config.reference.llmc

# llmconfig — Full Reference Config
# Every supported field is listed with its default value and a description.
# Copy this file into your llmconfig configs directory as <name>.yaml and
# remove what you don't need. Paths are shown as ~/.llmconfig/... for
# brevity; on Windows that resolves to %USERPROFILE%\.llmconfig\...
# Set LLMCONFIG_CONFIG_DIR to override the base directory.
#
# Minimal required fields: version, name, model.source, model.repo + model.file (or model.path / model.url)
# Backend-specific fields (whisper:, sd:) only apply when backend is set accordingly.

# ── Identity ──────────────────────────────────────────────────────────────────

version: 1                        # config schema version, always 1
name: my-model                    # unique name used in all CLI commands
description: ""                   # optional human-readable description
tags: []                          # optional labels, e.g. [instruct, 7b, gguf]

# ── Backend ───────────────────────────────────────────────────────────────────
# Which inference engine to run for this model.
#   llama   → text generation (llama.cpp) — default
#   sd      → image generation (stable-diffusion.cpp)
#   whisper → speech recognition (whisper.cpp)
backend: llama

meta:
  author: ""                      # optional author name
  created_at: ""                  # optional creation date (free-form string)
  updated_at: ""                  # optional last update date
  notes: ""                       # optional free-form notes

# ── Mode ──────────────────────────────────────────────────────────────────────

# server      → starts llama-server, exposes an OpenAI-compatible HTTP API
# interactive → starts llama-cli, opens a terminal chat session (no HTTP server)
mode: server                      # default: server

# ── Model ─────────────────────────────────────────────────────────────────────

model:
  # source controls where the model file comes from
  # huggingface → download from huggingface.co (requires repo + file)
  # url         → download from a direct URL (requires url)
  # local       → use a file already on disk (requires path)
  source: huggingface

  # --- HuggingFace (source: huggingface) ---
  repo: bartowski/google_gemma-4-E2B-it-GGUF   # HuggingFace repo in user/repo format
  file: google_gemma-4-E2B-it-Q4_K_M.gguf      # exact filename inside the repo

  # --- Direct URL (source: url) ---
  # url: https://example.com/model.gguf

  # --- Local file (source: local) ---
  # path: /absolute/path/to/model.gguf

  # Optional: sha256 checksum for integrity verification
  # Format: "sha256:<hex>"  e.g. sha256:abc123...
  checksum: ""

  download:
    resume: true                  # resume interrupted downloads (default: true)
    connections: 4                # parallel download connections (default: 4)
    verify_checksum: true         # verify checksum after download (default: true; set to false to skip)
    model_dir: ""                 # override models directory (default: ~/.llmconfig/models)

  # --- Speculative decoding (optional) ---
  # Requires a smaller draft model from the same family.
  # draft:
  #   source: huggingface
  #   repo: bartowski/some-draft-GGUF
  #   file: draft-model-Q4_K_M.gguf
  #   draft_n: 5                  # tokens to draft ahead (default: 5)
  #   draft_min: 0                # minimum draft tokens (default: 0)
  #   draft_p_min: 0.75           # minimum speculative decoding probability (default: 0.75)
  #   n_ctx: 0                    # draft model context size (0 = same as main model)
  #   n_gpu_layers: 0             # draft model GPU layers (0 = same as main model)
  #   devices: []                 # draft model GPU devices
  #   cache_type_k: f16           # draft model KV cache type K
  #   cache_type_v: f16           # draft model KV cache type V
  #   spec_replace_target: ""     # speculative decoding string replacement target
  #   spec_replace_draft: ""      # speculative decoding string replacement draft
  #   override_tensor: []         # tensor buffer overrides for draft model
  #   cpu_moe: false              # keep all MoE weights in CPU for draft (default: false)
  #   n_cpu_moe: 0                # keep first N layers MoE in CPU for draft (0 = not set)
  #   threads_draft: 0            # draft model generation threads (0 = same as --threads, server only)
  #   threads_batch_draft: 0      # draft model batch threads (0 = same as threads_draft, server only)
  #   spec_type: ""               # speculative decoding type: none|ngram-cache|ngram-simple|ngram-map-k|...
  #   spec_ngram_size_n: 0        # ngram lookup size N (0 = use default 12)
  #   spec_ngram_size_m: 0        # ngram draft size M (0 = use default 48)
  #   spec_ngram_min_hits: 0      # min hits for ngram-map (0 = use default 1)

