droboticflow/scripts/generate_operator_images.py

676 lines
39 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

#!/usr/bin/env python3
from __future__ import annotations
import argparse
import importlib.util
import json
import pathlib
import sys
from typing import Any
ROOT = pathlib.Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT / "scripts"))
from operator_release_manifest import release_image_reference
CATALOG_PATH = ROOT / "internal" / "api" / "builtin_operators_catalog.json"
OPERATOR_ROOT = ROOT / "backend" / "operators" / "builtin"
OPERATOR_REPO_ROOT = ROOT / "operator-repos" / "builtin"
IMAGE_ROOT = ROOT / "operator-images" / "builtin"
MANIFEST_PATH = IMAGE_ROOT / "operator-images.json"
DEFAULT_REGISTRY = "ccr-29eug8s3-pub.cnc.bj.baidubce.com/dataflow-poc"
IMAGE_TAG = "builtin"
DEFAULT_OPERATOR_BASE_IMAGE = f"{DEFAULT_REGISTRY}/droboticflow-operator-runner:{IMAGE_TAG}"
DEFAULT_GPU_BASE_IMAGE = f"{DEFAULT_REGISTRY}/droboticflow-operator-gpu-base:{IMAGE_TAG}"
DEFAULT_HOLOBRAIN_BASE_IMAGE = f"{DEFAULT_REGISTRY}/droboticflow-operator-holobrain-base:{IMAGE_TAG}"
DEFAULT_YOLO_BASE_IMAGE = f"{DEFAULT_REGISTRY}/droboticflow-operator-yolo-base:{IMAGE_TAG}"
COMMAND_ENTRYPOINT = ["python", "/app/backend/operator-runner/main.py"]
GPU_SLUGS = {
"act_policy_train",
"diffusion_policy_train",
"holobrain_export",
"holobrain_inference_test",
"holobrain_robotwin_eval",
"holobrain_train",
"openvla_finetune",
"openvla_inference",
"yolo_embodied_detection_train",
"yolo_video_target_clip",
}
HOLOBRAIN_SLUGS = {
"holobrain_data_prepare",
"holobrain_export",
"holobrain_inference_test",
"holobrain_robotwin_eval",
"holobrain_train",
}
HOLOBRAIN_RUNTIME_SLUGS = {
"holobrain_export",
"holobrain_inference_test",
"holobrain_robotwin_eval",
"holobrain_train",
}
HOLOBRAIN_ROBOTWIN_EVAL_RUNTIME_SLUGS = {
"holobrain_robotwin_eval",
}
LONG_RUNNING_SLUGS = {
"holobrain_data_prepare",
"holobrain_train",
}
REMOTE_GPU_SERVICE_SLUGS = {
"wilor_hand_pose",
}
COMMAND_WRAPPER_ALLOWED: set[str] = set()
CONTAINER_HANDLER_ALIASES = {
"builtin_source": "source",
"builtin_sim3d_asset_generate": "sim3d_asset_generate",
"builtin_vla_model_evaluate": "vla_model_evaluate",
"builtin_vla_model_inference": "vla_model_inference",
"builtin_vla_model_train": "vla_model_train",
"builtin_vla_sample_manifest_validate": "vla_sample_manifest_validate",
"builtin_lerobot_summary": "lerobot",
"builtin_rosbag_summary": "rosbag",
}
STRUCTURED_OPERATOR_OVERRIDES: dict[str, dict[str, Any]] = {
"operator-builtin-model-inference": {
"skillDefinition": "执行结构化模型推理节点;读取 MODEL_JSON 或 MODEL_PATH,支持 rule、linear、nearest_centroid 三类可解释模型;自定义命令请使用 Shell 命令算子,constant 等硬编码输出模型会被拒绝。",
},
"operator-builtin-act-policy-train": {
"skillDefinition": "ACT 策略训练算子;生产默认使用 Hugging Face LeRobot ACT 训练入口,读取 DATASET_REPO_ID 或 DATASET_PATH 指向的真实 LeRobot 数据集并输出 checkpoint 与 metrics。推荐 1 张 A800 80GB;仅本地单元测试可显式设置 TRAINING_BACKEND=torch_mlp。",
"hyperparams": [
{"key": "TRAINING_BACKEND", "value": "lerobot", "description": "生产训练后端;默认 lerobot,测试 fallback 才允许 torch_mlp", "type": "string"},
{"key": "DATASET_REPO_ID", "value": "", "description": "Hugging Face LeRobot 数据集 repo_id,例如 org/robot_dataset", "type": "string"},
{"key": "DATASET_PATH", "value": "", "description": "POC/PVC 中的真实 LeRobot 数据集根目录;与 DATASET_REPO_ID 至少提供一个", "type": "string"},
{"key": "POLICY_REPO_ID", "value": "", "description": "可选的 LeRobot policy repo_id,用于后续发布或复现实验", "type": "string"},
{"key": "MAX_STEPS", "value": "1000", "description": "短训最大 step 数,生产长训按数据规模调大", "type": "number"},
{"key": "BATCH_SIZE", "value": "8", "description": "训练 batch size", "type": "number"},
{"key": "LEARNING_RATE", "value": "0.0001", "description": "LeRobot optimizer.lr", "type": "number"},
{"key": "CHUNK_SIZE", "value": "100", "description": "ACT action chunk size", "type": "number"},
{"key": "LEROBOT_OVERRIDES_JSON", "value": "[]", "description": "额外 Hydra override 字符串数组,不允许 shell 命令", "type": "json"},
