b224c96db6
SetDefaults injected the llama.cpp server options cache_reuse (ApplyServingDefaults) and parallel (ApplyHardwareDefaults, re-applied per selected node by the distributed router) onto every model config regardless of backend. Every other backend ignores options it does not understand, so this was harmless until longcat-video, which strictly validates its options and fails LoadModel with "unknown model option(s): cache_reuse, parallel". Gate both injections behind a new UsesLlamaCppServingOptions allow-list (llama-cpp plus the empty/auto-detect case that resolves to llama.cpp from a GGUF file, mirroring how llamaCppDefaults is registered). This follows the existing UsesLlamaSamplerDefaults precedent for llama-only defaults. The typed NBatch field is deliberately left alone: it is a proto field every backend simply ignores, which is why batch never triggered the error. Also harden the longcat-video backend to warn-and-ignore unknown model options and request params through a testable select_known_options helper, matching the other LocalAI Python backends, so a future server-injected option cannot break loading again. Assisted-by: Claude:claude-opus-4-8 [Claude Code] Signed-off-by: Ettore Di Giacinto <mudler@localai.io> Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
LongCat Video backend
This backend serves Meituan's LongCat-Video and
LongCat-Video-Avatar-1.5 checkpoints through LocalAI's GenerateVideo
RPC. It supports:
- text-to-video and image-to-video with
LongCat-Video; - audio + text-to-avatar and portrait + audio-to-avatar with Avatar 1.5;
- multi-segment avatar continuation for speech longer than one segment;
- PyTorch SDPA when FlashAttention is unavailable, including CUDA 13 ARM64 systems such as NVIDIA DGX Spark.
Install the longcat-video or longcat-video-avatar-1.5 recipe from the
LocalAI Model Gallery. See the LongCat user guide
for Studio and API examples, hardware requirements, and manual configuration.
The upstream source is pinned in Makefile and patched at build time. The
patch adds only the missing SDPA attention branches; model and source licenses
remain MIT.
Model options
| Option | Default | Description |
|---|---|---|
attention_backend |
sdpa |
sdpa, auto, flash2, flash3, or xformers. The packaged backend guarantees only sdpa. |
use_distill |
true for Avatar, false for base |
Loads the checkpoint's fast distillation LoRA. |
use_int8 |
false |
Loads Avatar 1.5's INT8 DiT. BF16 has a lower load-time peak on unified-memory systems. |
base_model |
meituan-longcat/LongCat-Video |
Base components used by Avatar 1.5. |
max_segments |
8 |
Maximum avatar continuation segments accepted per request. |
resolution |
480p |
Image-conditioned generation resolution (480p or 720p). |
Per-request params may set num_segments, audio_guidance_scale,
offload_kv_cache, ref_img_index, mask_frame_range, and resolution.
Gallery and imported configs declare known_input_modalities and
known_output_modalities. Keep those declarations in manual configs as well;
they let model discovery distinguish base image-conditioned video from Avatar
audio conditioning without inspecting the backend or checkpoint name.
LongCat is CUDA-only and very large. Avatar 1.5 also loads tokenizer, text-encoder, and VAE components from the base checkpoint. Keep ample unified memory and storage available; no CPU or macOS backend image is published. The initial backend supports one GPU per process; tensor parallel sizes above one are rejected explicitly.