Issues

• New model family: plain Llama (text-only) on MLX

  ashhart/TensorFold#429 · Filter by author mikolaj92 opened on Oct 5, 2026

• Proposal: Fast startup by repacking model at import/initial launch

  ashhart/TensorFold#428 · Filter by author pmeenan opened on Oct 5, 2026

• Proposal: native Zig HIP backend and Qwen27B GGUF qualification on gfx1151

  ashhart/TensorFold#426 · Filter by author ThinkOffApp opened on Oct 5, 2026

• zig-preview CUDA: the box scripts need tools/triton_aot_manifest.py, which isn't in the branch

  ashhart/TensorFold#424 · Filter by author paragontasx opened on Oct 5, 2026

• GLM CUDA engine: _take_over clones every kept snapshot it is about to drop, so reserved memory spikes by about 15 GiB per rank when a new conversation follows a long one

  ashhart/TensorFold#420 · Filter by author m-naoki-m opened on Oct 5, 2026

• Flash Next CUDA two-rank decode: split_k's target is prompt-shaped, and the shapes a decode actually issues

  ashhart/TensorFold#417 · Filter by author greatyingzi opened on Oct 5, 2026

• CUDA Flash Next: one image in an agent chat makes every later turn re-prefill the whole conversation (~190k tokens, cached_tokens=0)

  ashhart/TensorFold#414 · Filter by author lijian1999 opened on Oct 5, 2026

• ROCm on RDNA4 (gfx1201, Radeon AI PRO R9700): a proposed stack on top of #144

  ashhart/TensorFold#412 · Filter by author millaguie opened on Oct 5, 2026
  1

• Proposal: optional Mojo kernels for the ROCm decode path (R9700: 6.6–10.5% faster decode, same bits)

  ashhart/TensorFold#406 · Filter by author jkuepker opened on Oct 5, 2026
  2

• Full GLM-5.3 (glm_moe_dsa) on three DGX Sparks, as layers of the GLM-5.3-Flash CUDA engine: how should it sit beside #159?

  ashhart/TensorFold#402 · Filter by author MiaAI-Lab opened on Oct 4, 2026

• GLM on CUDA: an opt-in FP8 or 4-bit latent cache (--kv-dtype fp8 / fp4), lossy, with paired quality results on full GLM-5.3

  ashhart/TensorFold#401 · Filter by author MiaAI-Lab opened on Oct 4, 2026
  1

• zig-preview: server fails to start on macOS 27 (Metal 4.1 rejects the mma_16x32 cooperative tensor in the _nax prefill kernels)

  ashhart/TensorFold#394 · Filter by author ThinkOffApp opened on Oct 4, 2026

• Mac / GLM-5.3 (0.6.5): screenshot turns are re-read from scratch, and stream errors arrive after the role chunk

  ashhart/TensorFold#388 · Filter by author Gogo6969 opened on Oct 4, 2026

• feat(mlx): support quantized KV cache (--kv-dtype int8/int4) on Apple Silicon

  ashhart/TensorFold#381 · Filter by author Benjamin-Wegener opened on Oct 4, 2026
  1

• CUDA glm5_next (TP / --parallel): kept prompt states never use the conversation spill tier (--spill-gib / --snapshot-dir are MLX-only and silently ignored)

  ashhart/TensorFold#373 · Filter by author vinicius-symetrix opened on Oct 4, 2026
  1

• Feature request: batch SSD n-gram gathers across concurrent Flash Next requests

  ashhart/TensorFold#368 · Filter by author optimisme opened on Oct 4, 2026
  4

• feat(vision): enable --vision for qwen3_5_moe (Qwen3.6-35B-A3B) on CUDA

  ashhart/TensorFold#366 · Filter by author 0mao0 opened on Oct 4, 2026

• GLM-5.3-Flash on CUDA: let a long prefill wait between chunks while any rank runs hot

  ashhart/TensorFold#339 · Filter by author Bizuayeu opened on Oct 4, 2026

• Flash Next CUDA --parallel: a growing conversation recaptures the lone-stream graphs at every context bucket and slot resize (single-stream decode −25…−60 % on those turns)

  ashhart/TensorFold#336 · Filter by author grearjake-star opened on Oct 4, 2026
  1

• Mac / GLM-5.3 on 0.6.4: a long prompt still waits (requests_waiting 1) until a running stream ends (follow-up to #228)

  ashhart/TensorFold#325 · Filter by author cwschroeder opened on Oct 3, 2026
  3

• Gemma 4 on Macs: the dense 31B, the 12B (gemma4_unified) with its MTP assistant, and oQ mixed widths?

  ashhart/TensorFold#322 · Filter by author gilby opened on Oct 3, 2026

• CUDA: may the bf16 head (b16) decode on the mma for K % 64 == 0? It changes the head's sum order

  ashhart/TensorFold#321 · Filter by author gilby opened on Oct 3, 2026
  1

• Flash Next (MLX 4-bit + MTP) on a single DGX Spark: v0.6.4 decodes ~24% slower than v0.6.3, with a single stream

  ashhart/TensorFold#317 · Filter by author rafafortes opened on Oct 3, 2026
  12

• Invoking vision will slow down the request.

  ashhart/TensorFold#314 · Filter by author xjqx2z opened on Oct 3, 2026

• Volta (sm_70) follow-ups to #264: NVFP4 checkpoints, Qwen3.6 MoE, two and four ranks

  ashhart/TensorFold#313 · Filter by author gilby opened on Oct 3, 2026