tao-run-on-brev v0.2.0B

Immutable. This exact content never changes and is served at /api/v1/blob/fc0cc42415f89d69.

---
name: tao-run-on-brev
description: Run a TAO training/evaluation/inference container on an NVIDIA Brev GPU instance. Instance provisioning (create/search/stop/delete/login) is delegated to the official brev-cli agent skill or the Brev MCP server; this skill covers only the TAO-specific part — running the container over `brev exec` via the four-verb docker contract. Trigger phrases include "run on Brev", "Brev GPU instance", "TAO on Brev", "submit job to Brev".
license: Apache-2.0
compatibility: Requires the brev CLI (https://github.com/brevdev/brev-cli) and an active brev login. Instance provisioning is handled by the official brev-cli agent skill or the Brev MCP server.
metadata:
  author: NVIDIA Corporation
  version: "0.2.0"
allowed-tools: Read Bash
tags:
- gpu
- compute
- instance-based
- brev
---

# Brev — TAO execution glue

> **Standalone install?** If this session was not initialized by the TAO skill bank plugin, run the `tao-setup` skill first (host preflight, credentials, cross-skill discovery).

NVIDIA Brev provides on-demand GPU instances (pre-loaded with NVIDIA drivers,
CUDA, Docker, and the NVIDIA Container Toolkit). Brev is **instance-based**: you
provision an instance, run commands on it over `brev exec`, and delete it when
done.

This skill is deliberately thin. **Provisioning and managing instances — create,
search by GPU/price, start/stop, delete, login — is owned by NVIDIA Brev's own
agent skill, not duplicated here.** This skill covers only the TAO-specific part:
running a TAO container on a reached instance through the **four-verb docker
contract**, deferring the container-how to `tao-run-on-docker` over `brev exec`.

## Provisioning: use the official Brev skill or MCP

NVIDIA Brev publishes an agent skill that manages instances in natural language
("create an A100 instance", "search for GPUs under $3/hr", "stop all my
instances"). Install it once — it self-registers into your agent's skills dir and
is discovered at runtime:

```bash
curl -fsSL https://raw.githubusercontent.com/brevdev/brev-cli/main/scripts/install-agent-skill.sh | bash
# installs to ~/.claude/skills/brev-cli/ , ~/.codex/skills/brev-cli/ , ~/.agents/skills/brev-cli/
```

Or connect the **Brev MCP server** (`https://docs.nvidia.com/brev/_mcp/server`).
Either one owns login/auth quirks, placement IDs, GPU search, and teardown flags.
It does **not** cover container execution on the instance — that is this skill.

**Preflight for this skill:** the `brev` CLI is on `PATH` and logged in (headless:
`brev login --token "$BREV_API_TOKEN"` before any other call), and you can reach a
target instance — poll with a **two-word** command until it succeeds before
issuing real work (a fresh instance reports `RUNNING` before sshd is up):

```bash
for i in $(seq 1 60); do brev exec <instance> "echo ok" 2>/dev/null | grep -qx ok && break; sleep 5; done
brev exec <instance> "echo ok" 2>/dev/null | grep -qx ok || { echo "instance not exec-ready"; exit 1; }
```

The probe must be **two words, quoted as one argument**. A single-token probe
(`brev exec <instance> -- true`) passes even when every real command is broken,
because `brev exec [instance...] <command>` treats only the LAST positional as
the command — so a lone `true` lands in the right slot by accident while
`docker run ...` does not. See *`brev exec` argument form* below.

Allow **≥ 600 s** for the first `brev exec` on a new instance (SSH bring-up +
first container pull); a 60–120 s wrapper timeout truncates startup and looks
like a spurious `exec failed`.

## Storage

No shared NFS/Lustre — storage tier **B/C** via `tao-data-io`: stage inputs from
S3 to the instance's local disk (or fetch in-container) and **upload results to S3
before deleting the instance**. Instance-local `~/` persists across stop/start but
**not** across delete/create, so the results upload must precede teardown.

## Execution — the four verbs (a compound over Docker)

Brev is a **compound consumer**: `submit` reaches an instance, then **defers the
container-how to the four docker verbs** (`tao-run-on-docker`) run over
`brev exec`. It is not a symmetric peer — teardown must additionally delete the
instance to stop billing. `$BANK` = `${TAO_SKILL_BANK_PATH}`.

