Issues

• SIGSEGV during model load on gfx1150 (Radeon 890M / RDNA 3.5 iGPU, HIP build)

  freriadoc/Avalon#54 · Filter by author 0n1cOn3 opened on Nov 13, 2026

• Eval bug: gfx1030: passes of 6 tokens or fewer diverge from larger passes (KLD ~0.25 vs ~0.04 noise)

  freriadoc/Avalon#50 · Filter by author bilboquet opened on Nov 12, 2026
  2

• Eval bug: RX 7900 XTX (gfx1100, ROCm 7.2.2): the prompt ring on HIP aborts — OOM in ggml_cuda_moe_block_grouped_host, then hipMemcpyAsync invalid argument once a reserve is set

  freriadoc/Avalon#49 · Filter by author yar3333 opened on Nov 12, 2026
  4

• Eval bug: Any prompt hangs at Processing (RX 7900 XT)

  freriadoc/Avalon#48 · Filter by author EmanuelBalaban opened on Nov 12, 2026
  2

• Eval bug: LLAMA_MOE_DOWN_SUM (made a default as-of commit f5a34ec) causes garbage output on Windows

  freriadoc/Avalon#47 · Filter by author ParadiceNZ opened on Nov 12, 2026
  6

• Eval bug: gfx1030 (RX 6800 XT): main aborts in the prompt with the default QSA tiles (GGML_CUDA_MK, indexer_sel_tiles)

  freriadoc/Avalon#46 · Filter by author bilboquet opened on Nov 12, 2026
  6

• GPU stays at 100% / ~150 W after a GPU-encoded image followed by a ~1.4k-token prompt (gfx1100, ROCm 7.2.1)

  freriadoc/Avalon#45 · Filter by author ninjaltd opened on Nov 12, 2026

• Research: Strata fork - Analysis and Research from Academic papers and other strata and llama forked projects

  freriadoc/Avalon#43 · Filter by author rocktboy1982 opened on Nov 9, 2026
  8

• Windows (CUDA) build available: prebuilt zip and launchers

  enhancement
  freriadoc/Avalon#42 · Filter by author newmilltechnology opened on Nov 8, 2026
  1

• Eval bug: MTP: first verification pass after a short first prompt hangs in moe_hosted_plan's defer wait

  freriadoc/Avalon#40 · Filter by author newmilltechnology opened on Nov 7, 2026
  1

• Windows + CUDA: server hangs at the first graph execution with the MoE expert cache on (GPU 100 %, /health 503); MOE_CACHE_MODE=off serves at 4.4 t/s

  freriadoc/Avalon#37 · Filter by author rueyke-lang opened on Nov 7, 2026
  2

• Misc. bug: general report on difficulties running on dual xtx 7900

  freriadoc/Avalon#36 · Filter by author kylebakerio opened on Nov 6, 2026
  3

• Misc. bug: Please consider to add support for Ivy Bridge / AVX | error: the CPU has no AVX2: the experts computed on the CPU need it

  freriadoc/Avalon#35 · Filter by author CycleAndCode opened on Nov 5, 2026

• Windows 11, RX 7900 XTX, 48 GB RAM: first request very slow, long prompts page heavily

  freriadoc/Avalon#34 · Filter by author The-Dude-2020 opened on Nov 5, 2026
  1

• gfx1030 (RX 6800 XT): hard system reset right after moe-cache borrow when a new, shorter prompt follows a ~23k one

  freriadoc/Avalon#27 · Filter by author bilboquet opened on Nov 4, 2026
  13

• Misc. bug: Two RX 7900 XT without P2P: crash found and fixed, speed limited by host syncs

  freriadoc/Avalon#26 · Filter by author NaifuKishi opened on Nov 4, 2026
  5

• Benchmark: 5900X + 7900 XTX, PCIe 3.0, 64 GB DDR4 3600, on ROCm 10.1

  freriadoc/Avalon#25 · Filter by author sa20001 opened on Nov 3, 2026
  5

• Feature Request: n-gram table residency (disk vs RAM) for the ngram-* speculative types

  enhancement
  freriadoc/Avalon#24 · Filter by author Creazy-chen opened on Nov 3, 2026

• Eval bug: gfx1030 (V620), ROCm 7.1, third-party GSQ-RCO quant

  freriadoc/Avalon#23 · Filter by author Cukrus-directive opened on Nov 3, 2026
  1

• Misc. bug: --fit and default load mode assert on an APU (gfx1151) because the ROCm device reports no mmap_support

  freriadoc/Avalon#22 · Filter by author pato6013 opened on Nov 3, 2026

• Eval bug: abort on NVIDIA 3090: GGML_ASSERT(!mask_tiles || tile_shape) failed

  freriadoc/Avalon#21 · Filter by author ping235 opened on Nov 3, 2026
  4

• CUDA abort: too many resources requested for launch on RTX 3090 with the README launch line and the MTP draft

  freriadoc/Avalon#19 · Filter by author mrq1911 opened on Nov 2, 2026
  1

• Research: Avalon with speculative decoding.

  freriadoc/Avalon#18 · Filter by author rocktboy1982 opened on Nov 2, 2026
  1

• llama-imatrix on a >RAM-experts model: livelock with the expert cache (any budget), segfault with --moe-cache off

  freriadoc/Avalon#17 · Filter by author yar3333 opened on Nov 2, 2026
  2

• Hosted MoE on multiple GPUs aborts (illegal memory access) on 9617344db, livelocks in cache-fill on a91372eb6

  freriadoc/Avalon#16 · Filter by author yar3333 opened on Nov 2, 2026
  2