  # --- LoRA adapters (optional) ---
  # lora: []                      # LoRA adapter file(s), e.g. ["./adapter.gguf"]
  # lora_scaled: []               # LoRA with scaling, format: ["./adapter.gguf:0.8"]

  # --- Control vectors (optional) ---
  # control_vector: []            # control vector file(s)
  # control_vector_scaled: []     # control vector with scaling, format: ["./vector.gguf:0.8"]
  # control_vector_layer_start: 0 # layer range start for control vector (default: not set)
  # control_vector_layer_end: 0   # layer range end for control vector (default: not set)

  # --- Model metadata overrides (advanced) ---
  # override_kv: []               # override model metadata, e.g. ["tokenizer.ggml.add_bos_token=bool:false"]

  # --- Multimodal projection (optional, for vision models) ---
  # mmproj:
  #   source: huggingface
  #   repo: user/repo-GGUF
  #   file: mmproj-model.gguf
  #   offload: true               # GPU offload for mmproj (default: true)

# ── Server ────────────────────────────────────────────────────────────────────
# Only applies when mode: server

server:
  host: 127.0.0.1                 # bind address (default: 127.0.0.1)
  port: 8080                      # listen port (default: 8080)
  api_key: ""                     # bearer token for Authorization header (default: none)
  # api_key_file: ""              # path to file containing API keys (one per line)
  parallel: 1                     # number of concurrent request slots (default: 1)
  cors_origins: []                # allowed CORS origins, e.g. ["https://myapp.com"]
  read_timeout: 600s              # llmconfig internal timeout (default: 600s)
  write_timeout: 600s             # llmconfig internal timeout (default: 600s)
  # timeout: 0                    # llama-server read/write timeout in seconds (0 = not set)
  # threads_http: -1              # HTTP request handler threads (-1 = auto)
  # reuse_port: false             # allow multiple sockets on same port (default: false)
  # path: ""                      # directory to serve static files from
  # api_prefix: ""                # API path prefix, without trailing slash

  # SSL / HTTPS
  # ssl_key_file: ""              # path to PEM-encoded SSL private key
  # ssl_cert_file: ""             # path to PEM-encoded SSL certificate

  # Prompt cache
  # cache_prompt: true            # prompt caching (set false to disable, default: true)
  # cache_reuse: 0                # min chunk size for KV-shift cache reuse (0 = disabled)

  # Slots
  # slot_prompt_similarity: 0.10  # slot matching threshold (0.0 = disabled, default: 0.10)
  # slot_save_path: ""            # directory to save slot KV cache
  # sleep_idle_seconds: 0         # sleep after idle seconds (-1 = disabled, 0 = not set)
  # lora_init_without_apply: false # load LoRA adapters without applying (apply via POST /lora-adapters)

  # KV
  # kv_unified: true              # unified KV buffer shared across slots (set false to disable)
  # clear_idle: true              # save and clear idle slots on new task (set false to disable)

  # Batching
  # cont_batching: true           # continuous batching / dynamic batching (set false to disable)

  # Model identification
  # alias: ""                     # model name aliases, comma-separated (used by API)
  # tags: ""                      # model tags, comma-separated (informational)

  # Web UI
  # webui: true                   # enable Web UI (set false to disable)
  # webui_config: ""              # JSON string for default WebUI settings
  # webui_config_file: ""         # path to JSON file for WebUI settings
  # webui_mcp_proxy: false        # experimental: MCP CORS proxy (do not enable in untrusted env)
  # tools: ""                     # experimental: built-in AI agent tools, comma-separated or "all"

  # Embeddings
  # pooling: ""                   # pooling type for embeddings: none | mean | cls | last | rank

  # Infill
  # spm_infill: false             # use Suffix/Prefix/Middle pattern (default: Prefix/Suffix/Middle)

  # Response behavior
  # prefill_assistant: true       # prefill assistant response if last message is assistant (set false to disable)

  # Media
  # media_path: ""                # directory for local media files accessible via file:// URLs

  # Router server (for multi-model serving)
  # models_dir: ""                # directory containing models
  # models_preset: ""             # path to INI file with model presets
  # models_max: -1                # max simultaneous models (-1 = not set, 0 = unlimited, default: 4)
  # models_autoload: true         # auto-load models (set false to disable)

