Backend Release Baselines

Backend Release Baselines

Audit completed: 2026-09-24. Every release published since the 2026-08-26 audit was checked: KoboldCpp v1.120 and v1.121, stable-diffusion.cpp master-832 through master-908, llama.cpp b10637 through b11157, whisper.cpp v1.9.4, and vLLM v0.29.0 and v0.30.0.

Backend	Source	Verified marker
KoboldCpp	https://github.com/LostRuins/koboldcpp/releases	v1.121
stable-diffusion.cpp	https://github.com/leejet/stable-diffusion.cpp/releases	master-908-88411ef
llama.cpp	https://github.com/ggml-org/llama.cpp/releases	b11157
whisper.cpp	https://github.com/ggml-org/whisper.cpp/releases/tag/v1.9.4	v1.9.4
vLLM	https://github.com/vllm-project/vllm/releases	v0.30.0

The listed markers are compatibility baselines for this router release, not a promise that they remain upstream's newest tags. whisper.cpp publishes Windows archives for v1.9.4 under the matching nightly tag b5130.

Launch flag requirements

• llama.cpp model loading uses --load-mode (auto, none, mmap, mlock, mmap+mlock, dio), available since b10105. llama.cpp b10875 removed --mlock, --mmap, --no-mmap and --direct-io, so the router no longer emits them. .kcpps files that only set usemmap and usemlock are translated to the matching --load-mode value.
• llama.cpp speculative drafts are selected with --spec-type (draft-dflash, draft-dspark, and the other draft and n-gram types). draft_dflash and draft_dspark map there, and the draft layer count maps to --spec-draft-ngl.
• llama.cpp --reasoning-preserve is an on/off flag pair. Reasoning preservation is on by default since b10763; setting preserve_reasoning through chat template kwargs is deprecated upstream.
• stable-diffusion.cpp --auto-fit takes on or off and is on by default since master-845-80bac2d. --stream-layers was removed in master-843-462d675; sdstreamlayers and sdstreaming are kept as legacy keys and no longer emitted.
• stable-diffusion.cpp flag names are matched exactly. The router emits --llm_vision, --clip_vision, --embeddings-connectors, --hires-upscalers-dir, --circularx and --circulary.
• whisper-server has no --no-context flag; the server never carries context between requests. whispercpp_no_context is a legacy key.

