# kromoss-image: an image model inside a capsule.
#
# What this example adds to model-for-opencode:
#   - the image.generate kind: m64's front lets POST /v1/images/generations
#     through, and only `prompt` and `seed` in its body;
#   - [params]: the settings the server bakes in, published so callers (and the
#     site) can see them; a caller may send the same value, never another;
#   - [routes] extra: one more path, /v1/config, that the front lets through;
#   - a Dockerfile on NVIDIA's PyTorch image, with every pip package pinned.
schema = "m64/capsule@1"
description = "An image model in a capsule: Kromoss with its 8-step DMD2 LoRA on diffusers, text to 1024x1024 PNG in about 19 s on a GB10. Shows the image.generate kind, [[artifacts]] for 40 GB of weights, [params] and an extra route."
kinds = ["image.generate"]
authors = ["motif42"]
license = "mit"

[hardware]
# Built and measured on an NVIDIA GB10 (DGX Spark class, ARM). The base image is
# also built for x86_64; on a big x86 card set platform and gpu to match it.
platform = "linux-aarch64"
gpu = "NVIDIA GB10"
gpu_count = 1
# ~10 GB transformer + ~17 GB Qwen3.5-9B text encoder (bf16) + fp32 VAE and LoRA.
# On unified memory (GB10) m64 adds ram_gb to this; here it is all in vram_gb.
vram_gb = 40
driver = ">=580"
disk_gb = 40

[runtime]
sandbox = "container"
# No image: m64 builds the Dockerfile (nvcr.io/nvidia/pytorch pinned by digest).
gpu = true
# PyTorch's data loaders and CUDA IPC use shared memory: give it more than the 1 GB default.
shm_gb = 8
internal_port = 9219
health = "/health"
# Reading ~40 GB of weights and moving them to the GPU takes minutes.
expected_load_seconds = 420
model_name = "kromoss-dmd2"
user = "65534:65534"

# The weights: one Hugging Face repo at a fixed commit, only the files the
# diffusers pipeline needs (the repo also has ComfyUI repackings and card images).
# m64 downloads them on the first serve (or `m64 install`), verifies them, and
# mounts them read-only at /weights/kromoss. With models_dir set
# (`m64 config set models_dir /mnt/models`), they live there, shared by capsules.
[[artifacts]]
type = "hf"
repo = "lodestones/Kromoss"
revision = "7f9bb7b06d7b6c27d294b0bab5a2b3056c379279"
include = [
  ".gitattributes", "LICENSE", "README.md", "model_index.json", "*.py",
  "8-steps-*.safetensors",
  "scheduler/*", "text_encoder/*", "tokenizer/*", "transformer/*", "vae/*",
]
mount = "/weights/kromoss"

# Baked into kromoss_server.py; returned as `params_used` with every image.
[params]
size = "1024x1024"
num_inference_steps = 8
guidance_scale = 1.0
shift = 5.0
lora = "8-steps-dmd2-lora"

# /v1/config describes the fixed settings and the accepted fields. The front
# lets it through; every other path that isn't a kind route is a 404.
[routes]
extra = ["/v1/config"]