{"key": "RECOMMENDED_GPU", "value": "1x NVIDIA A800 80GB", "description": "推荐 GPU 配置", "type": "string"},
],
},
"operator-builtin-diffusion-policy-train": {
"skillDefinition": "Diffusion Policy 训练算子;生产默认使用 real-stanford/diffusion_policy 的 Hydra train.py,读取真实 Zarr/轨迹数据并输出 .ckpt、logs 和 metrics。推荐 1 张 A800 80GB;仅本地单元测试可显式设置 TRAINING_BACKEND=torch_mlp。",
"hyperparams": [
{"key": "TRAINING_BACKEND", "value": "diffusion_policy", "description": "生产训练后端;默认 diffusion_policy,测试 fallback 才允许 torch_mlp", "type": "string"},
{"key": "DIFFUSION_POLICY_ROOT", "value": "/opt/diffusion_policy", "description": "real-stanford/diffusion_policy 仓库路径", "type": "string"},
{"key": "DATASET_PATH", "value": "", "description": "真实 Zarr/轨迹数据路径,会作为 task.dataset.zarr_path 传入", "type": "string"},
{"key": "CONFIG_NAME", "value": "image_pusht_diffusion_policy_cnn", "description": "Hydra config name 或 config yaml 名称", "type": "string"},
{"key": "TASK_NAME", "value": "", "description": "可选 task override", "type": "string"},
{"key": "EPOCHS", "value": "1", "description": "短训 epoch 数,生产长训按数据规模调大", "type": "number"},
{"key": "BATCH_SIZE", "value": "8", "description": "训练 batch size", "type": "number"},
{"key": "LEARNING_RATE", "value": "0.0001", "description": "optimizer.lr", "type": "number"},
{"key": "HORIZON", "value": "16", "description": "预测 horizon", "type": "number"},
{"key": "DIFFUSION_STEPS", "value": "32", "description": "扩散步数配置,用于训练记录和下游复现", "type": "number"},
{"key": "HYDRA_OVERRIDES_JSON", "value": "[]", "description": "额外 Hydra override 字符串数组,不允许 shell 命令", "type": "json"},
{"key": "RECOMMENDED_GPU", "value": "1x NVIDIA A800 80GB", "description": "推荐 GPU 配置", "type": "string"},
],
},
"operator-builtin-openvla-finetune": {
"skillDefinition": "OpenVLA 微调算子;生产默认使用 OpenVLA 官方 vla-scripts/finetune.py 的 PEFT/LoRA 路径,读取真实 RLDS 数据集并输出 LoRA adapter artifact 与 metrics。推荐 1 张 A800 80GB。",
"hyperparams": [
{"key": "TRAINING_BACKEND", "value": "openvla_lora", "description": "生产训练后端;默认 openvla_lora,兼容测试可显式 transformers_text", "type": "string"},
{"key": "OPENVLA_ROOT", "value": "/opt/openvla", "description": "OpenVLA 仓库路径", "type": "string"},
{"key": "BASE_MODEL", "value": "openvla/openvla-7b", "description": "OpenVLA 基础模型路径或 HF 模型名", "type": "string"},
{"key": "DATA_ROOT_DIR", "value": "", "description": "真实 RLDS 数据集根目录", "type": "string"},
{"key": "DATASET_NAME", "value": "", "description": "RLDS 数据集名称,例如 droid_wipe", "type": "string"},
{"key": "MAX_STEPS", "value": "1000", "description": "短训最大 step 数,生产长训按数据规模调大", "type": "number"},
{"key": "SAVE_STEPS", "value": "500", "description": "checkpoint 保存间隔", "type": "number"},
{"key": "BATCH_SIZE", "value": "8", "description": "LoRA fine-tune batch size", "type": "number"},
{"key": "LEARNING_RATE", "value": "0.0005", "description": "LoRA 学习率", "type": "number"},
{"key": "LORA_RANK", "value": "32", "description": "LoRA rank", "type": "number"},
{"key": "USE_QUANTIZATION", "value": "false", "description": "是否启用 4-bit quantization", "type": "boolean"},
{"key": "PRECISION", "value": "bf16", "description": "推荐精度配置", "type": "string"},
{"key": "RECOMMENDED_GPU", "value": "1x NVIDIA A800 80GB", "description": "推荐 GPU 配置", "type": "string"},
],
},
"operator-builtin-openvla-inference": {
"skillDefinition": "OpenVLA 推理算子;使用 MODEL_PATH 或 BASE_MODEL 加载真实 Transformers 权重,读取 PROMPT 或上游样本 instruction,并输出 inference-result.json。推荐 1 张 A800,量化模型可降低显存。",
"hyperparams": [
{"key": "INFERENCE_BACKEND", "value": "openvla_transformers", "description": "推理后端;默认从 MODEL_PATH 或上游 fine-tune artifact 加载 OpenVLA/Transformers 权重", "type": "string"},
{"key": "MODEL_PATH", "value": "", "description": "本地模型目录或 Hugging Face 模型名;为空时自动消费上游模型 artifact", "type": "string"},
{"key": "BASE_MODEL", "value": "", "description": "未提供 MODEL_PATH 时使用的基础模型", "type": "string"},
{"key": "PROMPT", "value": "", "description": "推理 prompt;为空则从上游样本读取 TEXT_FIELD", "type": "string"},
{"key": "TEXT_FIELD", "value": "instruction", "description": "上游样本指令字段名", "type": "string"},
