Full CLI Reference
Command-line interface for NeMo Platform.
Getting started:
- Browse documentation with
nemo docs --list - Run local platform services with
nemo services run --help - Read the Kubernetes deployment guide with
nemo docs set-up/helm/install
Examples:
Exit codes:
- 0: Success
- 1: Local or unexpected error
- 2: Command usage error
- 3: Remote/API error
Usage:
Global Options:
--base-url: Base URL for the NeMo Platform API--output-format, --output, -f <CHOICE>: Output format for how results are printed. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output--timestamp-format <CHOICE>: Timestamp format for table/markdown/csv output [possible values: relative, iso8601]--verbose, -v: Enable verbose messaging. This only impacts logs that are visible, it doesn’t change any data outputs.--agent-mode, -A: Enable agent-friendly output mode with extra context for coding agents.--no-telemetry: Disable anonymous usage telemetry for this invocation.
Help:
--version, -V: Show version information and exit.--install-completion: Install completion for the current shell.--show-completion: Show completion for the current shell, to copy it or customize the installation.--help, -h: Show this message and exit.
Setup
nemo setup
Set up NeMo Platform: connect or start services, configure a provider, install skills.
Uses an already-running platform, starts local services, or connects the CLI to an existing remote deployment. Then selects and registers an inference provider, picks default and fast agent models, installs coding agent skills, and optionally deploys a demo agent.
The active config context remembers the Platform URL. When a remote deployment is already reachable, setup asks whether to continue with it, start local services instead, or connect to a different remote URL.
To override the URL for one run only: nemo —base-url http://localhost:8080 setup
To persist a different URL: nemo config set —base-url http://localhost:8080
Requires an interactive terminal (TTY). In non-interactive contexts (CI, piped input), pass —auto to use environment variables instead.
Use —auto for non-interactive setup from environment variables (NEMO_DEFAULT_INFERENCE_KEY, NVIDIA_API_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY). Override the selected pair with NEMO_DEFAULT_MODEL and NEMO_FAST_MODEL.
Examples:
Usage:
Options:
--auto: Non-interactive mode: register provider from environment variables--workspace, -w: Target workspace [default: default]--start-services, --no-start-services: Start local platform services--install-skills, --no-install-skills: Install NeMo skills for coding agents--skills-agents: Comma-separated list of agents to install skills for (e.g. ‘codex,cursor’). Default: all detected. Only applied when —install-skills is set.--skills-scope <CHOICE>: Install scope for skills: ‘project’ (this repo) or ‘user’ (home). Default: project. Only applied when —install-skills is set. [possible values: project, user]--skills-from: Comma-separated list of skill sources to install from (e.g. ‘nemo-platform,nemo-evaluator-plugin’). Use ‘nemo-platform’ for the built-in set. Default: all sources. Only applied when —install-skills is set.--deploy-agent, --no-deploy-agent: Deploy the demo calculator agent--resume: Retry an interrupted setup using the normal idempotent setup path.--ready-timeout <INTEGER>: Seconds to wait for platform readiness (default: 240)
Help:
--help, -h: Show this message and exit.
nemo auth
Manage authentication for NeMo Platform.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
login: Authenticate with the NeMo Platform cluster.logout: Remove stored credentials for the current context.refresh: Refresh the current access token.token: Print the current access token (for use with SDK or curl).status: Show current authentication status.access-keys: Manage NeMo Platform Scoped Access Keys.
nemo auth login
Authenticate with the NeMo Platform cluster.
Uses device flow (browser) by default, or password grant when username and password are provided (e.g. for CI).
For quickstart, use --unsigned-token to generate an unsigned JWT.
Examples:
Usage:
Options:
--context: Context to use for this login command.--base-url: Set cluster base URL for the selected context before login--no-browser: Don’t open browser (device flow only)--scope: OAuth scopes to request (space-separated; quote for multiple, e.g. —scope “platform:read secrets:write”)--username: Username for password grant (CI / non-interactive)--password: Password for password grant (prefer env NMP_OIDC_PASSWORD)
Help:
--help, -h: Show this message and exit.
Unsigned Token Options:
--unsigned-token: Generate and save an unsigned JWT for local/testing authentication.--principal-id: Principal ID for the unsigned token (subclaim). Defaults to —email.--email: Email claim for the unsigned token (required with —unsigned-token).--group: Group claim value for unsigned token (repeat for multiple).--expires-in <INTEGER>: Unsigned token expiry in seconds from now. [default: 3600]--no-exp: Omit the exp claim from the unsigned token.--audience: Audience (aud) claim for unsigned token.--issuer: Issuer (iss) claim for unsigned token.
nemo auth logout
Remove stored credentials for the current context.
Usage:
Help:
--help, -h: Show this message and exit.
nemo auth refresh
Refresh the current access token.
This command uses the saved refresh token to obtain a new access token without requiring you to re-authenticate through the browser.
Usage:
Help:
--help, -h: Show this message and exit.
nemo auth token
Print the current access token (for use with SDK or curl).
By default this outputs the raw token to stdout, suitable for piping or capture.
Examples:
Usage:
Options:
--decode: Decode the JWT payload claims as JSON. This does not verify the token signature.
Help:
--help, -h: Show this message and exit.
nemo auth status
Show current authentication status.
Usage:
Help:
--help, -h: Show this message and exit.
nemo auth access-keys
Manage NeMo Platform Scoped Access Keys.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create a user-bound or service-bound Scoped Access Key.list: List Scoped Access Keys owned by the currently…revoke: Revoke a Scoped Access Key owned by the currently…suspend: Temporarily suspend a Scoped Access Key owned by the…unsuspend: Restore a suspended Scoped Access Key owned by the…
nemo auth access-keys create
Create a user-bound or service-bound Scoped Access Key.
Usage:
Options:
--name, -n: Optional human-readable label for the Scoped Access Key.--description, -d: Optional description for the Scoped Access Key.--expires-in: Scoped Access Key lifetime in seconds. Use ‘none’ to request no expiration.--service-account: Bind the key to a non-human service account (PlatformAdmin only).
Help:
--help, -h: Show this message and exit.
nemo auth access-keys list
List Scoped Access Keys owned by the currently authenticated user.
PlatformAdmins also see every service-bound Scoped Access Key, not just the ones they personally created.
Usage:
Options:
--page <INTEGER RANGE>: Page number to retrieve. [default: 1]--page-size <INTEGER RANGE>: Number of keys to retrieve per page. [default: 100]
Help:
--help, -h: Show this message and exit.
nemo auth access-keys revoke
Revoke a Scoped Access Key owned by the currently authenticated user.
Usage:
Arguments:
<JTI>: Stable ID of the Scoped Access Key to revoke.
Help:
--help, -h: Show this message and exit.
nemo auth access-keys suspend
Temporarily suspend a Scoped Access Key owned by the current user.
Unlike revocation, suspension is reversible until the key expires.
Usage:
Arguments:
<JTI>: Stable ID of the Scoped Access Key to suspend.
Help:
--help, -h: Show this message and exit.
nemo auth access-keys unsuspend
Restore a suspended Scoped Access Key owned by the current user.
Usage:
Arguments:
<JTI>: Stable ID of the Scoped Access Key to unsuspend.
Help:
--help, -h: Show this message and exit.
nemo services
Run platform services locally.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
run: Run platform services in the foreground.start: Start platform services in the background.stop: Stop running platform services.restart: Restart platform services.status: Show status of the platform services instance for this…ls: List service instances on this host.rm: Remove a stopped instance directory and its logs.prune: Remove all stopped instance directories on this host.logs: Show or locate the service log file.
nemo services run
Run platform services in the foreground. Ctrl-C to stop.
Usage:
Options:
--services: Comma-separated services to run, e.g. models,entities,jobs. Defaults to all available services.--service-group: Run a predefined service group. Cannot be combined with —services.--controllers: Comma-separated controllers to run, e.g. jobs,models.--controller-group: Run a predefined controller group. Cannot be combined with —controllers.--sidecars: Comma-separated sidecars to run, e.g. adapters,cache.--config: Path to a platform configuration YAML file.--host: Host to bind to. [default: 127.0.0.1]--port <INTEGER>: Port to bind to. [default: 8080]--keep-alive-timeout-seconds <INTEGER RANGE>: Seconds Uvicorn keeps idle HTTP connections open. [default: 5]--instance: Instance name. Defaults to a name derived from the working directory and port.
Help:
--help, -h: Show this message and exit.
nemo services start
Start platform services in the background.
Detaches the process, polls /status, then returns.
Examples:
Usage:
Options:
--services: Comma-separated services to run, e.g. models,entities,jobs.--service-group: Run a predefined service group. Cannot be combined with —services.--controllers: Comma-separated controllers to run, e.g. jobs,models.--controller-group: Run a predefined controller group. Cannot be combined with —controllers.--sidecars: Comma-separated sidecars to run, e.g. adapters,cache.--config: Path to a platform configuration YAML file.--host: Host to bind to. [default: 127.0.0.1]--port <INTEGER>: Port to bind to. [default: 8080]--keep-alive-timeout-seconds <INTEGER RANGE>: Seconds Uvicorn keeps idle HTTP connections open. [default: 5]--instance: Instance name. Defaults to a name derived from the working directory and port.
Help:
--help, -h: Show this message and exit.
nemo services stop
Stop running platform services.
Sends SIGTERM to the running service process and waits for it to exit.
Falls back to SIGKILL after a timeout. Foreground instances (started
with run) are protected; use --force to override.
Examples:
Usage:
Options:
--timeout <FLOAT>: Seconds to wait before SIGKILL. [default: 30.0]--instance: Instance name. Defaults to a name derived from the working directory and port.--port <INTEGER>: Port (used for scope computation if —instance not given). [default: 8080]--force: Stop even if the instance is running in the foreground.
Help:
--help, -h: Show this message and exit.
nemo services restart
Restart platform services.
Stops any running services and relaunches them. Without flags, preserves the service set from the previous run. Errors if no previously tracked instance exists for the computed scope; does not start a fresh instance.
Examples:
Usage:
Options:
--services: Comma-separated services to run. Overrides previous service set.--service-group: Run a predefined service group. Overrides previous setting.--controllers: Comma-separated controllers to run. Overrides previous controller set.--controller-group: Run a predefined controller group. Overrides previous setting.--sidecars: Comma-separated sidecars to run. Overrides previous setting.--config: Path to a platform configuration YAML file.--host: Host to bind to. Defaults to previous value or 127.0.0.1.--port <INTEGER>: Port to bind to. Defaults to previous value or 8080.--keep-alive-timeout-seconds <INTEGER RANGE>: Seconds Uvicorn keeps idle HTTP connections open. Defaults to the previous value or 5.--instance: Instance name. Defaults to a name derived from the working directory and port.
Help:
--help, -h: Show this message and exit.
nemo services status
Show status of the platform services instance for this scope.
Usage:
Options:
--instance: Instance name. Defaults to a name derived from the working directory and port.--port <INTEGER>: Port (used for scope computation if —instance not given). [default: 8080]
Help:
--help, -h: Show this message and exit.
nemo services ls
List service instances on this host.
By default shows running instances only. Use --all to include stopped
instance directories that still have logs on disk.
Examples:
Usage:
Options:
--all, -a: Include stopped instance directories (like docker ps -a).
Help:
--help, -h: Show this message and exit.
nemo services rm
Remove a stopped instance directory and its logs.
The scope must match a row from nemo services ls --all. Running
instances are refused; stop them first.
Unlike run/start, --instance here does not derive a scope from
cwd and port — it is an alternate spelling for the SCOPE argument.
Examples:
Usage:
Arguments:
<SCOPE>: Instance scope from ‘nemo services ls —all’.
Options:
--instance: Scope from ‘nemo services ls —all’ (same value as the SCOPE positional).
Help:
--help, -h: Show this message and exit.
nemo services prune
Remove all stopped instance directories on this host.
Stopped instance directories may include service logs from prior runs. Logs are deleted with the instance directory.
Examples:
Usage:
Options:
--force: Remove without confirmation.
Help:
--help, -h: Show this message and exit.
nemo services logs
Show or locate the service log file.
Examples:
Usage:
Options:
--path: Print the log file path instead of tailing.-n, --lines <INTEGER RANGE>: Number of lines to show from end of log. [default: 50]--instance: Instance name. Defaults to a name derived from the working directory and port.--port <INTEGER>: Port (used for scope computation if —instance not given). [default: 8080]
Help:
--help, -h: Show this message and exit.
nemo skills
Install AI agent skill files for Nemo.
Supported agents: claude, codex, cursor, opencode
Examples:
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
list: List available skills.show: Print skill content to stdout.install: Install Nemo skill files for an AI coding agent.
nemo skills list
List available skills.
The default table word-wraps long descriptions; use --no-truncate to let
descriptions fill the full terminal width. For structured output use
-f json|yaml|csv|markdown. When stdout is not a TTY (pipe/redirect),
JSON is the default so callers get parseable output.
Examples:
Usage:
Options:
--source: Filter to skills from a specific source (distribution / plugin name as shown in theSourcecolumn, e.g.nemo-platform,nemo-agents-plugin). Can be repeated to include multiple sources. Matching is case-insensitive.
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo skills show
Print skill content to stdout.
Without —agent, prints the raw skill content. With —agent, prints the agent-specific formatted version.
Examples:
Usage:
Arguments:
<NAME>: Skill name to show (use ‘nemo skills list’ to see available skills)
Options:
--agent, -a: Agent to format for. Supported: claude, codex, cursor, opencode
Help:
--help, -h: Show this message and exit.
nemo skills install
Install Nemo skill files for an AI coding agent.
By default, installs all skills to project scope. Use —skill to select specific skills, —user for user scope, or —project-dir to explicitly select the project install directory.
Examples:
Usage:
Options:
--agent, -a: Agent to install for (required). Supported: claude, codex, cursor, opencode--skill, -s: Install specific skill(s) only. Can be repeated.--user: Install to user scope (default: project scope)--project-dir, --project-root <DIRECTORY>: Project directory to install into (default: current working directory)
Help:
--help, -h: Show this message and exit.
CLI functions
nemo chat
Start an interactive chat session with a model.
By default, uses model entity routing where the model name should match what’s shown in ‘nemo models list’.
Use —provider for direct provider routing, where the model argument is passed directly to the provider’s API.
Passing PROMPT sends one message and exits unless —interactive is set. Omitting PROMPT in a TTY starts the interactive chat UI. In non-TTY contexts, PROMPT may also be piped on stdin. Piped stdin is read in full before sending. If both PROMPT and piped stdin are provided, PROMPT takes precedence.
Examples:
Usage:
Arguments:
<MODEL>: Model entity name (from ‘nemo models list’) or model ID when using —provider<PROMPT>: Prompt for one-shot mode. Takes precedence over piped stdin.
Options:
--provider: Provider name for direct provider routing (bypasses model entity routing)--workspace: Workspace name
Chat Options:
--interactive: Start the terminal chat UI; cannot be used with piped stdin. With PROMPT, send it first.
Help:
--help, -h: Show this message and exit.
Model Options:
--temperature <FLOAT>: Sampling temperature (0.0 to 2.0)--max-tokens <INTEGER>: Maximum tokens to generate--system-message: System message to set context for the conversation
Output Options:
--output-format, --format, -f <CHOICE>: Output format for one-shot responses. [possible values: text, json, raw]
nemo docs
Read NeMo Platform documentation.
Examples:
Usage:
Arguments:
<PATH>: Path to a doc topic (e.g., get-started/setup or set-up/helm/install). Omit to see available topics.
Options:
--list, -l: List available documentation topics.
Help:
--help, -h: Show this message and exit.
nemo wait
Wait for resources to reach a desired status.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
inference: Wait for inference resources
nemo wait inference
Wait for inference resources
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
deployment: Wait for a deployment to reach a desired status.provider: Wait for the inference gateway to be ready to route to a…
nemo wait inference deployment
Wait for a deployment to reach a desired status.
Polls the deployment status until it reaches the desired state or times out. For READY status, optionally verifies the gateway can route to the provider. For DELETED status, waits for the resource to be fully garbage collected.
Exit codes: 0: Desired status reached 1: Timeout or error
Examples:
Usage:
Arguments:
<NAME>: Name of the deployment to wait for
Options:
--workspace: Workspace name--status, -s <CHOICE>: Desired status to wait for [possible values: READY, DELETED, PENDING, ERROR; default: READY]--timeout, -t <INTEGER RANGE>: Maximum time to wait in seconds [default: 1200]--check-gateway, --no-check-gateway: When waiting for READY, also verify gateway can route to the provider--poll-interval <INTEGER RANGE>: Seconds between status checks [default: 3]
Help:
--help, -h: Show this message and exit.
nemo wait inference provider
Wait for the inference gateway to be ready to route to a provider.
Polls the gateway’s ready endpoint until it can route requests to the specified provider. This is useful after creating a deployment to ensure the gateway has refreshed its cache.
Exit codes: 0: Gateway is ready 1: Timeout
Examples:
Usage:
Arguments:
<NAME>: Name of the provider to wait for
Options:
--workspace: Workspace name--timeout, -t <INTEGER RANGE>: Maximum time to wait in seconds [default: 60]--poll-interval <INTEGER RANGE>: Seconds between status checks [default: 1]
Help:
--help, -h: Show this message and exit.
nemo agent
Commands for AI agent context and capability discovery.
Examples:
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
context: Dump everything an agent needs in one call.commands: List all available top-level CLI commands.
nemo agent context
Dump everything an agent needs in one call.
Outputs installed plugins, CLI commands, entry-point catalog, available skills, and quick-reference patterns. Runs without a connected cluster (metadata-only).
Examples:
Usage:
Help:
--help, -h: Show this message and exit.
nemo agent commands
List all available top-level CLI commands.
Outputs a flat list of commands with descriptions, useful for agent capability discovery.
Examples:
Usage:
Help:
--help, -h: Show this message and exit.
nemo plugins
Commands for plugin discovery.
Examples:
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
list: List installed plugins.
nemo plugins list
List installed plugins.
Discovers installed plugins from registered NeMo plugin entry points.
Examples:
Usage:
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
Core plugins
nemo files
Manage files.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
upload: Upload local files to a fileset.download: Download files from a fileset to a local path.list: List files in a fileset.delete: Delete a file from a fileset.filesets: Manage filesetsotlp: Otlp operations
nemo files upload
Upload local files to a fileset.
Supports uploading single files or directories. For directories, contents are uploaded recursively.
Examples:
# Upload a directory to a subdirectory in the fileset nemo files upload ./data/ my-fileset —remote-path uploads/
# Upload without specifying a fileset (auto-creates one) nemo files upload ./data.csv
Usage:
Arguments:
<LOCAL_PATH>: Local path to upload<FILESET>: Name of the fileset to upload to. If not provided, a new fileset is created.
Options:
--workspace--remote-path: Path within the fileset. Defaults to root. [default: ]
Help:
--help, -h: Show this message and exit.
nemo files download
Download files from a fileset to a local path.
Supports downloading single files or directories. For directories, contents are downloaded recursively.
Examples:
# Download a subdirectory from the fileset nemo files download my-fileset —remote-path data/ -o ./downloads/
Usage:
Arguments:
<FILESET>: Name of the fileset to download from
Options:
--workspace--remote-path: Path within the fileset. Defaults to root. [default: ]--output, -o <PATH>: Local path to download to.
Help:
--help, -h: Show this message and exit.
nemo files list
List files in a fileset.
Lists all files recursively from the specified path within the fileset.
Examples:
# List files in a subdirectory nemo files list my-fileset —remote-path data/
Usage:
Arguments:
<FILESET>: Name of the fileset to list files from
Options:
--workspace--remote-path: Path within the fileset. Defaults to root. [default: ]
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--no-truncate: Don’t truncate long values in table/markdown/csv output.
nemo files delete
Delete a file from a fileset.
Examples:
Usage:
Arguments:
<FILESET>: Name of the fileset containing the file
Options:
--workspace--remote-path: Path of the file to delete within the fileset
Help:
--help, -h: Show this message and exit.
nemo files filesets
Manage filesets
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create a new fileset.delete: Delete Fileset.list: List Filesets endpoint with filtering and pagination.get: Get Fileset by Workspace and Name.update: Update Fileset Metadata.
nemo files filesets create
Create a new fileset.
If no storage configuration is provided, the default storage backend will be used.
Required fields: name
Examples:
Usage:
Arguments:
<NAME>: The name of the fileset. Name must start with a lowercase letter, be 2-63 characters, and use lowercase letters, digits, hyphens, and dots (no consecutive hyphens, cannot end with a hyphen).
Options:
--workspace--cache: Cache all files after creation. Only applies to external storage.--custom-fields: Custom fields for the fileset. (JSON string)--description: The description of the fileset.--metadata: Tagged metadata container - the key indicates the type.
Example: metadata = FilesetMetadata( dataset=DatasetMetadataContent( schema={"columns": ["id", "name"]}, ) ) (JSON string)
--project: The name of the project associated with this fileset.--purpose <CHOICE>: The purpose of the fileset. [possible values: dataset, environment, generic, model]--storage: The storage configuration for the fileset. If not provided, uses default storage. (JSON string)--exist-ok: Do not raise an error if the resource already exists. Returns the existing resource.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo files filesets delete
Delete Fileset.
Permanently deletes an unreferenced fileset from the platform.
Referencing model or adapter entities must be relinked or deleted first. Returns metadata about the deleted fileset. For local storage backends, this also deletes the underlying files.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
nemo files filesets list
List Filesets endpoint with filtering and pagination.
Supports filtering by name, description, purpose, storage_type, created_at, and updated_at via query parameters. Returns paginated results with sorting options.
Usage:
Options:
--workspace--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort <CHOICE>: The field to sort by. To sort in decreasing order, use-in front of the field name. [possible values: -created_at, created_at, -updated_at, updated_at, -name, name]--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: created_at: {gte: str, lte: str} updated_at: {gte: str, lte: str}
Filter filesets by name, description, purpose, storage_type, created_at, and updated_at.
--filter.description--filter.name--filter.purpose--filter.storage-type
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo files filesets get
Get Fileset by Workspace and Name.
Returns the details of a specific fileset identified by its workspace and name.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo files filesets update
Update Fileset Metadata.
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--custom-fields: Custom fields for the fileset. (JSON string)--description: The description of the fileset.--metadata: Tagged metadata container - the key indicates the type.
Example: metadata = FilesetMetadata( dataset=DatasetMetadataContent( schema={"columns": ["id", "name"]}, ) ) (JSON string)
--project: The name of the project associated with this fileset.--purpose <CHOICE>: The purpose of the fileset. [possible values: dataset, environment, generic, model]
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo files otlp
Otlp operations
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
logs: Manage logs
nemo files otlp logs
Manage logs
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Upload OTLP logs to a specified fileset in JSON or…query: Query logs from parquet files in a fileset.
nemo files otlp logs create
Upload OTLP logs to a specified fileset in JSON or Protobuf format.
Supports both application/json and application/x-protobuf content types.
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--artifact-base-path: Folder inside the fileset to nest logs under
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo files otlp logs query
Query logs from parquet files in a fileset.
This is an internal endpoint that runs DuckDB queries with direct storage access.
Usage:
Arguments:
<NAME>
Options:
--workspace--artifact-base-path: Folder inside the fileset the logs were nested under (must match the value used on write)--filters: Key-value filters to apply to the query--limit <INTEGER>: Maximum number of results to return--page-cursor: Cursor for pagination--tail <INTEGER>: Number of newest log lines to return
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference
Inference operations.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
get-url: Print the OpenAI-compatible base URL for the inference…deployment-configs: Manage deployment_configsdeployments: Manage deploymentsgateway: Gateway operationsmodels: Manage modelsprompts: Manage promptsproviders: Manage providersvirtual-models: Manage virtual_models
nemo inference get-url
Print the OpenAI-compatible base URL for the inference gateway.
Examples:
Usage:
Options:
--workspace: Workspace to scope the URL to. Defaults to the CLI context workspace.--provider: Print the provider proxy route for this provider name.--virtual-model: Print the model entity proxy route for this virtual model name.
Help:
--help, -h: Show this message and exit.
nemo inference deployment-configs
Manage deployment_configs
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create a new ModelDeploymentConfig (version 1).delete: Delete all versions of a ModelDeploymentConfig.list: List ModelDeploymentConfigs for a specific workspace.get: Get the latest version of a ModelDeploymentConfig.update: Update a ModelDeploymentConfig (creates a new immutable…versions: Manage versions
nemo inference deployment-configs create
Create a new ModelDeploymentConfig (version 1).
Required fields: engine, executor_config, model_spec, name
Examples:
Usage:
Arguments:
<NAME>: Name of the deployment configuration. Name must start with a lowercase letter, be 2-63 characters, and use lowercase letters, digits, hyphens, and dots (no consecutive hyphens, cannot end with a hyphen).
