Nv Generate Ct Rflow

skill_manifest.yaml

# SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

id: medagent.nv_generate_ct_rflow
version: 0.1.0
upstream_refs:
  - kind: github_repo
    name: NVIDIA-Medtech/NV-Generate-CTMR
    repo_url: https://github.com/NVIDIA-Medtech/NV-Generate-CTMR
    git_commit: 61c4ec709b84cad468852243c48e250bec732074
    notes: Recommended default for a new checkout; an explicit caller-provided NV_GENERATE_ROOT is accepted.
  - kind: huggingface_repo
    name: nvidia/NV-Generate-CT
    repo_id: nvidia/NV-Generate-CT
    revision: 75ac080fb1083c403793563477724c038e7d430c
license: Apache-2.0
intended_use:
  summary: >
    Engineering-time wrapper around NVIDIA-Medtech/NV-Generate-CTMR's
    rectified-flow CT/mask synthesis pipeline (rflow-ct). Invokes
    `python -m scripts.inference --version rflow-ct` exactly as the upstream
    README documents; does not reimplement diffusion, sampling, or
    autoencoder decoding.
  scope: development
  not_for:
    - clinical deployment
    - clinical interpretation
    - autonomous diagnosis
    - regulatory submission
    - training data for production models

inputs:
  - name: config_infer_override
    type: file_path
    formats:
      - json
    description: >
      JSON file overriding upstream configs/config_infer.json (allowed keys:
      num_output_samples, body_region, anatomy_list,
      controllable_anatomy_size, output_size, spacing, num_inference_steps,
      and a handful of upstream-defined inference knobs). The sentinel value
      `default` selects the upstream config verbatim.

outputs:
  - name: synthetic_ct_volumes
    type: directory_path
    description: >
      Output directory containing sample_<timestamp>_image.nii.gz and
      sample_<timestamp>_label.nii.gz pairs.
  - name: result_json
    type: json
    schema: validators/output_schema.json

runtime:
  language: python
  python: ">=3.10"
  entrypoint: scripts/run_rflow_ct.py
  args:
    - "${python}"
    - "${script}"
    - "${fixture}"
    - "--output-dir"
    - "${out}/samples"
  dependencies:
    nibabel: ">=4.0"
    numpy: ">=1.24"
    typer: ">=0.9"
    torch: ">=2.1"
    monai: ">=1.5"
  side_effects:
    pip_packages:
      - nibabel>=4.0
      - numpy>=1.24
      - typer>=0.9
      - torch>=2.1
      - "monai>=1.5"
      - scipy>=1.10
      - scikit-image>=0.20
      - einops>=0.7
      - huggingface_hub>=0.20
      - matplotlib>=3.8
      - tqdm>=4.65
      - fire>=0.5
      - tensorboard>=2.14
      - PyYAML>=6.0
    local_writes:
      # Staged configs land under <output-dir>/_staged_configs/ so the
      # NV_GENERATE_ROOT clone is never mutated.
      - {path: "<caller-provided --output-dir>", approx_mb_max: 4000}
    home_writes:
      - {path: ~/.cache/huggingface/, approx_mb_max: 6000}
    network_endpoints:
      - https://huggingface.co
      - https://github.com
    requires_docker: false
    requires_gpu: cuda
    # rflow-ct is a 3D rectified-flow diffusion model: there is no CPU fallback
    # in the upstream code path. eval_engine should skip on CPU-only hosts.
    environment:
      clean_environment_required: false
      clean_environment_recommended: true
      modifies_active_python_environment: true
      user_environment_modification_ok: true
      recommended_isolation: fresh venv or container for benchmarks; caller-selected env for interactive use
      notes: >
        Runtime setup commands may install packages into the active Python
        environment. That is acceptable only when the caller chooses that
        environment; benchmark and evidence runs should use a fresh per-run
        environment.
    env_required:
      - NV_GENERATE_ROOT
  external_assets:
    - kind: upstream_repo
      repo_url: https://github.com/NVIDIA-Medtech/NV-Generate-CTMR
      install_path: ~/.cache/nvidia-skills/upstreams/NV-Generate-CTMR-61c4ec7
      install_command: >
        git clone https://github.com/NVIDIA-Medtech/NV-Generate-CTMR.git $NV_GENERATE_ROOT &&
        git -C $NV_GENERATE_ROOT checkout --detach 61c4ec709b84cad468852243c48e250bec732074 &&
        pip install -r $NV_GENERATE_ROOT/requirements.txt
      contains:
        - scripts/inference.py
        - scripts/download_model_data.py
        - configs/config_network_rflow.json
        - configs/config_infer.json
        - configs/environment_rflow-ct.json
    - kind: huggingface_repo
      repo_id: nvidia/NV-Generate-CT
      size_mb_approx: 5500
      install_path: $NV_GENERATE_ROOT/models/
      install_command: >
        cd $NV_GENERATE_ROOT && python -m scripts.download_model_data
        --version rflow-ct --root_dir ./
      contains:
        - models/autoencoder_v1.pt
        - models/diff_unet_3d_rflow-ct.pt
        - models/controlnet_3d_rflow-ct.pt
        - models/mask_generation_autoencoder.pt
        - models/mask_generation_diffusion_unet.pt
        - datasets/all_anatomy_size_conditions.json
        - datasets/candidate_masks_flexible_size_and_spacing_4000.json
        - datasets/all_masks_flexible_size_and_spacing_4000/