  # Lookup cache (for lookup decoding)
  # lookup_cache_static: ""       # path to static lookup cache (not updated by generation)
  # lookup_cache_dynamic: ""      # path to dynamic lookup cache (updated by generation)

  endpoints:
    metrics: false                # enable Prometheus metrics endpoint (default: false)
    slots: true                   # when false, disables slots monitoring endpoint
    embeddings: false             # restrict to embedding use case only (default: false)
    rerank: false                 # enable reranking endpoint (default: false)
    props: false                  # enable POST /props for dynamic property changes (default: false)

# ── Hardware Profiles ─────────────────────────────────────────────────────────
# llmconfig auto-selects the matching profile at runtime based on detected hardware.
# Override at startup with: llmconfig up <name> --profile cpu
#
# Profile selection order:
#   apple_silicon → macOS + ARM64
#   nvidia        → NVIDIA GPU detected (nvidia-smi)
#   amd           → AMD GPU detected (sysfs / wmic)
#   intel_gpu     → Intel GPU detected (sysfs)
#   cpu           → fallback

hardware_profiles:

  apple_silicon:
    n_gpu_layers: 99              # layers to offload to GPU (-1 = all, 0 = CPU only)
    metal: true                   # enable Apple Metal backend
    threads: 8                    # CPU threads for non-GPU layers
    threads_batch: 8              # CPU threads for batch processing

  nvidia:
    n_gpu_layers: 99
    cuda: true                    # enable CUDA backend
    threads: 8
    threads_batch: 8
    # devices: ["0"]              # specific CUDA device IDs (default: all)
    # tensor_split: [0.5, 0.5]   # VRAM split across multiple GPUs (must sum to 1.0)
    # split_mode: layer           # multi-GPU split: none | layer | row | tensor (default: layer)
    # main_gpu: 0                 # primary GPU index for split-mode none/row (default: 0)
    # priority: 0                 # process priority: -1=low, 0=normal, 1=medium, 2=high, 3=realtime
    # fit: on                     # auto-fit layers to device memory: "on" | "off" (default: on)
    # fit_target: [1024]          # target free margin per device in MiB (default: 1024)
    # fit_ctx: 4096               # minimum ctx size --fit can set (default: 4096)
    # override_tensor: []         # tensor buffer type overrides, e.g. ["blk\\..*\\.ffn=CPU"]
    # cpu_moe: false              # keep ALL MoE expert weights in CPU (default: false)
    # cpu_mask: ""               # CPU affinity mask (hex), e.g. "FF00"
    # cpu_mask_batch: ""         # CPU affinity mask for batch processing
    # cpu_range_batch: ""        # CPU range for batch affinity, e.g. "0-3"
    # cpu_strict_batch: false    # strict CPU placement for batch (default: false)
    # poll: 50                   # polling level 0-100 (0=no polling, default: 50)
    # poll_batch: 50             # polling level for batch (default: same as poll)
    # prio_batch: 0              # batch thread priority: 0=normal,1=medium,2=high,3=realtime
    # repack: true               # weight repacking optimization (default: true)
    # no_host: false             # bypass host buffer for extra device buffers (default: false)
    # op_offload: true           # offload host tensor ops to device (default: true)
    # rpc: ""                    # RPC servers for distributed inference, e.g. "host1:50052,host2:50052"
    # direct_io: false           # use DirectIO if available (default: false)

  amd:
    n_gpu_layers: 99
    rocm: true                    # enable ROCm backend
    threads: 8
    threads_batch: 8

  intel_gpu:
    n_gpu_layers: 99
    sycl: true                    # enable SYCL/oneAPI backend
    threads: 8
    threads_batch: 8

  cpu:
    n_gpu_layers: 0               # all layers on CPU
    threads: 8
    threads_batch: 8
    # cpu_range: "0-7"            # restrict to specific CPU cores (Linux)
    # cpu_strict: false           # strict CPU affinity (default: false)
    # numa: ""                    # NUMA strategy: distribute | isolate | numactl | mirror