Options:
--workspace--engine <CHOICE>: Inference engine selecting the compiler path for a deployment. The engine determines what command, image, and env a deployment compiles to. The fields a compiler consumes are not engine-specific; engines take the same inputs (model_spec + executor_config) and differ in what they do with them. [possible values: nim, vllm, generic]--executor-config: Compute + container settings shared by the docker and k8s executors. Both the docker and k8s executors run containers and share this shape. A future non-container executor (e.g. subprocess) would warrant turningexecutor_configinto a discriminated union. (JSON string)--model-spec: What model to serve and how — independent of the executor it runs on. Executor-invariant facts about the model. The compiler resolves the weight source per engine; serving fields override the model entity spec when set. (JSON string)--description: Optional description of the deployment configuration--model-entity-id: Optional reference to the base model entity ID for this deployment--project: The URN of the project associated with this deployment configuration--exist-ok: Do not raise an error if the resource already exists. Returns the existing resource.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference deployment-configs delete
Delete all versions of a ModelDeploymentConfig.
This operation will fail with 409 Conflict if any ModelDeployments currently reference this config and are not in DELETED status. Delete or wait for dependent deployments to reach DELETED status before deleting the config.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
nemo inference deployment-configs list
List ModelDeploymentConfigs for a specific workspace.
Returns only the latest version of each config.
Usage:
Options:
--workspace--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort: The field to sort by. To sort in decreasing order, use-in front of the field name.--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: created_at: {gte: str, lte: str} updated_at: {gte: str, lte: str}
Filter deployment configs by workspace, project, model_entity_id, name, description, created_at, and updated_at.
--filter.description--filter.model-entity-id--filter.name--filter.project--filter.workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo inference deployment-configs get
Get the latest version of a ModelDeploymentConfig.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference deployment-configs update
Update a ModelDeploymentConfig (creates a new immutable version).
Required fields: engine, executor_config, model_spec
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--engine <CHOICE>: Inference engine selecting the compiler path for a deployment. The engine determines what command, image, and env a deployment compiles to. The fields a compiler consumes are not engine-specific; engines take the same inputs (model_spec + executor_config) and differ in what they do with them. [possible values: nim, vllm, generic]--executor-config: Compute + container settings shared by the docker and k8s executors. Both the docker and k8s executors run containers and share this shape. A future non-container executor (e.g. subprocess) would warrant turningexecutor_configinto a discriminated union. (JSON string)--model-spec: What model to serve and how — independent of the executor it runs on. Executor-invariant facts about the model. The compiler resolves the weight source per engine; serving fields override the model entity spec when set. (JSON string)--description: Optional description of the deployment configuration--model-entity-id: Optional reference to the base model entity ID for this deployment
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference deployment-configs versions
Manage versions
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
delete: Delete a specific version of a ModelDeploymentConfig.list: List all versions of a ModelDeploymentConfig.get: Get a specific version of a ModelDeploymentConfig.
nemo inference deployment-configs versions delete
Delete a specific version of a ModelDeploymentConfig.
This operation will fail with 409 Conflict if any ModelDeployments currently reference this specific version and are not in DELETED status. Delete or wait for dependent deployments to reach DELETED status before deleting the config version.
Usage:
Arguments:
<NAME>
Options:
--workspace--config
Help:
--help, -h: Show this message and exit.
nemo inference deployment-configs versions list
List all versions of a ModelDeploymentConfig.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo inference deployment-configs versions get
Get a specific version of a ModelDeploymentConfig.
Usage:
Arguments:
<NAME>
Options:
--workspace--config
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference deployments
Manage deployments
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create a new ModelDeployment (version 1).delete: Delete all versions of a ModelDeployment.list: List ModelDeployments for a specific workspace.list-models: Get Latest ModelDeployment’s Model Entities from Entity…get: Get the latest version of a ModelDeployment.update: Update a ModelDeployment (creates a new immutable version).update-status: Update the status of a ModelDeployment (mutable operation).versions: Manage versions
nemo inference deployments create
Create a new ModelDeployment (version 1).
Required fields: config, name
Examples:
Usage:
Arguments:
<NAME>: Name of the deployment. Name must start with a lowercase letter, be 2-63 characters, and use lowercase letters, digits, hyphens, and dots (no consecutive hyphens, cannot end with a hyphen).
Options:
--workspace--config: Reference to the ModelDeploymentConfig name--config-version <INTEGER>: Reference to a specific ModelDeploymentConfig version. If not specified, uses latest.--project: The URN of the project associated with this deployment--exist-ok: Do not raise an error if the resource already exists. Returns the existing resource.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Lifecycle Options:
--wait: Wait for the created deployment to be up and running--watch: Watch the created deployment until it is stable while streaming status updates--timeout <INTEGER RANGE>: Maximum time to wait or watch in seconds [default: 1200]--poll-interval <INTEGER RANGE>: Seconds between status checks [default: 3]
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference deployments delete
Delete all versions of a ModelDeployment.
If the deployment is in any state other than DELETED, this will set its status to DELETING. The models controller will then:
- Delete the infrastructure (e.g., K8s NimService)
- Update the status to DELETED
If the deployment is already in DELETED status, calling delete again will permanently remove it from the database.
Returns:
- 202 Accepted: Deployment marked for deletion (status set to DELETING)
- 204 No Content: Deployment permanently removed from database (was already DELETED)
- 404 Not Found: Deployment doesn’t exist
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
nemo inference deployments list
List ModelDeployments for a specific workspace.
By default, returns only the latest version of each deployment.
Usage:
Options:
--workspace--all-versions: If true, return all versions of each deployment. If false (default), return only the latest version.--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort: The field to sort by. To sort in decreasing order, use-in front of the field name.--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: created_at: {gte: str, lte: str} updated_at: {gte: str, lte: str}
Filter deployments by workspace, project, status, config, model_provider_id, name, status_message, created_at, and updated_at.
--filter.config--filter.model-provider-id--filter.name--filter.project--filter.status--filter.status-message--filter.workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo inference deployments list-models
Get Latest ModelDeployment’s Model Entities from Entity Store.
This provides the API contract that NIMs expect from Entity Store today, for pulling LoRAs, but enables us to enforce AuthZ boundaries.
TODO: Implement model entity retrieval based on deployment config.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo inference deployments get
Get the latest version of a ModelDeployment.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference deployments update
Update a ModelDeployment (creates a new immutable version).
Required fields: config
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--config: Reference to the ModelDeploymentConfig name--config-version <INTEGER>: Reference to a specific ModelDeploymentConfig version. If not specified, uses latest.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference deployments update-status
Update the status of a ModelDeployment (mutable operation).
If version is not specified, updates the latest version.
Required fields: status
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--status <CHOICE>: Status enum for ModelDeployment objects. [possible values: UNKNOWN, CREATED, PENDING, READY, ERROR, DELETING, DELETED, LOST]--version--model-provider-id: Optional reference to the auto-created ModelProvider workspace/name (format: workspace/name)--status-message: Detailed status message
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference deployments versions
Manage versions
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
delete: Delete a specific version of a ModelDeployment.list: List all versions of a ModelDeployment.get: Get a specific version of a ModelDeployment.
nemo inference deployments versions delete
Delete a specific version of a ModelDeployment.
If the deployment is in any state other than DELETED, this will set its status to DELETING. The models controller will then:
- Delete the infrastructure (e.g., K8s NimService)
- Update the status to DELETED
If the deployment is already in DELETED status, calling delete again will permanently remove it from the database.
Returns:
- 202 Accepted: Deployment version marked for deletion (status set to DELETING)
- 204 No Content: Deployment version permanently removed from database (was already DELETED)
- 404 Not Found: Deployment version doesn’t exist
Usage:
Arguments:
<NAME>
Options:
--workspace--deployment
Help:
--help, -h: Show this message and exit.
nemo inference deployments versions list
List all versions of a ModelDeployment.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo inference deployments versions get
Get a specific version of a ModelDeployment.
Usage:
Arguments:
<NAME>
Options:
--workspace--deployment
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference gateway
Gateway operations
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
model: Manage modelopenai: Openai operationsprovider: Manage provider
nemo inference gateway model
Manage model
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
delete: Proxy requests to model entity inference endpoints.get: Proxy requests to model entity inference endpoints.patch: Proxy requests to model entity inference endpoints.post: Proxy requests to model entity inference endpoints.put: Proxy requests to model entity inference endpoints.
nemo inference gateway model delete
Proxy requests to model entity inference endpoints.
All inference requests must resolve to a VirtualModel. The platform’s provider
reconciler auto-creates an implicit autoprovisioned VirtualModel for every
served model entity (named after the entity, with default_model_entity set to
the entity ref) so this is the typical case; operators can also create custom
VirtualModels for routing, plugin chains, LoRA escape-hatches, etc. Requests for
which no VirtualModel can be found return 404.
Usage:
Arguments:
<TRAILING_URI>
Options:
--workspace--name
Help:
--help, -h: Show this message and exit.
nemo inference gateway model get
Proxy requests to model entity inference endpoints.
All inference requests must resolve to a VirtualModel. The platform’s provider
reconciler auto-creates an implicit autoprovisioned VirtualModel for every
served model entity (named after the entity, with default_model_entity set to
the entity ref) so this is the typical case; operators can also create custom
VirtualModels for routing, plugin chains, LoRA escape-hatches, etc. Requests for
which no VirtualModel can be found return 404.
Usage:
Arguments:
<TRAILING_URI>
Options:
--workspace--name
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference gateway model patch
Proxy requests to model entity inference endpoints.
All inference requests must resolve to a VirtualModel. The platform’s provider
reconciler auto-creates an implicit autoprovisioned VirtualModel for every
served model entity (named after the entity, with default_model_entity set to
the entity ref) so this is the typical case; operators can also create custom
VirtualModels for routing, plugin chains, LoRA escape-hatches, etc. Requests for
which no VirtualModel can be found return 404.
Required fields: name
Examples:
Usage:
Arguments:
<TRAILING_URI>
Options:
--workspace--name: (required)--body: JSON string
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference gateway model post
Proxy requests to model entity inference endpoints.
All inference requests must resolve to a VirtualModel. The platform’s provider
reconciler auto-creates an implicit autoprovisioned VirtualModel for every
served model entity (named after the entity, with default_model_entity set to
the entity ref) so this is the typical case; operators can also create custom
VirtualModels for routing, plugin chains, LoRA escape-hatches, etc. Requests for
which no VirtualModel can be found return 404.
Required fields: name
Examples:
Usage:
Arguments:
<TRAILING_URI><NAME>: (required)
Options:
--workspace--body: JSON string
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference gateway model put
Proxy requests to model entity inference endpoints.
All inference requests must resolve to a VirtualModel. The platform’s provider
reconciler auto-creates an implicit autoprovisioned VirtualModel for every
served model entity (named after the entity, with default_model_entity set to
the entity ref) so this is the typical case; operators can also create custom
VirtualModels for routing, plugin chains, LoRA escape-hatches, etc. Requests for
which no VirtualModel can be found return 404.
Required fields: name
Examples:
Usage:
Arguments:
<TRAILING_URI><NAME>: (required)
Options:
--workspace--body: JSON string
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference gateway openai
Openai operations
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
v1: V1 operations
nemo inference gateway openai v1
V1 operations
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
models: Manage models
nemo inference gateway openai v1 models
Manage models
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
get: Retrieve information about a specific OpenAI-compatible…list: This endpoint lists the routable VirtualModels in the…
nemo inference gateway openai v1 models get
Retrieve information about a specific OpenAI-compatible model.
Workspace is always taken from the URL path; name may be the VirtualModel name or workspace/name (workspace prefix is ignored). Resolves against routable VirtualModels, including custom ones, so this route agrees with the list route and the inference proxy.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference gateway openai v1 models list
This endpoint lists the routable VirtualModels in the requested workspace and returns them in OpenAI’s list models format. Each model ID is the VirtualModel identifier in format workspace/name. This includes both autoprovisioned VirtualModels (one per served model entity) and custom VirtualModels, keeping the catalog in agreement with the inference proxy, which also resolves VirtualModels scoped to the request workspace.
Usage:
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo inference gateway provider
Manage provider
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
delete: Proxy requests to provider inference endpoints.get: Proxy requests to provider inference endpoints.patch: Proxy requests to provider inference endpoints.post: Proxy requests to provider inference endpoints.put: Proxy requests to provider inference endpoints.ready: Check if a model provider is registered in the gateway’s…
nemo inference gateway provider delete
Proxy requests to provider inference endpoints.
Usage:
Arguments:
<TRAILING_URI>
Options:
--workspace--name
Help:
--help, -h: Show this message and exit.
nemo inference gateway provider get
Proxy requests to provider inference endpoints.
Usage:
Arguments:
<TRAILING_URI>
Options:
--workspace--name
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference gateway provider patch
Proxy requests to provider inference endpoints.
Required fields: name
Examples:
Usage:
Arguments:
<TRAILING_URI>
Options:
--workspace--name: (required)--body: JSON string
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference gateway provider post
Proxy requests to provider inference endpoints.
Required fields: name
Examples:
Usage:
Arguments:
<TRAILING_URI><NAME>: (required)
Options:
--workspace--body: JSON string
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference gateway provider put
Proxy requests to provider inference endpoints.
Required fields: name
Examples:
Usage:
Arguments:
<TRAILING_URI><NAME>: (required)
Options:
--workspace--body: JSON string
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference gateway provider ready
Check if a model provider is registered in the gateway’s cache.
This is a lightweight endpoint that only checks the gateway’s internal state, without making any requests to the actual provider backend. Use this to verify the gateway is ready to route requests to a provider after deployment.
Returns: 200 OK with provider info if the provider is registered 404 Not Found if the provider is not yet in the gateway’s cache
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference models
Manage models
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
get: Retrieve information about a specific OpenAI-compatible…list: This endpoint lists the routable VirtualModels in the…
nemo inference models get
Retrieve information about a specific OpenAI-compatible model.
Workspace is always taken from the URL path; name may be the VirtualModel name or workspace/name (workspace prefix is ignored). Resolves against routable VirtualModels, including custom ones, so this route agrees with the list route and the inference proxy.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference models list
This endpoint lists the routable VirtualModels in the requested workspace and returns them in OpenAI’s list models format. Each model ID is the VirtualModel identifier in format workspace/name. This includes both autoprovisioned VirtualModels (one per served model entity) and custom VirtualModels, keeping the catalog in agreement with the inference proxy, which also resolves VirtualModels scoped to the request workspace.
Usage:
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo inference prompts
Manage prompts
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create a new prompt.delete: Delete a prompt by workspace and name.list: List prompts for a specific workspace.get: Get a prompt by workspace and name.update: Update an existing prompt (full replacement of mutable…
nemo inference prompts create
Create a new prompt.
Required fields: name
Examples:
Usage:
Arguments:
<NAME>: Name of the prompt.
Options:
--workspace--description--inference-params: Parameters for model inference. Extra fields can be supplied for additional options applied to the inference request directly. Fields not supported by the model may cause inference errors during evaluation. (JSON string)--input-variables: Can be repeated for multiple values--messages: JSON string--project: The URN of the project associated with this prompt.--response-format: JSON string--tags: Can be repeated for multiple values--tool-choice: JSON string--tools: JSON string--exist-ok: Do not raise an error if the resource already exists. Returns the existing resource.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference prompts delete
Delete a prompt by workspace and name.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
nemo inference prompts list
List prompts for a specific workspace.
Usage:
Options:
--workspace--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort <CHOICE>: The field to sort by. To sort in decreasing order, use-in front of the field name. [possible values: -created_at, created_at, -updated_at, updated_at, -name, name]--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: created_at: {gte: str, lte: str} updated_at: {gte: str, lte: str}
Filter prompts by workspace, project, name, description, created_at, and updated_at.
--filter.description--filter.name--filter.project--filter.workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo inference prompts get
Get a prompt by workspace and name.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference prompts update
Update an existing prompt (full replacement of mutable fields).
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--description--inference-params: Parameters for model inference. Extra fields can be supplied for additional options applied to the inference request directly. Fields not supported by the model may cause inference errors during evaluation. (JSON string)--input-variables: Can be repeated for multiple values--messages: JSON string--project: The URN of the project associated with this prompt.--response-format: JSON string--tags: Can be repeated for multiple values--tool-choice: JSON string--tools: JSON string
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference providers
Manage providers
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create a new model provider.delete: Delete a model provider by workspace and name.list: List model providers for a specific workspace.get: Get a model provider by workspace and name.update: Create or update a model provider.update-status: Update status-related fields of a model provider.
nemo inference providers create
Create a new model provider.
Required fields: host_url, name
Examples:
Usage:
Arguments:
<NAME>: Name of the model provider. Name must start with a lowercase letter, be 2-63 characters, and use lowercase letters, digits, hyphens, and dots (no consecutive hyphens, cannot end with a hyphen).
Options:
--workspace--host-url: The network endpoint URL for the model provider--api-key-secret-name: Reference to an API key secret stored in the Secrets service. Create the secret first via secrets API, then pass the secret name here.--auth-header-format: Jinja2 template string controlling how the API key secret is sent to the upstream. Must contain exactly one variable namedauth_secret, which is substituted with the resolved secret value at request time. Example:'X-Api-Key: {{ auth_secret }}'. If not set, defaults to'Authorization: Bearer {{ auth_secret }}'.--default-extra-body: Default body parameters for inference requests. Can be overridden by user requests. (JSON string)--default-extra-headers: Default headers for inference requests. Can be overridden by user requests. (JSON string)--description: Optional description of the model provider--enabled-models: Optional list of specific models to enable from this provider (can be repeated)--model-deployment-id: Optional reference to the ModelDeployment ID if this provider is being auto-created for a deployment--project: The URN of the project associated with this model provider--required-extra-body: Required body parameters for inference requests. Cannot be overridden by user requests. (JSON string)--required-extra-headers: Required headers for inference requests. Cannot be overridden by user requests. (JSON string)--status <CHOICE>: Status enum for ModelProvider objects. [possible values: UNKNOWN, CREATED, PENDING, READY, ERROR, DELETING, DELETED, LOST]--status-message: Status message--exist-ok: Do not raise an error if the resource already exists. Returns the existing resource.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference providers delete
Delete a model provider by workspace and name.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
nemo inference providers list
List model providers for a specific workspace.
Usage:
Options:
--workspace--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort <CHOICE>: The field to sort by. To sort in decreasing order, use-in front of the field name. [possible values: name, -name, created_at, -created_at, updated_at, -updated_at, status, -status]--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: created_at: {gte: str, lte: str} updated_at: {gte: str, lte: str}
Filter model providers by workspace, project, status, model_deployment_id, name, description, host_url, created_at, and updated_at.
--filter.description--filter.host-url--filter.model-deployment-id--filter.name--filter.project--filter.status--filter.workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo inference providers get
Get a model provider by workspace and name.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference providers update
Create or update a model provider.
Required fields: host_url
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--host-url: The network endpoint URL for the model provider--api-key-secret-name: Reference to an API key secret stored in the Secrets service. Create the secret first via secrets API, then pass the secret name here.--auth-header-format: Jinja2 template string controlling how the API key secret is sent to the upstream. Must contain exactly one variable namedauth_secret, which is substituted with the resolved secret value at request time. Example:'X-Api-Key: {{ auth_secret }}'. If not set, defaults to'Authorization: Bearer {{ auth_secret }}'.--default-extra-body: Default body parameters for inference requests. Can be overridden by user requests. (JSON string)--default-extra-headers: Default headers for inference requests. Can be overridden by user requests. (JSON string)--description: Optional description of the model provider--enabled-models: Optional list of specific models to enable from this provider (can be repeated)--model-deployment-id: Optional reference to the ModelDeployment ID if this provider is associated with a deployment--project: The URN of the project associated with this model provider--required-extra-body: Required body parameters for inference requests. Cannot be overridden by user requests. (JSON string)--required-extra-headers: Required headers for inference requests. Cannot be overridden by user requests. (JSON string)--status <CHOICE>: Status enum for ModelProvider objects. [possible values: UNKNOWN, CREATED, PENDING, READY, ERROR, DELETING, DELETED, LOST]--status-message: Status message
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference providers update-status
Update status-related fields of a model provider.
This endpoint supports partial updates for fields managed by Models Controller:
- model_deployment_id
- served_models
- status
- status_message
If status is provided without status_message, status_message will be set to empty string.
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--model-deployment-id: Reference to the ModelDeployment ID if this provider is associated with a deployment--served-models: List of models served by this provider with routing information for IGW (JSON string)--status <CHOICE>: Status enum for ModelProvider objects. [possible values: UNKNOWN, CREATED, PENDING, READY, ERROR, DELETING, DELETED, LOST]--status-message: Status message. If status is provided without status_message, defaults to empty string.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference virtual-models
Manage virtual_models
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create a new VirtualModel in the given workspace.delete: Permanently delete a VirtualModel.list: List VirtualModels for the given workspace.patch: Partially update a VirtualModel.get: Get a VirtualModel by workspace and name.
nemo inference virtual-models create
Create a new VirtualModel in the given workspace.
A VirtualModel defines an ordered middleware pipeline that IGW executes when an
inference request arrives with model: "workspace/name" matching this entity.
Required fields: name
Examples:
Usage:
Arguments:
<NAME>: Name of the virtual model within the workspace. Must be unique per workspace.
Options:
--workspace--autoprovisioned: Marks this VirtualModel as controller-managed. The Models controller will delete it once no ModelProvider serves the matching entity. Setting this manually opts the VirtualModel into that cleanup behavior.--default-model-entity: Model entity to route to, in “workspace/name” format. Written into request[“model”] before the request middleware pipeline runs. If omitted, a request middleware plugin must handle backend routing itself. Set to null to clear an existing value.--models: Model entity references used by this VirtualModel. A per-entry backend_format overrides the referenced ModelEntity backend_format when IGW resolves the backend format for a request. (JSON string)--override-proxy: Plugin-provided proxy implementation for IGW to use instead of its default aiohttp proxy. Format: “plugin-name.proxy-name”. Leave unset to use the default IGW proxy. Set to null to clear an existing value.--post-response-middleware: Ordered list of middleware plugins invoked after the response has been returned to the caller. Intended for fire-and-forget work (logging, analytics) that must not block or modify the response. (JSON string)--request-middleware: Ordered list of middleware plugins applied before proxying to the backend. Each entry is a MiddlewareCall with a “name” (plugin identifier) and optional “config_type” and “config_id” fields that reference a stored plugin configuration. (JSON string)--response-middleware: Ordered list of middleware plugins applied after the backend response is received, before returning it to the caller. (JSON string)--exist-ok: Do not raise an error if the resource already exists. Returns the existing resource.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference virtual-models delete
Permanently delete a VirtualModel.
This does not affect any in-flight requests already being routed through this VirtualModel. IGW’s model cache is refreshed on its next polling cycle.
Usage:
Arguments:
<NAME>
Options:
--workspace--expected-db-version <INTEGER>: Optional database version for optimistic locking. Delete only succeeds if the VirtualModel still has this version.
Help:
--help, -h: Show this message and exit.
nemo inference virtual-models list
List VirtualModels for the given workspace.
Use workspace=- to list across all workspaces accessible to the caller.
Usage:
Options:
--workspace--exclude-autoprovisioned: When true, controller-managed (autoprovisioned) passthrough VirtualModels are excluded from the results.--page <INTEGER>: Page number (1-indexed).--page-size <INTEGER>: Number of results per page.--sort: Sort field. Prefix with-for descending order.--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: created_at: {gte: str, lte: str} updated_at: {gte: str, lte: str}
Filter virtual models by workspace, project, name, default_model_entity, guardrail_config, created_at, and updated_at.
--filter.default-model-entity--filter.guardrail-config--filter.name--filter.project--filter.workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo inference virtual-models patch
Partially update a VirtualModel.
Only fields present in the request body are modified. Fields absent from the request body retain their current values.
