Skip to content

Preview/agent sandbox rl - #2314

Open
Hyperion-shuo wants to merge 4 commits into
InternLM:agent-sandboox-rl-previewfrom
Hyperion-shuo:preview/agent-sandbox-rl
Open

Hyperion-shuo wants to merge 4 commits into
InternLM:agent-sandboox-rl-previewfrom
Hyperion-shuo:preview/agent-sandbox-rl

Conversation

@Hyperion-shuo

Copy link
Copy Markdown
Contributor

Summary

Fix several issues encountered during asynchronous agentic RL training:

  • Normalize null and missing message content before rendering trace keys, so trace insertion and lookup produce the same text.
  • Raise the SessionServer request-body limit from 1 MiB to 16 MiB for long-context requests, with a warning when the limit is exceeded.
  • Track PID and exit-code observation failures independently. Treat unavailable observations as unknown rather than a missing exit code, and bound recovery and diagnostic history.
  • Make the sampler seed configurable while retaining the default value of 1.

Validation

  • DeepSWE and process-monitor tests: 41 passed.
  • Session template and trace tests: 27 passed, plus 6 subtests.
  • HTTP proxy-chain check: a 2 MiB request passed through Lagent and SessionServer; 17 MiB requests were rejected at either endpoint.

The HTTP chain check used the companion Lagent change raising its request-body limit to 16 MiB.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant