NERSC Ollama Manager

For step-by-step CLI and TUI workflows, see QUICK_START.md.

Run personal Ollama servers in Perlmutter interactive allocations. Includes a CLI, an optional Textual dashboard, SSH tunnel reuse, model selection, and Codex launch. Requires Python 3.10+, a Python environment accessible on compute nodes, Slurm, SSH access to allocated compute nodes, and a separately installed Codex executable. This package targets any Perlmutter user, not arbitrary Slurm clusters.

Install into your environment

From a source checkout:

python -m pip install '.[tui]'
nersc-ollama tui

A built wheel can likewise be installed with python -m pip install /path/to/wheel. The distribution name is nersc-ollama-manager; the command is nersc-ollama. PyPI publication is pending. No environment manager is required. Activate your preferred environment before launching; the package never activates another one.

First launch and configuration

The first normal command creates a private per-user JSON configuration. Help and --version create nothing. Resolution order is:

1. --config PATH
2. NERSC_OLLAMA_CONFIG
3. $XDG_CONFIG_HOME/nersc-ollama/config.json
4. ~/.config/nersc-ollama/config.json

A missing explicitly selected file is an error; create it with setup. Relative XDG environment values are ignored. Runtime defaults to $XDG_DATA_HOME/nersc-ollama or ~/.local/share/nersc-ollama; Ollama storage is its ollama_root subdirectory. First-run configuration starts no downloads, servers, or allocations. Review available storage before downloading large models.

The TUI offers storage fields and an explicit Install Ollama action when Ollama is missing. Existing installations open the dashboard directly. CLI equivalent:

nersc-ollama setup --runtime /path/to/personal/runtime --root /path/to/personal/ollama
nersc-ollama setup --version 0.34.0
nersc-ollama doctor

Storage changes through setup/onboarding are allowed only before installation or server creation. Existing installations are never automatically moved. Releases are safely extracted under the root and activated with a current symlink. Recorded SHA-256 values provide provenance, not independent signature validation. Models are downloaded separately into ollama_root/models.

Already have an Ollama on NERSC some other way (a module, a shared install) and don't want this tool managing its own pinned copy? Point at it instead of installing:

nersc-ollama setup --ollama-binary /path/to/existing/ollama

(the TUI has the equivalent field in onboarding, next to Install Ollama). This can be changed any time, unlike --runtime/--root — it never moves or orphans files, just which binary this tool runs. setup --version (and the TUI's Install Ollama) refuse while ollama_binary is set, to avoid managing two Ollamas at once; remove it from your configuration first to switch back to a self-managed install.

Allocate and connect

Keep an allocation terminal open:

nersc-ollama allocate --profile cpu --name cpu
# GPU alternative: supply your actual charging account
nersc-ollama allocate --profile gpu --account YOUR_ACCOUNT_g --name gpu

In another login-node terminal using the same configuration:

nersc-ollama status
nersc-ollama models --server gpu pull YOUR_MODEL_TAG
nersc-ollama codex --server gpu --model YOUR_MODEL_TAG

CPU/GPU profiles default to one node, interactive QOS, 30 minutes, and a 65,536-token context. The GPU profile requests four GPUs and requires an account. CPU can use the Slurm default account. Edit profiles in your local JSON configuration. allocate --dry-run prints the command; allocation and stop require confirmation or explicit --yes authorization. No scheduler action occurs during installation.

The compute worker uses the Python interpreter running the manager, which must be accessible on compute nodes with the package installed. Codex resolves from the caller's PATH and inherits the environment and working directory. Arguments after -- pass through to Codex. Its own environment/security settings still apply. doctor shows the interpreter, resolved commands, config and storage paths.

The allocation terminal owns the server. Ctrl-C or terminal closure ends this foreground workflow. Codex exit leaves it running. In the TUI, Allocate suspends the dashboard; use another terminal to connect. There is no automatic renewal.

Status and cleanup

• status --json: scheduler/discovery state, unique IDs, remaining time.
• tunnel --server ID: verify or create a loopback SSH forward and print its URL.
• disconnect --server ID: close the owned tunnel, including after job expiry.
• stop --server ID: confirm before cancelling the allocation.
• logs ID: show recent logs, including for stopped servers.
• models list: list completed downloads in shared storage, without a server.
• models --server ID list: query the selected running server.
• models --server ID pull TAG: explicitly download through the selected server.
• forget ID / forget --all: remove an expired discovery record once Slurm confirms the allocation has ended (unlike the TUI's Dismiss session, this actually deletes the record, not just hides it from that dashboard).
• sessions: list ollama-* screen sessions on the login node (see allocate --screen below) — mainly useful while one is still queueing, before it has a discovery record.

