ollama2llama

Import locally installed Ollama models into a YAML configuration consumable by llama-swap (a model hot‑swap / lifecycle manager). It scans your Ollama manifests and (a) generates a fresh config or (b) incrementally augments an existing one without clobbering custom edits unless you explicitly request overrides.

Installation

From source (latest main):

cargo install --path .        # in a cloned repo

From Git (exact revision):

cargo install --git https://github.com/Exotik850/ollama2llama --rev <commit>

Using cargo-binstall (prebuilt if available):

cargo binstall ollama2llama

Quick Start

Generate a config for all discovered models:

ollama2llama --all-models --output-config llama-swap.yaml

This scans $OLLAMA_MODELS (or ~/.ollama/models) and writes llama-swap.yaml with one entry per model.

Dry run (no file writes):

ollama2llama --all-models --dry-run

Update an existing config in place (retains prior per‑model tweaks unless you pass template overrides):

ollama2llama -i llama-swap.yaml --all-models

Import only selected models:

ollama2llama --specify-models llama3:8b,phi3:3.8b --output-config llama-swap.yaml

Add aliases & filters while importing:

ollama2llama \
  --specify-models llama3:8b \
  --alias llama3:8b=llama3|meta-llama3 \
  --filter llama3:8b=family:llama|quant:Q4_K_M \
  --output-config llama-swap.yaml

Command Templates

Default launch template: {model_path} (the resolved blob path).

Override launch & stop commands (placeholders are substituted literally):

--cmd-template "./serve --model {model_path} --name {model_name} --port ${PORT}" \
--stop-cmd-template "pkill -f {model_name}"

Placeholders:

• {model_path} – Absolute path to the primary model blob.
• {model_name} – The Ollama model identifier (e.g. llama3:8b).

Only when you pass --cmd-template / --stop-cmd-template are existing cmd / cmd_stop values overwritten; otherwise they are preserved (idempotent incremental updates).

CLI Options

Sample Output

health_check_timeout: 120
log_level: info
models:
  llama3:8b:
    cmd: /home/user/.ollama/models/blobs/sha256-<truncated>
    aliases:
      - llama3
    filters:
      family: llama
      quant: Q4_K_M
groups:
  imported:
    swap: true
    members:
      - llama3:8b

YAML Schema Reference

Top‑level keys:

• health_check_timeout – u64 seconds (default 120)
• log_level – debug | info | warn | error (default info)
• start_port – u16; starting value for a downstream ${PORT} macro (if used by llama-swap)
• macros – map<string,string> of user macros
• models – map<string, ModelConfig>
• groups – map<string, GroupConfig>

ModelConfig:

• cmd – required launch command
• env – array of KEY=VALUE strings (manually editable)
• cmd_stop – optional graceful stop command
• proxy – optional proxy URL
• aliases – alternative names
• check_endpoint – readiness probe endpoint
• ttl – inactivity timeout seconds
• use_model_name – upstream name override
• filters – arbitrary metadata key/value pairs
• unlisted – hide from listing endpoints

GroupConfig:

• swap – only one member active at a time
• exclusive – starting a member unloads models in other groups
• persistent – prevent other groups from unloading these models
• members – required list of model names

also see here for more details about the configuration of llama-swap

Macros

Add / override macros:

ollama2llama -M PORT=5800,HOST=0.0.0.0 --all-models -o llama-swap.yaml

They appear under macros:; identical keys in an existing file are replaced.

Aliases & Filters

Aliases use = then | separated names:

--alias llama3:8b=llama3|meta-llama3

Filters accept k:v or k=v inside the right-hand side, separated by |:

--filter llama3:8b=family:llama|quant:Q4_K_M

Single Group Mode

Create a simple swapping group (mutual exclusion):

ollama2llama --all-models --single-group -o llama-swap.yaml

Produces a group named imported (or your override) with swap: true.

Recommended Workflow

1. Generate initial config for all (or selected) models.
2. Manually refine per-model cmd, add env, check_endpoint, or ttl as needed.
3. Re-run later with --specify-models to import new additions; existing customized entries remain unless you supply overriding templates.
4. Use version control to track manual refinements.

Idempotency & Safety

• Only template‑flagged fields (cmd, cmd_stop) are overwritten when you explicitly pass their flags.
• --macro-override always overwrites keys of same name.
• --no-clobber ensures you never accidentally overwrite a pre-existing output file.
• Missing blob paths are reported (with --verbose) and those models are skipped.

Design Notes

The tool focuses on discovery + structured augmentation, leaving nuanced runtime parameters (environment, health endpoints) for manual curation so you keep full control while still automating the repetitive parts.

Potential Future Enhancements

These are reasonable, non-breaking extensions you may consider (PRs welcome):

1. Exclusion flags: --exclude-models or pattern / glob filtering.
2. Pattern selection: --match <regex> to import models by regex.
3. Per-model overrides via files: --alias-file, --filter-file (YAML/CSV).
4. Bulk env injection: --env llama3:8b=KEY=VAL|KEY2=VAL2 syntax.
5. Default field flags: --ttl 1800, --check-endpoint /health applied to newly imported models only.
6. Grouping strategies: --group-by family to auto-create groups from a filter key or from name prefixes.
7. Sorting / stable ordering for deterministic diffs.
8. --diff mode to show changes vs existing config without writing.
9. JSON output option (--format json).
10. Exit codes signaling “no changes” (CI-friendly) with --quiet.

Contributing

Issues & PRs are welcome. Please keep additions minimally invasive and documented. For feature ideas see the list above; feel free to open a discussion first.

License

This crate is licensed under the MIT License