Nv Generate Mr

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_mr
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-MR
    repo_id: nvidia/NV-Generate-MR
    revision: b1680c66f2d3152a13a2ebecfd7155341d8fcff9
license: Apache-2.0
intended_use:
  summary: >
    Engineering-time wrapper around NVIDIA-Medtech/NV-Generate-CTMR's
    rflow-mr image-only synthesis workflow. Invokes
    `python -m scripts.diff_model_infer` with the upstream rflow-mr configs
    and summarizes the generated NIfTI volume.
  scope: development
  not_for:
    - clinical deployment
    - clinical interpretation
    - autonomous diagnosis
    - regulatory submission
    - training data for production models

inputs:
  - name: model_config_override
    type: file_path
    formats:
      - json
    description: >
      JSON file overriding the nested diffusion_unet_inference block in
      configs/config_maisi_diff_model_rflow-mr.json. The sentinel value
      `default` selects the upstream config plus CLI modality and random seed.

outputs:
  - name: synthetic_mr_volumes
    type: directory_path
    description: Output directory containing generated MR NIfTI volumes.
  - name: result_json
    type: json
    schema: validators/output_schema.json

paired_verifiers:
  - id: medagent.verifiers.mr_synthesis_quality_v1
    status: implemented
    consumes: evidence_pack_dir
    purpose: >
      Re-reads generated MR NIfTI image artifacts and checks source-pack
      success, official upstream entrypoint, model inventory, modality/version
      identity, sample count, shape and spacing consistency, finite
      nonconstant nonnegative voxel values, aggregate-flag honesty, runtime
      identity, and non-clinical scope disclosure. This is an engineering
      floor, not anatomical realism or training-data approval.

runtime:
  language: python
  python: ">=3.10"
  entrypoint: scripts/run_mr.py
  args:
    - "${python}"
    - "${script}"
    - "${fixture}"
    - "--output-dir"
    - "${out}/samples"
    - "--modality"
    - "mri_t1"
  dependencies:
    nibabel: ">=4.0"
    numpy: ">=1.23"
    typer: ">=0.9"
    torch: ">=2.1"
    monai: ">=1.5"
  side_effects:
    pip_packages:
      - nibabel>=4.0
      - numpy>=1.23
      - typer>=0.9
      - torch>=2.1
      - "monai>=1.5"
      - scipy>=1.10
      - scikit-image>=0.20
      - einops>=0.7
      - huggingface_hub>=0.20
      - tqdm>=4.65
      - fire>=0.5
      - tensorboard>=2.14
      - PyYAML>=6.0
    local_writes:
      - {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
    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: []
    env_optional:
      - 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/diff_model_infer.py
        - scripts/download_model_data.py
        - configs/config_network_rflow.json
        - configs/environment_maisi_diff_model_rflow-mr.json
        - configs/config_maisi_diff_model_rflow-mr.json
        - configs/modality_mapping.json
    - kind: huggingface_repo
      repo_id: nvidia/NV-Generate-MR
      install_path: $NV_GENERATE_ROOT/models/
      install_command: >
        cd $NV_GENERATE_ROOT && python -m scripts.download_model_data
        --version rflow-mr --root_dir ./ --model_only
      contains:
        - models/autoencoder_v2.pt
        - models/diff_unet_3d_rflow-mr.pt

limitations:
  - >
    This is a thin wrapper. Inference, sampling, and decoding are delegated
    entirely to NVIDIA-Medtech/NV-Generate-CTMR's `scripts.diff_model_infer`.
    Do not modify code under $NV_GENERATE_ROOT or the repo-local fallback at
    .workbench_data/upstreams/NV-Generate-CTMR.
  - >
    rflow-mr generates image-only synthetic MRI volumes. It does not emit
    paired segmentation masks.
  - >
    The upstream README recommends `rflow-mr-brain` instead for brain MRI
    synthesis; use `skills/nv-generate-mr-brain` for that path.
  - >
    NV-Generate-MR weights are listed by upstream as NVIDIA Non-Commercial.
    Do not use outputs as production training data without legal and quality
    review.

validation:
  expected_runtime_seconds:
    min: 1.0
    max: 1800.0
    inference_path: runtime.subprocess_seconds
  sanity_checks:
    - {path: skill, eq: nv_generate_mr}
    - {path: invocation.official_entrypoint, eq: "python -m scripts.diff_model_infer"}
    - {path: invocation.exit_code, eq: 0}
    - {path: invocation.model_inventory.all_present, eq: true}
    - {path: output.num_samples, gte: 1}
    - {path: output.all_images_readable, eq: true}
    - {path: output.all_shapes_match_requested, eq: true}
    - {path: output.all_spacing_match_requested, eq: true}
    - {path: output.all_images_finite, eq: true}
    - {path: output.all_images_nonconstant, eq: true}
    - {path: output.all_images_nonnegative, 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_mri_t1.json
    runs: 2
    reason: >
      End-to-end MR synthesis 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 artifact hashes.