{"key": "MAX_NEW_TOKENS", "value": "64", "description": "生成 token 数", "type": "number"},
{"key": "PRECISION", "value": "bf16", "description": "推荐精度配置", "type": "string"},
{"key": "RECOMMENDED_GPU", "value": "1x NVIDIA A800 80GB", "description": "推荐 GPU 配置", "type": "string"},
],
},
"operator-builtin-holobrain-data-prepare": {
"description": "校验真实 RoboTwin 2.0 数据并生成 HoloBrain manifest;支持 smoke 真实小数据和 full 完整数据模式。",
"skillDefinition": "HoloBrain 数据准备算子;TRAINING_MODE=smoke 使用真实小数据契约(Arrow 默认最多 1 个 episode),full 保持完整数据要求;两种模式都拒绝合成或空输入。",
"hyperparams": [
{"key": "TRAINING_MODE", "value": "full", "description": "full 保持完整训练数据行为;smoke 使用真实小数据完成端到端验证", "type": "string"},
{"key": "DATA_SOURCE", "value": "local", "description": "数据来源;仅支持 local,不支持运行时下载", "type": "string"},
{"key": "ROBOTWIN_SOURCE_PATH", "value": "", "description": "已预置的 RoboTwin/HoloBrain 真实数据根目录;必填", "type": "string"},
{"key": "EXPECTED_LMDB_DIRS_JSON", "value": "", "description": "可选覆盖 LMDB 目录;空值时 smoke 只要求 aloha_agilex_demo_clean,full 要求三个官方目录", "type": "json"},
{"key": "SMOKE_MAX_EPISODES", "value": "1", "description": "smoke 从真实 Arrow 数据转换的最大 episode 数", "type": "number"},
{"key": "ARROW_MAX_EPISODES", "value": "", "description": "显式 Arrow 转换 episode 上限;空值时由模式决定", "type": "number"},
],
},
"operator-builtin-holobrain-train": {
"description": "执行 HoloBrain 训练;full 保留官方长训默认,smoke 用真实小数据短训并输出可恢复 checkpoint 与结构化进度。",
"skillDefinition": "HoloBrain 训练算子;TRAINING_MODE=full 保留官方 100000 step 默认,smoke 默认在真实 RoboTwin 小数据上执行 2 step 并生成 checkpoint。失败重试默认恢复最新有效 Accelerate checkpoint;进度写入 workspace 并以 DROBOTICFLOW_PROGRESS {json} 输出;dtype 在启动前检查。",
"hyperparams": [
{"key": "TRAINING_MODE", "value": "full", "description": "full 或 smoke;full 不覆盖官方长训默认,smoke 使用短步数真实训练", "type": "string"},
{"key": "ROBOORCHARDLAB_ROOT", "value": "/opt/RoboOrchardLab", "description": "镜像内预置的 RoboOrchardLab 仓库路径;缺失时直接失败", "type": "string"},
{"key": "CONFIG_PATH", "value": "projects/holobrain/configs/config_holobrain_qwen_common.py", "description": "HoloBrain Qwen 训练配置", "type": "string"},
{"key": "ROBOTWIN_DATA_ROOT", "value": "", "description": "RoboTwin 数据根目录;为空时消费上游数据准备 manifest", "type": "string"},
{"key": "WORKSPACE", "value": "", "description": "训练 workspace;为空时使用节点输出目录下 holobrain-workspace", "type": "string"},
{"key": "LAUNCH_MODE", "value": "single", "description": "single 或 accelerate", "type": "string"},
{"key": "GPU_COUNT", "value": "1", "description": "accelerate 模式使用的进程/GPU 数", "type": "number"},
{"key": "CONFIG_KWARGS_JSON", "value": "{}", "description": "传给 train.py --kwargs 的 JSON;默认不覆盖 max_step=100000", "type": "json"},
{"key": "SMOKE_MAX_STEP", "value": "2", "description": "smoke 真实训练最大 step;默认 2", "type": "number"},
{"key": "SMOKE_BATCH_SIZE", "value": "1", "description": "smoke batch size", "type": "number"},
{"key": "SMOKE_NUM_WORKERS", "value": "0", "description": "smoke 数据加载进程数", "type": "number"},
{"key": "RESUME_POLICY", "value": "auto", "description": "auto 自动选择 workspace/checkpoints 下最新有效训练状态;explicit 要求 RESUME_FROM;never 禁用恢复", "type": "string"},
{"key": "RESUME_FROM", "value": "", "description": "显式 Accelerate 训练状态 checkpoint 目录", "type": "string"},
{"key": "INITIAL_CHECKPOINT_PATH", "value": "", "description": "可选初始模型权重文件;与训练状态恢复分开", "type": "string"},
{"key": "CHECKPOINT_ARTIFACT", "value": "latest_valid", "description": "latest_valid、explicit 或 none;控制训练报告暴露的模型产物", "type": "string"},
{"key": "OUTPUT_CHECKPOINT_PATH", "value": "", "description": "CHECKPOINT_ARTIFACT=explicit 时使用的模型 checkpoint", "type": "string"},
{"key": "PROGRESS_INTERVAL_SECONDS", "value": "15", "description": "结构化进度写入 droboticflow-progress.json、holobrain-progress.json 并输出 DROBOTICFLOW_PROGRESS {json};该值为最长刷新间隔", "type": "number"},
{"key": "MODEL_DTYPE", "value": "bfloat16", "description": "官方 HoloBrain Qwen 模型 dtype;不兼容值在启动前失败", "type": "string"},
{"key": "INPUT_DTYPE", "value": "float32", "description": "官方 RoboTwin processor 输入 tensor dtype;不兼容值在启动前失败", "type": "string"},
