CLI Reference¶
Use aegaeon or python -m aegaeon.
doctor¶
doctor checks the active Python environment and local host prerequisites. A configuration file adds model-path, profile, topology, and memory-budget checks.
aegaeon doctor
CUDA_VISIBLE_DEVICES=0 aegaeon doctor --check-cuda
| Option | Parser default | Meaning |
|---|---|---|
| --config | None | YAML file for local model, profile, topology, and memory checks |
| --models | None | Comma-separated startup models; requires --config |
| --tensor-parallel-size | None | Nonzero overrides YAML; 0 uses YAML; requires --config; the effective size must be 1 |
| --strict | false | Return a nonzero status for warnings as well as failures |
| --check-cuda | false | Run small PyTorch/CuPy CUDA operations, import vLLM's native extension, inspect visible GPUs, and test one-page host registration |
start¶
aegaeon start --config /srv/aegaeon/config.yaml --host 127.0.0.1 --port 8000 --ray-address 127.0.0.1:6789
| Option | Parser default | Meaning |
|---|---|---|
| --config | Required | YAML file |
| --host | 0.0.0.0 | Bind address |
| --port | 8000 | Port |
| --models | None | Comma-separated startup models; overrides startup_models; an empty string starts without models |
| --num-prefill-engines | 1 | Normal startup uses YAML ServerConfig |
| --num-decode-engines | 1 | Same |
| --num-engines | 0 | Same |
| --tensor-parallel-size | None | Explicit nonzero value overrides YAML; 0 falls back to YAML |
| --model-cache-size-gb | 20 | Fallback CPU weight capacity in GiB |
| --cpu-num-slabs | 15 | Fallback slab count |
| --cpu-slab-size-gb | 1 | Fallback size in GiB |
| --ray-address | None | Cluster address, ahead of environment variables |
Set topology and memory budgets in YAML configuration.
deploy¶
aegaeon deploy Qwen/Qwen3-4B --path /srv/models/Qwen3-4B --replicas 2
| Option | Default | Meaning |
|---|---|---|
| model | Required positional argument | Service model name |
| --path | None | Local weights; otherwise registered path, then download |
| --host | localhost | API host |
| --port | 8000 | API port |
| --max-model-len | None | Context override |
| --replicas | 1 | Minimum READY replica target for automatic placement |
| --nodes | None | Explicit node IDs; bypass automatic placement |
| --tool-parser | None | Supported vLLM tool-parser override for this model alias |
Partial operations exit with a nonzero status; see Model Management.
undeploy¶
aegaeon undeploy Qwen/Qwen3-4B --nodes node_1
| Option | Default | Meaning |
|---|---|---|
| model | Required positional argument | Service model name |
| --host | localhost | API host |
| --port | 8000 | API port |
| --nodes | All known replicas | Unload only the listed nodes |
Queries¶
Each command's help lists its options and parser defaults.
aegaeon --help
aegaeon doctor --help
aegaeon start --help
aegaeon deploy --help
aegaeon undeploy --help
For startup resources, YAML values and defaults take precedence over the legacy resource fallback options. --models and a nonzero --tensor-parallel-size override their YAML fields; --ray-address precedes AEGAEON_RAY_ADDRESS, RAY_ADDRESS, then auto.