msmodelslim tune Command-Line API Document¶

1. Function Description¶

msmodelslim tune is the auto-tuning command. It reads a tuning YAML containing strategy and evaluation, runs the loop "generate a quantization configuration → quantize → evaluate accuracy", and iteratively searches, based on the evaluation results, for a quantization configuration that meets the accuracy expectation, until the strategy generator is exhausted, the --timeout expires, or the maximum number of iterations is reached. Results such as the quantized model and the evaluation history are written to --save_path; after tuning finishes, the final best practice can be saved to a custom best-practice repository.

Command boundaries: this command does not expand on how to write specific tuning strategies and evaluation configurations; for field descriptions, see Auto Tuning Configuration Reference Navigation, whose strategy templates reference the quantization task protocol described in Quantization Task Configuration Reference Navigation; for operational steps, see Auto Tuning Usage Guide.

2. Command Format¶

msmodelslim tune --model_path <model_path> --save_path <save_path> --config <config> [--model_type <model_type>] [--device <device>] [--device_id <id> ...] [--timeout <timeout>] [--trust_remote_code [<BOOL>]] [--log_level <level>] [-v] [-q]

Symbols:

Values in angle brackets must be replaced; parameters in square brackets are optional.
--model_path, --save_path, and --config are required parameters.
--config is the required path of the tuning YAML; it has the same name as the --config of the one-click quantization command but a different meaning (the one-click quantization --config is optional).
--timeout is passed as an integer number of seconds (for example, 7200 means 2 hours); the old duration-string spellings (such as 2H and 3D4H) are kept only for backward compatibility, and new usages should use seconds uniformly.
--trust_remote_code is a Boolean parameter: passing it without a value enables it (equivalent to true), and an explicit true/false may also be passed (case-insensitive; the legacy True/False spellings are also accepted).
This command has no positional parameters. This syntax summary describes the parameter structure and is not a copy-ready command; for copy-ready commands, see Usage Examples.

3. Parameter List¶

Parameter	Alias	Type	Passing Form	Required/Optional	Default	Value Range or Format	Description
`--model_type`	None	`string`	Single value	Optional	`default`	Model type name, such as `Qwen2.5-7B-Instruct` or `Qwen3-32B`	Type of the model to tune, which must match a name in the Foundation Model Support Matrix; when it is not passed, the default adapter `default` is used.
`--model_path`	None	`string`	Single value	Required	None	Original model weight directory (must exist and be readable)	Weight directory of the model to tune.
`--save_path`	None	`string`	Single value	Required	None	Output directory (must be writable)	Directory in which the tuning results, the quantized model, and the history records are saved.
`--config`	`-c`	`string`	Single value	Required	None	Path to a readable YAML file	Tuning configuration YAML, containing the `strategy` and `evaluation` fields.
`--device`	None	`string`	Single value	Optional	`npu`	`npu`, `cpu`	Device type to run on; specify multi-device indices with `--device_id`.
`--device_id`	None	`list`	One or more values, space-separated	Optional	None	A list of non-negative integers, such as `0` or `0 1 2 3`	Device indices, used to specify multiple NPU devices.
`--timeout`	None	`int` (seconds)	Single value	Optional	None (no limit)	An integer number of seconds, such as `7200` (2 hours) or `3600` (1 hour); the old duration-string spelling (such as `1D2H30M15S`, with the units in the fixed order D/H/M/S) is kept only for backward compatibility	Tuning wall-clock timeout; when the timeout is reached, the current tuning run stops.
`--trust_remote_code`	None	`bool`	Boolean parameter: passing it without a value enables it; an explicit `true`/`false` value may also be given	Optional	`false`	`true`/`false` (case-insensitive; the legacy `True`/`False` spellings are also accepted)	Whether to trust and execute custom Python code in the model directory; enable it only when the code source is confirmed trustworthy.
`--log_level`	None	`string`	Single value	Optional	`info`	`debug`, `info`, `warning`, `error`	Log level.
`-v` / `--verbose`	None	`bool`	Boolean switch without a value	Optional	Off	Enabled when passed	Increase the level of output detail (equivalent to `--log_level debug`).
`-q` / `--quiet`	None	`bool`	Boolean switch without a value	Optional	Off	Enabled when passed	Suppress non-error output (equivalent to `--log_level error`).
`-h` / `--help`	None	`bool`	Boolean switch without a value	Optional	Off	Enabled when passed	Displays the help information and exits (for example, `msmodelslim tune -h`).
`-V` / `--version`	None	`bool`	Boolean switch without a value	Optional	Off	Enabled when passed	Displays the version information and exits (a top-level global parameter that can be used after any subcommand, for example, `msmodelslim tune --version`).

4. Parameter Relationships¶

This command loads the tuning YAML through the required --config.
When --timeout is not set, the tuning wall-clock time is not limited; once it is set, the current tuning run stops when the timeout is reached.
--device accepts only npu/cpu, and multi-device indices are specified through --device_id.
When --model_type is not passed, the default value default is used.

5. Referenced Configurations¶

Related Parameter	Configuration Name	Reference Relationship	Configuration Document
`--config`	Auto tuning configuration	Loads `strategy` and `evaluation`	Auto Tuning Configuration Reference Navigation

6. Environment Variables¶

Environment Variable	Type	Required/Optional	Default	Value Range or Format	Description
`MSMODELSLIM_CUSTOM_PRACTICE_REPO`	`string`	Optional	None (not set)	Directory path	After tuning finishes, the final best practice is saved to this custom best-practice repository; when it is not set, the command only records a "saving is not supported" notice.
`MSMODELSLIM_LOG_LEVEL`	`string`	Optional	`INFO`	`INFO`, `DEBUG`	Set the log level; once set, logs at that level and above are printed.

7. Usage Examples¶

7.1 Minimal Tuning Command¶

Contains only the minimum parameters required to complete one auto-tuning run:

msmodelslim tune \
  --model_path ${MODEL_PATH} \
  --save_path ${SAVE_PATH} \
  --config ${CONFIG_PATH}

${MODEL_PATH} is the weight directory, ${SAVE_PATH} is the tuning result directory, and ${CONFIG_PATH} is the tuning YAML (containing strategy and evaluation). When --model_type is not specified, the default adapter default is used; specifying it explicitly is recommended.

7.2 Explicitly Specifying the Model Type, the Device, and the Timeout¶

msmodelslim tune \
  --model_type ${MODEL_TYPE} \
  --model_path ${MODEL_PATH} \
  --save_path ${SAVE_PATH} \
  --config ${CONFIG_PATH} \
  --timeout 7200 \
  --device npu \
  --device_id 0 1 2 3

--timeout set to 7200 means running for at most 7200 seconds (2 hours); --device_id set to 0 1 2 3 means using 4 NPU devices.

8. Exit Codes and Exception Handling¶

Normal operation raises no exception (including when the policy iteration completes or the maximum number of iterations is reached); the results are found in the quantized model, the results, and the history records under ${SAVE_PATH}. On failure, an exception is thrown. Common causes: --config YAML validation failure, an invalid --timeout format, and failure to load the model adapter.

9. Security Notes¶

--trust_remote_code true executes custom Python code in the model directory; enable it only when the source is confirmed trustworthy.
--save_path writes the tuning results and history records; make sure the directory is writable and that unexpected data will not be overwritten.