{"key": "WANDB_MODE", "value": "offline", "description": "训练日志模式", "type": "string"},
{"key": "RECOMMENDED_GPU", "value": "1x NVIDIA A800 80GB", "description": "推荐 GPU 配置", "type": "string"},
],
},
"operator-builtin-holobrain-export": {
"description": "从显式、上游或训练 workspace 的最新有效 checkpoint 导出可部署 HoloBrain 模型。",
"skillDefinition": "HoloBrain 模型导出算子;按 CHECKPOINT_SELECTION 消费显式、上游或训练 workspace 中最新有效 checkpoint。smoke 模式没有真实训练 checkpoint 时直接失败。",
"hyperparams": [
{"key": "TRAINING_MODE", "value": "full", "description": "full 或 smoke;smoke 强制要求上游真实 checkpoint", "type": "string"},
{"key": "ROBOORCHARDLAB_ROOT", "value": "/opt/RoboOrchardLab", "description": "RoboOrchardLab 仓库路径", "type": "string"},
{"key": "CONFIG_PATH", "value": "projects/holobrain/configs/config_holobrain_qwen_common.py", "description": "HoloBrain Qwen 导出配置", "type": "string"},
{"key": "TRAINING_WORKSPACE", "value": "", "description": "训练 workspace;为空时消费上游训练报告", "type": "string"},
{"key": "CHECKPOINT_PATH", "value": "", "description": "显式 checkpoint 路径;为空时使用最新训练 checkpoint", "type": "string"},
{"key": "CHECKPOINT_SELECTION", "value": "latest_valid", "description": "latest_valid、upstream 或 explicit", "type": "string"},
{"key": "EXPORT_WORKSPACE", "value": "", "description": "导出目录;为空时使用节点输出目录下 holobrain-export", "type": "string"},
{"key": "CONFIG_KWARGS_JSON", "value": "{}", "description": "额外 export.py --kwargs JSON", "type": "json"},
{"key": "RECOMMENDED_GPU", "value": "1x NVIDIA A800 80GB", "description": "export reload-test 使用 CUDA", "type": "string"},
],
},
"operator-builtin-holobrain-robotwin-eval": {
"description": "调用官方 robotwin_eval.py 执行真实 HoloBrain RoboTwin 评测;smoke 为单任务单 episode。",
"skillDefinition": "HoloBrain RoboTwin 评测算子;full 保留官方多任务评测默认,smoke 固定执行 1 个真实 RoboTwin episode,必须由官方 evaluator 产生指标。",
"hyperparams": [
{"key": "TRAINING_MODE", "value": "full", "description": "full 或 smoke;smoke 使用专用单任务单 episode 参数", "type": "string"},
{"key": "ROBOORCHARDLAB_ROOT", "value": "/opt/RoboOrchardLab", "description": "RoboOrchardLab 仓库路径", "type": "string"},
{"key": "MODEL_DIR", "value": "", "description": "导出模型目录;为空时消费上游导出 manifest", "type": "string"},
{"key": "INFERENCE_PREFIX", "value": "robotwin2_0", "description": "导出 pipeline config prefix", "type": "string"},
{"key": "EVAL_MODE", "value": "ray", "description": "official eval mode:ray 或 local", "type": "string"},
{"key": "TASK_NAMES_JSON", "value": "[]", "description": "评测任务名数组;空数组使用官方默认 SEM_TASKS_16", "type": "json"},
{"key": "EPISODE_NUM", "value": "100", "description": "每任务 episode 数", "type": "number"},
{"key": "GPU_IDS_JSON", "value": "[0]", "description": "Ray 评测 GPU id 数组", "type": "json"},
{"key": "WORKERS_PER_GPU", "value": "1", "description": "每张 GPU 的并发 worker 数", "type": "number"},
{"key": "SAVE_VIDEO", "value": "false", "description": "是否保存评测视频", "type": "boolean"},
{"key": "SMOKE_TASK_NAME", "value": "beat_block_hammer", "description": "smoke 实际执行的 RoboTwin 任务", "type": "string"},
{"key": "SMOKE_EVAL_EPISODE_NUM", "value": "1", "description": "smoke episode 数", "type": "number"},
{"key": "SMOKE_EVAL_MODE", "value": "local", "description": "smoke evaluator 模式", "type": "string"},
{"key": "SMOKE_DEVICE", "value": "cuda", "description": "smoke 模型执行设备", "type": "string"},
{"key": "RECOMMENDED_GPU", "value": "1x NVIDIA A800 80GB", "description": "推荐 GPU 配置", "type": "string"},
],
},
"operator-builtin-holobrain-inference-test": {
"description": "加载导出模型并通过官方 RoboTwin evaluator 执行真实模型推理,不接受文件存在性或硬编码结果冒充成功。",
"skillDefinition": "HoloBrain 模型测试算子;加载导出的真实模型并通过官方 RoboTwin evaluator 执行至少 1 个 episode,不再用文件存在性或硬编码结果冒充推理。",
"hyperparams": [
{"key": "TRAINING_MODE", "value": "full", "description": "记录上游 full/smoke 模式", "type": "string"},
{"key": "ROBOORCHARDLAB_ROOT", "value": "/opt/RoboOrchardLab", "description": "RoboOrchardLab 仓库路径", "type": "string"},
{"key": "MODEL_DIR", "value": "", "description": "导出模型目录;为空时消费上游导出 manifest", "type": "string"},
