Issues

• Greedy output not reproducible when a turn resumes from a prefix-cache checkpoint

  Luce-Org/lucebox#802 · Filter by author davide221 opened on Oct 1, 2026

• tool_choice: "none" is not enforced (no token ban in decoding)

  Luce-Org/lucebox#801 · Filter by author davide221 opened on Oct 1, 2026

• [ROCm/gfx1201] PFlash scorer: FP stage ~45x slower than the CUDA reference (343s for S=120k vs 7.68s on a 3090), so PFlash is net-negative at long context

  Luce-Org/lucebox#752 · Filter by author aiworkeryuan-chuxi opened on Sep 22, 2026
  1

• 100% Single Thread Usage on Ryzen AI MAX 395 (gfx1151, Strix Halo)

  Luce-Org/lucebox#750 · Filter by author RandomLemon opened on Sep 22, 2026
  2

• Ds4HcMatvecPool: heap-use-after-free in dual-GPU MoE-TP decode path (fix included), plus a residual rare deadlock (unresolved)

  ds4flash
  Luce-Org/lucebox#682 · Filter by author chrisjohnson opened on Aug 29, 2026
  1

• R9700 (Thunderbolt eGPU) as hip:0 source: hipMemcpyPeer SIGSEGV, blocks dual-GPU DS4 expert parallelism on Strix Halo pairing

  Luce-Org/lucebox#681 · Filter by author chrisjohnson opened on Aug 28, 2026
  2

• any plans to upgrade to ROCm 7.14?

  rocm
  Luce-Org/lucebox#591 · Filter by author wszgrcy opened on Aug 7, 2026
  1

• DeepSeek-V4-Flash-0731-ROCmFPX-MIX-STRIX.gguf fails to load: tensor type 105 (unknown).

  Luce-Org/lucebox#588 · Filter by author Arkovski opened on Aug 6, 2026
  4

• DeepSeek V4 Flash on Strix Halo (gfx1151): 4 defects in the monolithic DS4 path, 2 validated patches attached

  ds4flash
  Luce-Org/lucebox#572 · Filter by author GRAVESIS opened on Aug 1, 2026
  1

• Feature Request: Support of DeepSeek V4 Flash 0731

  ds4flash
  Luce-Org/lucebox#568 · Filter by author Bushido76 opened on Jul 31, 2026
  1

• Add support draft model laguna-s-2.1-DFlash-BF16.gguf

  Luce-Org/lucebox#566 · Filter by author hitman249 opened on Jul 31, 2026

• Run Qwen3.6-35B-A3B MoE on GTX 1080 Ti

  spark
  Luce-Org/lucebox#553 · Filter by author SeanPedersen opened on Jul 21, 2026
  3

• Bug: ggml_set_rows faults on HIP/ROCm after many reuses of a cached single-token graph (Strix Halo / gfx1151)

  bug
  Luce-Org/lucebox#516 · Filter by author howard0su opened on Jul 13, 2026

• Run Ornith

  enhancementmodel
  Luce-Org/lucebox#463 · Filter by author mariuszkreft opened on Jun 29, 2026
  1

• gfx1100 (RX 7900 XTX, ROCm 6.4.1): dflash_server ~2x slower than llama.cpp, DDTree speculation never engages, SIGSEGV at long context

  Luce-Org/lucebox#457 · Filter by author L0GYKAL opened on Jun 27, 2026
  4

• Power Consumption in Idle on Strix Halo

  bugpowerstrix-halo
  Luce-Org/lucebox#427 · Filter by author Bushido76 opened on Jun 20, 2026
  2

• RTX 5060 Ti (Blackwell sm_120): draft OOM despite model sizes fitting — 3× driverInitFileInfo result=11 causes ~820 MiB of orphaned VRAM overhead

  Luce-Org/lucebox#407 · Filter by author EnricoFerrero64 opened on Jun 16, 2026

• Unable to get above 50/tks output Qwen 3.6 35 A3b Q4_K_M

  docker
  Luce-Org/lucebox#377 · Filter by author Kamalaja opened on Jun 12, 2026
  2

• GGUF tensor loader bug: blk.56.ssm_out.weight falsely reports 'overflows file' on valid Qwen3.6-27B Q4_K_M GGUF

  bugmodel
  Luce-Org/lucebox#318 · Filter by author batumilove opened on May 31, 2026

• pflash + dflash optimization on top of qwen35moe (PR #262)

  enhancement
  Luce-Org/lucebox#280 · Filter by author davide221 opened on May 26, 2026
  1

• MTP for Speculative Drafting

  enhancement
  Luce-Org/lucebox#206 · Filter by author momendo opened on May 17, 2026
  1

• Add support for Intel XPU

  enhancementintel
  Luce-Org/lucebox#201 · Filter by author ksp2000 opened on May 16, 2026
  1

• How is Qwen3.6-35B-A3B with PFlash+DFlash+DDTree?

  question
  Luce-Org/lucebox#161 · Filter by author TomLucidor opened on May 13, 2026
  4

• How to quickly obtain interface services in openai and claude formats

  openwebuiquestion
  Luce-Org/lucebox#157 · Filter by author Yelart opened on May 12, 2026
  7

• Does this support flash-moe for MoE model?

  enhancement
  Luce-Org/lucebox#123 · Filter by author hksfho opened on May 8, 2026
  1