Issues

• Misc. bug: Context checkpoints not working

  TheTom/llama-cpp-turboquant#396 · Filter by author jankais3r opened on Sep 29, 2026

• Permission to submit your turbo KV (HIP port) to halo-box/strix-llama.cpp, credited to you

  TheTom/llama-cpp-turboquant#394 · Filter by author BonkusCheemus opened on Sep 28, 2026
  2

• kv-stream: native turbo KV streaming - port offer with bench evidence (decode 2.42 -> 15 t/s at deep spill)

  TheTom/llama-cpp-turboquant#391 · Filter by author grogroningem-code opened on Sep 26, 2026
  6

• CUDA: SET_ROWS_TURBO3/TURBO4 tests never run because CPY from turbo to F32 is unsupported

  TheTom/llama-cpp-turboquant#385 · Filter by author TheTom opened on Sep 23, 2026

• Misc. bug: -40% decode speed regression with DeepSeek V4 Flash

  TheTom/llama-cpp-turboquant#379 · Filter by author Volunteer-1 opened on Sep 20, 2026

• speculative-simple: null draft context with a DFlash2 drafter (dflash requires ctx_other to be set)

  TheTom/llama-cpp-turboquant#376 · Filter by author jasstrong opened on Sep 20, 2026

• Vulkan MoE cache: 512 KiB min-expert default silently excludes small-expert MoE models (Qwen3.6-35B-A3B IQ2_XS, 296 KiB/expert)

  TheTom/llama-cpp-turboquant#365 · Filter by author Scintsz opened on Sep 9, 2026
  2

• SYCL: TurboQuant flash-attn decode (vec) kernels fail JIT on Arc Pro B60 (G21, 0xe211) with head_dim 256 models - pp OK, tg aborts

  TheTom/llama-cpp-turboquant#361 · Filter by author fantaskiss opened on Sep 7, 2026
  4

• Eval bug: First decode after prefill runs 2-45x slower: HIP CUDA-graph capture warmup (Qwen3.8-Flash-Next / qwen4exp QSA, gfx1201)

  stale
  TheTom/llama-cpp-turboquant#356 · Filter by author Draco-Lunaris opened on Sep 5, 2026

• CUDA: full test-backend-ops sweep segfaults in ggml_backend_cuda_graph_compute at ROLL on GB10 (passes with graphs disabled)

  bug
  TheTom/llama-cpp-turboquant#347 · Filter by author TheTom opened on Sep 3, 2026
  2

• Proposal: TurboAnchorKV cold prompt KV experiment

  stale
  TheTom/llama-cpp-turboquant#332 · Filter by author TheTom opened on Aug 31, 2026
  6

• Vulkan: Polaris wave64 incompatible with hard-coded subgroup size 32

  stale
  TheTom/llama-cpp-turboquant#330 · Filter by author gaigher opened on Aug 31, 2026
  1

• Eval bug: llama-server OOM Over Time

  TheTom/llama-cpp-turboquant#325 · Filter by author shizhx opened on Aug 28, 2026
  7

• turbo4 cache output corruption introduced by PR #197 — bisected: last good b9907, first bad b9914

  TheTom/llama-cpp-turboquant#311 · Filter by author poshih opened on Aug 19, 2026
  1

• Misc. bug: more VRAM usage use compared to older builds

  stale
  TheTom/llama-cpp-turboquant#308 · Filter by author aminya opened on Aug 18, 2026
  7

• Misc. bug: Llama Web UI is broken

  stale
  TheTom/llama-cpp-turboquant#307 · Filter by author Almazick opened on Aug 18, 2026
  1

• MTP bug: NVIDIA-Nemotron-3.5-Lightning crash

  TheTom/llama-cpp-turboquant#296 · Filter by author Piggidragon opened on Aug 13, 2026
  1

• HIP: turbo K-cache flash-attention spills 295-720 VGPRs at head size 256 on CDNA/RDNA2/RDNA3 (quality gate had never run)

  bug
  TheTom/llama-cpp-turboquant#294 · Filter by author TheTom opened on Aug 12, 2026
  15

• Metal: TQ3_1S and TQ4_1S MUL_MAT write past end of dst at large N (test-backend-ops sentinel mismatch)

  bug
  TheTom/llama-cpp-turboquant#285 · Filter by author TheTom opened on Aug 10, 2026
  1

• HIP: test-backend-ops FLASH_ATTN_EXT aborts on unsupported head size instead of skipping, so no full sweep can complete

  bug
  TheTom/llama-cpp-turboquant#280 · Filter by author TheTom opened on Aug 8, 2026
  2

• CPU prefill aborts on hybrid Nemotron-H ConfigI model: cannot copy tensors with different layouts

  bug
  TheTom/llama-cpp-turboquant#273 · Filter by author TheTom opened on Aug 7, 2026

• cuda/hip: AMD WMMA path has no ncols2==1 instance and no dispatcher fallback (latent abort)

  bug
  TheTom/llama-cpp-turboquant#251 · Filter by author TheTom opened on Aug 2, 2026
  2

• cuda: test-backend-ops segfaults in CUDA-graph path on GB10/sm_121 (pre-existing; graphs-disabled run passes 16311/16311)

  bug
  TheTom/llama-cpp-turboquant#247 · Filter by author TheTom opened on Aug 1, 2026

• metal: mixed turbo4-K / turbo3-V FA produces garbage output (pre-existing, both #244 and current head)

  bug
  TheTom/llama-cpp-turboquant#246 · Filter by author TheTom opened on Aug 1, 2026

• metal: TQ4_1S MUL_MAT out-of-bounds write (sentinel mismatch) at n=256 with offset

  bug
  TheTom/llama-cpp-turboquant#245 · Filter by author TheTom opened on Aug 1, 2026