{"key": "INFERENCE_PREFIX", "value": "robotwin2_0", "description": "导出 pipeline config prefix", "type": "string"},
{"key": "TASK_NAME", "value": "beat_block_hammer", "description": "实际执行的 RoboTwin 推理任务", "type": "string"},
{"key": "EPISODE_NUM", "value": "1", "description": "真实推理 episode 数,至少为 1", "type": "number"},
{"key": "INFERENCE_MODE", "value": "local", "description": "local 或 ray", "type": "string"},
{"key": "DEVICE", "value": "cuda", "description": "模型推理设备", "type": "string"},
{"key": "GPU_IDS_JSON", "value": "[0]", "description": "ray 模式 GPU id 数组", "type": "json"},
{"key": "WORKERS_PER_GPU", "value": "1", "description": "ray 模式每卡 worker 数", "type": "number"},
{"key": "RECOMMENDED_GPU", "value": "1x NVIDIA A800 80GB", "description": "完整推理测试推荐 GPU 配置", "type": "string"},
],
},
"operator-builtin-yolo-embodied-detection-train": {
"skillDefinition": "YOLO 具身目标检测训练算子;使用 Ultralytics API 读取真实 YOLO/COCO data.yaml,输出 best/last checkpoint 与 metrics.json。A800 对 YOLOv8/YOLOv10/YOLO11 中大模型训练充足,默认请求 1 卡。",
"hyperparams": [
{"key": "DATA_YAML", "value": "", "description": "真实 YOLO/COCO data.yaml 路径", "type": "string"},
{"key": "MODEL_SIZE", "value": "yolo11n.pt", "description": "Ultralytics 模型名或权重路径", "type": "string"},
{"key": "EPOCHS", "value": "10", "description": "训练轮数", "type": "number"},
{"key": "IMAGE_SIZE", "value": "640", "description": "输入图片尺寸", "type": "number"},
{"key": "BATCH_SIZE", "value": "8", "description": "训练 batch size", "type": "number"},
{"key": "WORKERS", "value": "0", "description": "数据加载 worker 数;K8s 默认共享内存较小时建议为 0", "type": "number"},
{"key": "CONF", "value": "0.25", "description": "后续检测推荐置信度", "type": "number"},
{"key": "TARGET_LABELS", "value": "", "description": "关注类别,逗号分隔", "type": "string"},
{"key": "RECOMMENDED_GPU", "value": "1x NVIDIA A800 80GB", "description": "推荐 GPU 配置", "type": "string"},
],
},
"operator-builtin-shell-command": {
"skillDefinition": "受控脚本算子;只允许执行算子目录 approved_scripts 中登记的生产验收脚本,不接受自由 COMMAND/SCRIPT。",
"hyperparams": [
{"key": "SCRIPT_ID", "value": "fixture_report", "description": "白名单脚本 ID", "type": "string"},
{"key": "SCRIPT_ARGS_JSON", "value": "{}", "description": "传给脚本的结构化 JSON 参数", "type": "json"},
{"key": "INPUT_PATHS_JSON", "value": "[]", "description": "允许脚本读取的输入路径 JSON 数组", "type": "json"},
{"key": "OUTPUT_CONTRACT_JSON", "value": "{}", "description": "脚本输出 artifact 契约", "type": "json"},
],
},
}
def load_repository_generator() -> Any:
path = ROOT / "scripts" / "generate_operator_repository_manifests.py"
spec = importlib.util.spec_from_file_location("operator_repository_generator", path)
if spec is None or spec.loader is None:
raise RuntimeError(f"cannot load {path}")
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return module
REPOSITORY_GENERATOR = load_repository_generator()
RUNTIME_DEPENDENCIES: dict[str, list[str]] = REPOSITORY_GENERATOR.RUNTIME_DEPENDENCIES
def load_json(path: pathlib.Path) -> Any:
return json.loads(path.read_text(encoding="utf-8"))
def dump_json(payload: Any) -> str:
return json.dumps(payload, ensure_ascii=False, indent=2, sort_keys=True) + "\n"
def operator_slug(operator_id: str) -> str:
return operator_id.removeprefix("operator-builtin-").replace("-", "_")
def container_handler(item: dict[str, Any], slug: str) -> str:
handler = item.get("runtimeHandler") or slug
return str(handler)
def image_name(slug: str, registry: str = DEFAULT_REGISTRY) -> str:
return f"{registry}/droboticflow-operator-{slug.replace('_', '-')}:{IMAGE_TAG}"
def docker_base(slug: str) -> str:
if slug in HOLOBRAIN_SLUGS:
return "${OPERATOR_HOLOBRAIN_BASE_IMAGE}"
if slug.startswith("yolo_"):
return "${OPERATOR_YOLO_BASE_IMAGE}"
if slug in GPU_SLUGS:
return "${OPERATOR_GPU_BASE_IMAGE}"
return "${OPERATOR_BASE_IMAGE}"
def dockerfile(slug: str, operator_id: str, handler: str) -> str:
if slug not in HOLOBRAIN_SLUGS and slug not in GPU_SLUGS and not slug.startswith("yolo_"):
return f"""# Generated by scripts/generate_operator_images.py. Do not edit by hand.