Per-feature minimums

• llama.cpp:
  • --n-cpu-ffn (n_cpu_ffn) needs b10645.
  • --video-fps, --video-timestamp-interval and --video-ffmpeg-dir need b10647. OpenAI video_url content and data: video URIs need b11136. The router passes --video-ffmpeg-dir when its own ffmpeg probe succeeded and the config has a multimodal projector.
  • --kv-unified-per-slot needs b10662.
  • --lazy-mode (lazy_mode) needs b10700; it replaced --tensor-read-lazy.
• stable-diffusion.cpp:
  • LTX-2.5 needs master-833-afd5306.
  • SenseNova U1.5 needs master-857-7f986a9.
  • Wan2.2 S2V (audio plus image to video) needs master-863-0bd72f0 and sdaudioencoder (--audio-encoder). sd-server has no HTTP audio input yet, so ComfyUI video requests with audio references are rejected on the llama_sdcpp backend.
  • External tokenizers (sdtokenizer, --tokenizer) need master-867-4964abd. PiD and Lens models require one.
  • Qwen Image 2.1 (RGBA input and output, qwen_image_2_1_prefix_cache model argument) needs master-883-137f740.
  • LLaDA-Image (llada_image scheduler) needs master-886-15f335d.
  • SageAttention (sdsageattention, --sage-attn) needs master-889-c678dfe.
  • image_preprocess needs master-899-28b454b. It replaces the removed auto_resize_ref_image request field together with ref_image_args (resize_before_vae=false).
  • Conditioning cache (sdconditioningcachesize) needs master-902-e6281b6.
• KoboldCpp:
  • usedirectio needs v1.120.
  • ffncpu, reasoningeffort: xhigh, MiniMax-H3 media references (several reference images plus audio clips) and video LoRAs need v1.121.
  • v1.121 renamed GET /sdapi/v1/get_last.png to GET /sdapi/v1/get_last.json, which returns the last generation as JSON. It also removed the reverse_refimg image field and added video_start_frame and video_end_frame.
  • From v1.121, image generation with keepalive, with keep_image_gen_on_disconnect, or behind Cloudflare (CF-Connecting-IP) answers with a close-delimited JSON body padded by whitespace lines. The router forwards the padding immediately.
• whisper.cpp v1.9.4 enables token timestamps only for verbose_json, max_len or split_on_word, and its language detection response includes language.
• vLLM:
  • POST /v1/messages/render and POST /cohere/v2/chat/render need 0.29.0; POST /v1/responses/render needs 0.30.0. From 0.30.0 the render endpoints, /derender and /inference/v1/generate are only registered when the model's serve_args include --enable-scale-out. The router proxies the three render endpoints and never /derender or /inference/v1/generate.
  • 0.29.0 deprecated python -m vllm.entrypoints.openai.api_server; the router launches vllm serve with one API server process.
  • 0.29.0 removed the Arctic, Chameleon, Cheers, Fairseq2Llama, FireRedLID, GritLM, HCXVision, MPT and PrithviGeoSpatialMAE architectures and the RWForCausalLM and StableLMEpochForCausalLM aliases.
  • 0.30.0 removed the VLLM_PREFIX_CACHE_RETENTION_INTERVAL and VLLM_MM_HASHER_ALGORITHM environment variables; 0.29.0 removed VLLM_TEST_FORCE_FP8_MARLIN and VLLM_ROCM_USE_AITER_FP4_ASM_GEMM.
• MiniMax-H3 video generation requires KoboldCpp v1.119 or stable-diffusion.cpp master-812-ea7f0c8 or newer. H3 is video-only: both backends reject it for still-image generation. sdaudiovae (--audio-vae) must be set for H3's audio track; KoboldCpp only preserves that audio in its MJPG-AVI video output, not its GIF output.
• minimax-01 GGUF architecture (MiniMaxText01ForCausalLM, MiniMaxM1ForCausalLM) requires llama.cpp b10437 or newer, or a KoboldCpp release built from it (v1.119 or newer). vLLM removed both architectures in 0.23.0.
• --model-vocoder and --model-talker were removed from llama.cpp on 2026-08-04 (TTS moved to the llama-tts CLI via mtmd). The router no longer emits either flag, and .kcpps fields that depended on them (ttsmodel, ttswavtokenizer, talkermodel, code2wavmodel) are rejected under backend_mode: llama_sdcpp. Use kobold or vllm for text-to-speech until a llama-tts integration exists. llama-server has no /v1/audio/speech endpoint.
• POST /v1/images/edits and GET /sdapi/v1/progress are KoboldCpp endpoints since v1.119.
• KoboldCpp v1.118.1 removed Row Split (--splitmode row on CUDA); v1.117.1 was the last version supporting it, and the RPC protocol changed to match upstream llama.cpp's.

Virtual Jinja kwargs profiles require KoboldCpp v1.114.1 or newer and chat_template_kwargs support from the listed llama.cpp baseline. The listed KoboldCpp baseline satisfies that minimum.

Known, not adopted

• llama.cpp --host with several addresses: the router binds each runtime to one loopback address.
• llama.cpp --log-jsonl: readiness and load error detection read the text log.
• llama.cpp LLAMA_ARG_TEMPERATURE and the other sampling environment variables: the router passes sampling settings on the command line.
• llama.cpp --spec-synth-len and --spec-synth-rates: benchmarking only.
• KoboldCpp --analyze for safetensors files, the reworked MusicUI, the Stable UI LoRA selector and the KCPP_LORA container variable: served or used by KoboldCpp itself.
• vLLM features are deferred to a later router release: admission control (--max-num-queued-reqs, --max-num-queued-tokens), per-request speculative decoding metrics, reasoning_tokens usage reporting, --engram-config, --grpc, and the new model families (DeepSeek-V4.1-Flash, GLM-5.3-Flash, K2-Horizon, Qwen3.8-Flash-Next). Flags without a dedicated .kcpps field can still be passed through vllm.serve_args.