26, 202614s
  More actions for workflow run 33

• release: 0.5.2

  completed successfully:
  gate#32:Pull request#231openedbyasher
  release-0.5.2

  Oct 26, 202616s
  More actions for workflow run 32

• feat(vlm): MiniCPM-V 4.6 and 4.7 vision, with DSpark support for the 35B-A3B

  completed successfully:
  gate#31:Pull request#230synchronizebyasher
  feat/minicpmv

  Oct 26, 202615s
  More actions for workflow run 31

• feat(vlm): MiniCPM-V 4.6 and 4.7 vision, with DSpark support for the 35B-A3B

  completed successfully:
  gate#30:Pull request#230openedbyasher
  feat/minicpmv

  Oct 26, 202616s
  More actions for workflow run 30

• feat(bench): qwen3.8-27b serve-bench --vs mtplx

  completed successfully:
  gate#29:Pull request#229openedbyasher
  bench-mtplx

  Oct 25, 202612s
  More actions for workflow run 29

• fix(systemone): letter options as A) on models other than OpenJev

  completed successfully:
  gate#28:Pull request#228openedbyasher
  systemone-plain-letters

  Oct 25, 202614s
  More actions for workflow run 28

• feat(pull): catalog of recommended models by memory class, with an interactive picker

  completed successfully:
  gate#27:Pull request#227synchronizebyasher
  pull-picker

  Oct 25, 202614s
  More actions for workflow run 27

• feat(pull): catalog of recommended models by memory class, with an interactive picker

  completed successfully:
  gate#26:Pull request#227openedbyasher
  pull-picker

  Oct 25, 202616s
  More actions for workflow run 26

• perf(qwen4exp): fused MoE on the f32 residual, confidence-sized MTP rounds

  completed successfully:
  gate#25:Pull request#225synchronizebyasher
  q4e-mtp-perf

  Oct 25, 202614s
  More actions for workflow run 25

• perf(qwen4exp): fused MoE on the f32 residual, confidence-sized MTP rounds

  completed successfully:
  gate#24:Pull request#225openedbyasher
  q4e-mtp-perf

  Oct 25, 202611s
  More actions for workflow run 24

• fix(memory): judge resident loads on memory not wired or reclaimable

  completed successfully:
  gate#23:Pull request#224openedbyasher
  q4e-mtp-decode

  Oct 25, 202618s
  More actions for workflow run 23

• fix(stream): case where ngram tables were counted towards expert size

  completed successfully:
  gate#22:Pull request#223openedbyasher
  mtp-subdir-drafter

  Oct 24, 202612s
  More actions for workflow run 22

• feat(bench): serve-bench --vs omlx arm

  completed successfully:
  gate#21:Pull request#222openedbyasher
  bench-omlx

  Oct 24, 202614s
  More actions for workflow run 21

• perf: full-attention kv rows append in place, flint decode up to 1.9x llama.cpp

  completed successfully:
  gate#20:Pull request#219synchronizebyasher
  flint-decode-perf

  Oct 24, 202612s
  More actions for workflow run 20

• feat(spec): gemma 4 targets run plain dflash drafters

  completed successfully:
  gate#19:Pull request#221openedbyasher
  feat/gemma4-dflash

  Oct 23, 202613s
  More actions for workflow run 19

• feat(vlm): quantized mmproj towers load packed for every family

  completed successfully:
  gate#18:Pull request#220openedbyasher
  feat/quantized-mmproj

  Oct 23, 202612s
  More actions for workflow run 18

• perf: full-attention kv rows append in place, flint decode up to 1.9x llama.cpp

  completed successfully:
  gate#17:Pull request#219synchronizebyasher
  flint-decode-perf

  Oct 23, 202616s
  More actions for workflow run 17

• perf: full-attention kv rows append in place, flint decode up to 1.9x llama.cpp

  completed successfully:
  gate#16:Pull request#219openedbyasher
  flint-decode-perf

  Oct 23, 202618s
  More actions for workflow run 16

• test(container): the other-thread signal test survives stalls, flaky issues quote the failure

  completed successfully:
  gate#15:Pull request#218openedbyasher
  fix/flaky-usr1-wait-test

  Oct 23, 202613s
  More actions for workflow run 15

• perf(stream): deepseek-v4.1 streamed decode and ttft improvements, per-phase command buffer caps, memory fixes

  completed successfully:
  gate#14:Pull request#217openedbyasher
  ds41-perf2

  Oct 22, 202613s
  More actions for workflow run 14

• feat(flint): RecursiveSI Flint support

  completed successfully:
  gate#13:Pull request#215synchronizebyasher
  flint

  Oct 22, 202613s
  More actions for workflow run 13