Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add nebius/nebius-physical-ai --skill ray-train-syntheticgit clone --depth 1 https://github.com/nebius/nebius-physical-aiWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/nebius/nebius-physical-ai/ray-train-synthetic)<a href="https://agentmods.dev/skills/nebius/nebius-physical-ai/ray-train-synthetic"><img src="https://agentmods.dev/badge/skills/nebius/nebius-physical-ai/ray-train-synthetic/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/nebius/nebius-physical-ai/ray-train-synthetic"><img src="https://agentmods.dev/badge/skills/nebius/nebius-physical-ai/ray-train-synthetic.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00033 | $0.00811 |
| Opus 5 | $0.00016 | $0.00405 |
| Sonnet 5 | $0.00007 | $0.00162 |
| Haiku 4.5 | $0.00003 | $0.00081 |
Grade A, and why
ray-train-synthetic scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
What it actually says
Native Ray Train reference
Use the reference guide.
This is a guarded application example: native Ray Jobs delivers source and owns
submission/status/logs/stop; SkyPilot owns the development hosts and service task.
It adds no NPA Jobs wrapper, controller, CLI group, image or workflow catalog.
Use npa.workflow for production composition beyond this single application.
Read health-preflight, gpu-selection, skypilot-workflows,
third-party-eula-preflight, and teardown-and-cost before live operation.
Preflight the selected project's exact workload bucket and prefix; do not use a
different project's Terraform-state or trajectory bucket as training storage.
The pins are application Ray 2.58.0 / Train V2, Torch 2.13.0+cu130 in the
digest-pinned upstream PyTorch image, Rerun 0.31.4 and Pillow 12.3.0. The driver and every new
or restarted worker enforce the Torch pin at run time. Keep application Ray
separate from SkyPilot's management environment and reserved ports. Pure Python
source changes use native Jobs --working-dir; native ABI changes require a
compatible prepared environment. There is no model or external dataset download.
Preparation uses the pinned image's root account and requires a fresh owned
mode-0700 /opt/npa-ray-train below an /opt without shared write access.
Keep its environment, exports, preparation receipt, and application Ray temporary
files inside that directory. Existing directories and symlinks fail closed;
preserve failed-attempt evidence and use fresh hosting pods rather than clearing
or adopting an unknown runtime. The service validates ownership and permissions
before it starts and scopes RAY_TMPDIR to this application directory.
Use two B200 hosts with one GPU each for the shipped profile. Require actual
rank/world-size, CUDA device and physical-host evidence before calling a result
multi-node. SPREAD and a requested GPU count alone are not placement evidence.
Ray Train V2 recovers through the same RunConfig(name, storage_path) and worker
get_checkpoint(). Do not use V1 resume_from_checkpoint, restore() or
trainer_resources. Every worker must call report equally often; only rank
zero uploads checkpoint bytes. Before injected failure, wait for native
CheckpointConsistencyMode.COMMITTED, then prove every worker resumed at the
next optimizer step with model and momentum state restored.
S3 credentials belong in the private hosting environment on every Ray host.
Pass no keys to S3FileSystem itself: its serialization otherwise carries them.
Keep credentials out of source delivery, Jobs runtime-env JSON and reports.
Store checkpoints with the native Arrow filesystem; upload the final export
with its checksum manifest last and require read-after-write verification.
Validate all optimizer steps and scalar values decoded from metrics.rrd
against metrics.json using inspect_results.py. Independently run
rerun rrd verify and rerun rrd print -vv. After compute teardown, download
again from S3 and reload the native model/optimizer checkpoint. Synthetic
regression proves training execution and recovery, not robot-policy quality.
npa/.venv/bin/python -m pytest npa/tests/workflows/test_ray_train_synthetic.py -q
The opt-in live suite is npa/tests/e2e/test_ray_train_synthetic_live.py.
Its private configuration selects an already-preflighted isolated Ray Jobs
endpoint and workload S3 prefix. Cancel exact Jobs before the Sky service task
and named development cluster; retain shared infrastructure and verified artifacts.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 3d ago First seen · 65 lines · 33 tokens per session scan A 9e3ad116380c
ray-train-synthetic is a skill published in the GitHub repository nebius/nebius-physical-ai (29 stars, last pushed today), licensed Apache-2.0. It adds 33 tokens to every session and 811 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-09.
Other skills, from other repositories
tensorrt-llm
High-throughput LLM inference on NVIDIA GPUs.
google-cloud-solution-guided-gke-ai-migration
Guides the migration of existing AI workloads (Cloud Run, Gemini API, Gemini Enterprise Agent Platform) to self-hosted GKE inference using gcloud and kubectl. Use when the user has an existing AI inference workload (on Cloud Run, the Gemini API, Gemini Enterprise Agent Platform, or a custom VM) and wants to move it to…
agent-platform-tuning
Agent Platform Model Tuning. Use when you need to fine-tune open models or Gemini models using Agent Platform infrastructure. Don't use for model training outside Agent Platform, model deployment to endpoints (use agent-platform-deploy), or managing serving endpoints (use agent-platform-endpoint-management).
modal
Modal is a serverless cloud platform for running Python on demand, including on-demand GPUs. Use when deploying or serving AI/ML models, running GPU-accelerated workloads (training, fine-tuning, inference), serving web endpoints, scheduling batch jobs, or scaling Python code to cloud containers with the Modal SDK.
agent-platform-endpoint-management
Manages Agent Platform serving endpoints. Use when you need to create, list, describe, update, or delete serving endpoints for model deployment on Agent Platform. Also use when troubleshooting endpoint permission, quota, or resource busy errors. Don't use for deploying models to endpoints or for running model…
gke-inference
Deploys and optimizes AI/ML inference workloads on GKE, using GPUs, TPUs, and model servers. Use when deploying GKE inference servers, configuring GKE GPU resources for inference, or deploying LLMs on GKE. Don't use for generic batch jobs or HPC task queues (use gke-batch-hpc instead).