# ── Context ───────────────────────────────────────────────────────────────────

context:
  n_ctx: 4096                     # context window size in tokens (default: 4096)
  n_batch: 512                    # prompt processing batch size (default: 512)
  n_ubatch: 512                   # physical batch size, <= n_batch (default: n_batch)
  n_keep: 0                       # tokens to keep from initial prompt on context overflow (default: 0)
  cache_type_k: f16               # key cache type: f32 | f16 | bf16 | q8_0 | q4_0 | q4_1 (default: f16)
  cache_type_v: f16               # value cache type: same options as cache_type_k (default: f16)
  mmap: true                      # memory-map model file (default: true)
  mlock: false                    # lock model in RAM, prevent swap (default: false)
  # flash_attention: on          # Flash Attention: "on" | "off" | "auto" (default: auto, recommended: on)
  n_cpu_moe: 0                    # CPU layers for MoE experts (0 = use n_gpu_layers logic)
  # n_predict: 0                  # max tokens to generate (-1 = infinity, 0 = not set)
  # context_shift: false          # enable context shift on infinite generation (default: false)
  # kv_offload: true              # KV cache GPU offload (set false to disable, default: true)
  # swa_full: false               # use full-size SWA cache (default: false)
  # cache_ram: 8192               # max RAM cache in MiB (-1 = no limit, 0 = not set, default: 8192)
  # image_min_tokens: 0           # minimum image tokens for dynamic resolution vision models
  # image_max_tokens: 0           # maximum image tokens for dynamic resolution vision models
  # check_tensors: false          # validate tensor data on load (default: false)
  # ctx_checkpoints: 32           # max context checkpoints per slot (default: 32)
  # checkpoint_every_n_tokens: 0  # checkpoint interval in tokens (-1 = disable, 0 = not set, default: 8192)

# ── Sampling ──────────────────────────────────────────────────────────────────

sampling:
  temperature: 0.8                # randomness; 0 = deterministic, 1 = creative (default: 0.8)
  top_k: 40                       # sample from top K tokens (default: 40)
  top_p: 0.95                     # nucleus sampling probability (default: 0.95)
  min_p: 0.05                     # minimum probability threshold (default: 0.05)
  repeat_penalty: 1.0             # penalize repeated tokens (1.0 = disabled, default: 1.0)
  repeat_last_n: 64               # tokens to look back for repeat penalty (default: 64)
  # presence_penalty: 0.0         # presence penalty — discourages any repeated token (default: 0.0)
  # frequency_penalty: 0.0        # frequency penalty — discourages frequent tokens (default: 0.0)
  # seed: 0                       # RNG seed for reproducible output (0 = random, default: 0)

  # DRY (Don't Repeat Yourself) sampling — suppresses repetitive patterns
  dry_multiplier: 0.0             # DRY penalty multiplier (0 = disabled, default: 0)
  dry_base: 1.75                  # DRY exponential base (default: 1.75)
  dry_allowed_length: 2           # token sequence length before DRY kicks in (default: 2)
  dry_penalty_last_n: -1          # DRY look-back window (-1 = full context, default: -1)

  # Mirostat adaptive sampling (alternative to top_k/top_p)
  # mirostat: 0                   # 0 = disabled, 1 = Mirostat v1, 2 = Mirostat v2
  # mirostat_tau: 5.0             # target entropy (default: 5.0)
  # mirostat_eta: 0.1             # learning rate (default: 0.1)

  # samplers: ""                  # sampler order override, e.g. "top_k;top_p;temp"

  # Dynamic temperature — varies temperature based on entropy
  # dynatemp_range: 0.0           # range around base temperature (0.0 = disabled)
  # dynatemp_exp: 1.0             # exponent for dynamic temperature curve

  # XTC (eXclude Top Choices) sampling — removes high-probability tokens to increase diversity
  # xtc_probability: 0.0          # probability of applying XTC (0.0 = disabled)
  # xtc_threshold: 0.1            # minimum token probability to be excluded

  # Locally typical sampling
  # typical: 0.0                  # typical-p value (0 = not set, 1.0 = disabled)

  # Top-n-sigma sampling
  # top_nsigma: 0.0               # sigma threshold (0 = not set, -1 = disabled)

  # Adaptive-p sampling — selects tokens near a target probability
  # adaptive_target: 0.0          # target probability (0 = not set, -1 = disabled)
  # adaptive_decay: 0.9           # decay rate for target adaptation (default: 0.9)

  # DRY custom sequence breakers (clears defaults: '\n', ':', '"', '*')
  # dry_sequence_breakers: []     # e.g. ["\n", ":"]

  # Sampler sequence shorthand
  # sampler_seq: ""              # simplified sampler order, e.g. "edskypmxt"
  # ignore_eos: false            # ignore end-of-stream token and continue generating (default: false)

