[AUTOREVERT] tune entry point for cli and event bridge, enabling providing custom running configurations

[jeanschmidt]

Changes lambda entry point for autorevert to enable better handling eventbridge parameters.

The changes are:

• Moved entry point from a global main to specific ones for CLI and lambda;
• Moved main execution logic to a reused function;
• Created a general config holder interface dataclass;

this enables to use the lambda with different configs from eventbridge like:

resource "aws_cloudwatch_event_rule" "pytorch_auto_revert_5m" {
  name                = "pytorch-auto-revert-5m"
  description         = "Every 5 minutes with config A"
  schedule_expression = "rate(5 minutes)"
}

resource "aws_cloudwatch_event_target" "pytorch_auto_revert_5m_target" {
  rule      = aws_cloudwatch_event_rule.pytorch_auto_revert_5m.name
  target_id = "pytorch-auto-revert-5m"
  arn       = aws_lambda_function.pytorch_auto_revert.arn

  input = jsonencode({
    mode      = "hud"
    dry_run   = false
    overrides = {
      bisection_limit    = 10
      clickhouse_database = "default"
    }
  })
}

Key reasonings on the decisions:

• created a single class for parsed configs that is distinct from default configs, still allowing validating and maintaining readability and single use (one is default, another is actual);
• Reused entry points from lambda and cli, so, mostly if not all, can be fully replicated via cli in case of firefight or debugging;
• broken entry point for cli and lambda: one parses opts and dotfiles, another parse eventbridge. Separation of concerns and easier to read;
• removed possibility of running without any parameters in CLI to replicate lambda - there will be multiple configurations, so, this does not make sense anymore;

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Preview	Updated (UTC)
torchci	Ignored	Preview	Dec 2, 2025 11:18pm

[izaitsevfb]

i think config.subcommand can never be None now?

[izaitsevfb]

Suggested change

	"SECRET_STORE_NAME", "pytorch-autorevert-secrets"
	"SECRET_STORE_NAME", ""

I'd suggest setting it to empty by default to make local CLI backward compatible.

Otherwise, if you try to run something like

python -m pytorch_auto_revert --dry-run autorevert-checker pull --hours 24  --hud-html

without setting --secret-store-name "" or SECRET_STORE_NAME=""
it would throw ERROR:root:Failed to retrieve secrets from AWS Secrets Manager which is confusing.