Issues

• Benchmark result 1.17 On MacBook Pro - M5 Pro Chip 48GB ram

  ashhart/TensorFold#1515 · Filter by author Deesha08 opened on Oct 28, 2026

• KV pool: host-RAM tier

  ashhart/TensorFold#1511 · Filter by author BlivionIaG opened on Oct 28, 2026
  2

• GLM CUDA: the KV page pool crowds out the prompt cache, so long chats prefill again every turn

  ashhart/TensorFold#1510 · Filter by author Bizuayeu opened on Oct 27, 2026

• HIP: a measured tuning table for gfx1201 (RDNA4)

  ashhart/TensorFold#1506 · Filter by author laudney opened on Oct 27, 2026
  2

• moss_image: the FLUX.2 image decoder as a native Metal forward

  ashhart/TensorFold#1504 · Filter by author BobClawblaw opened on Oct 27, 2026
  1

• GLM-5.3-Flash on CUDA: the KDA input projection falls off glm_mm at TP=3, and padding it is worth +24% at one stream

  ashhart/TensorFold#1497 · Filter by author BadAd84 opened on Oct 27, 2026
  3

• Add native Metal support for LFM3-1.7B and LFM3-VL-2B MLX 4/6/8-bit checkpoints

  ashhart/TensorFold#1491 · Filter by author BlivionIaG opened on Oct 27, 2026
  8

• GLM-5.3-Flash on CUDA: capture decode and verify rounds as CUDA graphs (one rank and TP=2)

  ashhart/TensorFold#1490 · Filter by author ashhart opened on Oct 27, 2026
  1

• Nemotron 3.5 Lightning CUDA prefill: one row departs from float64 after a near-tie routing pick in layer 6

  ashhart/TensorFold#1489 · Filter by author jschmied opened on Oct 27, 2026
  1

• LFM2.5-24B-A2B on CUDA: plain serve on top of #1377

  ashhart/TensorFold#1488 · Filter by author jschmied opened on Oct 27, 2026
  1

• Flash Next on Metal: wide-window dense projections, four rows a simdgroup and on the matrix units, bit-identical per row

  ashhart/TensorFold#1487 · Filter by author ashhart opened on Oct 27, 2026
  2

• Flash Next on Metal: a grouped gate/up that decodes an expert's weights once for two verified lanes

  ashhart/TensorFold#1486 · Filter by author ashhart opened on Oct 27, 2026
  2

• Flash Next: a feature dump for drafter work, and an offline acceptance tool

  ashhart/TensorFold#1485 · Filter by author ashhart opened on Oct 27, 2026
  2

• Flash Next: a shared DFlash2 drafter runtime, with Flash Next as a target (Metal first)

  ashhart/TensorFold#1484 · Filter by author ashhart opened on Oct 27, 2026
  2

• Benchmarks: expose why a complete v1 attached-native receipt remains pending and protocol_complete=false

  ashhart/TensorFold#1481 · Filter by author Maxrio9 opened on Oct 27, 2026

• RFC: fact_swiglu reads each mma A quad in one load (bit-equal, -8 % at 8k rows on GB10, flat at served sizes): call for measurements on other GPUs

  ashhart/TensorFold#1479 · Filter by author jschmied opened on Oct 27, 2026
  3

• Qwen3.8-27B on the M5 Ultra: bf16 prompts at 8k-32k run behind mlx-lm on the same checkpoint

  ashhart/TensorFold#1478 · Filter by author ashhart opened on Oct 27, 2026
  3

• Flash Next on Metal: --ple-on-ssd reads each forward's n-gram rows one at a time, 3.7 ms of GPU-idle time per decoded token

  ashhart/TensorFold#1461 · Filter by author paragontasx opened on Oct 27, 2026

• moss_image: the Qwen3.5 prompt encoder as a native Metal forward

  ashhart/TensorFold#1459 · Filter by author BobClawblaw opened on Oct 27, 2026
  2

• Qwen3.8-27B CUDA: whole-window prefill about 0.8% slower at 32k-64k since #1373

  ashhart/TensorFold#1458 · Filter by author ashhart opened on Oct 27, 2026
  1

• Lanes: grammar mask fills for many constrained streams run serially on the lane thread

  ashhart/TensorFold#1455 · Filter by author ashhart opened on Oct 27, 2026
  2

• Qwen3.8-27B NVFP4 (ModelOpt) fails with MissingLanePart on Jetson AGX Orin (sm_87)

  ashhart/TensorFold#1452 · Filter by author valacroix opened on Oct 27, 2026
  2

• App: integrate reviewed workflow improvements and engine 1.1.6

  ashhart/TensorFold#1450 · Filter by author ashhart opened on Oct 27, 2026
  5

• RFC: qualify nvcc 13.2 for the release fatbins (native FP4 decode, prompt16 −4.5 %, bit-equal)

  ashhart/TensorFold#1448 · Filter by author jschmied opened on Oct 27, 2026
  2

• GLM CUDA: the latent split across ranks (#1313 step 2), with a working reference at TP=2

  ashhart/TensorFold#1431 · Filter by author wojo opened on Oct 27, 2026
  1