ARG OPERATOR_BASE_IMAGE={DEFAULT_OPERATOR_BASE_IMAGE}
FROM ${{OPERATOR_BASE_IMAGE}}
ENV DROBOTICFLOW_OPERATOR_ID={operator_id} \\
DROBOTICFLOW_OPERATOR_SLUG={slug} \\
DROBOTICFLOW_OPERATOR_SOURCE=/app/operator-repos/builtin/{slug}/src/operator.py \\
DROBOTICFLOW_OPERATOR_REPOSITORY=/app/operator-repos/builtin/{slug} \\
DROBOTICFLOW_OPERATOR_HANDLER={handler}
COPY operator-repos/builtin/{slug} ./operator-repos/builtin/{slug}
ENTRYPOINT ["python", "/app/backend/operator-runner/main.py"]
"""
if slug in HOLOBRAIN_SLUGS:
base_arg = f"ARG OPERATOR_HOLOBRAIN_BASE_IMAGE={DEFAULT_HOLOBRAIN_BASE_IMAGE}"
elif slug.startswith("yolo_"):
base_arg = f"ARG OPERATOR_YOLO_BASE_IMAGE={DEFAULT_YOLO_BASE_IMAGE}"
else:
base_arg = f"ARG OPERATOR_GPU_BASE_IMAGE={DEFAULT_GPU_BASE_IMAGE}"
holobrain_shim_copy = (
"COPY operator-images/builtin/holobrain-shims/pytorch3d /app/pytorch3d\n"
if slug in HOLOBRAIN_RUNTIME_SLUGS
else ""
)
holobrain_asset_copy = (
"COPY operator-images/builtin/holobrain-assets/urdf /opt/RoboOrchardLab/projects/holobrain/urdf\n"
"COPY operator-images/builtin/holobrain-vendor/RoboOrchardLab/robo_orchard_lab/models/holobrain/processor.py /opt/RoboOrchardLab/robo_orchard_lab/models/holobrain/processor.py\n"
if slug in HOLOBRAIN_RUNTIME_SLUGS
else ""
)
holobrain_train_dependency_fix = ""
if slug in HOLOBRAIN_RUNTIME_SLUGS:
holobrain_train_dependency_fix = (
'RUN python -m pip install --no-cache-dir --force-reinstall "numpy>=2.0,<2.3" "scipy>=1.14,<1.16" "einops>=0.8,<0.9" && \\\n'
' python -m pip install --no-cache-dir --force-reinstall --no-deps \\\n'
' "transformers==4.57.6" "accelerate==1.14.0" "tokenizers==0.22.2" \\\n'
' "huggingface_hub==0.36.2" "safetensors==0.8.0" && \\\n'
' python -m pip install --no-cache-dir --no-deps "https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.6cxx11abiFALSE-cp311-cp311-linux_x86_64.whl"'
)
if slug in HOLOBRAIN_ROBOTWIN_EVAL_RUNTIME_SLUGS:
holobrain_train_dependency_fix += (
' && \\\n'
' python -m pip install --no-cache-dir \\\n'
' "mcap-protobuf-support>=0.5.3" "mcap-ros2-support" "mcap>=1.2.2" \\\n'
' "foxglove-schemas-protobuf>=0.3.0" "robo_orchard_schemas==0.2.0" "foxglove-sdk" \\\n'
' "sapien==3.0.3"'
)
holobrain_train_dependency_fix += "\n"
return f"""# Generated by scripts/generate_operator_images.py. Do not edit by hand.
{base_arg}
FROM {docker_base(slug)}
WORKDIR /app
ENV PYTHONUNBUFFERED=1 \\
PYTHONPATH=/app \\
DROBOTICFLOW_OPERATOR_ID={operator_id} \\
DROBOTICFLOW_OPERATOR_SLUG={slug} \\
DROBOTICFLOW_OPERATOR_SOURCE=/app/operator-repos/builtin/{slug}/src/operator.py \\
DROBOTICFLOW_OPERATOR_REPOSITORY=/app/operator-repos/builtin/{slug} \\
DROBOTICFLOW_OPERATOR_HANDLER={handler} \\
USER=droboticflow \\
LOGNAME=droboticflow \\
HOME=/tmp/droboticflow-home
RUN mkdir -p /tmp/droboticflow-home /workspace /outputs
COPY backend ./backend
{holobrain_train_dependency_fix}{holobrain_shim_copy}{holobrain_asset_copy}COPY operator-repos/builtin/{slug} ./operator-repos/builtin/{slug}
ENTRYPOINT ["python", "/app/backend/operator-runner/main.py"]
"""
def gpu_base_dockerfile() -> str:
return """# Generated by scripts/generate_operator_images.py. Do not edit by hand.
FROM pytorch/pytorch:2.6.0-cuda12.4-cudnn9-runtime
WORKDIR /app
ENV PYTHONUNBUFFERED=1 \\
PYTHONPATH=/app \\
USER=droboticflow \\
LOGNAME=droboticflow \\
HOME=/tmp/droboticflow-home
RUN apt-get update && apt-get install -y --no-install-recommends \\
build-essential \\
ffmpeg \\
git \\
libglib2.0-0 \\
libgl1 \\
linux-libc-dev \\
python3-dev \\
&& rm -rf /var/lib/apt/lists/*
RUN printf 'torch==2.6.0\\ntorchvision==0.21.0\\n' > /tmp/gpu-constraints.txt
RUN python -m pip install --no-cache-dir -c /tmp/gpu-constraints.txt transformers accelerate safetensors nvidia-ml-py
RUN python -m pip install --no-cache-dir -c /tmp/gpu-constraints.txt lerobot peft draccus wandb hydra-core omegaconf zarr numcodecs dill einops diffusers
RUN git clone --depth=1 https://github.com/openvla/openvla.git /opt/openvla && \
git clone --depth=1 https://github.com/real-stanford/diffusion_policy.git /opt/diffusion_policy
RUN mkdir -p /tmp/droboticflow-home /workspace /outputs
"""
def holobrain_base_dockerfile() -> str:
return """# Generated by scripts/generate_operator_images.py. Do not edit by hand.
