Implementation:Pytorch Serve Marsgen

Overview

Pytorch_Serve_Marsgen is the MAR file generation script that automates downloading model artifacts, constructing torch-model-archiver commands, and producing Model Archive files for the TorchServe test and example ecosystem.

Source

Property	Value
File	ts_scripts/marsgen.py
Lines	199
Language	Python
Key Functions	delete_model_store_gen_dir, gen_mar, generate_model, generate_mars, model_archiver_command_builder

Key Functions

delete_model_store_gen_dir (lines 16–23)

def delete_model_store_gen_dir():
    ...

Removes the generated model store directory to ensure a clean slate before regenerating MAR files. This prevents stale artifacts from polluting new builds.

gen_mar(model_store) (lines 29–43)

def gen_mar(model_store):
    ...

Top-level entry point for MAR generation. Reads the mar_config.json configuration file, iterates over defined models, and delegates to generate_mars for batch processing.

generate_model(model, model_store_dir) (lines 46–101)

def generate_model(model, model_store_dir):
    ...

Handles the generation of a single model archive. This function:

• Downloads the serialized model file if a URL is specified
• Resolves handler, model file, and extra file paths
• Manages symlinks for shared artifacts
• Invokes model_archiver_command_builder to assemble the archiver command
• Executes the command via subprocess

generate_mars(mar_config, model_store_dir) (lines 104–129)

def generate_mars(mar_config, model_store_dir):
    ...

Batch generation driver. Iterates over all model entries in the parsed mar_config dictionary and calls generate_model for each one. Handles error aggregation and reporting.

model_archiver_command_builder(...) (lines 132–180)

def model_archiver_command_builder(
    model_name, version, serialized_file, handler,
    model_file, extra_files, export_path, ...
):
    ...

Constructs the full torch-model-archiver CLI command as a list of arguments. This function maps the model configuration fields to their corresponding CLI flags:

Parameter	CLI Flag
model_name	--model-name
version	--version
serialized_file	--serialized-file
handler	--handler
model_file	--model-file
extra_files	--extra-files
export_path	--export-path

The --force flag is always appended to allow overwriting existing archives during generation.

Configuration: mar_config.json

The script reads from a mar_config.json file that declares each model's metadata, download URLs, handler paths, and archiver options. This JSON-driven approach decouples model definitions from the generation logic.

Relationship to Principles

This implementation directly supports the Pytorch_Serve_Model_Archiving principle. While ModelArchiverConfig defines what parameters an archive requires, marsgen.py automates the how -- orchestrating downloads, path resolution, and CLI invocation to produce MAR files at scale.

Metadata

Pytorch_Serve Pytorch_Serve_Model_Archiving 2026-02-13 18:52 GMT