Issues

• Flash Next FP4-Local on one GB10: measured gains over 1.1.5 we can offer as PRs

  ashhart/TensorFold#1367 · Filter by author Arminova opened on Oct 26, 2026

• Streamed chat: a call the whole reply keeps as text never reaches the client

  ashhart/TensorFold#1365 · Filter by author plotarmordev opened on Oct 26, 2026
  2

• Cloudflare Clef decisions on native CUDA: /v1/systemone through Clef's joint schema head (Clef-Flash 9B passes)

  ashhart/TensorFold#1360 · Filter by author gilby opened on Oct 26, 2026
  1

• CUDA/Qwen3.8: share W4A8 prompt quantization across projections (proposal and gate)

  ashhart/TensorFold#1348 · Filter by author yguanrong opened on Oct 26, 2026
  1

• Qwen3.8-27B CUDA: with TF_QWEN_KV_PAGES a batch takes the partial-only attention path, so its replies differ from the same replies alone (1.1.5, RTX 5090)

  ashhart/TensorFold#1344 · Filter by author BobClawblaw opened on Oct 26, 2026

• Qwen 27B on native CUDA: a chat's next turn never resumes at the previous turn, only on the 4,096-row grid

  ashhart/TensorFold#1334 · Filter by author gilby opened on Oct 26, 2026
  1

• Flash Next on Metal: shared rounds spend 7–17 ms outside the target forward at 4–9 streams

  ashhart/TensorFold#1333 · Filter by author gilby opened on Oct 26, 2026
  2

• Sage-A1.5 (qwen3_5 MoE 35B-A3B) on the XPU backend: decode, serve, vision and MTP

  ashhart/TensorFold#1329 · Filter by author teamlewis-bot opened on Oct 26, 2026

• xpu: EXL3 trellis weights on Intel GPUs (Xe2): kernels, Level Zero host, checks

  ashhart/TensorFold#1328 · Filter by author teamlewis-bot opened on Oct 26, 2026

• DeepSeek-V4.1-Flash at world 4 on four GB10s: main's deepseek_v41 boots, serves and keeps its equalities; tool calls leak, and it runs ~2× slower than a fork on the same weights

  ashhart/TensorFold#1324 · Filter by author ecohash-co opened on Oct 25, 2026

• GLM-5.3-Flash on CUDA cannot keep conversation states on disk: an 18.7k-token conversation evicted by one 60k filler recomputes its next turn in 15.6 s (0.49 s resumed), and 15.4 s after a restart

  ashhart/TensorFold#1312 · Filter by author wojo opened on Oct 25, 2026
  3

• Metal: KV pages from the core pool

  ashhart/TensorFold#1293 · Filter by author jschmied opened on Oct 25, 2026
  1

• Canary results on DGX Spark (GB10): memory ceiling forces --parallel 8 → 3, mixed latency/throughput vs vLLM

  ashhart/TensorFold#1292 · Filter by author fossouo opened on Oct 25, 2026
  1

• GLM-5.3-Flash at TP=3: its small all-gathers stay on NCCL, and fabric/tp's three-to-eight-rank RoCE (#1209) refuses a switchless ring with BadRecord. Which route should GLM take, and would per-peer devices be in scope?

  ashhart/TensorFold#1288 · Filter by author Bizuayeu opened on Oct 25, 2026
  1

• Nemotron CUDA copy drafts follow the latest match, often the wrong line of a list

  ashhart/TensorFold#1282 · Filter by author outcastofmusic opened on Oct 25, 2026
  2

• New model family: MiniMax-M3.1-Flash VL (minimax_m3_vl) EXL3, TP2-served

  ashhart/TensorFold#1280 · Filter by author styles01 opened on Oct 25, 2026
  1

• Drama 3: about 1 in 5 inputs stall after the first sentence unless the silence guard and redraws are on

  ashhart/TensorFold#1275 · Filter by author drowzeys opened on Oct 25, 2026
  1

• cosmos3: Cosmos3-Edge reasoner (image + text) as a native family on Metal and CUDA

  ashhart/TensorFold#1269 · Filter by author drowzeys opened on Oct 25, 2026
  1

• App: release beta.10 with TensorFold 1.1.4

  ashhart/TensorFold#1253 · Filter by author ashhart opened on Oct 24, 2026
  5

• 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
  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
  3

• Tensor parallel across Metal and CUDA ranks (Mac + NVIDIA): tracker

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

• 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
  10

• 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