  # Backend sampling (experimental)
  # backend_sampling: false       # enable backend sampling (default: false)

  # Structured output — constrains generation to a grammar or schema
  # grammar: ""                   # inline BNF-like grammar string
  # grammar_file: ""              # path to a grammar file
  # json_schema: ""               # inline JSON schema string
  # json_schema_file: ""          # path to a JSON schema file

# ── Chat ──────────────────────────────────────────────────────────────────────

chat:
  system_prompt: ""               # interactive mode only (llama-cli --system-prompt). llama-server dropped
                                  # this flag upstream (ggml-org/llama.cpp#9811) — for mode: server, send
                                  # the system prompt as a "system" role message in each chat request instead.
  template: ""                    # chat template name (leave empty to use model's built-in)
                                  # examples: llama3, chatml, gemma, mistral, phi3
  # jinja: true                   # true=--jinja, false=--no-jinja, omit=use llama.cpp default (enabled)
  template_kwargs: {}             # extra key-value pairs passed as JSON to --chat-template-kwargs
  # reasoning: auto               # thinking/reasoning mode: "on", "off", or "auto" (default: auto)
  # reasoning_budget: -1          # thinking token budget: -1 unlimited, 0 immediate end, N>0 budget
  # reasoning_budget_message: ""  # message injected before end-of-thinking tag when budget exhausted
  # reasoning_format: auto        # thought tag format: none | deepseek | deepseek-legacy (default: auto)
  # template_file: ""             # path to a jinja template file (overrides built-in template)
  # skip_chat_parsing: false      # force pure content parser, output everything in content (default: false)

# ── RoPE (Rotary Position Embedding) ─────────────────────────────────────────
# Only needed to extend context beyond the model's trained limit.

rope:
  scaling: ""                     # scaling type: linear | yarn | longrope (default: none)
  # scale: 0                      # context scaling factor, expands context by N (--rope-scale)
  freq_base: 0                    # RoPE frequency base (0 = use model default)
  freq_scale: 0                   # RoPE frequency scale factor, expands context by 1/N
  # YaRN-specific:
  yarn_ext_factor: -1.0           # extrapolation mix factor (-1 = auto, default: -1.0)
  yarn_attn_factor: 1.0           # attention scaling factor (default: 1.0)
  # yarn_beta_slow: 0             # YaRN high correction dim / alpha (default: -1.0)
  # yarn_beta_fast: 0             # YaRN low correction dim / beta (default: -1.0)
  yarn_orig_ctx: 0                # original context size the model was trained on

# ── Resources ─────────────────────────────────────────────────────────────────

resources:
  vram_limit: ""                  # VRAM cap, e.g. "8GB" (informational, not enforced by llama.cpp)
  vram_buffer: 512MB              # reserved VRAM headroom (default: 512MB)
  memory_limit: ""                # RAM cap, e.g. "16GB" (informational)
  cpu_limit: ""                   # CPU usage cap, e.g. "80%" (informational)
  cpu_priority: normal            # process priority: low | normal | high (default: normal)
  fallback_to_cpu: true           # fall back to CPU if GPU offload fails (default: true)
  request_timeout: 120s           # per-request timeout (default: 120s)
  max_concurrent: 0               # max concurrent requests (0 = unlimited)

# ── Logging ───────────────────────────────────────────────────────────────────

logging:
  level: info                     # log level: debug | info | warn | error (default: info)
  file: ""                        # log file path (default: ~/.llmconfig/logs/<name>.log)
  colors: auto                    # color output: auto | on | off (default: auto)
  # prefix: false                 # enable prefix in log messages (default: false)
  # timestamps: false             # enable timestamps in log messages (default: false)
  # verbosity: -1                 # verbosity threshold: 0=generic,1=error,2=warn,3=info,4=debug (-1 = not set)
  # show_timings: true            # show timing info after each response (set false to disable)

# ── Backend: whisper ──────────────────────────────────────────────────────────
# Only applies when backend: whisper. Remove this block for other backends.

