Skip to content

Import, Update & Regenerate

Three project lifecycle commands that let you work with existing endpoints, incrementally update configuration, and upgrade generator versions without starting from scratch.

Overview

Command Use case
mcc import Generate a project from a colleague's running endpoint
mcc update Change one or two fields without regenerating everything
mcc regenerate Upgrade to a new generator version with saved parameters

mcc import

Generate an operational project from a running SageMaker endpoint — even one you didn't create.

When to use

  • A colleague deployed an endpoint and you need to run do/test, do/benchmark, or do/adapter
  • You're diagnosing a production issue and need the operational tooling
  • You want to extend an existing endpoint with benchmarking or monitoring

Usage

mcc import <endpoint-arn> [--output-dir <path>] [--region <region>] [--dry-run]

Flags

Flag Description
--output-dir <path> Output directory (default: ./<endpoint-name>)
--region <region> AWS region override (default: extracted from ARN)
--dry-run Print reconstructed config without writing files

What gets generated

Included: do/config, do/test, do/benchmark, do/deploy, do/clean, do/logs, do/status, do/register, do/optimize, do/run, do/ic/<name>.conf

Not included: Dockerfile, do/build, do/push, do/submit, buildspec.yml, code/ directory

The generated project is a "no-build" project — the container already exists in ECR.

Endpoint status

Import works on any endpoint status: InService, Updating, OutOfService, or Failed. The DescribeEndpointConfig and DescribeInferenceComponent APIs are always accessible regardless of endpoint health.

Example

# Import from a running endpoint
mcc import arn:aws:sagemaker:us-east-1:123456789012:endpoint/llama3-prod

# Preview what would be generated
mcc import arn:aws:sagemaker:us-east-1:123456789012:endpoint/llama3-prod --dry-run

# Import to a specific directory
mcc import arn:aws:sagemaker:us-east-1:123456789012:endpoint/llama3-prod --output-dir ./my-project

After import:

cd llama3-prod
./do/test          # Test the endpoint
./do/benchmark     # Run performance benchmarks
./do/status        # Check endpoint health
./do/logs          # View CloudWatch logs

mcc update

Change specific configuration fields and regenerate only the affected files. Customizations to unrelated files are preserved.

When to use

  • Changing instance type (e.g., scaling up from ml.g5.xlarge to ml.g5.2xlarge)
  • Updating environment variables
  • Toggling LoRA support
  • Any single-field change that doesn't require full regeneration

Usage

mcc update [--field <key=value>...] [--dry-run] [--no-register]

Flags

Flag Description
--field <key=value> Set a field non-interactively (repeatable)
--dry-run Show affected files without writing
--no-register Skip do/register after update

Interactive vs non-interactive

Without --field: Interactive spacebar multi-select of fields, then prompted for new values.

With --field: Direct key=value assignment — no prompts.

# Non-interactive: change instance type
mcc update --field instanceType=ml.g5.4xlarge

# Multiple fields at once
mcc update --field instanceType=ml.g5.4xlarge --field icGpuCount=4

# Preview what would change
mcc update --field instanceType=ml.g5.4xlarge --dry-run

Affected files

The update command uses a dependency map to determine which template files are affected by each parameter change:

Parameter Affected files
instanceType do/config, do/ic/default.conf
deploymentConfig do/config, do/ic/default.conf, Dockerfile, do/build
baseImage Dockerfile, do/build
region do/config
icGpuCount do/ic/default.conf
enableLora do/ic/default.conf, Dockerfile

Only these files are rewritten. All other files remain untouched.

Update log

Every update appends an entry to do/update.log with timestamp, changed fields, and affected files — useful for auditing configuration drift.

mcc regenerate

Re-run project generation from saved parameters using the current generator version. Use this after upgrading MCC to get the latest template improvements.

When to use

  • After npm update @aws/ml-container-creator — get latest template fixes
  • After a generator release with new do/ script features
  • To reset generated files to a clean state (with --force)

Usage

mcc regenerate [--force] [--dry-run] [--no-register]

Flags

Flag Description
--force Regenerate even if version already matches
--dry-run Show what would change without writing
--no-register Skip do/register after regeneration

How it works

  1. Reads .mlcc-generation-params.json (written at generation time)
  2. Merges with live overrides from do/config (your manual edits win)
  3. Merges IC sizing from do/ic/default.conf
  4. Backs up current generated files to .mlcc-backup/<timestamp>/
  5. Runs full writeProject() with current templates
  6. Updates .mlcc-version

Guard on imported projects

If a project was created via mcc import (has .mlcc-import-source but no .mlcc-generation-params.json), regeneration is blocked:

❌ This is an imported project — regeneration requires original generation parameters.
   Use 'mcc update' to change specific fields instead.

Use mcc update for field-level changes on imported projects.

Example

# Check if regeneration is needed (version comparison)
mcc regenerate --dry-run

# Force regeneration regardless of version
mcc regenerate --force

# Regenerate without auto-registering
mcc regenerate --force --no-register

.mlcc-generation-params.json

Written automatically by writeProject() after every successful generation. Contains:

  • generatorVersion — version of MCC that created the project
  • generatedAt — ISO timestamp
  • answers — full configuration (secrets redacted)
  • bootstrapProfile — which bootstrap profile was active
  • catalogVersions — framework and model registry versions used

Not committed to git — listed in .gitignore by default. This is local project state.

Sensitive values (hfToken, ngcToken) are replaced with [REDACTED]. ARN references (hfTokenArn, ngcTokenArn) are preserved since they're not secret.