Discovery uses private atomic records with heartbeats and scheduler ownership checks. SSH control sockets live in a private host-local temporary directory. SSH authentication and host-key trust must already be configured; this package does not read credentials or modify SSH settings. Models require outbound network access for pulls. Cloud tags are rejected. Codex models must advertise tool support; actual tool reliability and memory requirements depend on model and context size. Multiple servers are supported without load balancing; choose an ID if ambiguous.

From outside NERSC

Think of this as two roles, not two locations: the NERSC server side (Slurm, the allocated compute node, Ollama itself — always on NERSC) and the Codex client side (wherever you actually type nersc-ollama and run Codex — a NERSC login node, some other node, or a laptop; --remote is what tells the client side it isn't already sitting on NERSC and needs an SSH hop to get there). "Local" vs. "remote" describes that hop, not which machine matters more — the server side is what does the work either way.

status, tunnel, logs, models, codex, disconnect, stop, forget, and sessions can run from any client machine with SSH access to a NERSC login node, so Codex itself runs on the client side — editing your own files — while only Ollama inference happens on the NERSC server side. allocate (without --screen) and serve still need to run directly on NERSC.

setup can too, but doesn't have to: --remote setup provisions the NERSC-side configuration directly from the client, no manual SSH login first — nersc-ollama --remote --remote-host saul.nersc.gov setup --runtime ... --root ... runs setup on the login node over SSH, exactly as if you'd typed it there yourself (omit --runtime/--root/--version to let the remote side pick its own defaults). The client's own local config, needed either way, comes from its own local nersc-ollama setup --runtime ... --root .... Ollama itself only needs to be installed locally for models pull; the generated Codex model catalog is written locally regardless. No ~/.ssh/config editing is required — every SSH call is built explicitly from the fields below plus a valid sshproxy key.

nersc-ollama --remote status
nersc-ollama --remote tunnel --server gpu
nersc-ollama --remote codex --server gpu --model YOUR_MODEL_TAG

Codex itself is a plain local process once launched — to keep it running if you close your terminal, wrap it in your own screen/tmux session; no special support is needed: screen -S mine nersc-ollama --remote codex --server gpu --model YOUR_MODEL_TAG.

--remote alone uses remote_login_host from config; --remote --remote-host HOST overrides it for one invocation. (Two separate flags, not one optional-valued flag, so a bare --remote never gets confused with the required subcommand after it.)

Optional configuration (existing configs need no migration):

{
  "remote_login_host": "saul.nersc.gov",
  "remote_user": "your_nersc_username",
  "remote_identity": "~/.ssh/nersc",
  "remote_python": "/path/to/venv/bin/python",
  "remote_config": null
}

Prefer a specific login node over the round-robin perlmutter.nersc.gov alias to avoid a different host key on every connection. remote_user may differ from the local account name. remote_identity defaults to ~/.ssh/nersc (the sshproxy-issued key, ~24h lifetime — an auth failure over --remote usually means it needs renewing, not a code bug). Set remote_python to the interpreter on the NERSC side that has nersc-ollama-manager installed: a non-interactive ssh host command often does not get interactive-shell PATH/henv activation, so without it a bare nersc-ollama is tried and fails with a clear error naming this field. remote_config overrides --config on the remote side only; omit it to let the remote side resolve its own default.

Surviving a dropped connection

A dropped SSH connection sends SIGHUP to salloc's foreground process tree, which releases the allocation (see salloc(1)) — with or without --remote. allocate --screen avoids this by running detached in a screen session on the login node instead, immune to the terminal that started it going away:

nersc-ollama allocate --profile gpu --account YOUR_ACCOUNT_g --screen
nersc-ollama --remote --remote-host saul.nersc.gov allocate --profile gpu --account YOUR_ACCOUNT_g --screen

--remote allocate is otherwise rejected (a foreground salloc cannot run over one SSH round-trip); --screen is the one exception, since it starts the session and returns immediately rather than blocking. status, logs ID, and --remote status remain the primary way to check on a screen-detached allocation. nersc-ollama peek NAME (session name is ollama-<name>) additionally snapshots the session's currently visible output — a one-shot, human-facing "is salloc still queueing?" check, not a substitute for logs. nersc-ollama sessions lists every ollama-* screen session currently on the login node — there's no registry, just what screen -ls reports right now — mainly useful before a discovery record exists yet, or to rediscover a --name you forgot. There's deliberately no "kill session" action: killing one kills the salloc ... process inside it, releasing the allocation much like Ctrl-C would rather than a safe, no-consequence cleanup — use stop (scancel) to end an allocation deliberately instead.