# whisper:
#   # --- Core ---
#   language: auto                 # auto | en | tr | de | fr | ... (default: en)
#   task: transcribe               # transcribe | translate (default: transcribe)
#   processors: 1                  # parallel processor count (default: 1)
#
#   # --- Timing & segmentation ---
#   offset_t: 0                    # time offset in milliseconds
#   offset_n: 0                    # segment index offset
#   duration: 0                    # audio duration to process in ms (0 = all)
#   max_context: -1                # max text context tokens (-1 = all)
#   max_len: 0                     # max segment length in characters
#   audio_ctx: 0                   # audio context size (0 = all)
#   split_on_word: false           # split on word rather than token
#
#   # --- Decoder thresholds ---
#   word_thold: 0.01               # word timestamp probability threshold
#   entropy_thold: 2.40            # entropy threshold for decoder fail
#   logprob_thold: -1.00           # log probability threshold
#   no_speech_thold: 0.60          # no speech threshold
#
#   # --- Sampling ---
#   beam_size: 5                   # beam search width (default: 5)
#   best_of: 5                     # best candidates to keep (default: 5)
#   temperature: 0.00              # sampling temperature (default: 0.0)
#   temperature_inc: 0.20          # temperature increment (default: 0.20)
#   no_fallback: false             # do not use temperature fallback
#
#   # --- Language & prompt ---
#   detect_language: false         # exit after detecting language
#   prompt: ""                     # initial prompt (max n_text_ctx/2 tokens)
#   carry_initial_prompt: false    # always prepend initial prompt (cli only)
#
#   # --- Diarization ---
#   diarize: false                 # stereo audio diarization
#   tinydiarize: false             # tdrz model diarization
#
#   # --- Word-level timestamps ---
#   word_timestamps: false         # shorthand: uses dtw=tiny if dtw is empty
#   dtw: ""                        # DTW model: tiny|base|small|medium|large-v1/v2/v3
#
#   # --- GPU ---
#   no_gpu: false                  # disable GPU
#   device: 0                      # GPU device ID
#   flash_attention: on          # set false to pass --no-flash-attn (default: enabled)
#
#   # --- Suppress ---
#   suppress_nst: false            # suppress non-speech tokens
#   suppress_regex: ""             # regex matching tokens to suppress (cli only)
#
#   # --- Grammar (cli only) ---
#   grammar: ""                    # GBNF grammar to guide decoding
#   grammar_rule: ""               # top-level grammar rule name
#   grammar_penalty: 100.0         # scales down nongrammar token logits
#
#   # --- OpenVINO ---
#   ov_e_device: CPU               # OpenVINO encode device (default: CPU)
#
#   # --- Output formats (cli only) ---
#   output_txt: false
#   output_vtt: false              # .vtt subtitle file
#   output_srt: false              # .srt subtitle file
#   output_lrc: false              # .lrc lyrics file
#   output_csv: false
#   output_json: false
#   output_json_full: false        # include more info in JSON
#   output_file: ""                # output path without extension
#   no_timestamps: false           # do not print timestamps
#
#   # --- Logging ---
#   no_prints: false               # suppress output except results (cli)
#   print_special: false
#   print_colors: false
#   print_confidence: false        # cli only
#   print_progress: false
#   print_realtime: false          # server only
#   log_score: false               # log decoder scores (cli only)
#   debug_mode: false
#
#   # --- VAD (Voice Activity Detection) ---
#   vad: false
#   vad_model: ""                  # path to VAD model file
#   vad_threshold: 0.50            # speech detection threshold
#   vad_min_speech_duration_ms: 250
#   vad_min_silence_duration_ms: 100
#   vad_max_speech_duration_s: 0   # 0 = unlimited
#   vad_speech_pad_ms: 30
#   vad_samples_overlap: 0.10
#
#   # --- Server-only ---
#   public_path: ""                # path to public folder (default: examples/server/public)
#   request_path: ""               # request path prefix
#   inference_path: ""             # inference path (default: /inference)
#   convert: false                 # convert audio to WAV via ffmpeg
#   tmp_dir: ""                    # temp dir for ffmpeg files
#   no_language_probabilities: false # exclude language probs from verbose_json

# ── Backend: sd ───────────────────────────────────────────────────────────────
# Only applies when backend: sd. Remove this block for other backends.

