Issues

• flashnext: CUDA copy windows stop at 16 rows

  ashhart/TensorFold#1251 · Filter by author outcastofmusic opened on Oct 24, 2026
  1

• Ling-3.0-tiny on a Qualcomm board: a Ling family on Metal first, then an ARM CPU backend

  ashhart/TensorFold#1247 · Filter by author ThinkOffApp opened on Oct 24, 2026

• server: LFM2.5-24B-A2B tool calls arrive as message content

  ashhart/TensorFold#1246 · Filter by author outcastofmusic opened on Oct 24, 2026
  1

• Qwen4-27B (dense) on native CUDA, GB10: shape parameterization ahead of the checkpoint (RFC)

  ashhart/TensorFold#1244 · Filter by author jschmied opened on Oct 24, 2026
  1

• server: a tool-call opener named in prose hides the rest of the reply

  ashhart/TensorFold#1242 · Filter by author outcastofmusic opened on Oct 24, 2026
  1

• Non-streaming reply puts <tool_call> markup in content when max_tokens cuts a call (streaming returns the call)

  ashhart/TensorFold#1238 · Filter by author laidick opened on Oct 24, 2026
  2

• Question: should one tensor-parallel group mix Metal and CUDA ranks?

  ashhart/TensorFold#1236 · Filter by author ThinkOffApp opened on Oct 24, 2026
  2

• Nemotron-Labs Dory 1.0 4B on native CUDA: a looped nGPT family (dory)

  ashhart/TensorFold#1233 · Filter by author jschmied opened on Oct 24, 2026
  1

• Retain completed prefill progress when a CUDA request is cancelled

  ashhart/TensorFold#1225 · Filter by author Sapphire-Rapids opened on Oct 24, 2026
  8

• Gemma 4 E4B (NVIDIA Nemotron 3.5 Bolt E4B): per-layer inputs and shared KV, Intel Arc first

  ashhart/TensorFold#1222 · Filter by author Mirrdhyn opened on Oct 23, 2026
  2

• RoCE transport above two ranks: roce.zig:148 refuses every world but 2, so four GB10s fall back to NCCL (rebased 2..16-rank branch, gate question)

  ashhart/TensorFold#1209 · Filter by author flashosophy opened on Oct 23, 2026

• Studio: carry #941 to #944 onto main as reviewed increments

  ashhart/TensorFold#1204 · Filter by author ashhart opened on Oct 22, 2026
  1

• Proposal: Qwen3.8-27B follow-up turns resume from slot 0 (Metal, small-memory Macs)

  ashhart/TensorFold#1197 · Filter by author RS91Dev opened on Oct 22, 2026
  1

• Decisions serving for qwen27 (Ornith-1.5-122B): score/decide passes, JointHead scoring, a free decision slot, GPU class profiling

  ashhart/TensorFold#1190 · Filter by author grantoverton opened on Oct 22, 2026

• New family: Recursive Flint (40B-A2.5B MoE) on Metal

  ashhart/TensorFold#1175 · Filter by author BobClawblaw opened on Oct 21, 2026

• CUDA: Qwen3.5/3.6 attention caches on the core KV pool (#781 follow-up)

  ashhart/TensorFold#1156 · Filter by author jschmied opened on Oct 21, 2026
  5

• Qwen3.5-MoE on CUDA: load ExLlamaV3 (EXL3) checkpoints

  ashhart/TensorFold#1149 · Filter by author MiaAI-Lab opened on Oct 21, 2026

• Proposal: Qwen3.8-27B ModelOpt NVFP4 checkpoints on the native CUDA engine (both precision modes)

  ashhart/TensorFold#1144 · Filter by author jschmied opened on Oct 21, 2026

• Proposal: GLM-5.3-Flash on CUDA takes its streams' KV pages from the core pool, and kept prompt states copy only their rows

  ashhart/TensorFold#1142 · Filter by author Bizuayeu opened on Oct 21, 2026

• Proposal: dense BF16 Qwen3.5 CUDA scoring

  ashhart/TensorFold#1136 · Filter by author akol1 opened on Oct 21, 2026

• CUDA kernel contract on nvcc 13.0: 656 pending launches have no 13.0 records, so kernel-contract and kernel-contract-limits fail on plain v1.1.1

  ashhart/TensorFold#1105 · Filter by author Bizuayeu opened on Oct 20, 2026
  1

• HIP/Nemotron contribution stack: review map and validation

  ashhart/TensorFold#1081 · Filter by author BlivionIaG opened on Oct 19, 2026
  1

• Flash Next 4-bit doesn't fit natively on a 128 GB Mac at default limits. Any plans for --ple-on-ssd on Metal?

  ashhart/TensorFold#1039 · Filter by author hudsonwa opened on Oct 18, 2026
  3

• Proposal: several models on one Mac under one memory budget

  ashhart/TensorFold#1027 · Filter by author cwschroeder opened on Oct 18, 2026
  1

• NVIDIA's GLM-5.3-Flash-FP4-Local on the native CUDA engine (tracker)

  ashhart/TensorFold#1015 · Filter by author Bizuayeu opened on Oct 17, 2026
  10