Issues

• Misc. bug: Numeric regression on qwen4exp

  halo-box/strix-llama.cpp#157 · Filter by author iggy opened on Oct 8, 2026

• Independent repro on full-power Ryzen AI Max+ 395 — prefill gains grow with context depth (opposite of the 70W Z13 report)

  halo-box/strix-llama.cpp#122 · Filter by author BaronMuchhumpin opened on Sep 30, 2026
  2

• Proposal: TurboQuant KV (turbo3/turbo4) with HIP kernels for gfx1151 (port of TheTom's CUDA kernels)

  halo-box/strix-llama.cpp#103 · Filter by author BonkusCheemus opened on Sep 28, 2026
  2

• vulkan : on gfx1150 (Strix Point) the q8_0/q6_K mat-vec chunking makes 3- and 5-column passes slower on Qwen3.8-Flash-Next, so draft-MTP n-max 2 gains nothing

  halo-box/strix-llama.cpp#88 · Filter by author rcfa opened on Sep 25, 2026

• qwen4exp: GGML_ABORT ("maskless selected-key attention was not taken by a QSA kernel") on KV-cache exhaustion under --kv-unified with concurrent slots

  halo-box/strix-llama.cpp#85 · Filter by author randomm opened on Sep 24, 2026
  2

• Research: How do you recommand to run this model ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF

  halo-box/strix-llama.cpp#80 · Filter by author xlazom00 opened on Sep 22, 2026
  1

• [BUG] llama-server SIGSEGV in qwen4exp_ple_prefetch → llama_ple_disk::page_cached() — this pointer is heap-corrupted

  halo-box/strix-llama.cpp#79 · Filter by author nanpaidashi opened on Sep 20, 2026

• Feature Request: Enable releases

  enhancement
  halo-box/strix-llama.cpp#77 · Filter by author xlazom00 opened on Sep 19, 2026
  2

• Eval bug: GGML_CUDA_ENABLE_UNIFIED_MEMORY causes garbled output on second token and onwards

  halo-box/strix-llama.cpp#54 · Filter by author ANTONBORODA opened on Sep 14, 2026

• vulkan : after the 2026-09-11 upstream sync (#47), batched mat-vec at 5..8 columns regresses on dense Qwen3.8-27B (pp8 79 -> 52 t/s)

  halo-box/strix-llama.cpp#51 · Filter by author c4software opened on Sep 13, 2026
  4

• vulkan : mat-vec chunking costs 3-7% at 5..8 columns on dense q8_0/q6_K tensors, worst at 7 (4+2+1)

  halo-box/strix-llama.cpp#50 · Filter by author c4software opened on Sep 13, 2026
  1

• Feature Request: support DeepSeek V4.1 Flash with bounded SSD expert caching

  halo-box/strix-llama.cpp#48 · Filter by author ajaxdude opened on Sep 12, 2026
  21

• benchmark numbers: ling-3.0-flash 124b q4_k_m (missing from the atlas)

  stale
  halo-box/strix-llama.cpp#28 · Filter by author julianmb opened on Sep 6, 2026
  1