schema = "m64/capsule@1"
description = "MOSS-Image-Base + data-archetype's 8-step DMD2 LoRA text-to-image on GB10 (diffusers). OpenMOSS's MOSS-Image-Base with data-archetype/moss_image_base_dmd2 (rank 128, applied unmerged) and the fresh-noise sampler it was trained for: 1024x1024, 8 steps, CFG 1, shift 4.4817, bf16. Runs fully offline."
kinds = ["image.generate"]
authors = ["dh7"]
license = "apache-2.0"

[hardware]
platform = "linux-aarch64"
gpu = "NVIDIA GB10"
gpu_count = 1
vram_gb = 40
driver = ">=580"
disk_gb = 41

[runtime]
sandbox = "container"
# No image: the Dockerfile builds on nvcr.io/nvidia/pytorch:25.09-py3 + diffusers 0.41.0, pinned by digest.
gpu = true
shm_gb = 8
internal_port = 9223
health = "/health"
expected_load_seconds = 600
model_name = "moss-image-base-dmd2"
user = "65534:65534"

[[artifacts]]
type = "hf"
repo = "OpenMOSS-Team/MOSS-Image-Base"
revision = "85975793d8b3d432f1b6e27f352abaa4c2a4c0d3"
include = [
  ".gitattributes",
  "LICENSE",
  "README.md",
  "model_index.json",
  "*.py",
  "scheduler/*",
  "text_encoder/*",
  "tokenizer/*",
  "transformer/*",
  "vae/*",
]
mount = "/weights/moss-image-base"

[[artifacts]]
type = "hf"
repo = "data-archetype/moss_image_base_dmd2"
revision = "969b5a36cc7eff434cea4b53993baa88218276c1"
include = [".gitattributes", "LICENSE", "README.md", "transformer/*"]
mount = "/weights/moss-base-dmd2"

# The server's own defaults (moss_image_server.py), which every request gets: callers can only send
# prompt and seed. Returned as params_used.
[params]
size = "1024x1024"
num_inference_steps = 8
guidance_scale = 1.0
shift = 4.4817
sampler = "fresh-noise"
lora = "moss_image_base_dmd2"

[routes]
extra = ["/v1/config"]