All pull requests

Pull requests

• Gemma 4 E4B on the Intel Arc Pro B70 (Xe2)

  #1552·Mirrdhyn opened Oct 28, 2026Contributor·

  ·

• server: keep the history mark inside what the next turn repeats after a tool reply

  #1551·Mirrdhyn opened Oct 28, 2026Contributor·

  ·

• cuda: flashnext_batch, flashnext_gdn and qwen4_gdn take tf_common::warp_sum, SASS unchanged

  #1550·jschmied opened Oct 28, 2026Contributor·

  ·

• rows: the write worker charges a block in flight once, and a save that fits the staging permit waits instead of being skipped

  #1548·wojo opened Oct 28, 2026Contributor·

  ·
  1

• glm cuda: TF_GLM_KDA_PAD packs the fused KDA input at TP=3 (opt-in, default off)

  #1545·BadAd84 opened Oct 28, 2026Contributor·

  ·
  1

• Native Vulkan backend for Qwen3.5-2B (opt-in): engine and kernel-contract registration (stacked on #1526)

  #1544·joshuaswarren opened Oct 28, 2026·

  ·

• rows: a state its family refuses while it scatters is forgotten after one recall

  #1543·wojo opened Oct 28, 2026Contributor·

  ·
  1

• build(hip): opt-in -Dhip-relax-branches for gfx1100 under clang 21

  #1542·AbrahamCete opened Oct 28, 2026·8 of 9 tasks·

  ·
  1

• ling: host-only family package from jschmied's ling-tiny

  #1541·ThinkOffApp opened Oct 28, 2026Contributor·5 of 9 tasks·

  ·

• metal nemotron: keys and values in the core KV pool's pages (TF_NEMOTRON_METAL_KV_PAGES)

  #1537·jschmied opened Oct 28, 2026Contributor·

  ·

• compact: leave short requests alone when the auto reserve exceeds the window

  #1536·ppardus opened Oct 28, 2026Contributor·

  ·
  2

• exl3 decode: read the trellis through ld.global.nc.L1::no_allocate

  #1534·BadAd84 opened Oct 28, 2026Contributor·

  ·
  2

• tools: cuda_sass_identity.py rebuilds a differing image and fails on an unstable build

  #1532·jschmied opened Oct 28, 2026Contributor·

  ·

• MOSS-Image: the denoiser on CUDA (GB10, RTX 5090)

  #1528·BobClawblaw opened Oct 28, 2026Contributor·

  ·
  1

• HIP: a measured tuning table for gfx1201, and the step that measures one

  #1527·laudney opened Oct 28, 2026·7 of 9 tasks·

  ·
  1

  2

• Vulkan base: runtime and the Q4 projection registered in the kernel contract (opt-in)

  #1526·joshuaswarren opened Oct 28, 2026·

  ·

• glm cuda: at three ranks or more, TF_GLM_ROCE_MULTI=1 sends the small all-gathers over fabric/tp's multirank RoCE

  #1524·Bizuayeu opened Oct 28, 2026Contributor·9 tasks done·

  ·
  2

• EXL3 Flash Next: host-resident K/V on the tier (6/8)

  #1522·BlivionIaG opened Oct 28, 2026Contributor·

  ·

• Native Vulkan backend for Qwen3.5-2B (opt-in)

  #1517·joshuaswarren opened Oct 28, 2026·

  ·
  4

• fabric/tp: multirank RoCE over a switchless ring, each peer's stripes on the HCAs that reach it

  #1516·Bizuayeu opened Oct 28, 2026Contributor·9 tasks done·

  ·
  1

• lfm3: serve the MLX text decoder on Metal

  #1514·BlivionIaG opened Oct 28, 2026Contributor·

  ·

• lfm3: admit official MLX text and VL checkpoints

  #1513·BlivionIaG opened Oct 28, 2026Contributor·

  ·

• MOSS-Image: the prompt encoder on Metal

  #1512·BobClawblaw opened Oct 28, 2026Contributor·9 tasks done·

  ·
  1

  2

• lfm3: serve LFM3-1.7B NVFP4 on CUDA, weight-only (stacked on #1513, #1492 and #1503)

  #1508·jschmied opened Oct 27, 2026Contributor·

  ·
  4

• qwen3_5 cuda: the DFlash2 drafter reads a dense EXL3 head over its two vocabulary spans

  #1507·diwakergupta opened Oct 27, 2026·9 tasks done·

  ·
  3