Issues

• qwen27 CUDA: the AOT capture harness never records a multi-stream _shared, so captured sets refuse multi-stream rounds

  ashhart/TensorFold#793 · Filter by author zweck opened on Oct 13, 2026

• tools/zig: one manifest → kernel set step for every family, reading a Recorder dump as well as a manifest

  ashhart/TensorFold#792 · Filter by author zweck opened on Oct 13, 2026

• aot.zig: a captured variant compiled without 16-divisibility refuses an aligned launch

  ashhart/TensorFold#791 · Filter by author zweck opened on Oct 13, 2026

• Qwen3.8-27B + DFlash2 with 4 concurrent sampled requests: native 1.0.8 yields 2.9 tokens/round against 0.6.5's 5.4, so a batch Q&A job takes 46% longer

  ashhart/TensorFold#790 · Filter by author appljd opened on Oct 13, 2026

• New family: MiniCPM-V 4.7 35B-A3B (Qwen3.6 MoE fine-tune + vision tower)

  ashhart/TensorFold#785 · Filter by author jkuepker opened on Oct 13, 2026
  1

• Universal model support, Living Weights with learn/unlearn, consumer GPU support, and Codex CLI compatibility

  ashhart/TensorFold#782 · Filter by author XAEAZii opened on Oct 13, 2026

• Proposal: streams take memory from one core pool as they grow, instead of a full window each at load

  ashhart/TensorFold#781 · Filter by author jschmied opened on Oct 13, 2026

• How to show prefill percentage ?

  ashhart/TensorFold#772 · Filter by author nps798 opened on Oct 13, 2026

• Benchmark log: Qwen3.8-Flash-Next NVFP4 on one DGX Spark (vs llama.cpp, three nights) — and GLM-5.3-Flash MLX 4-bit stream concurrency on an M3 Ultra — numbers, method, and two questions

  ashhart/TensorFold#754 · Filter by author ThefloorMiner opened on Oct 12, 2026

• GET /health has four different shapes across servers, backends and auth settings — could the schema be pinned in the docs?

  ashhart/TensorFold#753 · Filter by author ThefloorMiner opened on Oct 12, 2026

• Feature request: a llama.cpp-compatible GET /slots (per-stream state) — /health says how many wait, not why

  ashhart/TensorFold#752 · Filter by author ThefloorMiner opened on Oct 12, 2026

• Native server: which MLX checkpoints do the published binaries actually serve? (1.0.2/1.0.3 refuse the cards Python 0.6.6 serves — zig-preview GLM family not in any release)

  ashhart/TensorFold#751 · Filter by author ThefloorMiner opened on Oct 12, 2026

• CUDA on one DGX Spark (GB10, 128 GB): Qwen3.8-Flash-Next NVFP4 with --parallel 4 --context 131072 — host swaps 22 GB from 3 concurrent streams although the startup estimate was 81 GiB

  ashhart/TensorFold#750 · Filter by author ThefloorMiner opened on Oct 12, 2026

• Python server: default --max-tokens 4096 counts thinking tokens — "unlimited" thinking ends with finish_reason=length and an empty reply

  ashhart/TensorFold#749 · Filter by author ThefloorMiner opened on Oct 12, 2026

• Metal (MLX): loading 180 GB of weights from an SMB share stalls mx.eval past the Metal watchdog — GPU Timeout; the same files load from local SSD in under a minute

  ashhart/TensorFold#748 · Filter by author ThefloorMiner opened on Oct 12, 2026

• Proposal: restore single-Mac GLM-5.3-Flash on the M3 Ultra

  ashhart/TensorFold#743 · Filter by author harrisonfriia opened on Oct 12, 2026
  1

• Proposal: MiMo-V2.6-Flash on the CUDA native engine across two GPUs (EXL3, DFlash)

  ashhart/TensorFold#725 · Filter by author ajensenwaud opened on Oct 12, 2026
  1

• Support oQ/oQe mixed-precision quants in the Zig engine

  ashhart/TensorFold#716 · Filter by author grantoverton opened on Oct 12, 2026
  2

• 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

• 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
  2

• 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
  8

• 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
  3