Compatibility and releases

Version 0.2 ignores legacy python and henv fields without rewriting existing configuration. Existing paths, models, discovery records and external personal wrappers remain usable. The package no longer generates wrappers or accepts setup --henv; launch directly from your environment. Existing workers need no restart for this upgrade.

PYTHONPATH=src python -m unittest discover -s tests -v
python -m pip wheel --no-deps --no-build-isolation --wheel-dir dist .

Tests mock scheduler/SSH operations and run the dashboard headlessly. Live inference validation requires a separately approved allocation and selected model. Publication and license selection are separate release steps.

References: NERSC interactive jobs, Ollama Linux installation, Ollama with Codex.

Allocation time and download-only mode

The TUI allocation row includes an editable HH:MM:SS duration. Switching CPU/GPU profiles reloads that profile's default; edits apply to the next request without rewriting configuration or extending an existing job. Slurm enforces QOS limits.

Enter a model tag, select Download: DTN, and choose Download only. This is the default download backend: it uses SSH to an interactive NERSC data transfer node, starts a temporary loopback Ollama server solely for the pull, and stops it on completion, failure, or interruption. No Slurm allocation is requested; account and allocation-time fields do not apply. It publishes no inference discovery record or tunnel. Model data remains in the configured shared directory for GPU use later.

nersc-ollama download MODEL_TAG
nersc-ollama download MODEL_TAG --check  # DTN startup check; no model downloaded
nersc-ollama download MODEL_TAG --dry-run

The package source, Ollama binary, model storage and log paths must be accessible on the DTN. The worker is a standalone standard-library script using the DTN's Python 3.6+; it does not activate or import the login node's Python environment. SSH authentication and host-key trust use the user's normal SSH configuration. The remote PTY propagates Ctrl-C; disconnection stops the temporary worker. Re-run the same download to let Ollama reuse existing download data.

Optional configuration (existing configs need no migration):

"downloads": {
  "dtn_host": "dtn01.nersc.gov",
  "dtn_python": "python3"
}

Choose dtn01 through dtn04.nersc.gov. Transfer temporary files live alongside the model directory, never in DTN /tmp. Logs are stored under the runtime logs directory. No inference is performed on DTNs. See NERSC DTN restrictions.

For the earlier workflow, select Download: CPU job or run:

nersc-ollama download MODEL_TAG --backend cpu --time 01:00:00

CPU mode requests a temporary CPU allocation and releases it after the download. The TUI uses its displayed duration and, if the CPU profile is selected, its account field; otherwise it uses the configured CPU/default account. The exact Slurm request is displayed for confirmation. Neither backend automatically falls back to another.

Downloaded models in the dashboard

The model selector reads completed manifests from the configured shared model storage. It populates on startup, refreshes periodically and after downloads, and can be refreshed with Models. No allocation, Ollama server, SSH tunnel, or inference is needed for this list. The manager checks referenced blob sizes without reading or hashing large model files; partial or malformed manifests are omitted. The selected model is preserved across refreshes while it remains available.

To launch Codex, select a running server row and press Enter, select a downloaded model, and click Codex. The launcher then verifies the server and its model availability and establishes the tunnel. A visible downloaded model alone does not mean an allocation is running. With no allocation, model browsing and DTN downloads still work. Keep the allocation terminal open and use a second terminal for clients.

Startup timeout, live logs, and expired records

Compute-node Ollama startup now allows 300 seconds by default, including for existing configurations. Set top-level "startup_timeout": 600 in config.json to allow ten minutes. A profile's startup_timeout overrides that global value. Values are seconds, positive and at most 86400. This controls server readiness; it does not extend the Slurm allocation or change Ollama's model-load timeout.

After Slurm grants the allocation, Ollama logs stream to the allocation terminal and are also saved under the runtime logs directory. Startup errors include the log path. In a second dashboard, select the server record and click Logs to view recent saved output.

Select a row and click Dismiss session to hide it persistently from the dashboard. Jobs, records, logs and models are retained; no Slurm query is needed. To actually remove an expired record's file, use the CLI instead: nersc-ollama forget ID (or forget --all to sweep every currently-unavailable record at once); both refuse to touch anything Slurm still reports as running.

