Skip to content

Constants

constants

Defines deterministic constants and built-in default values for the server

Configuration Split

Environment-dependent configuration is resolved lazily by mirumoji.server.config.get_settings so that a .env loaded at startup is respected

Attributes:

Name Type Description
DB_URL str

Async SQLAlchemy database URL built from HOST_DB_PATH
DEFAULT_SRT_SYS_MSG str

Built-in default system message for SRT-fixing
DEFAULT_BREAKDOWN_SYS_MSG str

Built-in default system message for word-nuance breakdowns

CONVERSION_PRESETS = {'performance': {'x264': ('veryfast', '26'), 'nvenc': ('p2', '30')}, 'balanced': {'x264': ('medium', '23'), 'nvenc': ('p4', '26')}, 'quality': {'x264': ('slow', '20'), 'nvenc': ('p6', '22')}} module-attribute

FFmpeg encoder arguments used when converting video to MP4

A dictionary containing bundles of ffmpeg encoder arguments keyed by the name of a specific preset

Format

Each value is (speed, quality) for one encoder - libx264 -> (-preset, -crf) - h264_nvenc -> (-preset, -cq)

performance
  • Prioritises conversion speed over output quality
  • The fastest encoder presets paired with a lower quality target, for when throughput matters more than fidelity
balanced
  • Balances conversion speed with output quality
  • The default, with mid-range encoder presets and a moderate quality target that suits most playback
quality
  • Prioritises output quality
  • The slowest encoder presets paired with a higher quality target, for when fidelity matters more than encode time

resolution and target_bitrate options stay separate (output geometry and a rate ceiling), so a preset only trades encode speed against quality

DEFAULT_CONVERSION_PRESET = 'balanced' module-attribute

The conversion preset used when a request does not specify one

Names a key of CONVERSION_PRESETS

DEFAULT_MODAL_SCALEDOWN_WINDOW = 60 module-attribute

Default number of seconds an idle Modal GPU container is kept warm before it scales down

Matches Modal's own default. Raising it (via MIRUMOJI_MODAL_SCALEDOWN_WINDOW) keeps a container warm longer, so back-to-back jobs skip the cold start at the cost of more idle GPU time

MAX_LLM_CONCURRENCY = 4 module-attribute

Default number of LLM requests executed simultaneously when requesting an SRT Fix for a batch of files

Overridable at runtime with MIRUMOJI_MAX_LLM_CONCURRENCY

MODEL_DOWNLOAD_BACKOFF_BASE = 2.0 module-attribute

Base number of seconds for the exponential backoff between Whisper model download retries

MODEL_DOWNLOAD_RETRIES = 4 module-attribute

How many times to attempt loading the local Whisper model before giving up

The first load pulls the weights from the Hugging Face Hub, so transient network failures are retried (the partial download resumes between attempts)