Issues

• Compile bug: "Unresolved external symbol ggml_graph_view referenced in function "bool __cdecl ggml_cuda_compute_graph_nodes" on VS2022

  GenerelSchwerz/llama.cpp#154 · Filter by author NNG9y1PQdXl opened on Oct 8, 2026

• Generic hybrid: --moe-expert-profile crashes at first request (op not supported ffn_moe_gate-0 (MUL_MAT_ID) -> unspecified launch failure) on Qwen3.6-35B-A3B

  GenerelSchwerz/llama.cpp#151 · Filter by author Piggidragon opened on Oct 7, 2026

• Eval bug: grouped MoE refuses ffn_moe_gate-0 (MUL_MAT_ID) + illegal access on Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S (good: b46f7f7, bad: e69cf66)

  GenerelSchwerz/llama.cpp#120 · Filter by author larbouch opened on Oct 2, 2026
  1

• Feature Request: Please implement KVarN

  enhancement
  GenerelSchwerz/llama.cpp#113 · Filter by author Almazick opened on Sep 29, 2026
  1

• Eval bug: regression - CUDA error: an illegal memory access in function prepare_residency

  GenerelSchwerz/llama.cpp#112 · Filter by author NNG9y1PQdXl opened on Sep 29, 2026
  2

• Eval bug: CUDA error when running with high vram usage on Windows flash next

  GenerelSchwerz/llama.cpp#110 · Filter by author jdunne525 opened on Sep 26, 2026
  1

• Misc. bug: advisory MoE prefetch reduces short-context MTP decode speed

  bugCUDA
  GenerelSchwerz/llama.cpp#107 · Filter by author GenerelSchwerz opened on Sep 23, 2026

• Qwen3.8-Flash-Next stops mid-code before token limit, including upstream without MTP

  GenerelSchwerz/llama.cpp#103 · Filter by author GenerelSchwerz opened on Sep 23, 2026
  1

• 3060 tg speed drop

  GenerelSchwerz/llama.cpp#102 · Filter by author lavkalavkalavka opened on Sep 21, 2026
  1

• Unified KV microbatch splitting should preserve MTP verification spans

  buglow priority
  GenerelSchwerz/llama.cpp#100 · Filter by author GenerelSchwerz opened on Sep 20, 2026

• Research: Some 2x5070ti benchmarks

  GenerelSchwerz/llama.cpp#96 · Filter by author vlascik opened on Sep 16, 2026
  2

• MoE expert cache is pinned to CUDA device 0 (multi-GPU: 2nd GPU's VRAM unusable, ~10 GB idle)

  GenerelSchwerz/llama.cpp#90 · Filter by author lydstyl opened on Sep 13, 2026
  9

  

• server: unresolved parallel decode-overlap cancellation and slot-reuse validation

  GenerelSchwerz/llama.cpp#82 · Filter by author GenerelSchwerz opened on Sep 9, 2026

• Reduce unused intermediate workspace in MoE weighted-reduction fusion

  GenerelSchwerz/llama.cpp#80 · Filter by author GenerelSchwerz opened on Sep 9, 2026