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--autoprovisioned: Marks this VirtualModel as controller-managed. The Models controller will delete it once no ModelProvider serves the matching entity. Setting this manually opts the VirtualModel into that cleanup behavior.--default-model-entity: Model entity to route to, in “workspace/name” format. Written into request[“model”] before the request middleware pipeline runs. If omitted, a request middleware plugin must handle backend routing itself. Set to null to clear an existing value.--models: Model entity references used by this VirtualModel. A per-entry backend_format overrides the referenced ModelEntity backend_format when IGW resolves the backend format for a request. (JSON string)--override-proxy: Plugin-provided proxy implementation for IGW to use instead of its default aiohttp proxy. Format: “plugin-name.proxy-name”. Leave unset to use the default IGW proxy. Set to null to clear an existing value.--post-response-middleware: Ordered list of middleware plugins invoked after the response has been returned to the caller. Intended for fire-and-forget work (logging, analytics) that must not block or modify the response. (JSON string)--request-middleware: Ordered list of middleware plugins applied before proxying to the backend. Each entry is a MiddlewareCall with a “name” (plugin identifier) and optional “config_type” and “config_id” fields that reference a stored plugin configuration. (JSON string)--response-middleware: Ordered list of middleware plugins applied after the backend response is received, before returning it to the caller. (JSON string)
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo inference virtual-models get
Get a VirtualModel by workspace and name.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo jobs
Manage jobs.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
cancel: Cancel a platform job.create: Create a new platform job.delete: Delete a platform job.get-logs: Get paginated logs for a platform job.get-status: Get the status of a platform job.list: List platform jobs with filtering and pagination.list-execution-profiles: Get all currently configured execution profiles.pause: Pause a platform job.resume: Resume a paused platform job.get: Get a platform job by name.update-status-details: Update the status details of a platform job.watch: Watch a platform job until it reaches a terminal status.tail: Print the newest lines from a platform job log.results: Manage resultssteps: Manage stepstasks: Manage tasks
nemo jobs cancel
Cancel a platform job.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo jobs create
Create a new platform job.
Required fields: platform_spec, source, spec
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--platform-spec: Specification for a platform job, containing steps and secrets. (JSON string)--source: (required)--spec: JSON string--custom-fields: JSON string--description--output-location--ownership: JSON string--project
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Lifecycle Options:
--wait: Wait for the created job to reach a terminal state without streaming logs--watch: Watch the created job to a terminal state--timeout <INTEGER RANGE>: Maximum time to wait or watch in seconds--poll-interval <INTEGER RANGE>: Seconds between status checks [default: 3]
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo jobs delete
Delete a platform job.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
nemo jobs get-logs
Get paginated logs for a platform job.
Usage:
Arguments:
<NAME>
Options:
--workspace--attempt-id <INTEGER>: Filter logs by job attempt ID--limit <INTEGER>: Maximum number of logs to return--page-cursor: Page cursor--step-id: Filter logs by step name--tail <INTEGER>: Number of newest log lines to return--task-id: Filter logs by task ID--all-pages: Fetch all pages
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo jobs get-status
Get the status of a platform job.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo jobs list
List platform jobs with filtering and pagination.
Usage:
Options:
--workspace--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort <CHOICE>: The field to sort by. To sort in decreasing order, use-in front of the field name. [possible values: created_at, -created_at, updated_at, -updated_at, source, -source]--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: created_at: {gte: str, lte: str} updated_at: {gte: str, lte: str}
Filter jobs by workspace, project, name, status, source, created_at, and updated_at.
--filter.name--filter.project--filter.source--filter.status--filter.workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo jobs list-execution-profiles
Get all currently configured execution profiles.
Returns the capability-filtered merge from jobs config. In local standalone the controller may prune the shared list further after registry boot; in split topologies the API advertises its own merge result (not controller process memory).
Usage:
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo jobs pause
Pause a platform job.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo jobs resume
Resume a paused platform job.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo jobs get
Get a platform job by name.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo jobs update-status-details
Update the status details of a platform job.
Required fields: body
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--body: JSON string
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo jobs watch
Watch a platform job until it reaches a terminal status.
Usage:
Arguments:
<NAME>: Name of the platform job to watch
Options:
--workspace: Workspace containing the job--attempt-id <INTEGER>: Filter logs to an attempt ID--step-id: Filter logs to a step ID--task-id: Filter logs to a task ID--limit <INTEGER RANGE>: Maximum logs to fetch per page--timeout <INTEGER RANGE>: Maximum watch time in seconds--poll-interval <INTEGER RANGE>: Seconds between status checks [default: 3]--history, --no-history: Include logs already present before watching
Help:
--help, -h: Show this message and exit.
nemo jobs tail
Print the newest lines from a platform job log.
Usage:
Arguments:
<NAME>: Name of the platform job whose logs to tail
Options:
-n, --lines <INTEGER RANGE>: Number of lines to show [default: 100]--workspace: Workspace containing the job--attempt-id <INTEGER>: Filter logs to an attempt ID--step-id: Filter logs to a step ID--task-id: Filter logs to a task ID
Help:
--help, -h: Show this message and exit.
nemo jobs results
Manage results
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create a new result for a job.download: Download a job result file.list: List results for a job.get: Get a specific job result.
nemo jobs results create
Create a new result for a job.
Required fields: job, artifact_storage_type, artifact_url
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--job: (required)--artifact-storage-type <CHOICE>: (required) [possible values: fileset]--artifact-url: (required)
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo jobs results download
Download a job result file.
Usage:
Arguments:
<NAME>
Options:
--workspace--job--output-file, -o <PATH>: Output file path
Help:
--help, -h: Show this message and exit.
nemo jobs results list
List results for a job.
Usage:
Arguments:
<NAME>
Options:
--workspace--sort <CHOICE>: The field to sort by. [possible values: created_at, -created_at, updated_at, -updated_at]
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo jobs results get
Get a specific job result.
Usage:
Arguments:
<NAME>
Options:
--workspace--job
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo jobs steps
Manage steps
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
list: List job steps with pagination and filtering.get: Get a specific job step.update-status: Update a job step status.
nemo jobs steps list
List job steps with pagination and filtering.
Usage:
Arguments:
<NAME>
Options:
--workspace--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort <CHOICE>: The field to sort by. To sort in decreasing order, use-in front of the field name. [possible values: created_at, -created_at, updated_at, -updated_at]--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: status: [‘created’ | ‘pending’ | ‘active’ | ‘cancelled’ | ‘cancelling’ | ‘error’ | ‘completed’ | ‘paused’ | ‘pausing’ | ‘resuming’]
Filter steps by job, status, and source.
--filter.job--filter.source
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo jobs steps get
Get a specific job step.
Usage:
Arguments:
<NAME>
Options:
--workspace--job
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo jobs steps update-status
Update a job step status.
Required fields: job, status
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--job: (required)--status <CHOICE>: Enumeration of possible job statuses. This enum represents the various states a job can be in during its lifecycle, from creation to a terminal state. [possible values: created, pending, active, cancelled, cancelling, error, completed, paused, pausing, resuming]--error-details: Optional error details related to the status update. (JSON string)--status-details: Optional status details related to the status update. (JSON string)
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo jobs tasks
Manage tasks
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create-or-update: Update a job step task.list: List tasks for a job step.get: Get a specific job step task.
nemo jobs tasks create-or-update
Update a job step task.
Required fields: job, step
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--job: (required)--step: (required)--error-details: JSON string--error-stack--status <CHOICE>: Enumeration of possible job statuses. This enum represents the various states a job can be in during its lifecycle, from creation to a terminal state. [possible values: created, pending, active, cancelled, cancelling, error, completed, paused, pausing, resuming]--status-details: JSON string
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo jobs tasks list
List tasks for a job step.
Usage:
Arguments:
<NAME>
Options:
--workspace--job
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo jobs tasks get
Get a specific job step task.
Usage:
Arguments:
<NAME>
Options:
--workspace--job--step
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo models
Manage models.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create a new model entity.delete: Delete Model entity.list: List Models endpoint with filtering, pagination, and…get: Get Model by Workspace and Name.update: Update Model metadata.adapters: Manage adapters
nemo models create
Create a new model entity.
This endpoint creates a new Model Entity in the Models service database. The Model Entity will be registered for use within the platform.
Required fields: name
Examples:
Usage:
Arguments:
<NAME>: Name of the model entity. Name must start with a lowercase letter, be 2-63 characters, and use lowercase letters, digits, hyphens, and dots (no consecutive hyphens, cannot end with a hyphen).
Options:
--workspace--api-endpoint: Data about an inference endpoint. (JSON string)--backend-format <CHOICE>: Inference backend API wire formats understood by IGW and middleware plugins. [possible values: OPENAI_CHAT, ANTHROPIC_MESSAGES]--base-model: Link to another model which is used as a base for the current model--custom-fields: Custom fields for additional metadata (JSON string)--description: Optional description of the model--fileset: A set of checkpoint files, configs, and other auxiliary info associated with this model - expected format {workspace}/{fileset_name}--finetuning-type <CHOICE>: Finetuning types. [possible values: lora_merged, all_weights, last_layer, top_layers, gradual_unfreezing, bias_only, attention_only, lora, qlora, adalora, dora, lora_plus, prompt_tuning, prefix_tuning, p_tuning, p_tuning_v2, soft_prompt, ppo, dpo, cdpo, ipo, orpo, kto, rrhf, grpo]--model-providers: List of ModelProvider workspace/name resource names that provide inference for this Model Entity (can be repeated)--ownership: Ownership information for the model (JSON string)--project: The URN of the project associated with this model entity--prompt: Configuration for prompt engineering. (JSON string)--spec: Detailed specification for a model. (JSON string)--trust-remote-code: Whether to trust remote code for the checkpoint. Some models without support in certain libraries such as Transformers require additional custom Python code to execute. Due to security ramifications of running arbitrary code, this can only be set to true on one of the following conditions: (1) the model’s fileset’s source is pre-approved in the platform config, or (2) the user creating this model is an administrator.--exist-ok: Do not raise an error if the resource already exists. Returns the existing resource.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo models delete
Delete Model entity.
Permanently deletes a model entity from the platform.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
nemo models list
List Models endpoint with filtering, pagination, and sorting.
Supports filter parameters for various criteria (including peft, custom fields), pagination (page, page_size), sorting, and workspace filtering via query parameter.
Usage:
Options:
--workspace--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort <CHOICE>: The field to sort by. To sort in decreasing order, use-in front of the field name. [possible values: -created_at, created_at, -updated_at, updated_at, -name, name]--verbose: Whether to include full spec details--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: created_at: {gte: str, lte: str} updated_at: {gte: str, lte: str}
Filter models by name, project, workspace, base_model, adapters, finetuning_type, prompt, lora_enabled, description, created_at, and updated_at.
--filter.adapters--filter.base-model--filter.description--filter.fileset--filter.finetuning-type--filter.lora-enabled--filter.name--filter.project--filter.prompt--filter.workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo models get
Get Model by Workspace and Name.
Returns the details of a specific model entity identified by its workspace and name.
Usage:
Arguments:
<NAME>
Options:
--workspace--verbose: Whether to include full spec details
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo models update
Update Model metadata.
Updates the metadata of an existing model entity.
If the request body has an empty field, the old value is kept.
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--verbose: Whether to include full spec details--api-endpoint: Data about an inference endpoint. (JSON string)--backend-format <CHOICE>: Inference backend API wire formats understood by IGW and middleware plugins. [possible values: OPENAI_CHAT, ANTHROPIC_MESSAGES]--base-model: Link to another model which is used as a base for the current model--custom-fields: Custom fields for additional metadata (JSON string)--description: Optional description of the model--fileset: A set of checkpoint files, configs, and other auxiliary info associated with this model - expected format {workspace}/{fileset_name}--finetuning-type <CHOICE>: Finetuning types. [possible values: lora_merged, all_weights, last_layer, top_layers, gradual_unfreezing, bias_only, attention_only, lora, qlora, adalora, dora, lora_plus, prompt_tuning, prefix_tuning, p_tuning, p_tuning_v2, soft_prompt, ppo, dpo, cdpo, ipo, orpo, kto, rrhf, grpo]--model-providers: List of ModelProvider workspace/name resource names that provide inference for this Model Entity (can be repeated)--ownership: Ownership information for the model (JSON string)--prompt: Configuration for prompt engineering. (JSON string)--spec: Detailed specification for a model. (JSON string)--trust-remote-code: Whether to trust remote code for the checkpoint. Some models without support in certain libraries such as Transformers require additional custom Python code to execute. Due to security ramifications of running arbitrary code, this can only be set to true on one of the following conditions: (1) the model’s fileset’s source is pre-approved in the platform config, or (2) the user creating this model is an administrator.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo models adapters
Manage adapters
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Adds an Adapter to the Modeldelete: Delete Adapter from Model entity.update: Update Adapter deployment or description.
nemo models adapters create
Adds an Adapter to the Model
Required fields: fileset, finetuning_type, name
Examples:
Usage:
Arguments:
<MODEL_NAME><NAME>: Name of the adapter. Name must be unique in the workspace. Name must start with a lowercase letter, be 2-63 characters, and use lowercase letters, digits, hyphens, and dots (no consecutive hyphens, cannot end with a hyphen).
Options:
--workspace--fileset: Location where adapter files are stored - expected format {workspace}/{fileset_name}--finetuning-type <CHOICE>: Finetuning types. [possible values: lora_merged, all_weights, last_layer, top_layers, gradual_unfreezing, bias_only, attention_only, lora, qlora, adalora, dora, lora_plus, prompt_tuning, prefix_tuning, p_tuning, p_tuning_v2, soft_prompt, ppo, dpo, cdpo, ipo, orpo, kto, rrhf, grpo]--description: Optional description of the adapter--enabled: Whether to make this adapter available for inference post training--lora-config: Lora configuration specifics (JSON string)
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo models adapters delete
Delete Adapter from Model entity.
Permanently deletes an adapter from a model entity, if it was deployed, it will be cleaned up automatically.
Usage:
Arguments:
<ADAPTER>
Options:
--workspace--model-name
Help:
--help, -h: Show this message and exit.
nemo models adapters update
Update Adapter deployment or description.
Required fields: model_name
Examples:
Usage:
Arguments:
<ADAPTER>
Options:
--workspace--model-name: (required)--description: Optional description of the adapter--enabled: Whether to make this adapter available for inference post training--fileset: Updated fileset for the adapter
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo secrets
Manage secrets.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
access: Access the value of a secret.create: Create a new secret.delete: Delete a secret.list: List available secretsget: Retrieve a secret by its name.update: Update a secret’s metadata and/or value.admin: Manage admin
nemo secrets access
Access the value of a secret.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo secrets create
Create a new secret.
Examples:
Usage:
Arguments:
<NAME>: The name of the secret to create
Options:
--workspace--from-file: Path to file containing the secret value. Use ’-’ to read from stdin.--value: Secret value directly. Use —from-file for large or sensitive input.--description: An optional description of the secret
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo secrets delete
Delete a secret.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
nemo secrets list
List available secrets
Usage:
Options:
--workspace--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--all-pages: Fetch all pages
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo secrets get
Retrieve a secret by its name.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo secrets update
Update a secret’s metadata and/or value.
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--from-file: Path to file containing the secret value. Use ’-’ to read from stdin.--value: Secret value directly. Use —from-file for large or sensitive input.--description: An optional description of the secret
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo secrets admin
Manage admin
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
rotate-encryption-keys: Rotate encryption keys for all platform secrets.
nemo secrets admin rotate-encryption-keys
Rotate encryption keys for all platform secrets.
Usage:
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo workspaces
Manage workspaces.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create a new workspace.delete: Delete a workspace.list: List all workspaces with pagination.get: Get a specific workspace by ID.update: Update a workspace’s description.members: Manage members
nemo workspaces create
Create a new workspace.
The creator is automatically granted Admin role on the workspace. By default,
this endpoint waits for the Admin role to propagate before returning. Use
wait_role_propagation=false to skip waiting (useful for bulk operations).
Examples:
Required fields: name
Examples:
Usage:
Arguments:
<NAME>: Workspace name (unique identifier). Name must start with a lowercase letter, be 2-63 characters, and use lowercase letters, digits, hyphens, and dots (no consecutive hyphens, cannot end with a hyphen).
Options:
--wait-role-propagation: If true, wait for Admin role to propagate before returning (default: true). Set to false for bulk operations.--description: Optional description of the workspace--exist-ok: Do not raise an error if the resource already exists. Returns the existing resource.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo workspaces delete
Delete a workspace.
This marks the workspace for deletion and returns immediately. The workspace will no longer be accessible via the API. An asynchronous cleanup controller will handle deletion of all entities and external resources.
Role bindings are immediately deleted to revoke access.
Examples:
Usage:
Arguments:
<NAME>
Help:
--help, -h: Show this message and exit.
nemo workspaces list
List all workspaces with pagination.
Workspaces marked for deletion (non-null deletion_stage) are omitted so the list matches GET/DELETE, which treat those workspaces as not found.
When authentication is enabled, only workspaces the principal has access to are returned. Service principals and platform admins have access to all workspaces.
Query Parameters:
- page, page_size: Pagination
- sort: Sort field
- filter: Advanced filters (JSON, text, or bracket notation)
Examples:
Usage:
Options:
--filter: Query filter expression. Supports text and JSON syntaxes:
- Text: name:“value” AND status>500 with operators : ~ > >= < <= IN NOT IN AND OR and negation prefix -
- Object (JSON):
{"name":{"$like":"value"}}with operators$eq,$like,$lt,$lte,$gt,$gte,$in,$nin,$and,$or,$not
--page <INTEGER>: Page number--page-size <INTEGER>: Items per page--sort <CHOICE>: Sort field [possible values: -created_at, created_at, -updated_at, updated_at, -name, name]--all-pages: Fetch all pages
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo workspaces get
Get a specific workspace by ID.
Examples:
Usage:
Arguments:
<NAME>
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo workspaces update
Update a workspace’s description.
Examples:
Examples:
Usage:
Arguments:
<NAME>
Options:
--description: Updated description
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo workspaces members
Manage members
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Add a new member to the workspace with specified roles.delete: Remove a member from the workspace by revoking all their…list: List all members of a workspace with their roles.update: Update the roles for a workspace member.
nemo workspaces members create
Add a new member to the workspace with specified roles.
This creates role bindings for the specified principal with the given roles. By
default, this endpoint waits for the roles to propagate before returning. Use
wait_role_propagation=false to skip waiting (useful for bulk operations).
Examples:
Required fields: principal
Examples:
Usage:
Options:
--workspace--principal: The principal identifier (email, user ID, or group ID)--wait-role-propagation: If true, wait for roles to propagate before returning (default: true). Set to false for bulk operations.--roles: List of roles to grant to the principal (can be repeated)
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo workspaces members delete
Remove a member from the workspace by revoking all their roles.
This revokes all active role bindings for the principal in the workspace. By
default, this endpoint waits for all roles to be revoked before returning. Use
wait_role_propagation=false to skip waiting (useful for bulk operations).
Examples:
Usage:
Arguments:
<PRINCIPAL_ID>
Options:
--workspace--wait-role-propagation: If true, wait for roles to propagate before returning (default: true). Set to false for bulk operations.
Help:
--help, -h: Show this message and exit.
nemo workspaces members list
List all members of a workspace with their roles.
Returns a list of all principals with active role bindings in the workspace.
Examples:
Usage:
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo workspaces members update
Update the roles for a workspace member.
This will revoke existing roles not in the new list and add new roles. By
default, this endpoint waits for the roles to propagate before returning. Use
wait_role_propagation=false to skip waiting (useful for bulk operations).
Examples:
Required fields: roles
Examples:
Usage:
Arguments:
<PRINCIPAL_ID>
Options:
--workspace--roles: Updated list of roles for the principal (can be repeated)--wait-role-propagation: If true, wait for roles to propagate before returning (default: true). Set to false for bulk operations.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
Functional plugins
nemo agents
Plugin commands for agents.
Usage:
Help:
--install-completion: Install completion for the current shell.--show-completion: Show completion for the current shell, to copy it or customize the installation.--help, -h: Show this message and exit.
Commands:
Agent Resources (requires running cluster):
create: Register an agent on the platform.list: List agents on the platform.get: Get an agent by name.delete: Delete an agent from the platform.deploy: Deploy an agent on the platform.logs: Show logs for an agent deployment.undeploy: Stop and remove a deployment (or all deployments for an…deployments: Manage agent deployments.environment-specs: Manage agent environment specs.environments: Manage agent environments.compute-specs: Manage agent compute specs.
Deployed agent interaction (requires running cluster):
chat: Chat interactively with a new or existing deployed-agent…sessions: Discover and manage persisted deployed-agent sessions.
Jobs:
analyze: Analyze a batch of eval-suite results (clusters,…evaluate: Evaluate an agent workflow against a dataset as a…evaluate-suite: Run a directory of containerized eval tasks (Harbor or…execute: Execute an agent to completion as a scheduled platform job.optimize-skills: Optimize an agent’s skills against eval failures via a…package-agent: Build a container image for an agent stored on the platform.optimize: Optimize a Fabric agent workflow (numeric HPO).
Local commands:
invoke: Invoke an agent — locally (with —agent-config) or via…run: Run an agent locally as a persistent FastAPI server.leaderboard: Commands for usage leaderboard workflows.usage: Token-usage reports from nat_runner.py outputs.
Packaging (no platform required):
package: Package a NAT agent — render -> validate -> build ->…
Platform agents:
analyst: Analyze agent telemetry and record what the agent gets…experimentalist: NeMo Experimentalist commands.
nemo agents create
Register an agent on the platform.
Usage:
Options:
--name, -n: Agent name.--agent-config, -c <FILE>: Path to an agent YAML config file.--description: [default: ]--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents list
List agents on the platform.
Usage:
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
Output Options:
--format, --output-format, -o, -f <CHOICE>: Output format for the list of agents. [possible values: table, json, yaml, csv, markdown, raw]--no-truncate: Don’t truncate long values in table/markdown/csv output.
nemo agents get
Get an agent by name.
Usage:
Arguments:
<NAME>: Agent name.
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents delete
Delete an agent from the platform.
Usage:
Arguments:
<NAME>: Agent name.
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]--yes, -y: Skip confirmation prompt.
Help:
--help, -h: Show this message and exit.
nemo agents deploy
Deploy an agent on the platform.
Blocks until the deployment is running (exit 0) or failed /
timed out (exit 1) by default, so the exit code reflects the actual
outcome of runtime startup instead of merely the API call. Use
--no-wait to keep the previous fire-and-forget behaviour for
scripted pipelines that prefer to poll separately via nemo agents deployments wait.
Container modes (--mode docker|k8s) compile to the nemo-deployments
plugin. Requires a configured deployments executor (deployments.executors
/ agents.deployments.docker_executor or k8s_executor). Container
endpoint gateway routing and the full k8s runtime contract (in-cluster
inference gateway, wheel staging) are still evolving — docker mode is the
supported local path today.
Usage:
Options:
--agent, -a: Name of the agent to deploy.--name, -n: Deployment name (auto-generated if omitted).--mode: Runtime backend: subprocess (default), docker, or k8s. [default: subprocess]--image, -i: Container image for docker/k8s modes (falls back to deployments.default_image).--use-image-entrypoint: For docker/k8s modes, preserve the image ENTRYPOINT/CMD instead of injecting the platform-owned agent server command.--environment, -e: AgentEnvironment to deploy under, as a ‘workspace/name’ ref (e.g. ‘default/repo-research-ben’). Its EnvironmentSpec is merged into the agent config and its ComputeSpec/secret refs are snapshotted onto the deployment at create time.--wait, --no-wait: Wait for the deployment to reach a terminal status (running or failed) before returning. Exits 0 only on running; exits 1 with the failure reason if runtime startup fails or readiness times out. Pass —no-wait for fire-and-forget behaviour (the original default — returns the pending deployment immediately as JSON).--timeout, -t <INTEGER>: Maximum seconds to wait for a terminal status (only with —wait). [default: 300]--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents logs
Show logs for an agent deployment.
Reads the log file written by the local in-memory runner backend.
NAT subprocess deployments write process output there; Fabric-backed
deployments write validation/preparation entries there. The log file
location is the same convention the backend uses internally:
nmp_user_data_dir() / 'agents' / 'system' / \<deployment-name>.log
by default. This command is therefore only meaningful when the CLI runs
on the same host as the platform — once a remote backend lands, log
retrieval should move to a server-side endpoint.
With --follow (-f), this command behaves like tail -f and
streams new output until interrupted with Ctrl-C.
Usage:
Arguments:
<NAME>: Deployment name to print logs for. If omitted, pass —agent to look up the most recent deployment for that agent.
Options:
--agent, -a: Resolve the most recent deployment for this agent (bycreated_at), including failed ones — handy for post-mortem on a deploy that just died.--follow, -f: Tail the log file and stream new output as it is written.--tail, -n <INTEGER>: Print only the last N lines before exiting (or before following). Default: print full log.--path: Print only the absolute log file path and exit (useful for scripting).--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents undeploy
Stop and remove a deployment (or all deployments for an agent).
Usage:
Arguments:
<NAME>: Deployment name to remove.
Options:
--agent, --all, -a: Remove all deployments for this agent.--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]--yes, -y: Skip confirmation prompt.
Help:
--help, -h: Show this message and exit.
nemo agents deployments
Manage agent deployments.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
list: List deployments.get: Get a deployment by name.delete: Delete a deployment by name.wait: Wait for a deployment to reach ‘running’ or ‘failed’ status.
nemo agents deployments list
List deployments.
Usage:
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
Output Options:
--format, --output-format, -o, -f <CHOICE>: Output format for the list of deployments. [possible values: table, json, yaml, csv, markdown, raw]--no-truncate: Don’t truncate long values in table/markdown/csv output.
nemo agents deployments get
Get a deployment by name.