- **submit** — reach an instance (provision/reuse via the official Brev skill or
  MCP; reuse an existing instance by its `instance_id`; wait for readiness, above).
  Lint the assembled command, open the record to mint `$JOB_ID` **before** launch,
  then run the docker `submit` verb *inside* the instance and mark RUNNING:

  ```bash
  redact_secrets.py lint <<<"$REMOTE_CMD"     # no inline secrets; creds as -e VAR
  JOB_ID=$("$BANK/scripts/tao_job_record.py" open \
    --platform brev --image "$IMG" \
    --network-arch "$ARCH" --action "$ACTION" \
    --storage-tier "$TIER" --results-root "$RESULTS_ROOT")
  brev exec <instance> "docker inspect '$JOB_ID' >/dev/null 2>&1 && { echo '$JOB_ID already submitted'; exit 0; }; docker run -d --name '$JOB_ID' --label 'tao-job=$JOB_ID' ..."
  "$BANK/scripts/tao_job_record.py" mark "$JOB_ID" --state RUNNING \
    --backend-ref "<instance>/$JOB_ID"       # instance is part of the ref: the
                                             # container is unreachable without it
  ```
- **status / logs** — `brev exec <instance> "docker inspect $JOB_ID"` /
  `brev exec <instance> "docker logs $JOB_ID"`, mapped to the vocab exactly as
  the docker verbs do. Recover `<instance>` from the record's `backend-ref`.
- **cancel / teardown** — remove the container, then for an ephemeral instance
  delete it (stops billing), then mark the record. Never leave an ephemeral
  instance running:

  ```bash
  brev exec <instance> "docker rm -f $JOB_ID"
  brev delete <instance>                      # ephemeral instances only
  "$BANK/scripts/tao_job_record.py" mark "$JOB_ID" --state CANCELED --source agent
  ```

### `brev exec` argument form

The remote command is **one argument**. The CLI signature is
`brev exec [instance...] <command>`: every positional except the last is an
instance name, and `--` only ends flag parsing — it does not group the words
after it. So `brev exec <inst> -- docker inspect "$JOB_ID"` is read as
instances `<inst> docker inspect` plus command `"$JOB_ID"`, and fails with
`could not look up instance "docker"` / `ssh: illegal option -- -` (exit 255) —
an error that reads like an instance or SSH fault but is a syntax fault. Quote
every remote command as a single string, exactly as `brev exec --help` shows.

NGC auth once per instance — **never put `NGC_KEY` on argv** (it lands in the
remote process table); pipe it to `--password-stdin`:

```bash
IMG=nvcr.io/nvstaging/tao/tao-toolkit-pyt:7.2.0-rc-36-multiarch  # versions-key: images.tao_toolkit.pyt

# NGC auth (one-time per instance) — value never on argv.
# Single-quoted locally so $NGC_KEY expands in the instance's shell; export it
# there first (or pipe it in from the local shell, if the instance has no copy).
brev exec <instance> 'printf %s "$NGC_KEY" | docker login nvcr.io -u "$oauthtoken" --password-stdin'

# Verify auth without reading ~/.docker/config.json. Failure before a successful
# login = not authenticated; failure after = the key's org lacks entitlement.
brev exec <instance> "docker manifest inspect $IMG >/dev/null && echo AUTH_OK || echo AUTH_FAIL"

# Pull BEFORE the GPU run. `docker run` would pull implicitly, but the instance
# bills from boot, so a multi-GB first-time TAO pull is billed GPU-idle time.
# Pulling as its own step also separates a pull failure (auth/entitlement) from
# a training failure in the logs.
brev exec <instance> "docker image inspect $IMG >/dev/null 2>&1 || docker pull $IMG"

# Run a TAO job (the docker `submit` verb, over brev exec)
brev exec <instance> "docker inspect '$JOB_ID' >/dev/null 2>&1 && { echo '$JOB_ID already submitted'; exit 0; }; docker run -d --name '$JOB_ID' --label 'tao-job=$JOB_ID' --gpus all -v ~/data:/data -e NGC_KEY '$IMG' visual_changenet train -e /data/spec.yaml"
```

## Multi-GPU and multi-node

**Multi-node is not supported on Brev** — instance-based, no cross-instance
coordination. Multi-GPU **on a single instance** is supported (up to 8× H100 /
A100 / L40S); `torchrun --nproc-per-node=N` or PyTorch DDP work within the
instance.