# sd:
#   # --- Output (cli mode only) ---
#   output: ./output.png           # output path; use %d for sequences e.g. output_%03d.png
#   preview_path: ./preview.png    # preview image path
#   preview_interval: 1            # preview update interval in steps
#   preview: none                  # preview method: none | proj | tae | vae
#   mode: img_gen                  # run mode: img_gen | vid_gen | upscale | convert | metadata
#
#   # --- Server (server mode only) ---
#   serve_html_path: ""            # path to HTML file to serve at root
#
#   # --- Model components ---
#   clip_l: ""                     # path to clip-l text encoder
#   clip_g: ""                     # path to clip-g text encoder (SDXL)
#   clip_vision: ""                # path to clip-vision encoder
#   t5xxl: ""                      # path to t5xxl text encoder (SD3/Flux)
#   llm: ""                        # path to LLM text encoder (Flux2/QwenVL)
#   llm_vision: ""                 # path to LLM visual encoder
#   diffusion_model: ""            # path to standalone diffusion model (split model)
#   high_noise_diffusion_model: "" # path to high-noise diffusion model
#   vae: ""                        # path to standalone VAE model
#   taesd: ""                      # path to Tiny AutoEncoder (fast decode, low quality)
#   control_net: ""                # path to ControlNet model
#   embd_dir: ""                   # embeddings directory
#   lora_model_dir: ""             # LoRA model directory
#   tensor_type_rules: ""          # weight type per tensor pattern, e.g. "^vae\.=f16,model\.=q8_0"
#   photo_maker: ""                # path to PHOTOMAKER model
#   upscale_model: ""              # path to ESRGAN upscale model
#
#   # --- Hardware & weight type ---
#   type: ""                       # weight type: f32 | f16 | q4_0 | q4_1 | q8_0 | q2_K | ...
#   rng: cuda                      # RNG: std_default | cuda | cpu
#   sampler_rng: ""                # sampler RNG (default: same as --rng)
#   prediction: ""                 # prediction type: eps | v | edm_v | sd3_flow | flux_flow | flux2_flow
#   lora_apply_mode: auto          # LoRA apply mode: auto | immediately | at_runtime
#   offload_to_cpu: false          # offload weights to RAM, load to VRAM when needed
#   mmap: false                    # memory-map model file
#   control_net_cpu: false         # keep ControlNet in CPU
#   clip_on_cpu: false             # keep CLIP in CPU
#   vae_on_cpu: false              # keep VAE in CPU
#   flash_attention: false         # enable flash attention (--fa)
#   diffusion_fa: false            # flash attention in diffusion model only
#   circular: false                # circular padding for convolutions
#
#   # --- Generation defaults ---
#   width: 512                     # output width in pixels (default: 512)
#   height: 512                    # output height in pixels (default: 512)
#   steps: 20                      # diffusion steps (default: 20)
#   clip_skip: -1                  # CLIP layers to skip (-1 = auto: 1 for SD1.x, 2 for SD2.x)
#   batch_count: 1                 # number of images to generate
#   cfg_scale: 7.0                 # classifier-free guidance scale (default: 7.0)
#   guidance: 3.5                  # distilled guidance scale for Flux/SD3 (default: 3.5)
#   strength: 0.75                 # noise strength for img2img (default: 0.75)
#   control_strength: 0.9          # ControlNet strength (default: 0.9)
#   sampling_method: ""            # euler|euler_a|heun|dpm++2m|dpm++2s_a|lcm|tcd|... (default: auto)
#   scheduler: discrete            # sigma scheduler: discrete|karras|exponential|ays|gits|smoothstep|...
#   negative_prompt: ""            # default negative prompt
#   seed: -1                       # RNG seed (-1 = random, default: 42)
#
#   # --- VAE tiling (reduce VRAM for high-res) ---
#   vae_tiling: false              # process VAE in tiles
#   vae_tile_size: 32x32           # tile size, format WxH
#   vae_tile_overlap: 0.5          # tile overlap fraction (default: 0.5)
#
#   # --- SLG (Skip Layer Guidance, SD3.5) ---
#   slg_scale: 0.0                 # SLG scale (0 = disabled; ~2.5 for SD3.5 medium)
#   skip_layer_start: 0.01         # SLG enabling point (default: 0.01)
#   skip_layer_end: 0.2            # SLG disabling point (default: 0.2)
#   skip_layers: "[7,8,9]"         # layers to skip for SLG steps
#
#   # --- Cache acceleration ---
#   cache_mode: ""                 # easycache|ucache|dbcache|taylorseer|cache-dit|spectrum
#   cache_option: ""               # cache params, e.g. "threshold=0.25,reset=0"
#
#   # --- High noise stage (two-stage generation, e.g. Wan2.2 MoE) ---
#   high_noise_steps: -1           # high noise steps (-1 = auto)
#   high_noise_cfg_scale: 7.0      # high noise guidance scale
#   high_noise_guidance: 3.5       # high noise distilled guidance
#   high_noise_sampling_method: "" # high noise sampling method
#
#   # --- Logging ---
#   verbose: false
#   color: false