Usage:
Arguments:
<NAME>: Deployment name.
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents deployments delete
Delete a deployment by name.
Usage:
Arguments:
<NAME>: Deployment name.
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]--yes, -y: Skip confirmation prompt.
Help:
--help, -h: Show this message and exit.
nemo agents deployments wait
Wait for a deployment to reach ‘running’ or ‘failed’ status.
Polls the deployment until it is running (exit 0) or failed / timed out (exit 1). Prints a status line each time the status changes.
Provide either a deployment name directly or —agent to resolve the latest active deployment for that agent automatically.
Usage:
Arguments:
<NAME>: Deployment name to wait for.
Options:
--agent, -a: Wait for the latest active deployment of this agent (alternative to passing a name directly).--timeout, -t <INTEGER>: Maximum seconds to wait. [default: 300]--interval <FLOAT>: Poll interval in seconds. [default: 2.0]--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents environment-specs
Manage agent environment specs.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create an environment spec from a file or inline JSON.list: List environment specs.get: Get an environment spec by name.delete: Delete an environment spec by name.
nemo agents environment-specs create
Create an environment spec from a file or inline JSON.
Usage:
Arguments:
<NAME>: Unique environment-spec name.
Options:
--spec-file, -f <PATH>: Path to a JSON/YAML EnvironmentSpec body (without ‘name’).--spec: Inline EnvironmentSpec body as a JSON object string.--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents environment-specs list
List environment specs.
Usage:
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
Output Options:
--format, --output-format, -o, -f <CHOICE>: Output format. [possible values: table, json, yaml, csv, markdown, raw]--no-truncate: Don’t truncate long values.
nemo agents environment-specs get
Get an environment spec by name.
Usage:
Arguments:
<NAME>: Environment-spec name.
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents environment-specs delete
Delete an environment spec by name.
Usage:
Arguments:
<NAME>: Environment-spec name.
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]--yes, -y: Skip confirmation prompt.
Help:
--help, -h: Show this message and exit.
nemo agents environments
Manage agent environments.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create an AgentEnvironment.list: List environments.get: Get an environment by name.delete: Delete an environment by name.
nemo agents environments create
Create an AgentEnvironment.
Use —environment-spec / —compute-spec for the common ref case, or —spec-file / —spec for a fully inline body. The ref flags override the matching keys from a file/inline body.
Usage:
Arguments:
<NAME>: Unique environment name.
Options:
--environment-spec: ‘workspace/name’ ref to a stored AgentEnvironmentSpec.--compute-spec: ‘workspace/name’ ref to a stored AgentComputeSpec.--spec-file, -f <PATH>: Path to a JSON/YAML AgentEnvironment body (without ‘name’).--spec: Inline AgentEnvironment body as a JSON object string.--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents environments list
List environments.
Usage:
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
Output Options:
--format, --output-format, -o, -f <CHOICE>: Output format. [possible values: table, json, yaml, csv, markdown, raw]--no-truncate: Don’t truncate long values.
nemo agents environments get
Get an environment by name.
Usage:
Arguments:
<NAME>: Environment name.
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents environments delete
Delete an environment by name.
Usage:
Arguments:
<NAME>: Environment name.
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]--yes, -y: Skip confirmation prompt.
Help:
--help, -h: Show this message and exit.
nemo agents compute-specs
Manage agent compute specs.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create a compute spec from a file or inline JSON.list: List compute specs.get: Get a compute spec by name.delete: Delete a compute spec by name.
nemo agents compute-specs create
Create a compute spec from a file or inline JSON.
Usage:
Arguments:
<NAME>: Unique compute-spec name.
Options:
--spec-file, -f <PATH>: Path to a JSON/YAML ComputeSpec body (without ‘name’).--spec: Inline ComputeSpec body as a JSON object string.--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents compute-specs list
List compute specs.
Usage:
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
Output Options:
--format, --output-format, -o, -f <CHOICE>: Output format. [possible values: table, json, yaml, csv, markdown, raw]--no-truncate: Don’t truncate long values.
nemo agents compute-specs get
Get a compute spec by name.
Usage:
Arguments:
<NAME>: Compute-spec name.
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents compute-specs delete
Delete a compute spec by name.
Usage:
Arguments:
<NAME>: Compute-spec name.
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]--yes, -y: Skip confirmation prompt.
Help:
--help, -h: Show this message and exit.
nemo agents chat
Chat interactively with a new or existing deployed-agent session.
Usage:
Options:
--input, -i: Optional first message to send before prompting for the next turn.--agent-deployment, -d: Deployment to start a new persisted session against.--session, -s: Name of an existing persisted session to resume.--session-name: Name for a new session; valid only with —agent-deployment.--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]--timeout, -t <FLOAT RANGE>: Request timeout in seconds for each streamed agent turn. [default: 300] [env: NEMO_AGENTS_INVOKE_TIMEOUT=]
Help:
--help, -h: Show this message and exit.
nemo agents sessions
Discover and manage persisted deployed-agent sessions.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
list: List persisted sessions, newest first.get: Get a persisted session by name.close: Close a session and release its deployed runtime.
nemo agents sessions list
List persisted sessions, newest first.
Usage:
Options:
--agent-deployment, -d: Limit results to sessions bound to this deployment.--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
Output Options:
--format, --output-format, -o, -f <CHOICE>: Output format for the list of sessions. [possible values: table, json, yaml, csv, markdown, raw]--no-truncate: Don’t truncate long values in table/markdown/csv output.
nemo agents sessions get
Get a persisted session by name.
Usage:
Arguments:
<NAME>: Session name.
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents sessions close
Close a session and release its deployed runtime.
Usage:
Arguments:
<NAME>: Session name.
Options:
--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]--yes, -y: Skip confirmation prompt.
Help:
--help, -h: Show this message and exit.
nemo agents analyze
Analyze a batch of eval-suite results (clusters, regressions, hypotheses).
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--batch: Path to a batch directory produced by evaluate-suite.--mechanical-only: Skip the LLM analysis pass.--anthropic-api-key-secret: Name of a platform Secret holding the Anthropic API key. When set on a platform submission, the value is injected asANTHROPIC_API_KEYinto the dispatched subprocess so the LLM gap-analysis pass can call Anthropic. Ignored whenmechanical_only=True. Local (in-process) runs readANTHROPIC_API_KEYfrom the calling shell as before.--anthropic-base-url: Anthropic-compatible API base URL injected asANTHROPIC_BASE_URLfor platform submissions. Local (in-process) runs readANTHROPIC_BASE_URLfrom the calling shell as before.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the AnalyzeBatchConfig Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: format. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo agents evaluate
Evaluate an agent workflow against a dataset as a scheduled platform job.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--agent AGENT_REF | URL: Agent to evaluate against — either a platform-managed agent reference (e.g. ‘calculator’, ‘workspace/calculator’) or an HTTP(S) endpoint URL (e.g. ‘http://localhost:8080’). Bare names resolve to the platform gateway URL ‘{base_url}/apis/agents/v2/workspaces/{workspace}/agents/{name}/-’; URLs are passed through to ‘nat eval —endpoint’ verbatim. When omitted, the eval config must include an inline agent workflow.--eval-config: Path to the NAT evaluation YAML config file.--eval-config-fileset FILESET_REF: Optional fileset reference (nameorworkspace/name). When set, the runner downloads the fileset’s contents into a tempdir and resolveseval_configrelative to that dir. Local CLI runs leave thisNone.--output PATH | FILESET_REF: Where to write eval outputs — either a local directory (path-shaped: starts with ’/’, ’./’, ’../’, ’~/’) or a NeMo Platform fileset reference (‘name’ or ‘workspace/name’). Filesets are created on demand if missing. Defaults to <ctx.storage.persistent>/results (the platform-injected persistent volume) when not provided.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the EvaluateAgentSpec Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: workspace. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo agents evaluate-suite
Run a directory of containerized eval tasks (Harbor or NAT) against an agent.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--evals: Path to the directory of eval tasks.--agent: Agent root.--concurrency <INTEGER>: Parallel eval concurrency.--skip-build: Skip docker build (Harbor only).--output: Output dir for batch artifacts.--filter-glob: Glob filter on eval names.--repeats <INTEGER>: Trials per eval (median aggregation when >1).--anthropic-api-key-secret: Name of a platform Secret holding the Anthropic API key. When set on a platform submission, the value is injected asANTHROPIC_API_KEYinto the dispatched subprocess so the Harbor eval tasks’ LLM-judge calls reach Anthropic. Local (in-process) runs readANTHROPIC_API_KEYfrom the calling shell as before.--anthropic-base-url: Anthropic-compatible API base URL injected asANTHROPIC_BASE_URLfor platform submissions. Local (in-process) runs readANTHROPIC_BASE_URLfrom the calling shell as before.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the EvaluateSuiteSubmitConfig Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: runner, prefer. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo agents execute
Execute an agent to completion as a scheduled platform job.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--agent: Agent to execute: an Agent entity name or workspace/name ref, or an inline agent definition for a config composed at request time. This flag accepts the string form only; use —spec or —spec-file for the other union form(s).--input: Prompt to pass to the agent.--environment: AgentEnvironment to run under: a “workspace/name” ref to a stored AgentEnvironment, an inline environment, or None. Its EnvironmentSpec is merged into the agent config and its ComputeSpec/secret refs are snapshotted onto the job step at creation time. This flag accepts the string form only; use —spec or —spec-file for the other union form(s).--workdir.base-workdir: Optional Files reference for the initial working directory.--timeout-seconds <FLOAT>: Maximum time to wait for Fabric to return an execution result.--extension.kind: Trusted extension kind registered by an installed NeMo plugin.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the ExecuteAgentJobConfig Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: agent (other union forms), environment (other union forms), workdir.artifact_mounts, extension.config. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo agents optimize-skills
Optimize an agent’s skills against eval failures via a coding agent (Claude).
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--evals: Path to the directory of eval tasks.--agent: Agent root (where the loop runs from; skills live under it).--skills-path: Relative path inside agent where skills live.--filter-glob: Glob filter on eval names (e.g. ‘files-*’).--iterations <INTEGER>: Overrideiterationsin the spec.--concurrency <INTEGER>: Overrideconcurrencyin the spec.--state: Path to loop_state.json; default: <agent>/loop_state.json.--initial-batch: Existing batch dir to seed from.--full-verification: Overridefull_verificationin the spec.--open-pr: Overrideopen_prin the spec.--repeats <INTEGER>: Overriderepeatsin the spec.--analyze-only: Analyze-only mode: consume an existing —initial-batch, generate suggestions, exit. Skips worktree, apply, verify, MR. Works with any AUT (Harbor or NAT).--anthropic-api-key-secret: Name of a platform Secret holding the Anthropic API key. When set on a platform submission, the value is injected asANTHROPIC_API_KEYinto the dispatched subprocess so the loop’scheck_anthropic_api/ Claude analyzer preflights pass. Local (in-process) runs readANTHROPIC_API_KEYfrom the calling shell as before.--anthropic-base-url: Anthropic-compatible API base URL injected asANTHROPIC_BASE_URLfor platform submissions. Local (in-process) runs readANTHROPIC_BASE_URLfrom the calling shell as before.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the OptimizeSkillsConfig Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: trace_parser. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo agents package-agent
Build a container image for an agent stored on the platform.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--agent: Name of the Agent entity to package.--tag: Image name and optional tag, always nested under ‘nemo-agents/{workspace}/’. Defaults to ‘{agent_name}-{agent_id}:{agent_version}’.--base-image-url: Base image repository override (e.g. ‘nvcr.io/nvidia/base/ubuntu’).--base-image-tag: Base image tag override (e.g. ‘noble-20260217’).--python-version: Python version baked into the image (e.g. ‘3.13’).--uv-version: uv version baked into the image (e.g. ‘0.9.14’).--allow-root: Run the agent as root instead of the ‘agent’ user.--sandbox-runtime: Render an image compatible with a sandbox runtime (e.g. ‘openshell’).--agent-version: OCI image version label override.--agent-author: OCI image authors label override.--skip-validation: Bypass Fabric package validation before building.--registry: Push the built image to this registry (e.g. ‘nvcr.io/my-org’). The host executing the push must already be authenticated to it: the local machine for direct CLI invocation, or the platform’s job-execution host for remote execution or the REST API. Credentials are never accepted over this API.--push-tag: Fully-qualified remote tag. Defaults to ‘<registry>/<image>’, where <image> is the namespaced local reference ‘nemo-agents/{workspace}/{tag}’. Requires ‘registry’. Must start with ‘<registry>/nemo-agents/{workspace}/’ — Docker tags are daemon-global while the auth boundary here is the workspace, so an unscoped push_tag would let this workspace overwrite another workspace’s image, or redirect the push to a registry other than the one declared.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the PackageAgentInput Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
nemo agents optimize
Optimize a Fabric agent workflow (numeric HPO).
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--optimize-config: Location of the Fabric-native optimization YAML. With optimize_config_fileset set — required for remote submission — this is a path relative to the fileset root. Without it (programmatic local runs only) it is an absolute path on the host running the job.--optimize-config-fileset FILESET_REF: Fileset holding the optimization bundle: the config named by optimize_config plus every asset it references (Agent under Test package, dataset, eval.fabric.base_dir tree, hooks and MCP configs). Stage one withnemo agents optimize prepare-fileset. Required for remote submissions, where the job has no access to the client’s filesystem. Programmatic local runs may omit it and use an absolute host path for optimize_config.--agent: Optional platform agent reference (‘name’ or ‘workspace/name’). When omitted, the optimization config must include an inline Fabric agent package.--output PATH | FILESET_REF: Where to publish the study artifacts (optimized config, trials dataframe, pareto plots, ATIF evidence) once the study succeeds — either a local directory (path-shaped: starts with ’/’, ’./’, ’../’, ’~/’) or a NeMo Platform fileset reference (‘name’ or ‘workspace/name’). Filesets are created on demand if missing. This is in addition to the per-job artifacts thatctx.results.savealways registers; it gives remote clients a stable, addressable location to read from.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the OptimizeSubmitSpec Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: workspace. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
Commands:
prepare-fileset: Validate an optimize bundle and upload it to a fileset…
nemo agents optimize prepare-fileset
Validate an optimize bundle and upload it to a fileset for optimize.
Usage:
Options:
--source <DIRECTORY>: Directory holding the optimize config and every asset it references.--optimize-config: Optimize YAML, as a path relative to —source. This is the value to pass tooptimize --optimize-config.--fileset: Fileset to upload into (‘name’ or ‘workspace/name’). Created if missing.--workspace: Workspace for the fileset and for agent / model preflight. [default: default]--agent: Platform agent supplying the Agent under Test, for configs that carry only the optimizer and eval overlay.--check-models, --no-check-models: Also resolve the config’s models against the platform before uploading.--dry-run: Run preflight and print the result without uploading.--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents invoke
Invoke an agent — locally (with —agent-config) or via the platform (with —agent or —agent-deployment).
Usage:
Options:
--agent-config, -c <FILE>: Path to an agent YAML config file.--input, -i: Input query string for local invocation.--input-file <PATH>: JSON file containing a list of input queries for batch invocation.--agent, -a: Name of a platform-deployed agent to invoke (platform required).--agent-deployment, -d: Name of a specific deployment to invoke (platform required).--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]--timeout, -t <FLOAT>: Request timeout in seconds for platform invocation. [default: 300] [env: NEMO_AGENTS_INVOKE_TIMEOUT=]--no-progress: Suppress the stderr spinner while waiting for the response.
Help:
--help, -h: Show this message and exit.
nemo agents run
Run an agent locally as a persistent FastAPI server.
Usage:
Options:
--agent-config, -c <FILE>: Path to an agent YAML config file.--host: [default: 0.0.0.0]--port, -p <INTEGER>: [default: 8080]
Help:
--help, -h: Show this message and exit.
nemo agents leaderboard
Commands for usage leaderboard workflows.
Examples:
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
show: Show a ranked leaderboard from local usage report files.
nemo agents leaderboard show
Show a ranked leaderboard from local usage report files.
Examples:
Usage:
Arguments:
<PATHS...>: One or more usage report files or directories containing report files.
Help:
--help, -h: Show this message and exit.
Output Options:
--compact: Render the compact leaderboard table.
nemo agents usage
Token-usage reports from nat_runner.py outputs.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
show: Show a usage report for ref.
nemo agents usage show
Show a usage report for ref.
Usage:
Arguments:
<<PATH | FILESET_REF>>: Local path to a result.json / run dir / nat-jobs dir, or a NeMo Platform fileset reference.
Options:
--total-params <FLOAT>: Model’s total parameter count, in billions (e.g. 8.0 for Llama-3.1-8B, 70.0 for Llama-3.1-70B). When set with non-null tokens, compute_units = total_tokens × total_params. Closed-source models have no public number — leave unset and compute_units stays null.--workspace, -w: [default: default]--base-url: Platform base URL. Resolution order: (1) this —base-url flag or NEMO_BASE_URL; (2) shared CLI config (nemo config set --base-url) or NMP_BASE_URL; (3) http://localhost:8080 (default). [env: NEMO_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo agents package
Package a NAT agent — render -> validate -> build -> publish.
Progressive pipeline controlled by flags: —no-build emit Dockerfile + .dockerignore (no image) (default) render + validate + build —publish —registry … render + validate + build + push
Platform behavior:
- no —platform image built for the local daemon’s native platform
- one —platform image built for that platform (cross-arch via buildx)
- multi —platform rejected — multi-arch builds via buildx are not
yet wired up; build per-arch and combine with
docker buildx imagetools createuntil then.
Usage:
Options:
--agent, -c <FILE>: Path to a NAT workflow YAML config file.--pyproject <FILE>: Path to pyproject.toml (enables project mode).--no-build: Stop after render — emit Dockerfile + .dockerignore only (no image built).--publish: After building, tag and push to —registry.--format: Packaging format: ‘docker’ (Jinja2 Dockerfile). ‘whl’ is reserved for future wheel-based builds and is currently rejected. [default: docker]--dockerfile <FILE>: Use an existing Dockerfile instead of rendering (skips render stage).--tag, -t: Image tag. Defaults to ‘<agent-name>-<agent-id>:<agent-version>’.--platform: Target platform (e.g. ‘linux/amd64’ or ‘linux/arm64’). When omitted, defaults to the local daemon’s native platform. Multi-arch builds via buildx are not yet implemented; pass at most one value.--registry, -r: Remote registry URL (required when —publish is set).--push-tag: Fully-qualified remote tag. Defaults to ‘<registry>/<tag>’.--output, -o <PATH>: Output path for rendered Dockerfile (only used with —no-build). Defaults to ‘Dockerfile’ next to —pyproject when given (project root, so COPY statements resolve), otherwise next to the agent config.--base-image-url: [env: NEMO_AGENTS_BASE_IMAGE_URL=]--base-image-tag: [env: NEMO_AGENTS_BASE_IMAGE_TAG=]--python-version: [env: NEMO_AGENTS_PYTHON_VERSION=]--nat-version: NAT release to install (e.g. ‘1.7.0’). Strongly recommended: pin explicitly so image tags/labels/deps are reproducible. When omitted, a baked-in default is used and a warning is printed.--uv-version: [env: NEMO_AGENTS_UV_VERSION=]--allow-root: Disable non-root USER hardening in the rendered Dockerfile.--sandbox-runtime: Render an image compatible with a sandbox runtime (e.g. ‘openshell’). Discovers the runtime’s image profile and bakes in its required apt packages + users so the image can run inside that sandbox supervisor.--ignore, --no-ignore: Generate a .dockerignore file alongside the Dockerfile.--skip-validation: Bypass agent config validation before build.--agent-version: Override agent version OCI label.--agent-author: Override agent author OCI label.--template: Path to an external Jinja2 Dockerfile template.
Help:
--help, -h: Show this message and exit.
nemo agents analyst
Analyze agent telemetry and record what the agent gets wrong.
Usage:
Help:
--install-completion: Install completion for the current shell.--show-completion: Show completion for the current shell, to copy it or customize the installation.--help, -h: Show this message and exit.
Commands:
run: Run the analyst agent against a running NMP instance.doctor: Check whether the current profile is ready for analysis.
nemo agents analyst run
Run the analyst agent against a running NMP instance.
Builds the analyst agent with --agent (and optional --ethos)
formatted into its instructions and tools scoped
to --agent / --workspace / --base-url, runs it, and
prints whatever the agent returns. Insights are written to the
platform, and mirrored to --insights-file-output when given.
Usage:
Options:
--agent: Name of the agent (agent under test) the analyst should focus on.--ethos <PATH>: Path to a markdown file describing the agent under test (its Ethos).--workspace: Workspace the analyst should operate in.--base-url: Base URL of the running NMP instance the analyst’s tools should call.--profile <FILE>: Path to optimizer.yaml. Default: discovered by walking up from cwd.--insights-file-output <PATH>: Also write insights to this local YAML file. Insights always go to the platform first; the file mirrors what was stored, platform ids included, and each run merges into it.--verbose, -v: Stream the analyst’s tool calls and reasoning to stderr while it runs. Off by default so that stdout stays clean for piping the final answer.
Help:
--help, -h: Show this message and exit.
nemo agents analyst doctor
Check whether the current profile is ready for analysis.
Usage:
Options:
--profile <FILE>: Path to optimizer.yaml. Default: discovered by walking up from cwd.--base-url: Base URL of the running NMP instance to check.
Help:
--help, -h: Show this message and exit.
nemo agents experimentalist
NeMo Experimentalist commands.
Usage:
Help:
--install-completion: Install completion for the current shell.--show-completion: Show completion for the current shell, to copy it or customize the installation.--help, -h: Show this message and exit.
Commands:
run: Run offline optimization for a baseline agent (local dir…components: List the components this install can resolve by name.doctor: Diagnose Experimentalist setup: profile, artifacts,…
nemo agents experimentalist run
Run offline optimization for a baseline agent (local dir or git source).
Usage:
Options:
--agent: Baseline agent: a local directory or a git URL with an optional ref (e.g. ssh://git@host/group/repo.git@main). Optional in Mode 1 (the insight supplies the agent) and overrides the insight’s agent when given. A git source records provenance and enables —config storage.publish_winner to open a draft PR/MR for the winner against that ref.--ethos: URI of a markdown file describing the agent under test (its Ethos).--insight: Insight for Mode 1: a local insight file OR a platform insight id (surfaced in Studio). A path that exists on disk is read locally; otherwise it is fetched from the platform. Default: <profile-dir>/.nemo-optimizer/insights.yaml when it exists (wherenemo agents analyst runwrites by default).--insight-id: Selector for a local multi-insight file: exact id/title first; if none matches, a decimal value is a zero-based index.--no-insight: Disable insight use (Mode 2), including the shared profile insight file.--profile <FILE>: Path to optimizer.yaml. Default: discovered by walking up from the cwd.--train-dataset: Train dataset: local path or harbor registry ref. Falls back to the profile.--validation-dataset: Validation dataset: local path or harbor registry ref. Falls back to the profile.--task-template: Evaluator-specific task-template URI. Required with —insight.--experiment-dir, --output, --experiments-output, -o <DIRECTORY>: Local experiment directory; writes eval-and-optimize/ here. Default: <profile-dir>/.nemo-optimizer/experiments/<timestamp> when a profile governs the run, else ./tmp.--workspace: Intake/NMP workspace for traces and run/candidate metadata. Falls back to the profile.--base-url: Base URL of the running NMP instance. Default: NMP_BASE_URL (shell or profile-dir .env), else http://localhost:8080. [env: NMP_BASE_URL=]--config <FILE>: YAML or JSON configuration for the optimizer run.--framework-skills <DIRECTORY>: Path to a directory of framework skills to load into the optimizer agents. May be specified multiple times. [default: ]
Help:
--help, -h: Show this message and exit.
nemo agents experimentalist components
List the components this install can resolve by name.
Includes anything a pip installed package registered, which is how a
developer checks their own component was picked up.
Usage:
Options:
--role: Show only this role.
Help:
--help, -h: Show this message and exit.
nemo agents experimentalist doctor
Diagnose Experimentalist setup: profile, artifacts, models, platform, runtime.
Usage:
Options:
--insight: Optional insight ref to verify.--insight-id: Select an exact id/title or zero-based index from a local multi-insight file.--profile <PATH>: Path to optimizer.yaml.--base-url: Base URL of the running NMP instance. Default: NMP_BASE_URL (shell or profile-dir .env), else http://localhost:8080. [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo data-designer
Plugin commands for data-designer.
Usage:
Help:
--install-completion: Install completion for the current shell.--show-completion: Show completion for the current shell, to copy it or customize the installation.--help, -h: Show this message and exit.
Commands:
validate: Validate a Data Designer configuration.createpreviewpersonas: Manage Nemotron Personas datasetsretrieval: Build specs for Nemotron retrieval SDG commands.agent: Agent-only interface for dynamic Data Designer introspection
Functions:
retrieval-preview: Submit retrieval-preview over HTTP.
