chore: ⬆️ Update TheTom/llama-cpp-turboquant to e69af784add62d5d3b3321abc0e3068df41143e7

[localai-bot]

Changes: https://github.com/TheTom/llama-cpp-turboquant/compare/69d8e4be47243e83b3d0d71e932bc7aa61c644dc..e69af784add62d5d3b3321abc0e3068df41143e7

[mudler]

cc @TheTom FYI new changes triggers compilation issues on hipblas: https://github.com/mudler/LocalAI/actions/runs/25610872523/job/75180742246?pr=9740

[TheTom]

fixing

[TheTom]

Hey @mudler — owner of TheTom/llama-cpp-turboquant here. CI failure on the hipblas-turboquant target is on me, not on this bump. Root cause:

ld.lld: error: undefined symbol:
  void ggml_cuda_flash_attn_ext_vec_case<{64,128,256}, TURBO{2,3,4}_0, F16>(...)
clang++: error: linker command failed with exit code 1

In commit e69af784a I added (turbo*, F16) mixed-KV dispatch entries (and the matching template-instance .cu files) to fix a separate gfx1151 user crash, but only updated ggml/src/ggml-cuda/CMakeLists.txt. The parallel list in ggml/src/ggml-hip/CMakeLists.txt was missed, so the HIP build linked without those instantiations.

Just pushed the fix to feature/turboquant-kv-cache as 5aeb2fdbe (3-line CMakeLists mirror, no functional changes outside the build system). LocalAI nightly bot should pick it up automatically; if you want it sooner, bumping TURBOQUANT_VERSION in this PR to 5aeb2fdbe48... should make the hipblas job link and the rest of CI go green.

Also note: while we're here, my fork picked up one more fix this morning at 7e341660d (fix(perplexity): cast n_ctx * nv to size_t in KL logits save) — Qwen3 family at 16K context was crashing in the KL-divergence code path on int×int overflow. Worth grabbing both in a single bump if you re-roll.

Sorry for the breakage — the CUDA-only consumers (regular CUDA build, MLX, my own Mac builds) all linked clean, so this slipped through. LocalAI's hipblas job was the first to surface it.