All pull requests

Pull requests

• glm cuda: TF_GLM_L2PF reads a decode window's next weights into L2 on a side stream, on by default on a GB10

  #1580·Bizuayeu opened Oct 29, 2026Contributor·9 tasks done·

  ·

• kernel contract: the CUDA runner maps fp8_experts' image, so a run reaches its summary

  #1579·Bizuayeu opened Oct 29, 2026Contributor·9 tasks done·

  ·

• glm cuda: a prompt chunk's MTP pass runs its last row past the caches

  #1577·gilby opened Oct 29, 2026Contributor·8 of 9 tasks·

  ·
  1

• glm cuda: prompt chunks of 4,096 rows, and the DFlash2 context buffers sized to match

  #1576·gilby opened Oct 29, 2026Contributor·8 of 9 tasks·

  ·
  1

• cortheon: keep the task goal cut on a character boundary

  #1575·ppardus opened Oct 29, 2026Contributor·

  ·

• glm cuda: a logit gather broadcasts one block a rank, so TP=3 prompt_logprobs past ~100 tokens no longer hangs

  #1571·gilby opened Oct 28, 2026Contributor·8 of 9 tasks·

  ·
  1

• cuda: attn_train and train_attn compile one gated attention backward, attn_back.cuh, SASS unchanged

  #1565·jschmied opened Oct 28, 2026Contributor·

  ·

• cuda: glm_mm takes common/qmm_frag.cuh, and its 117-line copy glm_qmm_frag.cuh goes, SASS unchanged

  #1564·jschmied opened Oct 28, 2026Contributor·

  ·

• Qwen3.6 MTP drafter on CUDA: draft a shared round's streams together

  #1562·laudney opened Oct 28, 2026Contributor·

  ·
  1

• cuda: sigmoidf_, softplusf_, silu and silu_grad in common/act.cuh, SASS unchanged

  #1556·jschmied opened Oct 28, 2026Contributor·

  ·

• cuda: fp4pair, fp8pair and e4m3f in common/fp_pairs.cuh, SASS unchanged

  #1555·jschmied opened Oct 28, 2026Contributor·

  ·

• docs: Flash Next CUDA recipe, what TF_FLASHNEXT_CHUNK trades: faster prompts against longer decode waits

  #1553·jschmied opened Oct 28, 2026Contributor·

  ·

• 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·

  ·
  3

• 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·

  ·
  2

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

  #1537·jschmied opened Oct 28, 2026Contributor·

  ·

• 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