cellmap_flow.serving.launch

Starting an inference server for a model: the one launcher, and the command line it runs.

launch(model, data_path, ...) is what every launcher calls: cellmap_flow infer <type>, cellmap_flow yaml and add --run, the dashboard’s Models tab (its catalog, Hugging Face, zoo and Cellpose picks, Add a model, and a model ticked again) and the Training tab’s ranking. It applies the one data path and scale rule (config.yaml.resolve_data_path), falls back to the saved launcher settings for the queues and charge group, builds the command and starts the job (jobs.launch.start_hosts), which it returns.

The command is <SERVER_COMMAND> --model <entry> -d <data path>, where the entry is ModelConfig.launch_entry as JSON, which the server rebuilds as it rebuilds a YAML’s model entry, then --resample or --no-resample: always one of them, so a server reads the data as its launcher was told to (a YAML’s resample, infer --no-resample, the dashboard’s checkbox) whatever serve’s own default is.

The serve program is jobs.launch.SERVER_COMMAND, and deployments change it: the fileglancer deploy sets "pixi run cellmap_flow serve". So it is read each time a command is built, never copied at import (a copy misses the override), and split into words (quoted as one token, the shell looks for a program called “pixi run cellmap_flow serve”). cellmap_flow_server, the program before 0.3.0, takes --model too, so a deployment that still sets that works.

A model with an environment of its own (models.envs: its entry’s env, else its type’s default_env) is served from it instead: the entry’s env is taken off, and the program is that environment’s (pixi run --frozen --manifest-path <pixi.toml> -e <env> cellmap_flow serve, or <env>/bin/python -P -m cellmap_flow.cli.main serve) in place of SERVER_COMMAND, which names this deployment’s own.

The launchers submit it as a shell line, which jobs.spec.shell_join quotes: the entry’s JSON is full of double quotes and braces, and shlex.join would single-quote it, which LSF’s own quoting splits apart (jobs.spec.shell_quote says how).

Attributes

logger

Functions

`server_argv`(model_config, data_path[, resample]) → list	The server's argv for `model_config` reading `data_path`, resampling
`server_command`(model_config, data_path[, resample]) → str	`server_argv` as the shell line start_hosts takes.
`remembered`(queue, charge_group)	`(queue, charge_group)` for a command-line launch: the ones given,
`launch`(model, data_path, *[, queue, charge_group, ...])	Start an inference server for `model` on `data_path`, and return its job.

Module Contents

cellmap_flow.serving.launch.logger

cellmap_flow.serving.launch.server_argv(model_config, data_path: str, resample: bool = True) → list: The server’s argv for model_config reading data_path, resampling it to the model’s input voxel size when resample.

cellmap_flow.serving.launch.server_command(model_config, data_path: str, resample: bool = True) → str: server_argv as the shell line start_hosts takes.

cellmap_flow.serving.launch.remembered(queue, charge_group): (queue, charge_group) for a command-line launch: the ones given, else the saved launcher settings’; the ones given are saved for the next run and the dashboard. queue is -q’s text, one queue or several (queues.queue_list), and comes back a list.

cellmap_flow.serving.launch.launch(model, data_path: str, *, queue=None, charge_group=None, resample: bool = True, wait_for_host: bool = True)

Start an inference server for model on data_path, and return its job.

model is a model config, or a model entry (models.entry), which is built first (registry.build_model). The server reads data_path with the model’s scale applied (config.yaml.resolve_data_path), resampling it to the model’s voxel size when resample. queue, the GPU queues it may run on (a list; start_hosts tries them in the order queues.candidates gives), and charge_group default to the saved launcher settings (jobs.settings). The job is named after the model, and is returned once its server has announced its address, or at once when wait_for_host is off.

Raises:

ConfigError – model is an entry that does not describe a model.
JobStartError – no queue accepted the job, or its server never came up.