Skip to content

Doc build fix - #35

Merged
YannBerthelot merged 2 commits into
mainfrom
rl/baselines-harness
Sep 11, 2026
Merged

YannBerthelot merged 2 commits into
mainfrom
rl/baselines-harness

Conversation

@YannBerthelot

Copy link
Copy Markdown
Owner

No description provided.

YannBerthelot and others added 2 commits September 11, 2026 13:39
The changelog still listed both patrol variants as holding formation to roughly
139 m against a 60 m tolerance, pinned by six strict xfails. That was true
before the lead's turn rate was fed forward. The hardest case now settles at
2.4 m against a 3 m reward precision floor, both variants settle around 40 m at
the benchmark settings, and the strict xfail is gone.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012WdRJEyGAZwKDD1EFwwH3U
The docs-deploy render died on the runner with SIGTERM partway through the
nineteenth environment. The timings say why: 9 s per environment for the
aircraft, then 16, 19, 22, 24, 27, 29, 36, and 92 s for the cement kiln, with
hvac alone taking three and a half minutes. That climb is memory, not rendering.

It is not a matplotlib leak; render_kit.finish closes every figure. It is JAX:
each environment compiles its own step_env and XLA keeps those executables and
their device buffers alive for the life of the process, so twenty-one
environments accumulate twenty-one compilation caches.

A fresh process per environment releases all of it. Measured locally, peak
resident memory stays flat at 1.5-2.4 GB and per-environment time stays at
8-18 s with no climb, against a single process that reached 92 s and was killed.
The total comes in lower despite paying JAX start-up twenty-one times.

Each environment is wrapped in a ::group:: so the deploy log stays readable.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012WdRJEyGAZwKDD1EFwwH3U
@YannBerthelot
YannBerthelot merged commit 316fecd into main Sep 11, 2026
6 checks passed
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