Repository navigation
Expand file tree
/
Copy path.env.example
More file actions
88 lines (76 loc) · 4.87 KB
/
Copy path.env.example
File metadata and controls
88 lines (76 loc) · 4.87 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
# Settings for compose.yaml. Copy to .env and edit; .env is not committed (it can hold a token).
# A missing or empty value uses the default shown.
# --- Who the container runs as: your own user, so ./data and the model cache stay yours.
# Find yours with: id -u and id -g
STUDIO_UID=1000
STUDIO_GID=1000
# --- Where it listens. 0.0.0.0 = reachable from your LAN at http://<spark-address>:8080.
# 127.0.0.1 = only on the Spark itself (use an SSH tunnel or an NVIDIA Sync custom app).
STUDIO_BIND=0.0.0.0
STUDIO_PORT=8080
# Optional DNS-rebinding guard: the host names/addresses you browse to, comma-separated,
# e.g. spark-1a2b.local,192.168.1.20 (localhost always works). Empty = accept any.
STUDIO_ALLOWED_HOSTS=
# --- Model
STUDIO_MODEL=Qwen/Qwen-Image-2.1
# Hugging Face cache folder on the Spark (default: ~/.cache/huggingface, the usual shared one).
# HF_CACHE_DIR=/home/you/.cache/huggingface
# Only needed if the model repo is gated or private.
HF_TOKEN=
# Where model files come from. auto (the default) = use the local cache and touch the network only if files are
# missing from it, so after the first download nothing is fetched again, even if the model's authors publish an
# update. true = never contact Hugging Face (a missing file is an error). false = ask the hub on every load, as
# versions before 1.8 did (and download an update if there is one).
STUDIO_LOCAL_FILES_ONLY=auto
# --- Music (the Music tab, MiniMax-Music3; see docs/SPARK_TEST.md section 20)
# The first load downloads about 29 GB (up to 57 GB if the whole repository is fetched). Have 60 GB free.
STUDIO_MUSIC_MODEL=MiniMaxAI/MiniMax-Music3
# Refuse to load the music model unless this many GB are free (it needs about 24 GB). 40 is an estimate. 0 = no check.
STUDIO_MUSIC_MIN_FREE_GB=40
# The longest track you may ask for, in seconds (10-360, the model's own limit), and how many versions one request may make.
STUDIO_MUSIC_MAX_SECONDS=300
STUDIO_MUSIC_MAX_TRACKS=4
# --- Memory (the Spark's 128 GB is shared by the CPU, the GPU and everything else, Hermes included)
# Refuse to load the model unless at least this many GB are free (MemAvailable). 40 is an estimate,
# to be replaced by the number you measure (docs/SPARK_TEST.md). 0 = no check.
STUDIO_MIN_FREE_GB=40
# Unload the model after this many idle minutes, to give the memory back. 0 = right away.
STUDIO_IDLE_TIMEOUT_MIN=30
# true = keep parts of the model in CPU memory between uses (slower; rarely useful on the Spark).
STUDIO_CPU_OFFLOAD=false
# --- History
STUDIO_QUEUE_CAP=10
# Delete finished runs older than this many days, with their images; 0 = never. Checked at start-up
# and then once a day. Runs you press Keep on, and runs still waiting or running, are never deleted.
# A card warns when fewer than 7 days remain. Such runs, and runs you delete, go to the bin (below) first.
STUDIO_RETENTION_DAYS=30
# How many days a deleted run stays in the bin (the Deleted view) before it is deleted for good, with its files; you can restore it
# until then, or empty the bin yourself. 0 = no bin: Delete and the clean-up delete for good at once.
STUDIO_BIN_DAYS=30
# --- Drafts (the Draft button): a small, quick try of your prompt that jumps ahead of waiting full-size runs.
# The long side in pixels (256-1024, a multiple of 32), and the most steps a draft may use.
STUDIO_DRAFT_SIZE=512
STUDIO_DRAFT_STEPS=12
# --- Editing (the Edit mode on the page; scripts/edit_via_api.sh drives the same API from a terminal)
# How many images one edit may use, 1-10. Qwen allows ten; 4 until memory and time are measured on the Spark.
STUDIO_MAX_INPUT_IMAGES=4
# An uploaded image that no edit has used is deleted after this many hours.
STUDIO_UPLOAD_TTL_HOURS=24
# The largest single upload, in MB (an image may also have at most 16 megapixels).
STUDIO_MAX_UPLOAD_MB=20
# The page shows a "this edit is heavy" warning above this many units, where one unit is one image at 1K and
# the cost grows as images x (resolution / 1024)^2 (so 4 images at 2K is 16). 8 is a guess until the Spark has
# measured it (SPARK_TEST.md section 18); 0 turns the warning off.
STUDIO_EDIT_WARN_UNITS=8
# --- Enlarge (the page's Enlarge button: the same picture bigger and sharper, with an upscaler model; DESIGN.md section 28)
# The model file is yours to download, once (the studio never fetches it; the model has its own terms, read them):
# mkdir -p ~/.cache/huggingface/upscalers
# curl -L -o ~/.cache/huggingface/upscalers/RealESRGAN_x2plus.pth \
# https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth
# Leave this empty to use that file (the container sees it as /models/upscalers/RealESRGAN_x2plus.pth); or name another
# x2 upscaler file, as the container sees it.
STUDIO_UPSCALER_MODEL=
# auto (the GPU if there is one), cuda, or cpu (minutes per picture).
STUDIO_UPSCALER_DEVICE=auto
# --- Build: NVIDIA PyTorch release (nvcr.io/nvidia/pytorch:<tag>). 25.10 or newer for the Spark.
NGC_TAG=25.10-py3