Issues

• Proposal: native Metal support for Flash Next's 4-bit/group-32 checkpoint

  ashhart/TensorFold#706 · Filter by author cwschroeder opened on Oct 11, 2026

• Native Flash Next engine ignores thinking_budget (chat layer sets think_budget; flashnext_host.zig never passes it to generateFrom)

  ashhart/TensorFold#701 · Filter by author dinadur opened on Oct 11, 2026

• Proposal: Clef decision models on the native engine (/v1/systemone and joint-schema scoring for /v1/decisions)

  ashhart/TensorFold#692 · Filter by author cwschroeder opened on Oct 11, 2026

• Proposal: Living Weights: accept caller-supplied questions and answers

  ashhart/TensorFold#687 · Filter by author philip-pentatonic opened on Oct 11, 2026

• Proposal: experimental DeepSeek V4.1 Flash resident Metal backend (qualification blocked)

  ashhart/TensorFold#685 · Filter by author akol1 opened on Oct 11, 2026

• qwen3_5_moe: the CUDA 4-bit/group-64 check gates Metal serving, refusing 8-bit checkpoints the Mac kernels read

  ashhart/TensorFold#683 · Filter by author Moutonc opened on Oct 11, 2026
  1

• Living Weights: conflicting facts, updates and forgetting (measurements + proposal)

  ashhart/TensorFold#678 · Filter by author ThinkOffApp opened on Oct 11, 2026
  2

• Proposal: model-free calibrated Metal GPU counter probe

  ashhart/TensorFold#676 · Filter by author akol1 opened on Oct 11, 2026

• Proposal: standalone Metal EXL3 projection operator and qualification

  ashhart/TensorFold#677 · Filter by author akol1 opened on Oct 11, 2026

• Cooperative b4g32 lane projection fails its FP64 check on M5 Max

  ashhart/TensorFold#675 · Filter by author cwschroeder opened on Oct 11, 2026

• CUDA Nemotron: the prompt path is chunk-invariant, so the cache's 2,048 grid could go

  ashhart/TensorFold#670 · Filter by author jschmied opened on Oct 11, 2026

• Proposal: tensorfold merge, fold a LoRA into a quantized checkpoint, leaving every other byte identical

  ashhart/TensorFold#666 · Filter by author philip-pentatonic opened on Oct 11, 2026
  2

• Two-rank watchdog deadlines are asymmetric: a legal >300 s prefill kills both ranks (follower waits a fixed TF_TP_TIMEOUT_S while the leader's allowance scales with prompt length)

  ashhart/TensorFold#665 · Filter by author Jenpo opened on Oct 11, 2026
  1

• Vision on the MoE side: qwen3_5_moe (Qwen3.6-35B-A3B) gated out of --vision — the dense path ports over in ~70 lines

  ashhart/TensorFold#652 · Filter by author hennas-waifson opened on Oct 11, 2026

• Living Weights: lessons that pass recall (recalled: true) still overwrite real facts that share the question template

  ashhart/TensorFold#646 · Filter by author jimthedj65 opened on Oct 11, 2026
  2

• Qwen3.8-27B on native 1.0.5: M6 (GPU generation 18) refused, and memory is too large for 32 GB

  ashhart/TensorFold#643 · Filter by author RS91Dev opened on Oct 11, 2026
  3

• Deepseek Support for v4 flash 0731 and v4 flash vision

  ashhart/TensorFold#628 · Filter by author dschmitz opened on Oct 10, 2026
  1

• GB10: Nemotron decode is 1.4 % (code) / 3.0 % (prose) slower on 1.0.3 than on 1.0.2; on 1.0.5 the prose gap is still there

  ashhart/TensorFold#627 · Filter by author Nipale-ai opened on Oct 10, 2026
  3

• RFC: learned drafters for any family: one block drafter in the core, a recorder, and a recipe

  ashhart/TensorFold#620 · Filter by author jschmied opened on Oct 10, 2026

• Port Qwen3.8-27B (EXL3 3.00bpw + DFlash2) to the native CUDA engine on Ampere sm_86 (RTX 3090) — proposal and gate

  ashhart/TensorFold#603 · Filter by author briepala opened on Oct 10, 2026
  1

• GLM-5.3-Flash and full GLM-5.3 on the native CUDA engine (tracker)

  ashhart/TensorFold#593 · Filter by author ashhart opened on Oct 10, 2026
  6

• Native image and video inputs in the server (Python frontend's preparation, data and HTTPS URLs), before any family consumes them

  ashhart/TensorFold#565 · Filter by author BobClawblaw opened on Oct 9, 2026
  3

• DeepSeek-V4.1-Flash on the Zig engine (CUDA, 2 and 4 GB10): our port is yours to use, plus test time

  ashhart/TensorFold#559 · Filter by author bertholomus opened on Oct 9, 2026
  18

• Zig CUDA ports: shared layers below lanes.Backend, and one place to coordinate them

  ashhart/TensorFold#548 · Filter by author jschmied opened on Oct 9, 2026
  14

• v0.6.6 ：Prefix caching is permanently broken if the context contains images

  ashhart/TensorFold#510 · Filter by author senchacha opened on Oct 8, 2026
  1