FROM pytorch/pytorch:2.6.0-cuda12.4-cudnn9-runtime
WORKDIR /app
ENV PYTHONUNBUFFERED=1 \\
PYTHONPATH=/app:/opt/RoboOrchardLab \\
USER=droboticflow \\
LOGNAME=droboticflow \\
HOME=/tmp/droboticflow-home \\
ROBOORCHARDLAB_ROOT=/opt/RoboOrchardLab \\
HF_HOME=/opt/hf-cache \\
HF_HUB_CACHE=/opt/hf-cache/hub \\
TRANSFORMERS_CACHE=/opt/hf-cache/transformers \\
HOLOBRAIN_MODEL_DIR=/opt/models/HoloBrain_v0.0_Qwen
RUN apt-get update && apt-get install -y --no-install-recommends \\
build-essential \\
ffmpeg \\
git \\
libglib2.0-0 \\
libgl1 \\
linux-libc-dev \\
python3-dev \\
&& rm -rf /var/lib/apt/lists/*
RUN printf 'torch==2.6.0\\ntorchvision==0.21.0\\n' > /tmp/holobrain-constraints.txt
RUN python -m pip install --no-cache-dir -c /tmp/holobrain-constraints.txt \\
accelerate==0.34.2 \\
cloudpickle \\
datasets \\
deprecated \\
diffusers \\
"duckdb<=1.3.2" \\
duckdb-engine \\
flask \\
gevent \\
gymnasium \\
h5py \\
huggingface_hub \\
imageio[ffmpeg] \\
lmdb \\
"numpy>=2.0,<2.3" \\
numpydantic \\
nvidia-ml-py \\
opencv-python \\
ordered-set \\
pandas \\
peft \\
pydantic-settings \\
pytorch-kinematics \\
pyzstd \\
ray[default] \\
rtoml \\
safetensors \\
"scipy>=1.14,<1.16" \\
sortedcontainers \\
tensorboard \\
terminaltables \\
timeout-decorator \\
torchmetrics \\
"transformers<=5.8" \\
"websockets>=10.0"
COPY operator-images/builtin/holobrain-vendor/robo_orchard_core /opt/robo_orchard_core
COPY operator-images/builtin/holobrain-vendor/RoboOrchardLab /opt/RoboOrchardLab
RUN cd /opt/robo_orchard_core && \\
python -m pip install --no-cache-dir -e . --no-deps && \\
cd /opt/RoboOrchardLab && \\
python -m pip install --no-cache-dir -e . --no-deps
RUN python -m pip install --no-cache-dir accelerate==0.34.2
RUN mkdir -p /opt/hf-cache /opt/models/HoloBrain_v0.0_Qwen && \\
python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='HorizonRobotics/HoloBrain_v0.0_Qwen', repo_type='model', cache_dir='/opt/hf-cache/hub', local_dir='/opt/models/HoloBrain_v0.0_Qwen', allow_patterns=['pretrain/*']); snapshot_download(repo_id='Qwen/Qwen2.5-VL-3B-Instruct', repo_type='model', cache_dir='/opt/hf-cache/hub')"
ENV HF_HUB_OFFLINE=1 \\
TRANSFORMERS_OFFLINE=1 \\
HF_DATASETS_OFFLINE=1
RUN mkdir -p /opt/RoboOrchardLab/projects/holobrain/data /tmp/droboticflow-home /workspace /outputs && \\
chown -R 10001:10001 /opt/RoboOrchardLab/projects/holobrain/data /tmp/droboticflow-home && \\
chmod -R a+rX /opt/RoboOrchardLab /opt/hf-cache /opt/models
"""
def yolo_base_dockerfile() -> str:
return """# Generated by scripts/generate_operator_images.py. Do not edit by hand.
FROM ultralytics/ultralytics:latest
WORKDIR /app
ENV PYTHONUNBUFFERED=1 \\
PYTHONPATH=/app \\
USER=droboticflow \\
LOGNAME=droboticflow \\
HOME=/tmp/droboticflow-home
RUN python -m pip install --no-cache-dir nvidia-ml-py
RUN mkdir -p /tmp/droboticflow-home /workspace /outputs
"""
def contract_for(item: dict[str, Any], slug: str) -> dict[str, Any]:
dependencies = RUNTIME_DEPENDENCIES.get(slug, [])
gpu = slug in GPU_SLUGS
remote_gpu_service = slug in REMOTE_GPU_SERVICE_SLUGS
handler = container_handler(item, slug)
resource_contract: dict[str, Any] = {
"requiresGpu": gpu,
"recommendedGpuType": "NVIDIA A800 80GB" if gpu or remote_gpu_service else "",
"minGpuCount": 1 if gpu else 0,
}
if remote_gpu_service:
resource_contract["externalGpuService"] = True
contract = {
"schemaVersion": "droboticflow.operator.runtime/v1",
"operatorId": item["id"],
"operatorSlug": slug,
"handler": handler,
"entrypoint": "/app/backend/operator-runner/main.py",
"imagePolicy": {
"independentFinalImage": True,
"allowsSharedBaseLayer": True,
"forbidsSharedOperatorRunnerFinalImage": True,
},
"runtimeDependencies": dependencies,
"resources": resource_contract,
"parameters": {
"source": "hyperparams",
"structured": True,
"commandWrapperAllowed": False,
"controlledScriptAllowed": item["id"] == "operator-builtin-shell-command",
},
"artifacts": {
"requiresManifest": True,
"requiresChecksum": True,
"requiresLineage": True,
},
"acceptance": {
"requiresRealFixture": True,
"requiresFixtureChecksum": True,
"requiresArtifactChecksum": True,
"requiresWorkflowTransfer": True,
"requiresGpuEvidence": gpu or remote_gpu_service,
},
}
if slug in LONG_RUNNING_SLUGS:
contract["longRunning"] = True
return contract
def strip_disallowed_command_params(item: dict[str, Any]) -> dict[str, Any]:
if item["id"] in COMMAND_WRAPPER_ALLOWED:
return item
hyperparams = [
param
for param in item.get("hyperparams", [])
if str(param.get("key", "")).strip().upper() not in {"COMMAND", "SCRIPT"}
]
if len(hyperparams) != len(item.get("hyperparams", [])):
item = {**item, "hyperparams": hyperparams}
return item
def expected_catalog() -> list[dict[str, Any]]:
items = load_json(CATALOG_PATH)
next_items: list[dict[str, Any]] = []
for item in items:
if item.get("type") != "builtin":
next_items.append(item)
continue
slug = operator_slug(item["id"])
handler = item.get("runtimeHandler") or slug
next_item = strip_disallowed_command_params(dict(item))
overrides = STRUCTURED_OPERATOR_OVERRIDES.get(next_item["id"], {})
if "description" in overrides:
next_item["description"] = overrides["description"]
if "skillDefinition" in overrides:
next_item["skillDefinition"] = overrides["skillDefinition"]
if "hyperparams" in overrides:
next_item["hyperparams"] = overrides["hyperparams"]
next_item["image"], next_item["imageId"] = release_image_reference(slug)
next_item["containerCommand"] = COMMAND_ENTRYPOINT
next_item["containerArgs"] = []
next_item["runtimeSource"] = f"operator-repos/builtin/{slug}/src/operator.py"
next_item["runtimeContract"] = contract_for(next_item, slug)
next_items.append(next_item)
return next_items
def expected_files(catalog_items: list[dict[str, Any]]) -> dict[pathlib.Path, str]:
files: dict[pathlib.Path, str] = {}
manifest_items: list[dict[str, Any]] = []
files[IMAGE_ROOT / "gpu-base.Dockerfile"] = gpu_base_dockerfile()
files[IMAGE_ROOT / "holobrain-base.Dockerfile"] = holobrain_base_dockerfile()
files[IMAGE_ROOT / "yolo-base.Dockerfile"] = yolo_base_dockerfile()
for item in catalog_items:
if item.get("type") != "builtin":
continue
slug = operator_slug(item["id"])
metadata_path = OPERATOR_REPO_ROOT / slug / "operator.json"
if not metadata_path.exists():
raise RuntimeError(f"missing operator metadata for {item['id']}: {metadata_path.relative_to(ROOT)}")
handler = container_handler(item, slug)
files[OPERATOR_REPO_ROOT / slug / "Dockerfile"] = dockerfile(slug, item["id"], handler)
manifest_items.append(
{
"operatorId": item["id"],
"operatorSlug": slug,
"image": item["image"],
"imageId": item["imageId"],
"dockerfile": f"operator-repos/builtin/{slug}/Dockerfile",
"sourceRoot": f"operator-repos/builtin/{slug}",
"runtimeSource": f"operator-repos/builtin/{slug}/src/operator.py",
"runtimeHandler": handler,
"baseImage": docker_base(slug),
"dependencyProfile": RUNTIME_DEPENDENCIES.get(slug, []),
"requiresGpu": slug in GPU_SLUGS,
"containerCommand": COMMAND_ENTRYPOINT,
"containerArgs": [],
}
)
files[MANIFEST_PATH] = dump_json(
{
"schemaVersion": "droboticflow.operator-images/v1",
"registry": DEFAULT_REGISTRY,
"tag": IMAGE_TAG,
"items": sorted(manifest_items, key=lambda item: item["operatorId"]),
}
)
return files
def write_files(catalog_items: list[dict[str, Any]], files: dict[pathlib.Path, str]) -> None:
CATALOG_PATH.write_text(dump_json(catalog_items), encoding="utf-8")
for path, content in files.items():
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(content, encoding="utf-8")
def check_files(catalog_items: list[dict[str, Any]], files: dict[pathlib.Path, str]) -> int:
mismatches: list[str] = []
expected_catalog_text = dump_json(catalog_items)
if CATALOG_PATH.read_text(encoding="utf-8") != expected_catalog_text:
mismatches.append(f"outdated {CATALOG_PATH.relative_to(ROOT)}")
for path, content in files.items():
if not path.exists():
mismatches.append(f"missing {path.relative_to(ROOT)}")
continue
if path.read_text(encoding="utf-8") != content:
mismatches.append(f"outdated {path.relative_to(ROOT)}")
if mismatches:
for mismatch in mismatches:
print(mismatch, file=sys.stderr)
return 1
return 0
def main() -> int:
parser = argparse.ArgumentParser(description="Generate or check independent builtin operator images.")
mode = parser.add_mutually_exclusive_group()
mode.add_argument("--write", action="store_true", help="write catalog image contracts and Dockerfiles")
mode.add_argument("--check", action="store_true", help="check generated catalog image contracts and Dockerfiles")
args = parser.parse_args()
catalog_items = expected_catalog()
files = expected_files(catalog_items)
if args.check:
return check_files(catalog_items, files)
write_files(catalog_items, files)
return 0
if __name__ == "__main__":
raise SystemExit(main())