tao-run-on-brev

Brev

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 across multiple cloud providers. Instances come pre-loaded with NVIDIA drivers, CUDA, Docker, and NVIDIA Container Toolkit.

Brev is instance-based (not job-based). You create an instance, run commands on it via brev exec, and delete it when done. The TAO SDK's BrevHandler wraps this into the standard job interface.

Preflight

This skill needs the brev CLI and an active login. Check before proceeding:

# 1. brev CLI installed
command -v brev >

If any non-pip step fails, the agent prompts the user to authorize the fix via Bash, then re-runs the preflight before continuing. The TAO SDK is not required for Brev — brev exec <instance> "docker run …" is sufficient. Reach for the SDK only if you want Job handles, S3 I/O wrapping via script_runner, or state persistence; nvidia-tao-sdk is on public PyPI; install missing SDK requirements automatically from the pinned Brev extra: python -m pip install "nvidia-tao-sdk[brev]==7.1.0rc42" . When going the SDK route, read tao-skill-bank:tao-run-platform for the BrevSDK kwarg reference, build_entrypoint, and ActionWorkflow patterns.

Authentication

Two options:

1. Automated (recommended): Get an API token from the Brev console settings page. Set BREV_API_TOKEN as an environment variable (e.g., export BREV_API_TOKEN=... in your shell). The handler auto-authenticates via brev login --token on first use.

2. Manual: Run brev login (opens browser). Tokens expire hourly — the handler refreshes automatically.

S3 credentials (ACCESS_KEY, SECRET_KEY) are needed separately for data transfer.

Headless / non-interactive

In a CI shell, container, or agent session with no controlling TTY, always run brev login --token "$BREV_API_TOKEN" before any other brev call — even when the token is exported. Otherwise the CLI prompts on stdin and returns an EOF auth error on commands like brev ls, brev create, or brev exec. Re-run the token login if a call returns auth-EOF; a single refresh is usually enough.

Launch Preflight

Before generating scripts or submitting jobs:

1. Verify BREV_API_TOKEN is set.
2. Verify the brev CLI is installed and can list instances, for example brev ls --json. If needed, authenticate with brev login --token.
3. For s3:// datasets/results, verify ACCESS_KEY and SECRET_KEY are set and the exact paths are readable with aws s3 ls.
4. Do not accept local /path inputs for Brev unless the user has proven those paths exist on the target Brev instance or are mounted into it.
5. Verify model-specific credentials such as HF_TOKEN before launch.

Instance Lifecycle

The agent controls instance lifecycle:

• Reuse: Pass instance_id in backend_details to run multiple jobs on the same instance. Efficient for multi-step workflows.
• Ephemeral: Omit instance_id — the handler creates a new instance per job. Clean but slower (instance boot ~2-5 min).

Creating an instance — placement info

For accounts with more than one cloud credential or workspace group, plain brev create rejects the call with a placement error. Pass the account-specific IDs explicitly:

Discover the values once and export them in your shell before launching:

When using the SDK, pass them through backend_details:

Multi-GPU and multi-node

Multi-node is not supported on Brev. Brev is instance-based — one job runs on one instance, with no cross-instance coordination.

Multi-GPU on a single instance is supported (instances available with up to 8× H100 / A100 / L40S). gpu_count maps to the GPU count on the instance; torchrun --nproc-per-node=N or PyTorch DDP work within the instance.

GPU Types

Available via brev search:

• L40S, A100 80GB, H100 (availability varies by provider)
• Use --gpu-name to filter, --min-vram for memory requirements

Storage

No shared NFS/Lustre. All data flows through S3 via the script_runner's fsspec integration. Instance-local disk at ~/ persists across stop/start but not across delete/create.

Docker on Brev

VM Mode instances have Docker pre-installed.

brev exec syntax — the remote command is ONE quoted string

The CLI signature is brev exec [instance...] <command>: every positional except the last is treated as an instance name, and -- only terminates flag parsing — it does not bundle the words after it into a single command. The multi-token form brev exec <instance> -- docker run --gpus all … therefore makes the CLI treat docker, run, … as instance names and fails with could not look up instance "docker" / illegal option -- - (exit 255) — an error that reads like an SSH/instance fault but is a syntax fault. A single-token probe (brev exec <instance> -- true) "works" only because the lone token is parsed as the command; never take it as proof that exec is healthy. Always pass the whole remote command as one quoted string:

NGC auth + running TAO containers

Wait for instance readiness before the first brev exec

A freshly created instance reports RUNNING long before sshd, hostname resolution, and the user shell are ready. The first brev exec against an unsettled instance fails with hostname not resolvable, Connection refused, or a silent timeout. Always poll until a trivial exec succeeds before issuing real work. The probe must be a multi-word quoted command: it proves both SSH readiness and the single-string exec form that every real command depends on (a bare true probe passes even when every multi-token command would fail):

brev exec timeout for cold-start workloads

brev exec inherits no default timeout, but anything that wraps it (the SDK handler, CI step wrappers, timeout shell builtins) must allow time for both the SSH bring-up window and the container pull on a fresh instance. Use ≥ 600 s (10 min) for the first exec on a new instance; the previous 60–120 s default truncates remote startup and surfaces as a spurious exec failed even though the remote command is still progressing.

Cleanup

The CLI does not accept --yes / -y; passing it errors with unknown flag: --yes. brev delete <instance> is already non-interactive on recent CLIs, so no confirmation flag is needed.

Error Patterns

brev CLI not found: Install from https://docs.nvidia.com/brev/.

brev ls returns auth EOF even with BREV_API_TOKEN set: Headless shell has no stdin for the interactive auth prompt. Run brev login --token "$BREV_API_TOKEN" first, then retry. If the failure persists across a single retry, the token itself is stale — mint a fresh one.

Token expired: Handler auto-refreshes via brev login --token. If persistent, run brev login manually.

brev create rejected with placement error (cloudCredId / workspaceGroupId required): Multi-credential or multi-workspace accounts must pass --cloud-cred-id and/or --workspace-group-id. See Creating an instance — placement info above.

brev exec fails with hostname not resolvable or Connection refused right after create: Instance reports RUNNING before sshd is up. Use the readiness-wait loop in Wait for instance readiness before the first brev exec before issuing the real command.

could not look up instance "<word>" / ssh: illegal option -- - / exit 255 on a multi-word brev exec: The remote command was passed as separate tokens (brev exec <instance> -- docker run …). The CLI parses every positional except the last as an instance name, so docker, run, … are looked up as instances and stray --flags leak to ssh. Pass the whole remote command as ONE quoted string: brev exec <instance> "docker run …". See brev exec syntax above.

SDK exec timeout / exec failed on a fresh instance: The SDK's brev exec wrapper timed out before remote startup finished. Raise the timeout to ≥ 600 s for cold-start runs (see brev exec timeout for cold-start workloads).

brev delete --yes: unknown flag: --yes: The CLI has no confirmation flag. Use plain brev delete <instance>.

Instance stuck in provisioning: Some GPU types have limited availability. Try a different --gpu-name or provider.

Docker pull / docker manifest inspect fails on nvcr.io: Two distinct causes — distinguish them by login state:

• No prior Login Succeeded on the instance: not authenticated — NGC_KEY unset, expired, or the login never ran. Re-run the stdin login from NGC auth + running TAO containers with a valid key.
• Login succeeded but pull/inspect still returns Access Denied / unauthorized: the key authenticates but its NGC org lacks entitlement for the image's org (e.g. a pre-release staging org). Request org access or select an image from an org the key can read — re-running login will not help.