Issues

• qwen3 test model: a one-token ubatch aborts under -sm tensor (meta split state of the K norm)

  Piggidragon/llama.cpp#45 · Filter by author Piggidragon opened on Oct 19, 2026

• Experiment: GPU sampling of the verify positions under -sm tensor

  enhancementggml
  Piggidragon/llama.cpp#42 · Filter by author Piggidragon opened on Oct 18, 2026
  1

• Experiment: time to the first token of short prompts (150-450 tokens)

  enhancementggml
  Piggidragon/llama.cpp#38 · Filter by author Piggidragon opened on Oct 18, 2026
  2

• Unified KV cache (#32): what is left

  Piggidragon/llama.cpp#33 · Filter by author Piggidragon opened on Oct 18, 2026

• minimax-01: wrong logits on CUDA for ubatches with several sequences

  Piggidragon/llama.cpp#30 · Filter by author Piggidragon opened on Oct 18, 2026

• llama-server: first requests of a new conversation re-prefill full prompt instead of reusing cached prefix

  Piggidragon/llama.cpp#29 · Filter by author Piggidragon opened on Oct 17, 2026
  1

• kimi-linear: context creation aborts when n_ubatch does not divide by n_seq_max

  Piggidragon/llama.cpp#27 · Filter by author Piggidragon opened on Oct 17, 2026

• test-llama-archs: maion-coder fails at random (NMSE, mixed batch, state range)

  Piggidragon/llama.cpp#26 · Filter by author Piggidragon opened on Oct 17, 2026