Jobs:
retrieval-generate: Generate retrieval Q&A JSONL from a document corpus…retrieval-prepare: Convert retrieval SDG output to eval_beir and training…retrieval-run: Chain retrieval generate then prepare as a multi-step…
nemo data-designer validate
Validate a Data Designer configuration.
Usage:
Arguments:
<CONFIG_SOURCE>: Path or URL to a config file (.yaml/.yml/.json), or a local Python module (.py) that defines a load_config_builder() function.
Options:
--workspace: Workspace used to resolve provider references and seed sources (remote pass). Defaults to the SDK’s configured workspace, or ‘default’.--output <CHOICE>: Output format. ‘json’ suppresses the human-formatted blocks. [possible values: text, json; default: text]
Help:
--help, -h: Show this message and exit.
nemo data-designer create
Usage:
Arguments:
<CONFIG_SOURCE>: Path or URL to a config file (.yaml/.yml/.json), or a local Python module (.py) that defines a load_config_builder() function.
Options:
--num-records, -n <INTEGER RANGE>: [default: 10]--workspace, -w: [default: default]--profile--cluster--base-url-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>
Help:
--help, -h: Show this message and exit.
nemo data-designer preview
Usage:
Arguments:
<CONFIG_SOURCE>: Path or URL to a config file (.yaml/.yml/.json), or a local Python module (.py) that defines a load_config_builder() function.
Options:
--num-records, -n <INTEGER RANGE>: [default: 10]--workspace, -w: [default: default]--cluster--base-url--request-id--non-interactive: Display all records at once instead of browsing interactively. Ignored when —save-results is used.--save-results: Save preview artifacts to disk (dataset.parquet, per-record HTML, analysis report) instead of displaying records in the terminal.--artifact-path, -o: Directory for saved results (used with —save-results). Defaults to ./artifacts.
Help:
--help, -h: Show this message and exit.
nemo data-designer personas
Manage Nemotron Personas datasets
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
make-fileset: Create the system fileset for one Nemotron Personas locale.
nemo data-designer personas make-fileset
Create the system fileset for one Nemotron Personas locale.
Usage:
Options:
--locale <CHOICE>: Locale fileset to create. [possible values: en_IN, en_SG, en_US, fr_FR, hi_Deva_IN, hi_Latn_IN, ja_JP, ko_KR, pt_BR]--api-key-secret: Fully qualified NGC API key secret reference (WORKSPACE/NAME).--api-key-env-var: Environment variable containing an NGC API key to create at —api-key-secret.
Help:
--help, -h: Show this message and exit.
nemo data-designer retrieval
Build specs for Nemotron retrieval SDG commands. These helpers print the auto-generated retrieval-generate, retrieval-prepare, or retrieval-preview command to submit.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
generate: Build a spec for the auto-generated…prepare: Build a spec for the auto-generatedretrieval-prepare…preview: Build a spec for the auto-generatedretrieval-preview…
nemo data-designer retrieval generate
Build a spec for the auto-generated retrieval-generate job command.
Usage:
Options:
--corpus: Corpus fileset ref or hf:// URI.--provider: Inference Gateway provider (workspace/name).--chat-model: Chat model for artifact extraction, Q&A, and judging.--embed-model: Embedding model.--workspace, -w: [default: default]--print-spec: Print JSON spec instead of a submission command.
Help:
--help, -h: Show this message and exit.
nemo data-designer retrieval prepare
Build a spec for the auto-generated retrieval-prepare job command.
Usage:
Options:
--sdg-input: Stage 0 fileset, generation_result.json, or hf:// URI.--train-input-file--mine, --no-mine: Run GPU hard-negative mining after conversion. Conversion-only is the default.--workspace, -w: [default: default]
Help:
--help, -h: Show this message and exit.
nemo data-designer retrieval preview
Build a spec for the auto-generated retrieval-preview function command.
Usage:
Options:
--corpus: Corpus fileset ref or hf:// URI.--provider: Inference Gateway provider (workspace/name).--chat-model: Chat model for artifact extraction, Q&A, and judging.--embed-model: Embedding model.--workspace, -w: [default: default]
Help:
--help, -h: Show this message and exit.
nemo data-designer agent
Agent-only interface for dynamic Data Designer introspection
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
state: Return current local state relevant to agentscontext: Prints output from all agent subcommands to bootstrap…types: Type names, descriptions, and source files for one or all…
nemo data-designer agent state
Return current local state relevant to agents
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
model-aliases: Model aliases and usability status.persona-datasets: Persona locales and install status.
nemo data-designer agent state model-aliases
Model aliases and usability status.
Usage:
Help:
--help, -h: Show this message and exit.
nemo data-designer agent state persona-datasets
Persona locales and install status.
Usage:
Help:
--help, -h: Show this message and exit.
nemo data-designer agent context
Prints output from all agent subcommands to bootstrap context.
Usage:
Help:
--help, -h: Show this message and exit.
nemo data-designer agent types
Type names, descriptions, and source files for one or all families.
Usage:
Help:
--help, -h: Show this message and exit.
nemo data-designer retrieval-preview
Submit retrieval-preview over HTTP.
Usage:
Function Spec:
--generate.corpus: Fileset ref (workspace/fileset[#subdir]) or hf:// URI.--generate.provider: Default Inference Gateway provider (name or workspace/name).--generate.chat-provider: Optional provider override for artifact extraction, Q&A generation, and quality judging.--generate.embed-provider: Optional provider override for embedding calls.--generate.corpus-id: Overridegenerate.corpus_idin the spec.--generate.dataset-name: Overridegenerate.dataset_namein the spec.--generate.min-text-length <INTEGER>: Overridegenerate.min_text_lengthin the spec.--generate.sentences-per-chunk <INTEGER>: Overridegenerate.sentences_per_chunkin the spec.--generate.num-sections <INTEGER>: Overridegenerate.num_sectionsin the spec.--generate.num-files <INTEGER>: Overridegenerate.num_filesin the spec.--generate.max-artifacts-per-type <INTEGER>: Overridegenerate.max_artifacts_per_typein the spec.--generate.num-pairs <INTEGER>: Overridegenerate.num_pairsin the spec.--generate.min-hops <INTEGER>: Overridegenerate.min_hopsin the spec.--generate.max-hops <INTEGER>: Overridegenerate.max_hopsin the spec.--generate.min-complexity <INTEGER>: Overridegenerate.min_complexityin the spec.--generate.similarity-threshold <FLOAT>: Overridegenerate.similarity_thresholdin the spec.--generate.buffer-size <INTEGER>: Overridegenerate.buffer_sizein the spec.--generate.num-records <INTEGER>: Overridegenerate.num_recordsin the spec.--generate.artifact-extraction-model: Chat model for artifact extraction.--generate.qa-generation-model: Chat model for question/answer generation.--generate.quality-judge-model: Chat model for quality judging.--generate.embed-model: Embedding model for retrieval SDG.--generate.hf-token-secret: Overridegenerate.hf_token_secretin the spec.--num-records <INTEGER>: Overridenum_recordsin the spec.
Help:
--help, -h: Show this message and exit.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace path segment used in the submit URL. [default: default]--request-id: Set the X-Request-ID header (echoed back in ctx.request_id).
Function Spec flags are generated from the RetrievalPreviewSpec Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: generate.file_extensions, generate.query_counts, generate.reasoning_counts, generate.resume. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo data-designer retrieval-generate
Generate retrieval Q&A JSONL from a document corpus (Nemotron Stage 0).
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--corpus: Fileset ref (workspace/fileset[#subdir]) or hf:// URI.--provider: Default Inference Gateway provider (name or workspace/name).--chat-provider: Optional provider override for artifact extraction, Q&A generation, and quality judging.--embed-provider: Optional provider override for embedding calls.--corpus-id: Overridecorpus_idin the spec.--dataset-name: Overridedataset_namein the spec.--min-text-length <INTEGER>: Overridemin_text_lengthin the spec.--sentences-per-chunk <INTEGER>: Overridesentences_per_chunkin the spec.--num-sections <INTEGER>: Overridenum_sectionsin the spec.--num-files <INTEGER>: Overridenum_filesin the spec.--max-artifacts-per-type <INTEGER>: Overridemax_artifacts_per_typein the spec.--num-pairs <INTEGER>: Overridenum_pairsin the spec.--min-hops <INTEGER>: Overridemin_hopsin the spec.--max-hops <INTEGER>: Overridemax_hopsin the spec.--min-complexity <INTEGER>: Overridemin_complexityin the spec.--similarity-threshold <FLOAT>: Overridesimilarity_thresholdin the spec.--buffer-size <INTEGER>: Overridebuffer_sizein the spec.--num-records <INTEGER>: Overridenum_recordsin the spec.--artifact-extraction-model: Chat model for artifact extraction.--qa-generation-model: Chat model for question/answer generation.--quality-judge-model: Chat model for quality judging.--embed-model: Embedding model for retrieval SDG.--hf-token-secret: Overridehf_token_secretin the spec.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the RetrievalGenerateJobConfig Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: file_extensions, query_counts, reasoning_counts, resume. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo data-designer retrieval-prepare
Convert retrieval SDG output to eval_beir and training JSONL (Nemotron Stage 1).
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--sdg-input: Fileset or hf:// URI to Stage 0 output or generation_result.json.--train-input-file: Fileset containing a pre-converted wrapped train.json; skips conversion.--corpus-id: Overridecorpus_idin the spec.--quality-threshold <FLOAT>: Overridequality_thresholdin the spec.--train-ratio <FLOAT>: Overridetrain_ratioin the spec.--val-ratio <FLOAT>: Overrideval_ratioin the spec.--seed <INTEGER>: Overrideseedin the spec.--max-pos-docs <INTEGER>: Overridemax_pos_docsin the spec.--use-group-id-in-eval: Overrideuse_group_id_in_evalin the spec.--enable-mining: When true, run GPU hard-negative mining after conversion. Conversion-only is the default.--model: Platform model entity whose fileset contains the mining encoder and tokenizer.--hard-negatives-to-mine <INTEGER>: Overridehard_negatives_to_minein the spec.--hard-neg-margin <FLOAT>: Overridehard_neg_marginin the spec.--mining-batch-size <INTEGER>: Overridemining_batch_sizein the spec.--query-prefix: Overridequery_prefixin the spec.--passage-prefix: Overridepassage_prefixin the spec.--query-max-length <INTEGER>: Overridequery_max_lengthin the spec.--passage-max-length <INTEGER>: Overridepassage_max_lengthin the spec.--add-bos-token: Overrideadd_bos_tokenin the spec.--add-eos-token: Overrideadd_eos_tokenin the spec.--dist-backend: Overridedist_backendin the spec.--dist-timeout-minutes <INTEGER>: Overridedist_timeout_minutesin the spec.--mining.query-embedding-batch-size <INTEGER>: Overridemining.query_embedding_batch_sizein the spec.--mining.document-embedding-batch-size <INTEGER>: Overridemining.document_embedding_batch_sizein the spec.--mining.corpus-chunk-size <INTEGER>: Overridemining.corpus_chunk_sizein the spec.--mining.load-embeddings-from-cache: Overridemining.load_embeddings_from_cachein the spec.--mining.use-negatives-from-file: Overridemining.use_negatives_from_filein the spec.--hf-token-secret: Overridehf_token_secretin the spec.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the RetrievalPrepareJobConfig Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: split_strategy, attn_implementation, mining.hard_neg_margin_type. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo data-designer retrieval-run
Chain retrieval generate then prepare as a multi-step jobs-service workflow.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--generate.corpus: Fileset ref (workspace/fileset[#subdir]) or hf:// URI.--generate.provider: Default Inference Gateway provider (name or workspace/name).--generate.chat-provider: Optional provider override for artifact extraction, Q&A generation, and quality judging.--generate.embed-provider: Optional provider override for embedding calls.--generate.corpus-id: Overridegenerate.corpus_idin the spec.--generate.dataset-name: Overridegenerate.dataset_namein the spec.--generate.min-text-length <INTEGER>: Overridegenerate.min_text_lengthin the spec.--generate.sentences-per-chunk <INTEGER>: Overridegenerate.sentences_per_chunkin the spec.--generate.num-sections <INTEGER>: Overridegenerate.num_sectionsin the spec.--generate.num-files <INTEGER>: Overridegenerate.num_filesin the spec.--generate.max-artifacts-per-type <INTEGER>: Overridegenerate.max_artifacts_per_typein the spec.--generate.num-pairs <INTEGER>: Overridegenerate.num_pairsin the spec.--generate.min-hops <INTEGER>: Overridegenerate.min_hopsin the spec.--generate.max-hops <INTEGER>: Overridegenerate.max_hopsin the spec.--generate.min-complexity <INTEGER>: Overridegenerate.min_complexityin the spec.--generate.similarity-threshold <FLOAT>: Overridegenerate.similarity_thresholdin the spec.--generate.buffer-size <INTEGER>: Overridegenerate.buffer_sizein the spec.--generate.num-records <INTEGER>: Overridegenerate.num_recordsin the spec.--generate.artifact-extraction-model: Chat model for artifact extraction.--generate.qa-generation-model: Chat model for question/answer generation.--generate.quality-judge-model: Chat model for quality judging.--generate.embed-model: Embedding model for retrieval SDG.--generate.hf-token-secret: Overridegenerate.hf_token_secretin the spec.--prepare.sdg-input: Fileset or hf:// URI to Stage 0 output or generation_result.json.--prepare.train-input-file: Fileset containing a pre-converted wrapped train.json; skips conversion.--prepare.corpus-id: Overrideprepare.corpus_idin the spec.--prepare.quality-threshold <FLOAT>: Overrideprepare.quality_thresholdin the spec.--prepare.train-ratio <FLOAT>: Overrideprepare.train_ratioin the spec.--prepare.val-ratio <FLOAT>: Overrideprepare.val_ratioin the spec.--prepare.seed <INTEGER>: Overrideprepare.seedin the spec.--prepare.max-pos-docs <INTEGER>: Overrideprepare.max_pos_docsin the spec.--prepare.use-group-id-in-eval: Overrideprepare.use_group_id_in_evalin the spec.--prepare.enable-mining: When true, run GPU hard-negative mining after conversion. Conversion-only is the default.--prepare.model: Platform model entity whose fileset contains the mining encoder and tokenizer.--prepare.hard-negatives-to-mine <INTEGER>: Overrideprepare.hard_negatives_to_minein the spec.--prepare.hard-neg-margin <FLOAT>: Overrideprepare.hard_neg_marginin the spec.--prepare.mining-batch-size <INTEGER>: Overrideprepare.mining_batch_sizein the spec.--prepare.query-prefix: Overrideprepare.query_prefixin the spec.--prepare.passage-prefix: Overrideprepare.passage_prefixin the spec.--prepare.query-max-length <INTEGER>: Overrideprepare.query_max_lengthin the spec.--prepare.passage-max-length <INTEGER>: Overrideprepare.passage_max_lengthin the spec.--prepare.add-bos-token: Overrideprepare.add_bos_tokenin the spec.--prepare.add-eos-token: Overrideprepare.add_eos_tokenin the spec.--prepare.dist-backend: Overrideprepare.dist_backendin the spec.--prepare.dist-timeout-minutes <INTEGER>: Overrideprepare.dist_timeout_minutesin the spec.--prepare.mining.query-embedding-batch-size <INTEGER>: Overrideprepare.mining.query_embedding_batch_sizein the spec.--prepare.mining.document-embedding-batch-size <INTEGER>: Overrideprepare.mining.document_embedding_batch_sizein the spec.--prepare.mining.corpus-chunk-size <INTEGER>: Overrideprepare.mining.corpus_chunk_sizein the spec.--prepare.mining.load-embeddings-from-cache: Overrideprepare.mining.load_embeddings_from_cachein the spec.--prepare.mining.use-negatives-from-file: Overrideprepare.mining.use_negatives_from_filein the spec.--prepare.hf-token-secret: Overrideprepare.hf_token_secretin the spec.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the RetrievalRunJobConfig Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: generate.file_extensions, generate.query_counts, generate.reasoning_counts, generate.resume, prepare.split_strategy, prepare.attn_implementation, prepare.mining.hard_neg_margin_type. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo guardrail
Manage guardrails.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
check: Chat completion for the provided conversation.configs: Manage configs
nemo guardrail check
Chat completion for the provided conversation.
Required fields: messages, model
Examples:
Usage:
Options:
--workspace--messages: A list of messages comprising the conversation so far (JSON string)--model: The model to use for completion. Must be one of the available models.--frequency-penalty <FLOAT>: Positive values penalize new tokens based on their existing frequency in the text.--function-call: Deprecated in favor of tool_choice. ‘none’ means the model will not call a function and instead generates a message. ‘auto’ means the model can pick between generating a message or calling a function. Specifying a particular function via {‘name’: ‘my_function’} forces the model to call that function. (JSON string)--guardrails: Guardrails specific options for the request. (JSON string)--ignore-eos: Ignore the eos when running--logit-bias: Modify the likelihood of specified tokens appearing in the completion. Maps token IDs (as strings) to bias values from -100 to 100. (JSON string)--logprobs: Whether to return log probabilities of the output tokens or not. If true, returns the log probabilities of each output token returned in the content of message--max-completion-tokens <INTEGER>: An upper bound for the number of tokens that can be generated for a completion, including visible output tokens and reasoning tokens. Preferred over max_tokens for reasoning models.--max-tokens <INTEGER>: The maximum number of tokens that can be generated in the chat completion.--n <INTEGER>: How many chat completion choices to generate for each input message.--presence-penalty <FLOAT>: Positive values penalize new tokens based on whether they appear in the text so far.--reasoning-effort: Constrains effort on reasoning for reasoning models. Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response.--response-format: Format of the response. Use {‘type’: ‘json_object’} for JSON mode or {‘type’: ‘json_schema’, ‘json_schema’: {…}} for structured outputs. (JSON string)--seed <INTEGER>: If specified, attempts to sample deterministically.--stop: Up to 4 sequences where the API will stop generating further tokens. (JSON string)--stream: If set, partial message deltas will be sent, like in ChatGPT.--stream-options: Options for streaming response. Only set this when stream=True. Supports include_usage to receive token usage in the final stream chunk. (JSON string)--temperature <FLOAT>: What sampling temperature to use, between 0 and 2.--tool-choice: Controls which (if any) tool is called by the model. ‘none’ means no tool is called, ‘auto’ lets the model decide, ‘required’ forces a tool call. (JSON string)--tools: A list of tools the model may call. Each tool is an object with a ‘type’ field and a ‘function’ definition. (JSON string)--top-logprobs <INTEGER>: The number of most likely tokens to return at each token position.--top-p <FLOAT>: An alternative to sampling with temperature, called nucleus sampling.--user: A unique identifier representing your end-user, used by some providers for abuse monitoring.--vision: Whether this is a vision-capable request with image inputs.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo guardrail configs
Manage configs
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create a new guardrail config.delete: Delete a guardrail config.list: List available guardrail configs.get: Get info about a guardrail configuration.update: Update model metadata.
nemo guardrail configs create
Create a new guardrail config.
Required fields: name
Examples:
Usage:
Arguments:
<NAME>: The name of the guardrail config
Options:
--workspace--data: Guardrail configuration data (JSON string)--description: Description of the guardrail config--exist-ok: Do not raise an error if the resource already exists. Returns the existing resource.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo guardrail configs delete
Delete a guardrail config.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
nemo guardrail configs list
List available guardrail configs.
Lists guardrail configs for a specific workspace.
Usage:
Options:
--workspace--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort <CHOICE>: The field to sort by. To sort in decreasing order, use-in front of the field name. [possible values: created_at, -created_at, updated_at, -updated_at, name, -name]--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: created_at: {gte: str, lte: str} updated_at: {gte: str, lte: str}
Filter guardrail configs by name, description, project, created_at, and updated_at.
--filter.description--filter.name--filter.project
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo guardrail configs get
Get info about a guardrail configuration.
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo guardrail configs update
Update model metadata.
If the request body has an empty field, keep the old value.
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--data: Guardrail configuration data (JSON string)--description: Description of the guardrail config
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo auditor
Plugin commands for auditor.
Usage:
Help:
--install-completion: Install completion for the current shell.--show-completion: Show completion for the current shell, to copy it or customize the installation.--help, -h: Show this message and exit.
Commands:
info: Print the current plugin status.configs: Manage audit configurations.targets: Manage audit targets.
Jobs:
audit: Run an auditor scan against a configured target.
nemo auditor info
Print the current plugin status.
Usage:
Help:
--help, -h: Show this message and exit.
nemo auditor configs
Manage audit configurations.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
createlistgetupdatedelete
nemo auditor configs create
Usage:
Arguments:
<NAME>: Config name.
Options:
--data-file, -f <FILE>: JSON file with the config body (without name/workspace).--data, -d: Inline JSON body for the config.--workspace, -w: [default: default]--base-url: [default: http://localhost:8080] [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo auditor configs list
Usage:
Options:
--workspace, -w: [default: default]--base-url: [default: http://localhost:8080] [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo auditor configs get
Usage:
Arguments:
<NAME>: Config name.
Options:
--workspace, -w: [default: default]--base-url: [default: http://localhost:8080] [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo auditor configs update
Usage:
Arguments:
<NAME>: Config name.
Options:
--data-file, -f <FILE>: JSON file with the new config body.--data, -d: Inline JSON body.--workspace, -w: [default: default]--base-url: [default: http://localhost:8080] [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo auditor configs delete
Usage:
Arguments:
<NAME>: Config name.
Options:
--workspace, -w: [default: default]--base-url: [default: http://localhost:8080] [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo auditor targets
Manage audit targets.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
createlistgetupdatedelete
nemo auditor targets create
Usage:
Arguments:
<NAME>: Target name.
Options:
--data-file, -f <FILE>: JSON file with the target body (without name/workspace).--data, -d: Inline JSON body for the target.--workspace, -w: [default: default]--base-url: [default: http://localhost:8080] [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo auditor targets list
Usage:
Options:
--workspace, -w: [default: default]--base-url: [default: http://localhost:8080] [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo auditor targets get
Usage:
Arguments:
<NAME>: Target name.
Options:
--workspace, -w: [default: default]--base-url: [default: http://localhost:8080] [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo auditor targets update
Usage:
Arguments:
<NAME>: Target name.
Options:
--data-file, -f <FILE>: JSON file with the new target body.--data, -d: Inline JSON body.--workspace, -w: [default: default]--base-url: [default: http://localhost:8080] [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo auditor targets delete
Usage:
Arguments:
<NAME>: Target name.
Options:
--workspace, -w: [default: default]--base-url: [default: http://localhost:8080] [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo auditor audit
Run an auditor scan against a configured target.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
run: Run locally, in-process.submit: Submit to a cluster.explain: Show input/output schemas.
nemo auditor audit run
Run locally, in-process.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--max-probe-retries <INTEGER>: Overridemax_probe_retriesin the spec.--fail-job-on-retries-exhausted: Overridefail_job_on_retries_exhaustedin the spec.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Job Spec flags are generated from the AuditInputSpec Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: config, target. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo auditor audit submit
Submit to a cluster.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--max-probe-retries <INTEGER>: Overridemax_probe_retriesin the spec.--fail-job-on-retries-exhausted: Overridefail_job_on_retries_exhaustedin the spec.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the AuditInputSpec Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: config, target. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo auditor audit explain
Show input/output schemas.
Usage:
Options:
--profile: Annotate the bundle with this profile. Profile metadata lands in MR 1.4b.--cluster: Accepted for forward compatibility; unused in MR 1.4a.
Help:
--help, -h: Show this message and exit.
nemo anonymizer
Plugin commands for anonymizer.
Usage:
Help:
--install-completion: Install completion for the current shell.--show-completion: Show completion for the current shell, to copy it or customize the installation.--help, -h: Show this message and exit.
Commands:
validate: Validate an AnonymizerConfig against the model selection.
Functions:
preview: Submit preview over HTTP.
Jobs:
run: Anonymize a dataset of records
nemo anonymizer validate
Validate an AnonymizerConfig against the model selection.
Usage:
Options:
--config <PATH>: Path to AnonymizerConfig YAML.--model-configs <PATH>
Help:
--help, -h: Show this message and exit.
nemo anonymizer preview
Submit preview over HTTP.
