Issues

• Upgrade to torch 2.14.1, Triton 3.8 and sglang-kernel 0.4.9

  FlashML-org/FreeToken#629 · Filter by author jason-fxz opened on Oct 7, 2026

• Pre-sm70: remaining launch-config fixes on the Qwen3.8-Flash-Next path

  FlashML-org/FreeToken#628 · Filter by author gdevenyi opened on Oct 7, 2026

• Size Triton's autotune flush buffer from the device L2 instead of a fixed 256 MB

  FlashML-org/FreeToken#627 · Filter by author gdevenyi opened on Oct 7, 2026

• CPU MoE: place expert rows per NUMA node and schedule tiles on their node

  FlashML-org/FreeToken#626 · Filter by author gdevenyi opened on Oct 7, 2026

• CPU MoE: AVX2 int8-activation NVFP4 kernel for CPUs without VNNI

  FlashML-org/FreeToken#625 · Filter by author gdevenyi opened on Oct 7, 2026

• Load-time MXFP8 requantization of bf16 dense, hyper-connection and lm_head weights

  FlashML-org/FreeToken#624 · Filter by author gdevenyi opened on Oct 7, 2026

• mxfp8_linear dequantizes the whole weight past the GEMV cap (2.4 GiB transient for a vocab-sized projection)

  FlashML-org/FreeToken#623 · Filter by author gdevenyi opened on Oct 7, 2026

• QSA host KV tier: compact selected pages with a column bitmap and scan (7.5x)

  FlashML-org/FreeToken#622 · Filter by author gdevenyi opened on Oct 7, 2026

• QSA host KV tier: give each layer its own slot map (decode at 231K 6.8 -> 15.6 tok/s)

  FlashML-org/FreeToken#621 · Filter by author gdevenyi opened on Oct 7, 2026

• GLM-5.3-Flash-NVFP4 CUDA OOM during load on RTX 5070 Ti 16 GB (sm_120)

  FlashML-org/FreeToken#611 · Filter by author ogino opened on Oct 6, 2026

• Windows wheels resolve unpinned torch dependency to CPU-only build: 'cannot use CUDA device 0: only 0 device(s) visible'

  FlashML-org/FreeToken#607 · Filter by author pablosilva87 opened on Oct 5, 2026
  1

• DeepSeek-V4-Flash-0731-NVFP4 (nvidia) fails on 0.1.3+gfa0e9537d: WeightLoadError KeyError 'layers.0.ffn.experts.0.w1.scale'

  FlashML-org/FreeToken#606 · Filter by author pablosilva87 opened on Oct 5, 2026
  1

• [Engine] Qwen3.8-Flash-Next-FP8 exhausts host RAM on RTX 5090 + 128 GB after serial loading fallback

  buglinux
  FlashML-org/FreeToken#600 · Filter by author yunwei37 opened on Oct 4, 2026
  1

• [Windows][Desktop] WeightLoadError: invalid python storage with Qwen3.6 NVFP4 and gpt-oss-20b MXFP4

  bugDesktopwindows
  FlashML-org/FreeToken#597 · Filter by author Shahrad-gh opened on Oct 4, 2026
  2

• The computer on the local area network cannot be accessed

  bugDesktopwindows
  FlashML-org/FreeToken#584 · Filter by author albertjacksoon opened on Sep 30, 2026

• An offloaded prefill streams every expert of each layer, however short the chunk

  FlashML-org/FreeToken#579 · Filter by author YevheniiKotyrlo opened on Sep 30, 2026

• On Windows the engine sizes its pools as if no other process held any of the card

  windows
  FlashML-org/FreeToken#570 · Filter by author YevheniiKotyrlo opened on Sep 29, 2026

• A hybrid model's cached prefix cannot be resumed far back: no GDN snapshot survives below a prompt's last chunk

  windows
  FlashML-org/FreeToken#569 · Filter by author YevheniiKotyrlo opened on Sep 29, 2026

• Serving Qwen3.8-Flash-Next's 1M-token window on one 24 GB card

  FlashML-org/FreeToken#568 · Filter by author YevheniiKotyrlo opened on Sep 29, 2026

• A stream that is silent through a long prefill is dropped by Node clients and prefilled for nobody

  windows
  FlashML-org/FreeToken#567 · Filter by author YevheniiKotyrlo opened on Sep 29, 2026

• Bug: AssertionError: Not enough memory for KV cache when serving Qwen3.8-27B-NVFP4 on Windows

  bugwindows
  FlashML-org/FreeToken#561 · Filter by author hulonghe opened on Sep 28, 2026
  2

• Feature: --api-key flag for serve mode (token auth on the HTTP API)

  FlashML-org/FreeToken#557 · Filter by author santanu20 opened on Sep 27, 2026

• [Feature] Cumulative prefill/decode seconds on /v1/stats so pollers can compute speeds

  FlashML-org/FreeToken#552 · Filter by author mgeldi opened on Sep 25, 2026

• Qwen3.8-Flash-Next-GSQ-RCO-GGUF not loading.

  buglinux
  FlashML-org/FreeToken#549 · Filter by author DarkReaperBoy opened on Sep 25, 2026
  1

• Show model sampling defaults and allow per-request overrides in the Windows client

  Desktopfeature
  FlashML-org/FreeToken#547 · Filter by author dai-ych opened on Sep 24, 2026
  1