paired_verifiers:
  - id: medagent.verifiers.ct_synthesis_quality_v1
    status: implemented
    consumes: evidence_pack_dir
    purpose: >
      Reads the produced image/mask pairs and checks geometry consistency,
      CT-HU plausibility (image has negative-HU air voxels and bone-range
      bright voxels, image is non-constant), mask label-set sanity (at
      least one foreground class present, all label ids in [0, 132]),
      declared output-label coverage, and lung-lobe HU plausibility. For
      controllable generation, the saved paired label map uses upstream's
      local 1..N output-label ids; wrapper evidence maps those ids back to
      MAISI label ids such as lung tumor=23.

limitations:
  - >
    This is a thin wrapper. Inference, sampling, and decoding are delegated
    entirely to NVIDIA-Medtech/NV-Generate-CTMR's `scripts.inference`. Do not
    modify code under $NV_GENERATE_ROOT.
  - >
    rflow-ct requires CUDA and ≈ 16 GB VRAM minimum for the default 256³
    output_size. Larger output_size (e.g. 512×512×768) needs an A100/H100.
  - >
    Output volumes are synthetic. They are not safe to use as training data
    for production medtech models without an independent quality review.

validation:
  expected_runtime_seconds:
    min: 1.0
    max: 1800.0
    inference_path: runtime.subprocess_seconds
  sanity_checks:
    - {path: invocation.exit_code, eq: 0}
    - {path: invocation.model_inventory.all_present, eq: true}
    - {path: output.num_samples, gte: 1}
    - {path: output.all_pairs_readable, eq: true}
    - {path: output.all_geometry_consistent, eq: true}
    - {path: output.any_foreground_present, eq: true}
    - {path: output.all_images_nonconstant, eq: true}
    - {path: output.all_images_hu_like, eq: true}
    - {path: output.all_effective_anatomy_labels_present, eq: true}
  expected_cost:
    wall_seconds:        {max: 1800}
    cpu_seconds:         {max: 3600}
    rss_mb_peak:         {min: 500, max: 32000}
    gpu_seconds:         {max: 1800}
    gpu_memory_mb_peak:  {max: 48000}
  reproducibility:
    mode: preflight
    fixture: fixtures/default_config_infer.json
    runs: 2
    reason: >
      End-to-end synthetic CT repeatability requires CUDA, the
      NV-Generate-CTMR checkout, and downloaded model weights. The repository
      audit repeats the declared config/env boundary check; promoted evidence
      packs must compare generated image and mask artifact hashes.