Issues

• CUDA on integrated GPUs (GB10): keep prompt states and weights in place instead of copying them through pinned host memory

  ashhart/TensorFold#2043 · Filter by author jschmied opened on Nov 3, 2026

• CI never reaches 246 of the 393 named test steps, and nine of them fail on main

  ashhart/TensorFold#2031 · Filter by author jishnuvenugopal opened on Nov 3, 2026
  1

• RFC: a layer-range (pipeline) layout across mixed nodes: one hidden row per token over Ethernet, GB10 and Mac in one model

  ashhart/TensorFold#2005 · Filter by author jschmied opened on Nov 3, 2026
  7

• CUDA: the decode-window L2 prefetch from #1580 on Qwen 3.5/3.6 and Nemotron

  ashhart/TensorFold#2003 · Filter by author ashhart opened on Nov 3, 2026
  1

• Agree the gate for Drama 3 model-maximum native speech context on Metal and CUDA

  ashhart/TensorFold#2002 · Filter by author akol1 opened on Nov 3, 2026
  2

• GLM CUDA at TP=3: under TF_GLM_CHUNK_SOLO=4096, TF_GLM_ROCE_MULTI costs a cold 128K prompt about 7.5%, not 1 to 2%

  ashhart/TensorFold#1994 · Filter by author Bizuayeu opened on Nov 3, 2026
  1

• Qwen3-8B embedding and reranker checkpoints (BF16) on native HIP

  ashhart/TensorFold#1977 · Filter by author laudney opened on Nov 3, 2026
  3

• Bug: Cannot change the model away from planning-only mode

  ashhart/TensorFold#1976 · Filter by author abandaru opened on Nov 3, 2026
  3

• Streamed chat: GLM and DSML tool calls arrive in one piece, only XML calls stream their arguments

  ashhart/TensorFold#1975 · Filter by author nuk3s opened on Nov 3, 2026
  3

• roce_multi on a switchless ring of four: start-up warm-up and the relay

  ashhart/TensorFold#1971 · Filter by author ashhart opened on Nov 2, 2026
  1

• Living Weights: locate-then-place, shrink-after-keep, corrections and a pocket audit (ideas from weight-space programming)

  ashhart/TensorFold#1968 · Filter by author ashhart opened on Nov 2, 2026

• Qwen3.5 BF16 on HIP: the publisher's checkpoint as a decision backbone (Clef-Flash), prompt rows through rocBLAS

  ashhart/TensorFold#1952 · Filter by author laudney opened on Nov 2, 2026
  3

• Proposal: benchmark receipts record the Mac's product family (Mac mini, MacBook Pro, Mac Studio), not only the chip

  ashhart/TensorFold#1950 · Filter by author pauldodd123 opened on Nov 2, 2026
  3

• Flash Next EXL3: elastic device memory between expert lines and K/V, graphs kept (opt-in)

  ashhart/TensorFold#1948 · Filter by author BlivionIaG opened on Nov 2, 2026
  4

• Proposal: one KV admission ledger for unpaged streams: a request reserves its prompt and reply, not a whole window per --parallel stream (#781 follow-up)

  ashhart/TensorFold#1943 · Filter by author laudney opened on Nov 2, 2026
  8

• Native OCR models on HIP: dots.mocr, GLM-OCR, LightOnOCR-3 and Nemotron Parse, CUDA to follow (tracker)

  ashhart/TensorFold#1927 · Filter by author laudney opened on Nov 2, 2026
  2

• Native GGUF on Apple Silicon: GGUF blocks loaded exactly onto the Metal affine kernels (tracker)

  enhancement
  ashhart/TensorFold#1920 · Filter by author ashhart opened on Nov 2, 2026
  1

• glm 0.6.0 (Python, TP=2): a returning conversation re-prefills from 0 after pool eviction, and the released image's --spill-gib never spills

  ashhart/TensorFold#1912 · Filter by author ss-cong opened on Nov 2, 2026
  1

• glm: the released Python engine renders the --no-thinking prompt 6-7 tokens shorter than the checkpoint template; native renders it verbatim

  ashhart/TensorFold#1911 · Filter by author ss-cong opened on Nov 2, 2026
  3

• glm cuda (native): copy drafts are missing, and the 48% acceptance the Python lane gets from them is what keeps it ahead

  ashhart/TensorFold#1910 · Filter by author ss-cong opened on Nov 2, 2026
  1

• Prompt cache: bounded intermediate checkpoints for long histories that are later compacted

  ashhart/TensorFold#1909 · Filter by author eliteworkstation94-ai opened on Nov 2, 2026
  6

• CUDA (TP=2, GB10): cached_tokens_total grows by ~2x the prompt for 337k-370k-token prompts, exceeding prompt_tokens_total

  ashhart/TensorFold#1899 · Filter by author cluncc opened on Nov 2, 2026
  2

• Proposal: POST /v1/rerank, and Qwen3-Reranker-0.6B scoring on Metal

  ashhart/TensorFold#1895 · Filter by author grantoverton opened on Nov 2, 2026
  1

• Flash Next on Metal: structured output and required/named tool_choice

  ashhart/TensorFold#1894 · Filter by author dinadur opened on Nov 2, 2026
  3

• Flash Next on Metal: --parallel 2+ still loses prompt-cache reuse (shared-session repair held since 1.1.0)

  ashhart/TensorFold#1893 · Filter by author dinadur opened on Nov 2, 2026
  2