Codex model metadata

The launcher generates a private model catalog under runtime/codex-models for the exact selected model tag. It supplies that catalog through invocation-specific Codex configuration, along with the context limit and a 90% compaction threshold. Text/image capability comes from Ollama's model information. The context limit uses the server record's configured context (or the profile for older records), capped by the model's advertised maximum when available. New workers record their context at startup, so later configuration edits do not change that recorded value.

Catalogs are named by content hash to avoid conflicts between simultaneous sessions. Global Codex settings are not rewritten. Restart Codex through the manager to pick up the generated catalog; an existing Codex session is unaffected. This addresses the missing-model-metadata warning, not GPU discovery or cold-load latency. No model weights are downloaded or inference requests made to generate it.

In the TUI session table, use the arrow keys to highlight a session and press d (or click Dismiss session) to hide it persistently. Dismissal makes no Slurm request and leaves jobs, discovery records, logs, and models intact. Hidden sessions are skipped by dashboard refreshes, but remain available to CLI commands. Preferences are stored as individual files in <runtime>/dismissed-sessions/; removing a session's preference file restores it. The shortcut applies only while the table has focus.

To open a shell on a job node, highlight a live session row, then click SSH next to Codex. Type exit (or press Ctrl-D) to return to the dashboard. SSH uses your existing authentication and does not extend the allocation. Codex remains available as a separate action.

Server actions use the highlighted row; Enter is optional. For Codex, also choose a model from Downloaded model. An empty model selection shows a validation message. SSH does not require a model.

Codex directly on the allocated node

Install the package in your environment to put codex-local on PATH. On the job node, run codex-local -m MODEL, codex-local exec "your task", or codex-local resume --last. Arguments are forwarded unchanged to Codex. The resolver uses the local loopback server and supplies the same model metadata as the dashboard. It never starts an allocation, tunnel, or model download. Set NERSC_OLLAMA_CONFIG for a non-default manager config. The model defaults to NERSC_OLLAMA_MODEL, the config's codex_model, or the sole installed model. With multiple local servers, set NERSC_OLLAMA_SERVER to the exact server ID. The caller's working directory and environment are preserved.

The TUI's SSH button writes a unique private setup file under <runtime>/ssh-sessions/session-*.bash and runs bash --rcfile FILE -i on the selected node. It loads .bashrc, verifies the node, and defines codex-local with the selected configuration, server and model. All arguments are forwarded.

Each shell gets a private /tmp/nersc-codex.* directory on that node. The function sets CODEX_HOME, CODEX_SQLITE_HOME, and TMPDIR only for the launched command, avoiding shared-home SQLite and temporary-file locking problems. The existing ~/.codex database, configuration, credentials and plugins are not copied, changed or loaded as the user Codex home. Project configuration still applies. This is an isolated local-model client; user-home customizations are not automatically inherited.

Repeated codex-local calls in the same shell share history. Separate SSH shells have independent state; node-local history may disappear when the allocation ends. Setup files are retained for inspection, contain no credentials, and check the selected node before use. No shell startup file is modified.

nersc-ollama-tui is an installed shortcut for nersc-ollama tui. It also accepts global options, for example nersc-ollama-tui --config /path/to/config.json.

Release automation and PyPI setup are documented in RELEASE.md.

The SSH setup exports its session directory. The Python codex-local entry point validates the directory ownership, permissions, node, server and configuration before applying local Codex storage. This also works in child scripts and after henv -x activation; the interactive function is not required. If a bound session is missing or invalid, the launcher stops instead of falling back to shared home. After upgrading, open a new TUI SSH shell to receive the exported binding. Shell scripts should begin with #!/usr/bin/env bash.

Select the GPU profile and toggle Spread across GPUs to save the Ollama placement preference, or from the CLI, add --gpu-spread/--no-gpu-spread to allocate --profile gpu (only valid with the GPU profile; persisted the same way as the TUI switch, right after you confirm the Slurm request — a --dry-run never persists it, only previews the resulting command). New GPU workers set OLLAMA_SCHED_SPREAD from profiles.gpu.sched_spread in the manager configuration; no shell export is needed. When enabled, the Slurm step also uses --gpu-bind none so the single Ollama process can see the full GPU allocation. Turning it off permits Ollama to use one GPU when the model fits. Existing servers are unchanged. Without a saved preference, the environment variable remains the fallback. Spreading does not guarantee faster generation.