quartermaster-generate.example.yaml

Rama main · réplica de lectura del repositorio original.

# Quartermaster autogen control file — Go port of Generate-Config.ps1 params + $Overrides.
# settings: global tuning knobs (CLI --models-dir overrides settings.modelsRoot).
# overrides: per-model tweaks gguf metadata can't supply; first match wins.
#
# Copy this to quartermaster-generate.yaml and edit modelsRoot / serverExe for
# your machine, then add overrides for your own models. The server auto-discovers
# every *.gguf under modelsRoot; overrides are optional tweaks on top.
settings:
  # Folder scanned recursively for *.gguf models. Leave blank to start with an
  # empty catalog and pick a folder later in the setup UI.
  modelsRoot: ""
  # Path to llama.cpp's llama-server executable.
  serverExe: llama-server
  # targetVramGB: the VRAM a model's whole footprint (weights + KV + compute
  # buffers + overhead) must fit under. Left UNSET on purpose: with no value
  # here the budget is measured from your card's free VRAM, by the setup wizard
  # on first run and at every startup otherwise, so a card of any size is used
  # in full. Uncomment it only to pin a ceiling BELOW what the card has - a
  # number written here is read as your deliberate choice and is never
  # re-measured behind your back.
  # targetVramGB: 7
  # autoVram: measure free VRAM at startup and cap targetVramGB at it. The
  # reading is used in full - vramOverheadGB is charged inside the plan, not
  # deducted again here. Only ever tightens: a smaller static targetVramGB above
  # stays a deliberate ceiling. Forces a regen each boot to keep the snapshot
  # current. Leave false to use the static targetVramGB above.
  autoVram: false
  # vramOverheadGB: small slack held back from targetVramGB for allocator
  # fragmentation / CUDA spikes. targetVramGB already accounts for system/desktop
  # VRAM, so this does NOT need to reserve room for the OS — keep it small.
  vramOverheadGB: 0.5
  # maxRamGB: the RAM ceiling a plan may spill CPU-offloaded weights and KV into.
  # Unset and measured exactly like targetVramGB above, from the memory actually
  # available on this machine. A shipped 24 was worse than no value at all on a
  # 16 GB box: a ceiling the machine cannot honour never binds, so the guard that
  # is supposed to stop a plan from swapping stopped nothing. Uncomment only to
  # pin a lower ceiling.
  # maxRamGB: 24
  moeCtxTarget: 65536
  # denseCtxLadder: the TOP rung is the context ceiling the sizer may grow a dense
  # model to; within it the largest window whose KV cache fits the VRAM budget
  # wins. Keep the top at or above the trained window of your largest-context
  # model or that model is pinned below its own ceiling with VRAM left unused.
  denseCtxLadder: [1048576, 524288, 262144, 131072, 65536, 32768]
  denseMinCtx: 32768
  threads: 7
  ttlSec: 600
  healthCheckTimeout: 300
  # groups: split emitted models across named groups, each optionally bound to a
  # separate listen address. Membership is by model-name glob (first match wins,
  # so list specific groups before a "*" catch-all). Models matching nothing go
  # to an implicit "default" group (no listener). All groups are exclusive swap
  # groups, so loading on any port evicts the others (one GPU, VRAM-exclusive).
  # Omit entirely for one group on one port. Example — variants on a second port:
  # groups:
  #   - { name: variants,  listen: 0.0.0.0:1251, match: ["*-game", "*-judge"] }
  #   - { name: assistant, listen: 0.0.0.0:1250, match: ["*"] }
  # extraModels: hand-written model blocks for a server autogen has no backend
  # kind for (any OpenAI-compatible engine, e.g. Strata). Same shape as a
  # config.yaml model entry, copied into the generated config as written so a
  # regen keeps it. Nothing is sized for you: cmd and estVramGB (or estRamGB for
  # a CPU-only server) are required, since the estimate is all the scheduler
  # knows about it. Entries join the swap groups like any discovered model. A
  # bad entry (unknown key, name already taken) is skipped with a
  # "# SKIPPED" comment in config.yaml instead of breaking startup.
  # extraModels:
  #   strata-qwen3-32b:
  #     cmd: >
  #       D:/strata/strata-server.exe --port ${PORT}
  #       --model D:/models/Qwen3-32B
  #     proxy: http://127.0.0.1:${PORT}
  #     checkEndpoint: /v1/models
  #     estVramGB: 20

# Per-model overrides — optional. Each entry matches discovered gguf paths by
# glob (first match wins). Fields: quant (pick a quant when several exist),
# aliases (extra model IDs), spec (e.g. draft-mtp), reasoningFmt, ctxVariants
# (emit extra context-size variants), skip (exclude a match). Examples:
# overrides:
#   - { match: "*-draft-*", skip: true }
#   - { match: "*Qwen3-8B-GGUF*", quant: Q6_K }
#   - { match: "*Gemma-12B-it-GGUF*", quant: Q4_K_M, reasoningFmt: auto, ctxVariants: [4096, 32768, 65536, 131072] }
overrides: []