Issues

• HIP Flash Next on gfx1151: results and approach of a complete branch, as a reference for the open series

  ashhart/TensorFold#1661 · Filter by author davidtai opened on Oct 30, 2026
  2

• RFC: --profile stable | fast-exact | fast, one table for the opt-in switches

  ashhart/TensorFold#1660 · Filter by author jschmied opened on Oct 30, 2026
  1

• CUDA: one paged prompt store for every family on the core KV pool (Gemma 4, Kolibri-1, Spark2.5 first)

  ashhart/TensorFold#1657 · Filter by author jschmied opened on Oct 30, 2026
  2

• Qwen3.6-35B-A3B on CUDA: serve nvidia/Qwen3.6-35B-A3B-NVFP4 (ModelOpt NVFP4 MoE) as downloaded

  ashhart/TensorFold#1656 · Filter by author laudney opened on Oct 30, 2026
  2

• Nemotron 3.5 Bolt on CUDA (GB10): deterministic "BoltForward" 500s under sustained concurrency at TensorFold v1.1.6 — deterministic repro data from a benchmark harness

  ashhart/TensorFold#1654 · Filter by author DaviddevIDM opened on Oct 30, 2026

• HIP Flash Next on gfx1151: defaults for prompt step and MTP drafts

  ashhart/TensorFold#1649 · Filter by author gilby opened on Oct 30, 2026

• Metal: 100% GPU residency and ~115 W persist indefinitely after one realistic request; co-tenant engines starve; --keep-warm 0 does not disable the loop

  ashhart/TensorFold#1647 · Filter by author astro-15-alive opened on Oct 30, 2026

• Chat-template rendering appears quadratic in message count: a 1,999-message / 765k-token prompt maps ~66 GiB of host memory before any model work

  ashhart/TensorFold#1633 · Filter by author fche-hermes opened on Oct 30, 2026
  2

• LFM3-Omni-2B on Metal and CUDA: speech in, streamed text and speech out (gate and split)

  ashhart/TensorFold#1627 · Filter by author drowzeys opened on Oct 29, 2026
  1

• RFE: timestamp log lines

  ashhart/TensorFold#1626 · Filter by author fche opened on Oct 29, 2026

• DeepSeek-V4.1: prompt cache resumes only at 2,048-token boundaries and ignores --prompt-cache-gib

  ashhart/TensorFold#1625 · Filter by author ashhart opened on Oct 29, 2026

• Qwen3.8-27B on pre-M5 Macs: 32 rows a pass for one stream

  ashhart/TensorFold#1621 · Filter by author ashhart opened on Oct 29, 2026
  2

• Flash Next on HIP: which tensors an oQ checkpoint can narrow for bandwidth-bound decode

  ashhart/TensorFold#1617 · Filter by author gilby opened on Oct 29, 2026

• HIP Flash Next: the prompt's n-gram gather is a single-threaded host copy

  ashhart/TensorFold#1606 · Filter by author gilby opened on Oct 29, 2026

• HIP Flash Next: about 850 small kernels per verify round cost a fifth of decode

  ashhart/TensorFold#1603 · Filter by author gilby opened on Oct 29, 2026

• HIP Flash Next: the first request after load decodes at half speed

  ashhart/TensorFold#1602 · Filter by author gilby opened on Oct 29, 2026

• HIP: a 16/32-row routed-expert prompt tile for many-expert MoE (Flash Next)

  ashhart/TensorFold#1601 · Filter by author gilby opened on Oct 29, 2026

• HIP: the prompt GEMM pays a group-32 scale and bias fold every K stage

  ashhart/TensorFold#1600 · Filter by author gilby opened on Oct 29, 2026

• HIP Flash Next: a shared expert wider than the routed experts cannot load

  ashhart/TensorFold#1599 · Filter by author gilby opened on Oct 29, 2026

• HIP Flash Next: the loader requires the MTP head under language_model.mtp, so oMLX checkpoints fail

  ashhart/TensorFold#1598 · Filter by author gilby opened on Oct 29, 2026

• HIP Flash Next: WMMA sparse attention for prompt chunks on gfx11

  ashhart/TensorFold#1597 · Filter by author gilby opened on Oct 29, 2026
  2

• MCDMA over a ConnectX link, and serve --cluster on real hosts (transport only)

  ashhart/TensorFold#1594 · Filter by author ThinkOffApp opened on Oct 29, 2026

• Living Weights: lessons are refused when the server runs with --no-thinking

  ashhart/TensorFold#1587 · Filter by author ashhart opened on Oct 29, 2026

• glm cuda: load an Apache-2.0 DFlash2 drafter (canada-quant DFlash2-G: 8 full-attention layers, 9 taps)?

  ashhart/TensorFold#1583 · Filter by author gilby opened on Oct 29, 2026
  2

• d1-2B (LFM3-VL decisions) on native CUDA: bf16 text decoder, LFM3-VL tower, set attention

  ashhart/TensorFold#1581 · Filter by author drowzeys opened on Oct 29, 2026
  6