Usage:
Function Spec:
--config.detect.gliner-threshold <FLOAT>: GLiNER detection confidence threshold (0.0-1.0).--config.detect.validation-max-entities-per-call <INTEGER>: Maximum number of candidate entities included in a single validator LLM call. When a row has more candidates than this, validation is split into chunks that are dispatched (round-robin) across the validator pool.--config.detect.validation-excerpt-window-chars <INTEGER>: Number of characters to include before and after a chunk’s entity span when building the text excerpt sent to the validator. Bounds the prompt context the validator sees per chunk; it is NOT the LLM’s context window limit.--config.rewrite.privacy-goal.protect: What to protect (e.g. direct identifiers, quasi-identifiers).--config.rewrite.privacy-goal.preserve: What to preserve (e.g. utility, semantic meaning).--config.rewrite.instructions: Additional instructions for the rewrite LLM.--config.rewrite.risk-tolerance: Preset controlling repair thresholds and review flagging.--config.rewrite.max-repair-iterations <INTEGER>: Maximum repair rounds. Set to 0 to disable repair.--config.rewrite.strict-entity-protection: If True, requires every entity to receive a protective disposition during sensitivity analysis.--config.emit-telemetry: Whether to emit anonymous Anonymizer telemetry events. See the Telemetry section in the README for what is collected and how to opt out at the environment or CLI level.--data.source: HTTP(S) URL or fileset reference for a CSV/Parquet input file.--data.text-column: Column containing text to anonymize.--data.id-column: Optional column to use as record identifier.--data.data-summary: Short description of the data.--num-records <INTEGER>: Overridenum_recordsin the spec.
Help:
--help, -h: Show this message and exit.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace path segment used in the submit URL. [default: default]--request-id: Set the X-Request-ID header (echoed back in ctx.request_id).
Function Spec flags are generated from the PreviewRequest Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: config.detect.entity_labels, config.replace, model_configs, selected_models.detection, selected_models.replace, selected_models.rewrite. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo anonymizer run
Anonymize a dataset of records
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--config.detect.gliner-threshold <FLOAT>: GLiNER detection confidence threshold (0.0-1.0).--config.detect.validation-max-entities-per-call <INTEGER>: Maximum number of candidate entities included in a single validator LLM call. When a row has more candidates than this, validation is split into chunks that are dispatched (round-robin) across the validator pool.--config.detect.validation-excerpt-window-chars <INTEGER>: Number of characters to include before and after a chunk’s entity span when building the text excerpt sent to the validator. Bounds the prompt context the validator sees per chunk; it is NOT the LLM’s context window limit.--config.rewrite.privacy-goal.protect: What to protect (e.g. direct identifiers, quasi-identifiers).--config.rewrite.privacy-goal.preserve: What to preserve (e.g. utility, semantic meaning).--config.rewrite.instructions: Additional instructions for the rewrite LLM.--config.rewrite.risk-tolerance: Preset controlling repair thresholds and review flagging.--config.rewrite.max-repair-iterations <INTEGER>: Maximum repair rounds. Set to 0 to disable repair.--config.rewrite.strict-entity-protection: If True, requires every entity to receive a protective disposition during sensitivity analysis.--config.emit-telemetry: Whether to emit anonymous Anonymizer telemetry events. See the Telemetry section in the README for what is collected and how to opt out at the environment or CLI level.--data.source: HTTP(S) URL or fileset reference for a CSV/Parquet input file.--data.text-column: Column containing text to anonymize.--data.id-column: Optional column to use as record identifier.--data.data-summary: Short description of the data.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the AnonymizerRequest Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: config.detect.entity_labels, config.replace, model_configs, selected_models.detection, selected_models.replace, selected_models.rewrite. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo evaluator
Plugin commands for evaluator.
Usage:
Help:
--install-completion: Install completion for the current shell.--show-completion: Show completion for the current shell, to copy it or customize the installation.--help, -h: Show this message and exit.
Commands:
info: Print the current plugin status.metric-types: Print available evaluator metric names or a metric JSON…
Jobs:
agent-evaluate: Run agent evaluation over tasks against a model, agent,…evaluate: Run evaluator SDK metrics against inline rows or…
nemo evaluator info
Print the current plugin status.
Usage:
Help:
--help, -h: Show this message and exit.
nemo evaluator metric-types
Print available evaluator metric names or a metric JSON schema.
Usage:
Arguments:
<<metric-name>>
Help:
--help, -h: Show this message and exit.
nemo evaluator agent-evaluate
Run agent evaluation over tasks against a model, agent, or runner.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
run: Run locally, in-process.submit: Submit to a cluster.explain: Show input/output schemas.
nemo evaluator agent-evaluate run
Run locally, in-process.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--max-concurrent-tasks <INTEGER>: Maximum number of tasks evaluated concurrently. Distinct from a target’sparams.parallelism, which bounds concurrent inference requests within trial generation.--fail-fast: Stop the run on the first scoring failure when True.--publication.intake.evaluation-id: Name of the existing Evaluation to publish under. Must already exist; the job does not create it.--publication.intake.agent-name: Agent name recorded on each published trajectory. Derived from the target when it names one; required otherwise.--publication.intake.agent-version: Agent version recorded on each published trajectory. Neither a Model nor an Agent carries a version, so this defaults to ‘unknown’ unless the submitter supplies one.--publication.intake.required: Fail the job when publication fails. Defaults to True so a run that asked to publish does not report success with nothing in Experiments. The result bundle is saved before publication runs, so a failed job still leaves the results intact to re-publish. Set False to keep the job successful and report the failure in its output instead.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Job Spec flags are generated from the AgentEvalInputSpec Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: target, trials, labels, tasks. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo evaluator agent-evaluate submit
Submit to a cluster.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--max-concurrent-tasks <INTEGER>: Maximum number of tasks evaluated concurrently. Distinct from a target’sparams.parallelism, which bounds concurrent inference requests within trial generation.--fail-fast: Stop the run on the first scoring failure when True.--publication.intake.evaluation-id: Name of the existing Evaluation to publish under. Must already exist; the job does not create it.--publication.intake.agent-name: Agent name recorded on each published trajectory. Derived from the target when it names one; required otherwise.--publication.intake.agent-version: Agent version recorded on each published trajectory. Neither a Model nor an Agent carries a version, so this defaults to ‘unknown’ unless the submitter supplies one.--publication.intake.required: Fail the job when publication fails. Defaults to True so a run that asked to publish does not report success with nothing in Experiments. The result bundle is saved before publication runs, so a failed job still leaves the results intact to re-publish. Set False to keep the job successful and report the failure in its output instead.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the AgentEvalInputSpec Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: target, trials, labels, tasks. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo evaluator agent-evaluate explain
Show input/output schemas.
Usage:
Options:
--profile: Annotate the bundle with this profile. Profile metadata lands in MR 1.4b.--cluster: Accepted for forward compatibility; unused in MR 1.4a.
Help:
--help, -h: Show this message and exit.
nemo evaluator evaluate
Run evaluator SDK metrics against inline rows or FilesetRef datasets.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
run: Run locally, in-process.submit: Submit to a cluster.explain: Show input/output schemas.
nemo evaluator evaluate run
Run locally, in-process.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--prompt-template: Optional prompt template for online target generation. This flag accepts the string form only; use —spec or —spec-file for the other union form(s).--publication.intake.evaluation-id: Name of the existing Evaluation to publish under. Must already exist; the job does not create it.--publication.intake.agent-name: Agent name recorded on each published trajectory. Derived from the target when it names one; required otherwise.--publication.intake.agent-version: Agent version recorded on each published trajectory. Neither a Model nor an Agent carries a version, so this defaults to ‘unknown’ unless the submitter supplies one.--publication.intake.required: Fail the job when publication fails. Defaults to True so a run that asked to publish does not report success with nothing in Experiments. The result bundle is saved before publication runs, so a failed job still leaves the results intact to re-publish. Set False to keep the job successful and report the failure in its output instead.--publication.intake.test-case-id-field: Dataset column identifying each row, recorded as the published test case id. Defaults to a hash of the row’s content, which keeps a row’s id stable across dataset reorderings and revisions so the same test case can be compared run over run; editing a row makes it a new test case. Name a column here when the dataset has a real identifier — it reads better and survives content edits. Values must be unique per row; the run is rejected if they are not.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Job Spec flags are generated from the EvaluateInputSpec Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: dataset, params, target, prompt_template (other union forms), field_mapping.input, field_mapping.output, field_mapping.context, field_mapping.reference, field_mapping.trajectory, field_mapping.messages, field_mapping.tool_calls, field_mapping.tools, field_mapping.custom, metrics. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo evaluator evaluate submit
Submit to a cluster.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--prompt-template: Optional prompt template for online target generation. This flag accepts the string form only; use —spec or —spec-file for the other union form(s).--publication.intake.evaluation-id: Name of the existing Evaluation to publish under. Must already exist; the job does not create it.--publication.intake.agent-name: Agent name recorded on each published trajectory. Derived from the target when it names one; required otherwise.--publication.intake.agent-version: Agent version recorded on each published trajectory. Neither a Model nor an Agent carries a version, so this defaults to ‘unknown’ unless the submitter supplies one.--publication.intake.required: Fail the job when publication fails. Defaults to True so a run that asked to publish does not report success with nothing in Experiments. The result bundle is saved before publication runs, so a failed job still leaves the results intact to re-publish. Set False to keep the job successful and report the failure in its output instead.--publication.intake.test-case-id-field: Dataset column identifying each row, recorded as the published test case id. Defaults to a hash of the row’s content, which keeps a row’s id stable across dataset reorderings and revisions so the same test case can be compared run over run; editing a row makes it a new test case. Name a column here when the dataset has a real identifier — it reads better and survives content edits. Values must be unique per row; the run is rejected if they are not.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the EvaluateInputSpec Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: dataset, params, target, prompt_template (other union forms), field_mapping.input, field_mapping.output, field_mapping.context, field_mapping.reference, field_mapping.trajectory, field_mapping.messages, field_mapping.tool_calls, field_mapping.tools, field_mapping.custom, metrics. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo evaluator evaluate explain
Show input/output schemas.
Usage:
Options:
--profile: Annotate the bundle with this profile. Profile metadata lands in MR 1.4b.--cluster: Accepted for forward compatibility; unused in MR 1.4a.
Help:
--help, -h: Show this message and exit.
nemo customization
Plugin commands for customization.
Usage:
Help:
--install-completion: Install completion for the current shell.--show-completion: Show completion for the current shell, to copy it or customize the installation.--help, -h: Show this message and exit.
Commands:
automodel: Automodel training jobs (SFT, distillation).rl: NeMo-RL training on a Ray cluster: DPO (preference pairs)…unsloth: Unsloth GPU fine-tuning (container submit).
Jobs:
automodel.jobs: Automodel SFT, retrieval, and knowledge-distillation…rl.jobs: NeMo-RL DPO and GRPO training jobs on the platform…unsloth.jobs: Unsloth SFT (LoRA / full / merged) training jobs on the…
nemo customization automodel
Automodel training jobs (SFT, distillation).
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
explain: Show input/output schemas.submit
nemo customization automodel explain
Show input/output schemas.
Usage:
Options:
--profile: Annotate the bundle with this profile. Profile metadata lands in MR 1.4b.--cluster: Accepted for forward compatibility; unused in MR 1.4a.
Help:
--help, -h: Show this message and exit.
nemo customization automodel submit
Usage:
Arguments:
<JOB_JSON>: Path to Automodel job JSON (AutomodelJobInput schema).
Options:
--workspace, -w: Target workspace. [default: default]--profile--cluster--base-url: Override platform API host. If omitted: —cluster, then CLI context, then$NMP_BASE_URL, then http://localhost:8080.-o: Backend option override, ‘backend.key=value’. [default: ]--options-file <PATH>
Help:
--help, -h: Show this message and exit.
nemo customization rl
NeMo-RL training on a Ray cluster: DPO (preference pairs) or GRPO (NeMo Gym environment). Set training.type in the job JSON. Remote Kubernetes only.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
explain: Show input/output schemas.submit
nemo customization rl explain
Show input/output schemas.
Usage:
Options:
--profile: Annotate the bundle with this profile. Profile metadata lands in MR 1.4b.--cluster: Accepted for forward compatibility; unused in MR 1.4a.
Help:
--help, -h: Show this message and exit.
nemo customization rl submit
Usage:
Arguments:
<JOB_JSON>: Path to NeMo-RL job JSON (RlJobInput schema).
Options:
--workspace, -w: Target workspace. [default: default]--profile--cluster--base-url: Override platform API host. If omitted: —cluster, then CLI context, then$NMP_BASE_URL, then http://localhost:8080.-o: Backend option override, ‘backend.key=value’. [default: ]--options-file <PATH>
Help:
--help, -h: Show this message and exit.
nemo customization unsloth
Unsloth GPU fine-tuning (container submit). SFT only.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
explain: Show input/output schemas.submit
nemo customization unsloth explain
Show input/output schemas.
Usage:
Options:
--profile: Annotate the bundle with this profile. Profile metadata lands in MR 1.4b.--cluster: Accepted for forward compatibility; unused in MR 1.4a.
Help:
--help, -h: Show this message and exit.
nemo customization unsloth submit
Usage:
Arguments:
<JOB_JSON>: Path to Unsloth job JSON (UnslothJobInput schema).
Options:
--workspace, -w: Target workspace. [default: default]--profile--cluster--base-url: Override platform API host. If omitted: —cluster, then CLI context, then$NMP_BASE_URL, then http://localhost:8080.-o: Backend option override, ‘backend.key=value’. [default: ]--options-file <PATH>
Help:
--help, -h: Show this message and exit.
nemo customization automodel.jobs
Automodel SFT, retrieval, and knowledge-distillation training jobs.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
run: Run locally, in-process.submit: Submit to a cluster.explain: Show input/output schemas.
nemo customization automodel.jobs run
Run locally, in-process.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--name: Overridenamein the spec.--model: Overridemodelin the spec.--dataset.training: Training fileset as ‘name’ or ‘workspace/name’.--dataset.validation: Overridedataset.validationin the spec.--dataset.prompt-template: Overridedataset.prompt_templatein the spec.--training.lora.rank <INTEGER>: Overridetraining.lora.rankin the spec.--training.lora.alpha <INTEGER>: Overridetraining.lora.alphain the spec.--training.lora.dropout <FLOAT>: LoRA dropout probability for regularization.--training.lora.merge: Overridetraining.lora.mergein the spec.--training.lora.use-triton: Use the optimized Triton LoRA kernel.--training.max-seq-length <INTEGER>: Overridetraining.max_seq_lengthin the spec.--training.execution-profile: Overridetraining.execution_profilein the spec.--training.teacher-model: Overridetraining.teacher_modelin the spec.--training.distillation-ratio <FLOAT>: Overridetraining.distillation_ratioin the spec.--training.distillation-temperature <FLOAT>: Overridetraining.distillation_temperaturein the spec.--training.offload-teacher: Overridetraining.offload_teacherin the spec.--training.embedding.train-n-passages <INTEGER>: Overridetraining.embedding.train_n_passagesin the spec.--training.embedding.eval-negative-size <INTEGER>: Overridetraining.embedding.eval_negative_sizein the spec.--training.embedding.do-gradient-checkpointing: Overridetraining.embedding.do_gradient_checkpointingin the spec.--training.embedding.query-max-length <INTEGER>: Overridetraining.embedding.query_max_lengthin the spec.--training.embedding.passage-max-length <INTEGER>: Overridetraining.embedding.passage_max_lengthin the spec.--training.embedding.query-prefix: Collator-side prefix; BiEncoderCollator adds a space.--training.embedding.passage-prefix: Collator-side prefix; BiEncoderCollator adds a space.--schedule.epochs <INTEGER>: Overrideschedule.epochsin the spec.--schedule.max-steps <INTEGER>: Overrideschedule.max_stepsin the spec.--schedule.val-check-interval <FLOAT>: Overrideschedule.val_check_intervalin the spec.--schedule.seed <INTEGER>: Overrideschedule.seedin the spec.--schedule.progress-reporting.min-report-interval-seconds <FLOAT>: Least number of seconds between progress reports reaching the Jobs service. Every metric the training library logs is still recorded at full resolution — this only buffers them in memory and decides how often the accumulated set is sent, which is what the reporting actually costs. 0 sends a report for every step the library logs. Raising it reduces the time training spends blocked on reporting, at the cost of a progress bar and charts that update less often.--batch.global-batch-size <INTEGER>: Overridebatch.global_batch_sizein the spec.--batch.micro-batch-size <INTEGER>: Overridebatch.micro_batch_sizein the spec.--batch.sequence-packing: Overridebatch.sequence_packingin the spec.--batch.sequence-packing-max-samples <INTEGER>: Samples analyzed to estimate the optimal pack size when packing is enabled.--optimizer.learning-rate <FLOAT>: Overrideoptimizer.learning_ratein the spec.--optimizer.min-learning-rate <FLOAT>: Minimum learning rate for the cosine decay schedule.--optimizer.weight-decay <FLOAT>: Overrideoptimizer.weight_decayin the spec.--optimizer.adam-beta1 <FLOAT>: Adam optimizer beta1.--optimizer.adam-beta2 <FLOAT>: Adam optimizer beta2.--optimizer.warmup-steps <INTEGER>: Overrideoptimizer.warmup_stepsin the spec.--optimizer.adam-eps <FLOAT>: Adam/AdamW epsilon for numerical stability.--parallelism.num-nodes <INTEGER>: Overrideparallelism.num_nodesin the spec.--parallelism.num-gpus-per-node <INTEGER>: Overrideparallelism.num_gpus_per_nodein the spec.--parallelism.tensor-parallel-size <INTEGER>: Overrideparallelism.tensor_parallel_sizein the spec.--parallelism.pipeline-parallel-size <INTEGER>: Overrideparallelism.pipeline_parallel_sizein the spec.--parallelism.context-parallel-size <INTEGER>: Overrideparallelism.context_parallel_sizein the spec.--parallelism.expert-parallel-size <INTEGER>: Overrideparallelism.expert_parallel_sizein the spec.--parallelism.sequence-parallel: Enable sequence parallelism.--output.name: Overrideoutput.namein the spec.--output.description: Overrideoutput.descriptionin the spec.--integrations.wandb.project: W&B project name (groups related runs). Defaults to output.name if not set.--integrations.wandb.name: W&B run name. Defaults to job_id if not provided.--integrations.wandb.entity: W&B entity (team or username).--integrations.wandb.notes: W&B notes/description for the run.--integrations.wandb.base-url: Base URL for self-hosted W&B server (e.g., ‘https://wandb.mycompany.com’). If not provided, uses the default W&B cloud service.--integrations.wandb.api-key-secret.root: Reference to a secret. Format: ‘secret_name’ (uses request workspace) or ‘workspace/secret_name’ (explicit workspace).--integrations.mlflow.experiment-name: MLflow experiment name (groups related runs). Defaults to output.name if not set.--integrations.mlflow.name: MLflow run name. Defaults to job_id if not provided.--integrations.mlflow.description: MLflow run description.--integrations.mlflow.tracking-uri: MLflow tracking server URI (e.g., ‘http://mlflow.mycompany.com:5000’). Can also be set via MLFLOW_TRACKING_URI environment variable.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Job Spec flags are generated from the AutomodelJobInput Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: training.training_type, training.recipe, training.finetuning_type, training.lora.target_modules, training.lora.exclude_modules, training.precision, training.attn_implementation, training.teacher_precision, schedule.progress_reporting.time_series_metrics, optimizer.optimizer, optimizer.lr_decay_style, integrations.wandb.tags, integrations.mlflow.tags. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo customization automodel.jobs submit
Submit to a cluster.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--name: Overridenamein the spec.--model: Overridemodelin the spec.--dataset.training: Training fileset as ‘name’ or ‘workspace/name’.--dataset.validation: Overridedataset.validationin the spec.--dataset.prompt-template: Overridedataset.prompt_templatein the spec.--training.lora.rank <INTEGER>: Overridetraining.lora.rankin the spec.--training.lora.alpha <INTEGER>: Overridetraining.lora.alphain the spec.--training.lora.dropout <FLOAT>: LoRA dropout probability for regularization.--training.lora.merge: Overridetraining.lora.mergein the spec.--training.lora.use-triton: Use the optimized Triton LoRA kernel.--training.max-seq-length <INTEGER>: Overridetraining.max_seq_lengthin the spec.--training.execution-profile: Overridetraining.execution_profilein the spec.--training.teacher-model: Overridetraining.teacher_modelin the spec.--training.distillation-ratio <FLOAT>: Overridetraining.distillation_ratioin the spec.--training.distillation-temperature <FLOAT>: Overridetraining.distillation_temperaturein the spec.--training.offload-teacher: Overridetraining.offload_teacherin the spec.--training.embedding.train-n-passages <INTEGER>: Overridetraining.embedding.train_n_passagesin the spec.--training.embedding.eval-negative-size <INTEGER>: Overridetraining.embedding.eval_negative_sizein the spec.--training.embedding.do-gradient-checkpointing: Overridetraining.embedding.do_gradient_checkpointingin the spec.--training.embedding.query-max-length <INTEGER>: Overridetraining.embedding.query_max_lengthin the spec.--training.embedding.passage-max-length <INTEGER>: Overridetraining.embedding.passage_max_lengthin the spec.--training.embedding.query-prefix: Collator-side prefix; BiEncoderCollator adds a space.--training.embedding.passage-prefix: Collator-side prefix; BiEncoderCollator adds a space.--schedule.epochs <INTEGER>: Overrideschedule.epochsin the spec.--schedule.max-steps <INTEGER>: Overrideschedule.max_stepsin the spec.--schedule.val-check-interval <FLOAT>: Overrideschedule.val_check_intervalin the spec.--schedule.seed <INTEGER>: Overrideschedule.seedin the spec.--schedule.progress-reporting.min-report-interval-seconds <FLOAT>: Least number of seconds between progress reports reaching the Jobs service. Every metric the training library logs is still recorded at full resolution — this only buffers them in memory and decides how often the accumulated set is sent, which is what the reporting actually costs. 0 sends a report for every step the library logs. Raising it reduces the time training spends blocked on reporting, at the cost of a progress bar and charts that update less often.--batch.global-batch-size <INTEGER>: Overridebatch.global_batch_sizein the spec.--batch.micro-batch-size <INTEGER>: Overridebatch.micro_batch_sizein the spec.--batch.sequence-packing: Overridebatch.sequence_packingin the spec.--batch.sequence-packing-max-samples <INTEGER>: Samples analyzed to estimate the optimal pack size when packing is enabled.--optimizer.learning-rate <FLOAT>: Overrideoptimizer.learning_ratein the spec.--optimizer.min-learning-rate <FLOAT>: Minimum learning rate for the cosine decay schedule.--optimizer.weight-decay <FLOAT>: Overrideoptimizer.weight_decayin the spec.--optimizer.adam-beta1 <FLOAT>: Adam optimizer beta1.--optimizer.adam-beta2 <FLOAT>: Adam optimizer beta2.--optimizer.warmup-steps <INTEGER>: Overrideoptimizer.warmup_stepsin the spec.--optimizer.adam-eps <FLOAT>: Adam/AdamW epsilon for numerical stability.--parallelism.num-nodes <INTEGER>: Overrideparallelism.num_nodesin the spec.--parallelism.num-gpus-per-node <INTEGER>: Overrideparallelism.num_gpus_per_nodein the spec.--parallelism.tensor-parallel-size <INTEGER>: Overrideparallelism.tensor_parallel_sizein the spec.--parallelism.pipeline-parallel-size <INTEGER>: Overrideparallelism.pipeline_parallel_sizein the spec.--parallelism.context-parallel-size <INTEGER>: Overrideparallelism.context_parallel_sizein the spec.--parallelism.expert-parallel-size <INTEGER>: Overrideparallelism.expert_parallel_sizein the spec.--parallelism.sequence-parallel: Enable sequence parallelism.--output.name: Overrideoutput.namein the spec.--output.description: Overrideoutput.descriptionin the spec.--integrations.wandb.project: W&B project name (groups related runs). Defaults to output.name if not set.--integrations.wandb.name: W&B run name. Defaults to job_id if not provided.--integrations.wandb.entity: W&B entity (team or username).--integrations.wandb.notes: W&B notes/description for the run.--integrations.wandb.base-url: Base URL for self-hosted W&B server (e.g., ‘https://wandb.mycompany.com’). If not provided, uses the default W&B cloud service.--integrations.wandb.api-key-secret.root: Reference to a secret. Format: ‘secret_name’ (uses request workspace) or ‘workspace/secret_name’ (explicit workspace).--integrations.mlflow.experiment-name: MLflow experiment name (groups related runs). Defaults to output.name if not set.--integrations.mlflow.name: MLflow run name. Defaults to job_id if not provided.--integrations.mlflow.description: MLflow run description.--integrations.mlflow.tracking-uri: MLflow tracking server URI (e.g., ‘http://mlflow.mycompany.com:5000’). Can also be set via MLFLOW_TRACKING_URI environment variable.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the AutomodelJobInput Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: training.training_type, training.recipe, training.finetuning_type, training.lora.target_modules, training.lora.exclude_modules, training.precision, training.attn_implementation, training.teacher_precision, schedule.progress_reporting.time_series_metrics, optimizer.optimizer, optimizer.lr_decay_style, integrations.wandb.tags, integrations.mlflow.tags. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo customization automodel.jobs explain
Show input/output schemas.
