Issues

• [Desktop] weights failed to load: Gemma-4-12b-it NVFP4

  new-model
  FlashML-org/FreeToken#847 · Filter by author Samo007 opened on Nov 1, 2026

• [WIP] Speculative decoding: MTP and DFlash (tracking)

  featurespec decode
  FlashML-org/FreeToken#828 · Filter by author jason-fxz opened on Oct 26, 2026

• error when installing engine update

  bugDesktopwindows
  FlashML-org/FreeToken#826 · Filter by author gotenks917 opened on Oct 26, 2026
  1

• Performance regression from v0.1.3 to v0.2.0 on RTX4060 8GB

  bugwindows
  FlashML-org/FreeToken#822 · Filter by author moro-hide opened on Oct 24, 2026
  6

• Support nvidia/Qwen3.8-Flash-Next-FP4-Local (Qwen4Exp, MIXED_PRECISION CSF NVFP4 + MXFP8 + FP4 PLE table)

  FlashML-org/FreeToken#817 · Filter by author chrisqianz opened on Oct 24, 2026

• "Mamba State Slot" is easy to misread — suggest "session snapshots (GDN state)" with a tooltip about the VRAM budget

  Desktopfeature
  FlashML-org/FreeToken#800 · Filter by author Domir6 opened on Oct 22, 2026
  1

• Engine crash with high KV profile

  buglinux
  FlashML-org/FreeToken#798 · Filter by author lolmc0587 opened on Oct 22, 2026
  2

• GGUF vision towers (mmproj): survey and plan

  gguf
  FlashML-org/FreeToken#794 · Filter by author jason-fxz opened on Oct 22, 2026

• Bound repeated disk PLE row reads with an opt-in host cache

  FlashML-org/FreeToken#791 · Filter by author wenlive opened on Oct 22, 2026

• Context clipping mutates sampling parameters shared by offline requests

  FlashML-org/FreeToken#787 · Filter by author wenlive opened on Oct 22, 2026

• [Desktop] weights failed to load: Qwen3.6-35B-A3B NVFP4

  duplicate
  FlashML-org/FreeToken#784 · Filter by author mehedih985 opened on Oct 21, 2026
  2

• perf(qsa): prefill indexer scoring pads 4 heads to 16 per row; 16-row tiles are ~5x faster and bitwise identical

  FlashML-org/FreeToken#783 · Filter by author gdevenyi opened on Oct 21, 2026

• [Desktop] weights failed to load: Qwen3.6-35B-A3B NVFP4

  bug
  FlashML-org/FreeToken#782 · Filter by author srvhbkking-collab opened on Oct 21, 2026
  2

• [Perf] Fuse the BF16 attention output gate in short gfx1201 decode graphs

  amd
  FlashML-org/FreeToken#780 · Filter by author zihaomu opened on Oct 21, 2026

• Capture-time cuGraphUpload (#697) widens inter-kernel gaps in replayed decode graphs: ~1-2% slower decode

  FlashML-org/FreeToken#775 · Filter by author gdevenyi opened on Oct 21, 2026

• Proposal: isolate logprob reads (max_tokens 1) from batch composition and the prefix cache so scoring is reproducible

  FlashML-org/FreeToken#772 · Filter by author gdevenyi opened on Oct 20, 2026

• Proposal: load an FTW with fp16 NVFP4 global scales (upcast with a warning) instead of refusing it after #739

  FlashML-org/FreeToken#769 · Filter by author gdevenyi opened on Oct 20, 2026

• Proposal: vLLM scoring fields on /v1/completions (allowed_token_ids, return_tokens_as_token_ids, token-id prompts) for classifier-style clients

  FlashML-org/FreeToken#768 · Filter by author gdevenyi opened on Oct 20, 2026

• /v1/completions: prompt: [] returns 200 with no choices, and both completions paths drop the error code (context_length_exceeded)

  FlashML-org/FreeToken#767 · Filter by author gdevenyi opened on Oct 20, 2026

• An FTW checkpoint can never serve the W4A4 NVFP4 MoE prefill, and the refusal tells you to reconvert with the option that produced it

  FlashML-org/FreeToken#766 · Filter by author gdevenyi opened on Oct 20, 2026

• [Perf] Offloaded decode: overlap the expert-miss gather with the same layer's shared expert, event-ordered inside the CUDA graph (bit-exact implementation)

  FlashML-org/FreeToken#753 · Filter by author SUKUNA-AI opened on Oct 19, 2026

• [Perf] Offloaded prefill: keep buffer invalidation off the prefill copy stream and give that stream top priority (short-prompt H2D 21 -> 42 GB/s under Nsight)

  FlashML-org/FreeToken#752 · Filter by author SUKUNA-AI opened on Oct 19, 2026

• [Perf] Offloaded decode: gather missed experts with a few TMA bulk-copy CTAs (cp.async.bulk, .shared::cta) instead of 48 x 1024 LDG threads

  FlashML-org/FreeToken#751 · Filter by author SUKUNA-AI opened on Oct 19, 2026

• [Perf] Layer-major prefill for offloaded MoE: stream each expert layer once per prompt, not once per chunk (qwen3_5_moe)

  FlashML-org/FreeToken#750 · Filter by author SUKUNA-AI opened on Oct 19, 2026

• [Bug] Context overflow is checked only after tokenization in the worker: a streamed request gets 200 plus an in-stream error, and an oversized max_tokens is silently clamped

  FlashML-org/FreeToken#749 · Filter by author SUKUNA-AI opened on Oct 19, 2026