spiritbuun/buun-llama-cpp is a fork of TheTom/llama-cpp-turboquant that adds
two independent features on top: DFlash block-diffusion speculative decoding
(via a dedicated DFlashDraftModel GGUF arch) and two extra TCQ KV-cache
variants (turbo2_tcq, turbo3_tcq) on top of TurboQuant's turbo2/turbo3/turbo4.
Follows the turboquant thin-wrapper pattern — reuses backend/cpp/llama-cpp
grpc-server sources verbatim, patches only the build copy to extend the KV
allow-list and wire up buun-exclusive tree_budget / draft_topk options.
DraftModel is already wired end-to-end (proto field 39 → params.speculative),
so DFlash activation only needs the existing options passthrough
(spec_type:dflash) plus the drafter path in draft_model.
CacheTypeOptions now surfaces the five turbo* values so the React UI dropdown
shows them — benefits turboquant too (previously users had to type them in
YAML manually).
Assisted-by: Claude:Opus-4.7 [Read] [Edit] [Bash] [WebFetch]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
whisperx has no upstream AMD GPU support and its core transcription path
(faster-whisper -> ctranslate2) falls back to CPU on AMD since the PyPI
ctranslate2 is CUDA-only. The torch rocm wheels would accelerate only the
alignment/diarization stages, producing a misleadingly half-working image.
Drop the hipblas variant rather than shipping a partially accelerated build
users can't distinguish from the real thing. AMD hosts now fall through
the capability map to cpu-whisperx / cpu-whisperx-development.
Also removes the now-dangling rocm-whisperx assertion from
pkg/system/capabilities_test.go and the ROCm mention from the whisperx
row in docs/content/reference/compatibility-table.md.
Assisted-by: Claude Code:claude-opus-4-7
openai-functions.md used to claim LocalAI tool calling worked only on
llama.cpp-compatible models. That was true when it was written; it's
not true now — vLLM (since PR #9328) and MLX/MLX-VLM both extract
structured tool calls from model output.
- openai-functions.md: new 'Supported backends' matrix covering
llama.cpp, vllm, vllm-omni, mlx, mlx-vlm, with the key distinction
that vllm needs an explicit tool_parser: option while mlx auto-
detects from the chat template. Reasoning content (think tags) is
extracted on the same set of backends. Added setup snippets for
both the vllm and mlx paths, and noted the gallery importer
pre-fills tool_parser:/reasoning_parser: for known families.
- compatibility-table.md: fix the stale 'Streaming: no' for vllm,
vllm-omni, mlx, mlx-vlm (all four support streaming now). Add
'Functions' to their capabilities. Also widen the MLX Acceleration
column to reflect the CPU/CUDA/Jetson L4T backends that already
exist in backend/index.yaml — 'Metal' on its own was misleading.
* docs: Update model compatibility documentation with missing backends
Added the following backends to README.md and compatibility-table.md:
- vllm-omni: Multimodal vLLM with vision and audio support
- nemo: NVIDIA NeMo framework for speech models
- outetts: OuteTTS with voice cloning capabilities
- faster-qwen3-tts: Faster Qwen3 TTS implementation
- qwen-asr: Qwen automatic speech recognition
- voxcpm: VoxCPM speech understanding model
- whisperx: Enhanced Whisper with word-level transcription
These backends exist in the codebase (backend/index.yaml) but were missing
from the documentation. This update ensures accurate reflection of currently
supported backends in LocalAI.
* Apply suggestion from @mudler
Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
---------
Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
Co-authored-by: localai-bot <localai-bot@example.com>
Co-authored-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
exllama2 development has stalled and only old architectures are
supported. exllamav3 is still in development, meanwhile cleaning up
exllama2 from the gallery.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