Usage:
Options:
--profile: Annotate the bundle with this profile. Profile metadata lands in MR 1.4b.--cluster: Accepted for forward compatibility; unused in MR 1.4a.
Help:
--help, -h: Show this message and exit.
nemo customization rl.jobs
NeMo-RL DPO and GRPO training jobs on the platform Kubernetes GPU cluster (Ray).
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
run: Run locally, in-process.submit: Submit to a cluster.explain: Show input/output schemas.
nemo customization rl.jobs run
Run locally, in-process.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--name: Overridenamein the spec.--model: Model entity reference (‘name’ or ‘workspace/name’).--dataset: Dataset fileset reference. DPO: preference JSONL (training.jsonl + validation.jsonl). GRPO: Gym JSONL (training.jsonl required).--environment: Environment fileset reference (required when training.type is grpo).--integrations.wandb.project: W&B project name (groups related runs). Defaults to output.name if not set.--integrations.wandb.name: W&B run name. Defaults to job_id if not provided.--integrations.wandb.entity: W&B entity (team or username).--integrations.wandb.notes: W&B notes/description for the run.--integrations.wandb.base-url: Base URL for self-hosted W&B server (e.g., ‘https://wandb.mycompany.com’). If not provided, uses the default W&B cloud service.--integrations.wandb.api-key-secret.root: Reference to a secret. Format: ‘secret_name’ (uses request workspace) or ‘workspace/secret_name’ (explicit workspace).--integrations.mlflow.experiment-name: MLflow experiment name (groups related runs). Defaults to output.name if not set.--integrations.mlflow.name: MLflow run name. Defaults to job_id if not provided.--integrations.mlflow.description: MLflow run description.--integrations.mlflow.tracking-uri: MLflow tracking server URI (e.g., ‘http://mlflow.mycompany.com:5000’). Can also be set via MLFLOW_TRACKING_URI environment variable.--output.name: Overrideoutput.namein the spec.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Job Spec flags are generated from the RlJobInput Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: training, integrations.wandb.tags, integrations.mlflow.tags. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo customization rl.jobs submit
Submit to a cluster.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--name: Overridenamein the spec.--model: Model entity reference (‘name’ or ‘workspace/name’).--dataset: Dataset fileset reference. DPO: preference JSONL (training.jsonl + validation.jsonl). GRPO: Gym JSONL (training.jsonl required).--environment: Environment fileset reference (required when training.type is grpo).--integrations.wandb.project: W&B project name (groups related runs). Defaults to output.name if not set.--integrations.wandb.name: W&B run name. Defaults to job_id if not provided.--integrations.wandb.entity: W&B entity (team or username).--integrations.wandb.notes: W&B notes/description for the run.--integrations.wandb.base-url: Base URL for self-hosted W&B server (e.g., ‘https://wandb.mycompany.com’). If not provided, uses the default W&B cloud service.--integrations.wandb.api-key-secret.root: Reference to a secret. Format: ‘secret_name’ (uses request workspace) or ‘workspace/secret_name’ (explicit workspace).--integrations.mlflow.experiment-name: MLflow experiment name (groups related runs). Defaults to output.name if not set.--integrations.mlflow.name: MLflow run name. Defaults to job_id if not provided.--integrations.mlflow.description: MLflow run description.--integrations.mlflow.tracking-uri: MLflow tracking server URI (e.g., ‘http://mlflow.mycompany.com:5000’). Can also be set via MLFLOW_TRACKING_URI environment variable.--output.name: Overrideoutput.namein the spec.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the RlJobInput Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: training, integrations.wandb.tags, integrations.mlflow.tags. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo customization rl.jobs explain
Show input/output schemas.
Usage:
Options:
--profile: Annotate the bundle with this profile. Profile metadata lands in MR 1.4b.--cluster: Accepted for forward compatibility; unused in MR 1.4a.
Help:
--help, -h: Show this message and exit.
nemo customization unsloth.jobs
Unsloth SFT (LoRA / full / merged) training jobs on the platform GPU cluster.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
run: Run locally, in-process.submit: Submit to a cluster.explain: Show input/output schemas.
nemo customization unsloth.jobs run
Run locally, in-process.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--name: Overridenamein the spec.--model.name: Model entity reference. Accepts ‘name’ (uses the job’s workspace) or ‘workspace/name’. The plugin’s run resolves this to a local path before training.--model.max-seq-length <INTEGER>: Overridemodel.max_seq_lengthin the spec.--model.load-in-4bit: bitsandbytes 4-bit. Mutex with load_in_8bit. Default for Unsloth’s headline path.--model.load-in-8bit: Overridemodel.load_in_8bitin the spec.--model.trust-remote-code: Overridemodel.trust_remote_codein the spec.--dataset.path: Training fileset reference: ‘name’ (uses the job’s workspace) or ‘workspace/name’. Resolved to a local path by the plugin run.--dataset.text-field: Row field consumed by SFTTrainer.--dataset.apply-chat-template: If True, expects rows with a ‘messages’ field and applies tokenizer.apply_chat_template at training time.--dataset.validation-path: Optional validation fileset reference (same format as ‘path’). Downloaded under the same scheme.--dataset.packing: trl. SFTTrainer packing flag.--training.lora.rank <INTEGER>: LoRA rank.--training.lora.alpha <INTEGER>: LoRA scaling factor (alpha).--training.lora.dropout <FLOAT>: Overridetraining.lora.dropoutin the spec.--training.lora.use-rslora: Overridetraining.lora.use_rslorain the spec.--training.lora.random-state <INTEGER>: Overridetraining.lora.random_statein the spec.--training.lora.use-dora: DoRA (weight-decomposed LoRA). Improves quality at low ranks; adds training overhead.--training.lora.layers-to-transform <INTEGER>: Restrict LoRA to specific layer index(es). None applies to all layers. This flag accepts the integer form only; use —spec or —spec-file for the other union form(s).--training.lora.init-lora-weights: LoRA weight init scheme. True = PEFT default; ‘pissa’/‘olora’/‘loftq’ for advanced inits. This flag accepts the boolean form only; use —spec or —spec-file for the other union form(s).--schedule.epochs <INTEGER>: Overrideschedule.epochsin the spec.--schedule.max-steps <INTEGER>: Overrideschedule.max_stepsin the spec.--schedule.warmup-steps <INTEGER>: Overrideschedule.warmup_stepsin the spec.--schedule.warmup-ratio <FLOAT>: Overrideschedule.warmup_ratioin the spec.--schedule.logging-steps <INTEGER>: Overrideschedule.logging_stepsin the spec.--schedule.save-steps <INTEGER>: Overrideschedule.save_stepsin the spec.--schedule.eval-steps <INTEGER>: Overrideschedule.eval_stepsin the spec.--schedule.progress-reporting.min-report-interval-seconds <FLOAT>: Least number of seconds between progress reports reaching the Jobs service. Every metric the training library logs is still recorded at full resolution — this only buffers them in memory and decides how often the accumulated set is sent, which is what the reporting actually costs. 0 sends a report for every step the library logs. Raising it reduces the time training spends blocked on reporting, at the cost of a progress bar and charts that update less often.--schedule.seed <INTEGER>: Overrideschedule.seedin the spec.--batch.per-device-train-batch-size <INTEGER>: Overridebatch.per_device_train_batch_sizein the spec.--batch.gradient-accumulation-steps <INTEGER>: Overridebatch.gradient_accumulation_stepsin the spec.--optimizer.learning-rate <FLOAT>: Overrideoptimizer.learning_ratein the spec.--optimizer.weight-decay <FLOAT>: Overrideoptimizer.weight_decayin the spec.--optimizer.adam-beta1 <FLOAT>: Adam/AdamW beta1.--optimizer.adam-beta2 <FLOAT>: Adam/AdamW beta2.--optimizer.adam-epsilon <FLOAT>: Adam/AdamW epsilon for numerical stability.--optimizer.max-grad-norm <FLOAT>: Gradient-clipping max norm (TRL default 1.0).--optimizer.label-smoothing-factor <FLOAT>: Label smoothing for the cross-entropy loss. 0.0 disables.--optimizer.neftune-noise-alpha <FLOAT>: NEFTune embedding-noise alpha (quality boost). None disables.--hardware.gpus: Comma-separated GPU indices (‘0’ or ‘0,1’) for CUDA_VISIBLE_DEVICES. Selection, not reservation.--integrations.wandb.project: W&B project name (groups related runs). Defaults to output.name if not set.--integrations.wandb.name: W&B run name. Defaults to job_id if not provided.--integrations.wandb.entity: W&B entity (team or username).--integrations.wandb.notes: W&B notes/description for the run.--integrations.wandb.base-url: Base URL for self-hosted W&B server (e.g., ‘https://wandb.mycompany.com’). If not provided, uses the default W&B cloud service.--integrations.wandb.api-key-secret.root: Reference to a secret. Format: ‘secret_name’ (uses request workspace) or ‘workspace/secret_name’ (explicit workspace).--integrations.mlflow.experiment-name: MLflow experiment name (groups related runs). Defaults to output.name if not set.--integrations.mlflow.name: MLflow run name. Defaults to job_id if not provided.--integrations.mlflow.description: MLflow run description.--integrations.mlflow.tracking-uri: MLflow tracking server URI (e.g., ‘http://mlflow.mycompany.com:5000’). Can also be set via MLFLOW_TRACKING_URI environment variable.--output.name: Overrideoutput.namein the spec.--output.description: Overrideoutput.descriptionin the spec.--deployment-config: Deployment configuration for auto-deploying the model after training. Pass a string to reference an existing ModelDeploymentConfig by name (‘my-config’ or ‘workspace/my-config’). An object provides inline NIM deployment parameters. Omit to skip deployment. This flag accepts the string form only; use —spec or —spec-file for the other union form(s).
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Job Spec flags are generated from the UnslothJobInput Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: model.dtype, model.device_map, model.rope_scaling, training.training_type, training.finetuning_type, training.lora.target_modules, training.lora.bias, training.lora.loftq_config, training.lora.modules_to_save, training.lora.layers_to_transform (other union forms), training.lora.layer_replication, training.lora.init_lora_weights (other union forms), training.use_gradient_checkpointing, schedule.lr_scheduler_type, schedule.progress_reporting.time_series_metrics, schedule.lr_scheduler_kwargs, optimizer.optim, hardware.precision, integrations.wandb.tags, integrations.mlflow.tags, output.save_method, deployment_config (other union forms). Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo customization unsloth.jobs submit
Submit to a cluster.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--name: Overridenamein the spec.--model.name: Model entity reference. Accepts ‘name’ (uses the job’s workspace) or ‘workspace/name’. The plugin’s run resolves this to a local path before training.--model.max-seq-length <INTEGER>: Overridemodel.max_seq_lengthin the spec.--model.load-in-4bit: bitsandbytes 4-bit. Mutex with load_in_8bit. Default for Unsloth’s headline path.--model.load-in-8bit: Overridemodel.load_in_8bitin the spec.--model.trust-remote-code: Overridemodel.trust_remote_codein the spec.--dataset.path: Training fileset reference: ‘name’ (uses the job’s workspace) or ‘workspace/name’. Resolved to a local path by the plugin run.--dataset.text-field: Row field consumed by SFTTrainer.--dataset.apply-chat-template: If True, expects rows with a ‘messages’ field and applies tokenizer.apply_chat_template at training time.--dataset.validation-path: Optional validation fileset reference (same format as ‘path’). Downloaded under the same scheme.--dataset.packing: trl. SFTTrainer packing flag.--training.lora.rank <INTEGER>: LoRA rank.--training.lora.alpha <INTEGER>: LoRA scaling factor (alpha).--training.lora.dropout <FLOAT>: Overridetraining.lora.dropoutin the spec.--training.lora.use-rslora: Overridetraining.lora.use_rslorain the spec.--training.lora.random-state <INTEGER>: Overridetraining.lora.random_statein the spec.--training.lora.use-dora: DoRA (weight-decomposed LoRA). Improves quality at low ranks; adds training overhead.--training.lora.layers-to-transform <INTEGER>: Restrict LoRA to specific layer index(es). None applies to all layers. This flag accepts the integer form only; use —spec or —spec-file for the other union form(s).--training.lora.init-lora-weights: LoRA weight init scheme. True = PEFT default; ‘pissa’/‘olora’/‘loftq’ for advanced inits. This flag accepts the boolean form only; use —spec or —spec-file for the other union form(s).--schedule.epochs <INTEGER>: Overrideschedule.epochsin the spec.--schedule.max-steps <INTEGER>: Overrideschedule.max_stepsin the spec.--schedule.warmup-steps <INTEGER>: Overrideschedule.warmup_stepsin the spec.--schedule.warmup-ratio <FLOAT>: Overrideschedule.warmup_ratioin the spec.--schedule.logging-steps <INTEGER>: Overrideschedule.logging_stepsin the spec.--schedule.save-steps <INTEGER>: Overrideschedule.save_stepsin the spec.--schedule.eval-steps <INTEGER>: Overrideschedule.eval_stepsin the spec.--schedule.progress-reporting.min-report-interval-seconds <FLOAT>: Least number of seconds between progress reports reaching the Jobs service. Every metric the training library logs is still recorded at full resolution — this only buffers them in memory and decides how often the accumulated set is sent, which is what the reporting actually costs. 0 sends a report for every step the library logs. Raising it reduces the time training spends blocked on reporting, at the cost of a progress bar and charts that update less often.--schedule.seed <INTEGER>: Overrideschedule.seedin the spec.--batch.per-device-train-batch-size <INTEGER>: Overridebatch.per_device_train_batch_sizein the spec.--batch.gradient-accumulation-steps <INTEGER>: Overridebatch.gradient_accumulation_stepsin the spec.--optimizer.learning-rate <FLOAT>: Overrideoptimizer.learning_ratein the spec.--optimizer.weight-decay <FLOAT>: Overrideoptimizer.weight_decayin the spec.--optimizer.adam-beta1 <FLOAT>: Adam/AdamW beta1.--optimizer.adam-beta2 <FLOAT>: Adam/AdamW beta2.--optimizer.adam-epsilon <FLOAT>: Adam/AdamW epsilon for numerical stability.--optimizer.max-grad-norm <FLOAT>: Gradient-clipping max norm (TRL default 1.0).--optimizer.label-smoothing-factor <FLOAT>: Label smoothing for the cross-entropy loss. 0.0 disables.--optimizer.neftune-noise-alpha <FLOAT>: NEFTune embedding-noise alpha (quality boost). None disables.--hardware.gpus: Comma-separated GPU indices (‘0’ or ‘0,1’) for CUDA_VISIBLE_DEVICES. Selection, not reservation.--integrations.wandb.project: W&B project name (groups related runs). Defaults to output.name if not set.--integrations.wandb.name: W&B run name. Defaults to job_id if not provided.--integrations.wandb.entity: W&B entity (team or username).--integrations.wandb.notes: W&B notes/description for the run.--integrations.wandb.base-url: Base URL for self-hosted W&B server (e.g., ‘https://wandb.mycompany.com’). If not provided, uses the default W&B cloud service.--integrations.wandb.api-key-secret.root: Reference to a secret. Format: ‘secret_name’ (uses request workspace) or ‘workspace/secret_name’ (explicit workspace).--integrations.mlflow.experiment-name: MLflow experiment name (groups related runs). Defaults to output.name if not set.--integrations.mlflow.name: MLflow run name. Defaults to job_id if not provided.--integrations.mlflow.description: MLflow run description.--integrations.mlflow.tracking-uri: MLflow tracking server URI (e.g., ‘http://mlflow.mycompany.com:5000’). Can also be set via MLFLOW_TRACKING_URI environment variable.--output.name: Overrideoutput.namein the spec.--output.description: Overrideoutput.descriptionin the spec.--deployment-config: Deployment configuration for auto-deploying the model after training. Pass a string to reference an existing ModelDeploymentConfig by name (‘my-config’ or ‘workspace/my-config’). An object provides inline NIM deployment parameters. Omit to skip deployment. This flag accepts the string form only; use —spec or —spec-file for the other union form(s).
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the UnslothJobInput Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: model.dtype, model.device_map, model.rope_scaling, training.training_type, training.finetuning_type, training.lora.target_modules, training.lora.bias, training.lora.loftq_config, training.lora.modules_to_save, training.lora.layers_to_transform (other union forms), training.lora.layer_replication, training.lora.init_lora_weights (other union forms), training.use_gradient_checkpointing, schedule.lr_scheduler_type, schedule.progress_reporting.time_series_metrics, schedule.lr_scheduler_kwargs, optimizer.optim, hardware.precision, integrations.wandb.tags, integrations.mlflow.tags, output.save_method, deployment_config (other union forms). Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo customization unsloth.jobs explain
Show input/output schemas.
Usage:
Options:
--profile: Annotate the bundle with this profile. Profile metadata lands in MR 1.4b.--cluster: Accepted for forward compatibility; unused in MR 1.4a.
Help:
--help, -h: Show this message and exit.
nemo insights
Plugin commands for insights.
Usage:
Help:
--install-completion: Install completion for the current shell.--show-completion: Show completion for the current shell, to copy it or customize the installation.--help, -h: Show this message and exit.
Commands:
analysis: Manage periodic agent analysis opt-in state.analysis-runs: Submit and inspect on-demand analysis runs.
Jobs:
analyze-job: Run the insights analyst once for a single agent.
nemo insights analysis
Manage periodic agent analysis opt-in state.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
enable: Enable periodic analysis for an agent.disable: Disable periodic analysis for an agent.status: Show periodic analysis opt-in state.
nemo insights analysis enable
Enable periodic analysis for an agent.
Usage:
Options:
--agent: Name of the agent to opt in to periodic analysis.--workspace: Workspace the agent belongs to. [default: default]--base-url: Base URL of the running NMP instance. [default: http://localhost:8080] [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo insights analysis disable
Disable periodic analysis for an agent.
Usage:
Options:
--agent: Name of the agent to opt out of periodic analysis.--workspace: Workspace the agent belongs to. [default: default]--base-url: Base URL of the running NMP instance. [default: http://localhost:8080] [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo insights analysis status
Show periodic analysis opt-in state.
Usage:
Options:
--agent: Optional agent name. Omit to list all analysis configs.--workspace: Workspace to inspect. [default: default]--base-url: Base URL of the running NMP instance. [default: http://localhost:8080] [env: NMP_BASE_URL=]
Help:
--help, -h: Show this message and exit.
nemo insights analysis-runs
Submit and inspect on-demand analysis runs.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Submit an analysis run for an agent.list: List analysis runs.get: Get one analysis run, joined with the live state of its…
nemo insights analysis-runs create
Submit an analysis run for an agent.
The run is backed by an agents.execute job that shares its name. With —wait, exits non-zero if that job does not complete.
Usage:
Options:
--agent: Name of the agent whose telemetry should be analyzed.--workspace: Workspace the agent belongs to. [default: default]--base-url: Base URL of the running NMP instance. [default: http://localhost:8080] [env: NMP_BASE_URL=]--default-model: Model Entity ref for analysis work. Default: the configured default model.--fast-model: Model Entity ref for context summarization. Default: the configured fast model.--since: ISO-8601 lower bound enforced on the analyst’s trace/span reads.--ethos <PATH>: Path to the agent’s Ethos Markdown. Its contents are sent with the run.--evaluation-id: Restrict the run to spans from one evaluation.--timeout-seconds <FLOAT>: Timeout applied to the backing execute-agent job.--wait: Poll the run until its backing job reaches a terminal state.--poll-timeout <FLOAT>: How long —wait polls before giving up. [default: 900.0]--poll-interval <FLOAT>: Seconds between —wait polls. [default: 5.0]
Help:
--help, -h: Show this message and exit.
nemo insights analysis-runs list
List analysis runs. Job state is not joined — read one run to get it.
Usage:
Options:
--agent: Only list runs that analyzed this agent.--workspace: Workspace to inspect. [default: default]--base-url: Base URL of the running NMP instance. [default: http://localhost:8080] [env: NMP_BASE_URL=]--page <INTEGER>: Page number (1-indexed). [default: 1]--page-size <INTEGER>: Items per page. [default: 20]--sort: Sort field; prefix with ’-’ for descending. [default: -created_at]
Help:
--help, -h: Show this message and exit.
nemo insights analysis-runs get
Get one analysis run, joined with the live state of its backing job.
A null job means submission never landed: no job exists under the run’s name, and the run can be resubmitted.
Usage:
Arguments:
<NAME>: Name of the analysis run.
Options:
--workspace: Workspace the run belongs to. [default: default]--base-url: Base URL of the running NMP instance. [default: http://localhost:8080] [env: NMP_BASE_URL=]--wait: Poll until the run’s backing job reaches a terminal state.--poll-timeout <FLOAT>: How long —wait polls before giving up. [default: 900.0]--poll-interval <FLOAT>: Seconds between —wait polls. [default: 5.0]
Help:
--help, -h: Show this message and exit.
nemo insights analyze-job
Run the insights analyst once for a single agent.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
run: Run locally, in-process.submit: Submit to a cluster.explain: Show input/output schemas.
nemo insights analyze-job run
Run locally, in-process.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--agent: Agent under test.--ethos: Optional Ethos Markdown for the agent under test.--base-url: Optional platform base URL. Unset uses the active platform context.--insights-output: Optional local YAML path mirroring the Insights the platform stored. Container-local unless it points at mounted storage.--update-analysis-config: Update the matching AnalysisRunStatus with run metadata.--default-model: Workspace-qualified default Model Entity ID selected during setup.--fast-model: Workspace-qualified fast Model Entity ID selected during setup.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Job Spec flags are generated from the AnalyzeSpec Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: since. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo insights analyze-job submit
Submit to a cluster.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--agent: Agent under test.--ethos: Optional Ethos Markdown for the agent under test.--insights-output: Optional local YAML path mirroring the Insights the platform stored. Container-local unless it points at mounted storage.--update-analysis-config: Update the matching AnalysisRunStatus with run metadata.--default-model: Workspace-qualified default Model Entity ID selected during setup.--fast-model: Workspace-qualified fast Model Entity ID selected during setup.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the AnalyzeSpec Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: base_url, since. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo insights analyze-job explain
Show input/output schemas.
Usage:
Options:
--profile: Annotate the bundle with this profile. Profile metadata lands in MR 1.4b.--cluster: Accepted for forward compatibility; unused in MR 1.4a.
Help:
--help, -h: Show this message and exit.
nemo safe-synthesizer
Plugin commands for safe-synthesizer.
Usage:
Help:
--install-completion: Install completion for the current shell.--show-completion: Show completion for the current shell, to copy it or customize the installation.--help, -h: Show this message and exit.
Commands:
Jobs:
generate: Generate synthetic data using Safe Synthesizer.
nemo safe-synthesizer generate
Generate synthetic data using Safe Synthesizer.
Usage:
Help:
--help, -h: Show this message and exit.
