/physical-ai-video-data-augmentation

Video Data Augmentation — Setup

Table of Contents

• What you get
• Prerequisites
• Credential check
• Path A — OSMO cache workflow (recommended)
• Path B — Manual cache publication
• URL layout
• Wiring into VDA flows
• Troubleshooting

One-time bootstrap for VDA execution: credentials, storage prefix, model caches, and submit prerequisites. Runtime workflows assume this setup exists.

What you get

Prerequisites

1. OSMO CLI installed and authenticated:

   command -v osmo && osmo version
   osmo profile show

2. OSMO credentials available:

   • hf_token (GENERIC) for Hugging Face downloads.
   • active DATA credential/profile matching target backend.
   • nvcr_io (REGISTRY) is optional and used when you want to refresh/store registry credentials explicitly.

3. Backend-native storage root known (s3://..., azure://..., swift://..., etc.). This root becomes storage_url in submit commands.

Credential check

Run before setup and before workflow submission:

bash scripts/preflight_credentials.sh --workflow assets/configs/osmo/<mode>.yaml

Restricted egress:

bash scripts/preflight_credentials.sh --no-probe --workflow assets/configs/osmo/<mode>.yaml

If output contains USER_INPUT_REQUIRED:, ask one concise unblock question and stop. Do not continue with submit-time interpolation until this gate passes. Pass --workflow to validate exact workflow image refs via registry probe. If rotated secrets are supplied in env, preflight refreshes existing OSMO credentials automatically. To force overwrite without new env secrets:

bash scripts/preflight_credentials.sh --workflow assets/configs/osmo/<mode>.yaml --refresh

Path A — OSMO cache workflow (recommended)

Use the built-in cache workflow and publish to backend-native storage:

osmo workflow submit assets/configs/osmo/setup_model_cache.yaml \
  --set-string storage_url=<backend-prefix> path=data

This creates:

• {{storage_url}}/data/models/cosmos_transfer
• {{storage_url}}/data/models/auto_labeling

Expected runtime is typically tens of minutes depending on outbound bandwidth.

Path B — Manual cache publication

Use only when Path A is unavailable in the environment. The final object storage layout must still match:

<storage_url>/data/models/cosmos_transfer
<storage_url>/data/models/auto_labeling

If manual publishing deviates from this layout, pass explicit overrides at submit time (append to the same --set-string list):

cosmos_model_cache_url=<custom-cosmos-cache-url> \
auto_labeling_model_cache_url=<custom-auto-labeling-cache-url>

URL layout

Default root:

storage_url=<backend-prefix>

Always derive storage_url from the actual dataset/upload target for that run. Do not silently keep stale s3:// prefixes on non-S3 backends.

Wiring into VDA flows

Before submit:

python3 scripts/pre_submit_guard.py --workflow assets/configs/osmo/<mode>.yaml

If guard reports cache validation failure, default behavior is to run Path A, then rerun guard and proceed only after it passes.

Required submit variables across flows:

storage_url, dataset, run_id, gpu_platform, skills_dir, video

Optional (defaults supplied by YAML):

cookbook=city_traffic
vlm_url=http://qwen3-vl.osmo-nims.svc.cluster.local:8000/v1
llm_url=http://qwen25-14b.osmo-nims.svc.cluster.local:8000/v1

Troubleshooting