Issues

• HIP: prompt attention's chunk-start test fails on gfx1151 with the wide 64-row tile

  ashhart/TensorFold#2317 · Filter by author ashhart opened on Nov 6, 2026
  1

• Finished tool call before a later </think> stays in reasoning_content

  ashhart/TensorFold#2310 · Filter by author AdrianBinDC opened on Nov 6, 2026
  2

• GLM CUDA: kept states under TF_KV_SHARE sit outside the startup plan

  ashhart/TensorFold#2307 · Filter by author Bizuayeu opened on Nov 6, 2026
  2

• kernel contract: qwen27 decode control edits text that 9be323a0b changed, so M5 runs stop with ControlEdit

  ashhart/TensorFold#2301 · Filter by author BobClawblaw opened on Nov 6, 2026

• Ripple-65B (ricursive) on CUDA: the publisher's NVFP4 release

  ashhart/TensorFold#2300 · Filter by author jrb-ai opened on Nov 6, 2026

• deepseek_v41: three ranks on three GB10s (TP3), on the shared multi-rank layer

  ashhart/TensorFold#2270 · Filter by author Defilan opened on Nov 5, 2026
  7

• nvfp4 experts cuda: decode kernels stall on load latency; a cp.async ring makes GLM-5.3 decode up to 24% faster, same bits

  ashhart/TensorFold#2265 · Filter by author eleqtrizit opened on Nov 5, 2026

• Qwen3 embedding and reranker checkpoints (BF16) on the native CUDA engine

  ashhart/TensorFold#2261 · Filter by author Arminova opened on Nov 5, 2026
  1

• Ripple-65B (ricursive) on Metal: the publisher's DFlash2 drafter

  ashhart/TensorFold#2259 · Filter by author BobClawblaw opened on Nov 5, 2026
  1

• metal: unify duplicated kernels (attention partials, norms, gemm_nax wrappers, glue, RoPE)

  ashhart/TensorFold#2251 · Filter by author jschmied opened on Nov 5, 2026
  1

• CI: zig and zig-windows spend most of their time rerunning tests that can't have changed

  ashhart/TensorFold#2242 · Filter by author ashhart opened on Nov 5, 2026
  3

• Parakeet-TDT-0.6B-v3 as a second ASR family on the transcription route

  ashhart/TensorFold#2235 · Filter by author devopsfortunato opened on Nov 5, 2026
  2

• Test the computer-use key table and request parser on Linux

  ashhart/TensorFold#2217 · Filter by author AdrianBinDC opened on Nov 5, 2026

• Kimi K3.1 Pro on the native engine

  ashhart/TensorFold#2192 · Filter by author ashhart opened on Nov 4, 2026
  2

• GLM on CUDA: a kept prompt state carries the dead half of the double-buffered KDA state (71 MB a rank)

  ashhart/TensorFold#2180 · Filter by author volfco opened on Nov 4, 2026
  4

• Ripple-65B (ricursive) on native Metal: the publisher's BF16 release

  ashhart/TensorFold#2171 · Filter by author BobClawblaw opened on Nov 4, 2026
  1

• DeepSeek-V4.1: our prompts differ from DeepSeek's encoder on tool schemas, earlier reasoning and reminders

  ashhart/TensorFold#2141 · Filter by author ashhart opened on Nov 4, 2026

• Capabilities, /v1/models and real behaviour agree; bounded phase telemetry

  ashhart/TensorFold#2139 · Filter by author ashhart opened on Nov 4, 2026

• One reason code for every wait or refusal, and an audit of queue limits

  ashhart/TensorFold#2135 · Filter by author ashhart opened on Nov 4, 2026
  4

• Audit each family's persistence keys

  ashhart/TensorFold#2131 · Filter by author ashhart opened on Nov 4, 2026

• Memory pressure on Linux: a reading behind the same levels as macOS

  ashhart/TensorFold#2129 · Filter by author ashhart opened on Nov 4, 2026

• RFC: one host package per family, backends bind the kernels (Drama 3's codec first)

  ashhart/TensorFold#2118 · Filter by author jschmied opened on Nov 4, 2026
  1

• Use TensorFold from Hermes Agent: a provider plugin with model selection and server stats

  ashhart/TensorFold#2109 · Filter by author ashhart opened on Nov 4, 2026
  2

• Qwen3-ASR (0.6B and 1.7B) as a native CUDA family on the audio request path

  ashhart/TensorFold#2105 · Filter by author shuuul opened on Nov 4, 2026
  6

• CUDA on GB10: a synchronous DtoH copy into pageable memory can miss pages (driver 580.178.04)

  ashhart/TensorFold#2102 · Filter by author shuuul opened on Nov 4, 2026
  1