Job Spec:
--data-source: The data source for the job.--config.data.group-training-examples-by: Column to group training examples by. This is useful when you want the model to learn inter-record correlations for a given grouping of records.--config.data.order-training-examples-by: Column to order training examples by. This is useful when you want the model to learn sequential relationships for a given ordering of records. If you provide this parameter, you must also providegroup_training_examples_by.--config.data.max-sequences-per-example <INTEGER>: If specified, adds at most this number of sequences per example. Supports ‘auto’, which resolves to 1 when differential privacy is enabled, None for time-series mode (each example fills the context window), and 10 otherwise. If set to None, fills up the context window. Required for DP to limit contribution of each example. This flag accepts the integer form only; use —spec or —spec-file for the other union form(s).--config.data.holdout <FLOAT>: Amount of records to hold out for evaluation. If this is a float between 0 and 1, that ratio of records is held out. If an integer greater than 1, that number of records is held out. If the value is equal to zero, no holdout will be performed. Must be >= 0.--config.data.max-holdout <INTEGER>: Maximum number of records to hold out. Overrides any behavior set byholdout. Must be >= 0.--config.data.random-state <INTEGER>: Random state for holdout split to ensure reproducibility.--config.evaluation.mia-enabled: Enable membership inference attack evaluation for privacy assessment.--config.evaluation.aia-enabled: Enable attribute inference attack evaluation for privacy assessment.--config.evaluation.sqs-report-columns <INTEGER>: Number of columns to include in statistical quality reports.--config.evaluation.sqs-report-rows <INTEGER>: Number of rows to include in statistical quality reports.--config.evaluation.mandatory-columns <INTEGER>: Number of mandatory columns that must be used in evaluation.--config.evaluation.enabled: Enable or disable evaluation.--config.evaluation.quasi-identifier-count <INTEGER>: Number of quasi-identifiers to sample for privacy attacks.--config.evaluation.pii-replay-enabled: Enable PII Replay detection.--config.training.num-input-records-to-sample <INTEGER>: Number of records the model will see during training. This parameter is a proxy for training time. For example, if its value is the same size as the input dataset, this is like training for a single epoch. If its value is larger, this is like training for multiple (possibly fractional) epochs. If its value is smaller, this is like training for a fraction of an epoch. Supports ‘auto’ where a reasonable value is chosen based on other config params and data. This flag accepts the integer form only; use —spec or —spec-file for the other union form(s).--config.training.batch-size <INTEGER>: The batch size per device for training. Must be >= 1.--config.training.gradient-accumulation-steps <INTEGER>: Number of update steps to accumulate the gradients for, before performing a backward/update pass. This technique increases the effective batch size that will fit into GPU memory. Must be >= 1.--config.training.weight-decay <FLOAT>: The weight decay to apply to all layers except all bias and LayerNorm weights in the AdamW optimizer. Must be in (0, 1).--config.training.warmup-ratio <FLOAT>: Ratio of total training steps used for a linear warmup from 0 to the learning rate. Must be > 0.--config.training.lr-scheduler: The scheduler type to use. See the HuggingFace documentation ofSchedulerTypefor all possible values.--config.training.learning-rate <FLOAT>: The initial learning rate forAdamWoptimizer. Must be in (0, 1). Setting to ‘auto’ uses a model-specific default if one exists. This flag accepts the number form only; use —spec or —spec-file for the other union form(s).--config.training.lora-r <INTEGER>: The rank of the LoRA update matrices. Lower rank results in smaller update matrices with fewer trainable parameters. Must be > 0.--config.training.lora-alpha-over-r <FLOAT>: The ratio of the LoRA scaling factor (alpha) to the LoRA rank. Empirically, this parameter works well when set to 0.5, 1, or 2. Must be in [0.5, 3].--config.training.rope-scaling-factor <INTEGER>: Scale the base LLM’s context length by this factor using RoPE scaling. Must be >= 1 or ‘auto’. This flag accepts the integer form only; use —spec or —spec-file for the other union form(s).--config.training.validation-ratio <FLOAT>: The fraction of the training data used for validation. Must be in [0, 1]. If set to 0, no validation will be performed. If set larger than 0, validation loss will be computed and reported throughout training.--config.training.validation-steps <INTEGER>: The number of steps between validation checks for the HF Trainer arguments. Must be > 0.--config.training.pretrained-model: Pretrained model to use for fine-tuning. Defaults to SmolLM3. May be a Hugging Face model ID (loaded from the Hugging Face Hub or cache) or a local path. See security note in docs before using untrusted sources.--config.training.quantize-model: Whether to quantize the model during training. This can reduce memory usage and potentially speed up training, but may also impact model accuracy.--config.training.quantization-scheme: Quantization scheme to use whenquantize_model=True. Acceptsbnb-4bit,bnb-8bit,fp8,nvfp4, ormxfp4. If unset, falls back toquantization_bitsfor backward compatibility. Non-bitsandbytes schemes are incompatible withpeft_implementation='loftq'.--config.training.peft-implementation: The PEFT (Parameter-Efficient Fine-Tuning) implementation to use. Options: ‘lora’ for Low-Rank Adaptation, ‘QLORA’ for Quantized LoRA.--config.training.max-vram-fraction <FLOAT>: The fraction of the total VRAM to use for training. Modify this to allow longer sequences. Must be in [0, 1].--config.training.attn-implementation: The attention implementation to use for model loading. Default uses ‘sdpa’ (PyTorch scaled dot product attention) for broad compatibility. Other common values: ‘flash_attention_2’ (requires flash-attn pip package), ‘flash_attention_3’ (requires flash-attn-3 support), ‘eager’ (standard PyTorch). Custom HuggingFace Kernels Hub paths (e.g. ‘kernels-community/flash-attn2’) are also supported.--config.generation.num-records <INTEGER>: Number of records to generate.--config.generation.temperature <FLOAT>: Sampling temperature for controlling randomness (higher = more random).--config.generation.repetition-penalty <FLOAT>: The value used to control the likelihood of the model repeating the same token. Must be > 0.--config.generation.top-p <FLOAT>: Nucleus sampling probability for token selection. Must be in (0, 1].--config.generation.patience <INTEGER>: Number of consecutive generations where theinvalid_fraction_thresholdis reached before stopping generation. Must be >= 1.--config.generation.invalid-fraction-threshold <FLOAT>: The fraction of invalid records that will stop generation after thepatiencelimit is reached. Must be in [0, 1].--config.generation.structured-generation.enabled: Whether to use structured generation for better format control.--config.generation.structured-generation.use-single-sequence: Whether to use a regex that matches exactly one sequence or record ifmax_sequences_per_exampleis 1.--config.generation.enforce-timeseries-fidelity: Enforce time-series fidelity by enforcing order, intervals, start and end times of the records.--config.generation.validation.group-by-accept-no-delineator: Whether to accept completions without both beginning and end of sequence delineators as a single sequence.--config.generation.validation.group-by-ignore-invalid-records: Whether to ignore invalid records in a sequence and proceed with the valid records.--config.generation.validation.group-by-fix-non-unique-value: Whether to automatically fix non-unique group-by values in a sequence by using the first unique value for all records.--config.generation.validation.group-by-fix-unordered-records: Whether to automatically fix unordered records in a sequence by sorting the records.--config.generation.attention-backend: The attention backend for the vLLM engine. Common values: ‘FLASHINFER’, ‘FLASH_ATTN’, ‘TRITON_ATTN’, ‘FLEX_ATTENTION’. IfNoneor ‘auto’, vLLM will auto-select the best available backend.--config.privacy.dp-enabled: Enable differentially-private training with DP-SGD.--config.privacy.epsilon <FLOAT>: Target privacy budget — lower values provide stronger privacy. Must be > 0.--config.privacy.delta <FLOAT>: Probability of accidentally leaking information. Should be much smaller than 1/n where n is the number of training records. Setting to ‘auto’ uses delta of 1/n^1.2. Must be in [0, 1) or ‘auto’. This flag accepts the number form only; use —spec or —spec-file for the other union form(s).--config.privacy.per-sample-max-grad-norm <FLOAT>: Maximum L2 norm for per-sample gradient clipping. Must be > 0.--config.time-series.is-timeseries: Whether to treat the dataset as time series. When enabled, eithertimestamp_columnortimestamp_interval_secondsis required. For grouped time series,group_training_examples_byneeds to be set.--config.time-series.timestamp-column: Name of the column containing timestamps used to order records whenis_timeseriesisTrue. Required only whenis_timeseriesisTrueandtimestamp_interval_secondsis not provided.--config.time-series.timestamp-interval-seconds <INTEGER>: Positive whole-number interval in seconds between timestamps. If not provided, the timestamp column will be used to infer the interval.--config.time-series.timestamp-format: Format of the timestamp column. Accepts either: (1) Python strftime format codes for string timestamps (e.g., ‘%Y-%m-%d %H:%M:%S’, ‘%m/%d/%Y’), or (2) ‘elapsed_seconds’ for numeric (int/float) timestamps representing seconds as an increasing counter (e.g., 0, 60, 120 for 1-minute intervals). If not provided, the format will be inferred from the data.--config.replace-pii.globals.seed <INTEGER>: Optional random seed.--config.replace-pii.globals.classify.enable-classify: Enable column classification.--config.replace-pii.globals.classify.num-samples <INTEGER>: Number of column values to sample for classification.--config.replace-pii.globals.classify.classify-model-provider: Name of the model provider in the Inference Gateway for column classification. The job compiler will resolve this to the appropriate endpoint URL.--config.replace-pii.globals.ner.ner-threshold <FLOAT>: NER model threshold.--config.replace-pii.globals.ner.enable-regexps: Enable NER regular expressions (experimental).--config.replace-pii.globals.ner.gliner.enable-gliner: Enable GLiNER NER module.--config.replace-pii.globals.ner.gliner.enable-batch-mode: Enable GLiNER batch mode.--config.replace-pii.globals.ner.gliner.batch-size <INTEGER>: GLiNER batch size.--config.replace-pii.globals.ner.gliner.chunk-length <INTEGER>: GLiNER batch chunk length in characters.--config.replace-pii.globals.ner.gliner.gliner-model: GLiNER model name.--config.emit-telemetry: Whether to emit anonymous Safe Synthesizer telemetry events. Defaults from NEMO_TELEMETRY_ENABLED when unset.--config.enable-synthesis: Whether to run LLM training and generation phases. When false the task only performs PII replacement and returns the processed data.--config.enable-replace-pii: Whether to run the default PII replacement pipeline before synthesis.--hf-token-secret: Name of platform secret containing the HuggingFace token. Must exist in the same workspace as the job.--pretrained-model-job: Optional previous NSS job whose stored adapter artifact is reused for generation-only synthesis. Accepts either ‘<job>’ in the current workspace or ‘<workspace>/<job>’. The plugin resolves the prior job’s ‘adapter’ result from Files.--enable-synthesis: Whether to run LLM training and generation phases. When False the task only performs PII replacement and returns the processed data.
Spec Source:
--spec: Spec as a JSON string. [default: {}]--spec-file <PATH>: Path to a YAML or JSON spec file (used as base; per-flag values override).
Submission:
-o: Backend option override, ‘backend.key=value’ (repeatable). [default: ]--options-file <PATH>: Path to a YAML or JSON options file (nested by backend name).--profile: Execution profile (operator-configured). Required unless the active cluster has a ‘default’ profile.--cluster: Configured cluster name to resolve via the NeMo CLI config.--base-url: Explicit plugin-service base URL (overrides —cluster and all other submit host resolution).--workspace: Workspace scope for the submission. [default: default]
Job Spec flags are generated from the SafeSynthesizerJobConfig Pydantic schema. Precedence: —spec-file (base) → —spec JSON (overlay) → per-flag values (top).
Some spec fields cannot be represented as CLI flags: config.data.max_sequences_per_example (other union forms), config.evaluation.pii_replay_entities, config.evaluation.pii_replay_columns, config.training.num_input_records_to_sample (other union forms), config.training.learning_rate (other union forms), config.training.lora_target_modules, config.training.rope_scaling_factor (other union forms), config.training.quantization_bits, config.generation.structured_generation.backend, config.generation.structured_generation.schema_method, config.privacy.delta (other union forms), config.time_series.start_timestamp, config.time_series.stop_timestamp, config.replace_pii.globals.locales, config.replace_pii.globals.classify.entities, config.replace_pii.globals.ner.ner_entities, config.replace_pii.globals.lock_columns, config.replace_pii.steps, config.preflight.disabled_checks. Supply them with —spec or —spec-file. Individual flags override values from the supplied spec.
nemo experiments
Manage experiments.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create Experimentdelete: Delete Experimentlist: List Experimentsget: Get Experimentupdate: Update Experiment
nemo experiments create
Create Experiment
Required fields: name
Examples:
Usage:
Arguments:
<NAME>: Workspace-unique experiment name.
Options:
--workspace--column-layout: A saved table layout for a group’s evaluations list: column order and which columns are hidden. Column ids are Studio’s and cannot be enumerated here — the table builds a column per evaluator and metadata key found in the rows — so ids are stored and echoed back unvalidated. Visibility is stored as the hidden ids rather than a map over every column, so a column that appears later (a new evaluator, a new metadata key) shows up by default. (JSON string)--default-sort: Default sort for this experiment’s evaluations list, as asort-param string: a comma-separated, ordered list of fields where the first is the primary sort and the rest break ties (leading ’-’ on a field = descending), e.g. ‘-evaluators.reward.mean,cost_usd.mean’. Defaults to ‘-created_at’. Accepts any field the evaluations listsortparam does; clients apply it as the listsortparam.--description: Human-readable purpose of the experiment.--insight-id: Reference to an external insight that seeded this experiment, if any.--is-favorite: Whether this Experiment is marked as a favorite. Defaults to false on create; omit on update to preserve the existing value.--metadata: Free-form producer metadata for the experiment. (JSON string)--pareto: Default X/Y metrics for a group’s cost-vs-accuracy Pareto view. Metric ids use the same vocabulary as the evaluations list sort/filter fields —cost_usd,latency_ms, orevaluators.<name>. Defaults to cost (x) vs latency (y): both exist for every group, so the chart always has something to render before anyone customizes it. (JSON string)--show-evaluations-over-time: Whether Studio should display this Experiment’s Evaluation results over time. Defaults to false on create; omit on update to preserve the existing value.--summary: Human- or agent-authored summary of the experiment’s findings.--exist-ok: Do not raise an error if the resource already exists. Returns the existing resource.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo experiments delete
Delete Experiment
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
nemo experiments list
List Experiments
Usage:
Options:
--workspace--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort <CHOICE>: Sort field; prefix with ’-’ for descending. [possible values: -created_at, created_at, -updated_at, updated_at, -name, name]--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: metadata: dict[str, str]
Filter experiments by name, insight_id, is_favorite, show_evaluations_over_time, baseline_evaluation_name, is_deleted, or a metadata key/value (filter[metadata.<key>]=<value>). Pass is_deleted=true to return only soft-deleted experiments; omit to see only live ones.
--filter.baseline-evaluation-name--filter.insight-id--filter.is-deleted--filter.is-favorite--filter.name--filter.show-evaluations-over-time
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo experiments get
Get Experiment
Usage:
Arguments:
<NAME>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo experiments update
Update Experiment
Required fields: body_name
Examples:
Usage:
Arguments:
<PATH_NAME>
Options:
--workspace--body-name: Workspace-unique experiment name.--baseline-evaluation-name: Name of this Experiment’s baseline Evaluation. The Evaluation must already be a live member of the Experiment. Set null to clear the selected baseline.--column-layout: A saved table layout for a group’s evaluations list: column order and which columns are hidden. Column ids are Studio’s and cannot be enumerated here — the table builds a column per evaluator and metadata key found in the rows — so ids are stored and echoed back unvalidated. Visibility is stored as the hidden ids rather than a map over every column, so a column that appears later (a new evaluator, a new metadata key) shows up by default. (JSON string)--default-sort: Default sort for this experiment’s evaluations list, as asort-param string: a comma-separated, ordered list of fields where the first is the primary sort and the rest break ties (leading ’-’ on a field = descending), e.g. ‘-evaluators.reward.mean,cost_usd.mean’. Defaults to ‘-created_at’. Accepts any field the evaluations listsortparam does; clients apply it as the listsortparam.--description: Human-readable purpose of the experiment.--insight-id: Reference to an external insight that seeded this experiment, if any.--is-favorite: Whether this Experiment is marked as a favorite. Defaults to false on create; omit on update to preserve the existing value.--metadata: Free-form producer metadata for the experiment. (JSON string)--pareto: Default X/Y metrics for a group’s cost-vs-accuracy Pareto view. Metric ids use the same vocabulary as the evaluations list sort/filter fields —cost_usd,latency_ms, orevaluators.<name>. Defaults to cost (x) vs latency (y): both exist for every group, so the chart always has something to render before anyone customizes it. (JSON string)--show-evaluations-over-time: Whether Studio should display this Experiment’s Evaluation results over time. Defaults to false on create; omit on update to preserve the existing value.--summary: Human- or agent-authored summary of the experiment’s findings.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo intake
Intake operations.
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
annotations: Manage annotationsevaluator-results: Manage evaluator_resultsingest: Ingest operationssessions: Manage sessionsspans: Manage spanstraces: Manage traces
nemo intake annotations
Manage annotations
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create annotations.delete: Delete Annotationlist: List Annotationsget: Get Annotation
nemo intake annotations create
Create annotations.
Required fields: kind, session_id
Examples:
Usage:
Arguments:
<NAME>
Options:
--workspace--kind <CHOICE>: (required) [possible values: feedback, note, metadata, label]--session-id: (required)--value--span-id--text--metadata: JSON string--value-type <CHOICE>: [possible values: text, numeric]--exist-ok
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo intake annotations delete
Delete Annotation
Usage:
Arguments:
<ANNOTATION_ID>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
nemo intake annotations list
List Annotations
Usage:
Options:
--workspace--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort <CHOICE>: [possible values: created_at, -created_at]--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: created_at: {gte: str, lte: str} value_numeric: {gte: float, lte: float}
Filter annotations by span_id, session_id, kind, name, created_by, and created_at range.
--filter.created-by--filter.kind--filter.name--filter.session-id--filter.span-id--filter.value-text
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo intake annotations get
Get Annotation
Usage:
Arguments:
<ANNOTATION_ID>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo intake evaluator-results
Manage evaluator_results
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Create Evaluator Resultlist: List Evaluator Resultsget: Get Evaluator Result
nemo intake evaluator-results create
Create Evaluator Result
Required fields: data_type, name, session_id, span_id
Examples:
Usage:
Arguments:
<NAME>: Evaluator / metric identity (e.g. ‘faithfulness/v1’).
Options:
--workspace--data-type <CHOICE>: Discriminator for which of value / string_value carries the payload. [possible values: NUMERIC, CATEGORICAL, BOOLEAN, TEXT]--session-id: Session id the target span belongs to. Denormalized so session-scoped reads stay fast.--span-id: Target span id. Not validated against existing spans (loose target policy).--comment: Free-text rationale or explanation.--string-value: String value. Required when data_type is CATEGORICAL or TEXT.--value <FLOAT>: Numeric value. Required when data_type is NUMERIC or BOOLEAN (0|1).--exist-ok: Do not raise an error if the resource already exists. Returns the existing resource.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo intake evaluator-results list
List Evaluator Results
Usage:
Options:
--workspace--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort <CHOICE>: [possible values: created_at, -created_at, value, -value]--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: created_at: {gte: str, lte: str} value: {gte: float, lte: float}
Filter evaluator results by span_id, session_id, name, data_type, created_by, value range, and created_at range.
--filter.created-by--filter.data-type--filter.name--filter.session-id--filter.span-id
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo intake evaluator-results get
Get Evaluator Result
Usage:
Arguments:
<EVALUATOR_RESULT_ID>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo intake ingest
Ingest operations
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
atif: Manage atifchat-completions: Manage chat_completionsspans: Manage spans
nemo intake ingest atif
Manage atif
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Ingest Atif
nemo intake ingest atif create
Ingest Atif
Required fields: agent, schema_version
Examples:
Usage:
Options:
--workspace--agent: JSON string--schema-version <CHOICE>: (required) [possible values: ATIF-v1.0, ATIF-v1.1, ATIF-v1.2, ATIF-v1.3, ATIF-v1.4, ATIF-v1.5, ATIF-v1.6, ATIF-v1.7]--continued-trajectory-ref--evaluation-context: Identifies the Evaluation and optional test case associated with ingested telemetry. (JSON string)--extra: JSON string--final-metrics: JSON string--notes--session-id--steps: JSON string--subagent-trajectories: JSON string--trajectory-id
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo intake ingest chat-completions
Manage chat_completions
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Ingest Chat Completion
nemo intake ingest chat-completions create
Ingest Chat Completion
Required fields: request, response
Examples:
Usage:
Options:
--workspace--request: Flexible captured chat-completions request. (JSON string)--response: Flexible captured chat-completions response. (JSON string)--cost-details: Additional estimated cost breakdown fields in USD. (JSON string)--cost-input-usd <FLOAT>: Estimated input-token cost of this model call in USD.--cost-output-usd <FLOAT>: Estimated output-token cost of this model call in USD.--cost-usd <FLOAT>: Total estimated cost of this model call in USD. This matches ATIF step metrics; Intake stores it as semantic cost_total_usd on spans.--evaluation-context: Identifies the Evaluation and optional test case associated with ingested telemetry. (JSON string)--provider--session-id: Groups related chat-completions calls without forcing them into the same trace.--trace-id: Opt into joining an existing trace built via OTel or ATIF. This is not a grouping mechanism for chat-completions calls; use session_id to group related calls.
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo intake ingest spans
Manage spans
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
create: Ingest Spans
nemo intake ingest spans create
Ingest Spans
Required fields: source, spans
Examples:
Usage:
Options:
--workspace--source: Stable name for the source trace store, such aslangsmithormlflow.--spans: JSON string
Help:
--help, -h: Show this message and exit.
Input Options:
--input-file: Path to JSON file (use ’-’ for stdin)--input-data: Input data for the request (JSON or YAML)
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo intake sessions
Manage sessions
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
get: Get Session
nemo intake sessions get
Get Session
Usage:
Arguments:
<ID>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo intake spans
Manage spans
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
list: List Spansget: Get Spanevaluator-results: Manage evaluator_resultsgroups: Manage groups
nemo intake spans list
List Spans
Usage:
Options:
--workspace--mode <CHOICE>: Response mode. summary omits payloads and raw attributes; preview includes input and output truncated to 300 characters; detailed returns full payloads and raw attributes. [possible values: summary, preview, detailed]--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort <CHOICE>: [possible values: started_at, -started_at]--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: started_at: {gte: str, lte: str}
Filter spans by session_id, trace_id, parent_span_id, project, evaluation_name, test_case_name, source, kind, status, model, tool_name, provider, agent_id, agent_name, and started_at. Every field takes one exact value, except started_at, which takes gte and lte.
--filter.agent-id--filter.agent-name--filter.evaluation-id--filter.evaluation-name--filter.kind--filter.model--filter.parent-span-id--filter.project--filter.provider--filter.session-id--filter.source--filter.status--filter.test-case-id--filter.test-case-name--filter.tool-name--filter.trace-id
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo intake spans get
Get Span
Usage:
Arguments:
<SPAN_ID>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo intake spans evaluator-results
Manage evaluator_results
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
list: List Evaluator Results For Span
nemo intake spans evaluator-results list
List Evaluator Results For Span
Usage:
Arguments:
<SPAN_ID>
Options:
--workspace
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo intake spans groups
Manage groups
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
list: List Span Groups
nemo intake spans groups list
List Span Groups
Usage:
Options:
--workspace--by: Comma-separated span fields to group by, e.g.--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort <CHOICE>: Sort groups by size or by start time. Use -started_at for the traces or sessions that began most recently, which answers ‘what ran lately’ in one call instead of paging through spans. A group’s time is its earliest matching span, so this orders by when work started and not by when it was last active. [possible values: span_count, -span_count, started_at, -started_at]--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: started_at: {gte: str, lte: str}
Filter spans by the same fields as the span list endpoint, then group matching spans by the comma-separated fields in the by query parameter.
--filter.agent-id--filter.agent-name--filter.evaluation-id--filter.evaluation-name--filter.kind--filter.model--filter.parent-span-id--filter.project--filter.provider--filter.session-id--filter.source--filter.status--filter.test-case-id--filter.test-case-name--filter.tool-name--filter.trace-id
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo intake traces
Manage traces
Usage:
Help:
--help, -h: Show this message and exit.
Commands:
get-metrics: Get Trace Metricslist: List Tracesget: Get Trace
nemo intake traces get-metrics
Get Trace Metrics
Usage:
Options:
--workspace--bucket <CHOICE>: Time bucket granularity. [possible values: total, hour, day, week, month]--timezone: IANA timezone the buckets are aligned to, e.g. America/Los_Angeles.
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: started_at: {gte: str, lte: str}
Filter the traces the metrics are computed over. Accepts the same fields as the traces list, so agent_name scopes the rollup to one agent. Without a started_at lower bound the rollup covers the last 7 days.
--filter.id--filter.agent-name--filter.evaluation-id--filter.evaluation-name--filter.session-id--filter.status--filter.test-case-id--filter.test-case-name
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]
nemo intake traces list
List Traces
Usage:
Options:
--workspace--mode <CHOICE>: Response mode. summary returns root-span fields without payloads or rollups; preview adds token, cost, and span-count rollups plus 300-character input/output previews; detailed returns rollups and full payloads. [possible values: summary, preview, detailed]--page <INTEGER>: Page number.--page-size <INTEGER>: Page size.--sort <CHOICE>: [possible values: started_at, -started_at]--all-pages: Fetch all pages
Filter Options:
--filter FILTER_JSON: Use —filter with JSON for complex/nested queries, or —filter. FIELD options for simple fields. Both can be combined, with field options taking precedence. JSON-only fields: started_at: {gte: str, lte: str}
Filter root-span-backed traces by id, session_id, root status, root span started_at, evaluation_name, test_case_name, and agent_name.
--filter.id--filter.agent-name--filter.evaluation-id--filter.evaluation-name--filter.session-id--filter.status--filter.test-case-id--filter.test-case-name
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for the list of results. [possible values: table, json, yaml, markdown, csv, raw, code]--no-truncate: Don’t truncate long values in table/markdown/csv output.--output-columns, -c: Columns to display: ‘default’, ‘all’, or comma-separated names. Only affects table/csv/markdown formats.--stream: Emit newline-delimited JSON, one record per line. Requires JSON or raw output.
nemo intake traces get
Get Trace
Usage:
Arguments:
<ID>
Options:
--workspace--mode <CHOICE>: Response mode. [possible values: summary, preview, detailed]
Help:
--help, -h: Show this message and exit.
Output Options:
--output-format, --output, -f <CHOICE>: Output format for an entity. [possible values: json, yaml, raw, code]