Repository navigation
refactor(code-gen): replace legacy Ray execution with sandboxes - #4958
Conversation
Signed-off-by: Igor Gitman <igitman@nvidia.com>
Signed-off-by: Igor Gitman <igitman@nvidia.com>
|
/review |
|
/ok to test 92ff1e0 |
Signed-off-by: Igor Gitman <igitman@nvidia.com>
|
Merged upstream main |
Signed-off-by: Igor Gitman <igitman@nvidia.com>
|
Merged upstream main 77 code-generation component tests pass (1 optional Docker test skipped), plus 28 Ray inventory/opt-out tests. The sandbox entrypoint remains independent of parent Ray, and patch targets follow main’s new module layout. The old terminal-shell failure was unrelated to this patch.
|
|
/ok to test 5a22e53 |
|
/review |
yaoyu-33
left a comment
There was a problem hiding this comment.
Thanks. A separate sandbox_app.py is the simplest way to get past the launcher's static Ray inventory. It also refuses to start without sandbox_provider, so a config slip can't quietly fall back to Ray. Deferring the Ray import to the Ray branch and declaring ReverifyMode.STATELESS both look right, and the 1362/1362 OpenSandbox regrade is good evidence.
One nit inline. CI is green apart from the Autobot check.
|
💬 The non-blocking harness check failed: Kilocode’s |
|
💬 The core/server tests and Exact-main reproduction remains unverified: main’s CI skipped the probes, and the managed Gym test profile doesn’t expose this conformance suite. That blocks a validated integration candidate. The next step is to incorporate the existing fix and validate conformance on the resulting head; rerunning the unchanged head wouldn’t address the missing code. |
Sounds good to me, no concerns from my end |
Signed-off-by: Igor Gitman <igitman@nvidia.com>
|
Updated to main Removed the old Ray execution branch completely, including its wrapper, unused batch-metric helpers and dependency. New-head CI is pending and may require qualified-maintainer trust authorization; local checks are not a claim that CI has passed. |
|
💬 I verified that |
|
/ok to test b799bc8 |
|
✅ CI is green on |
|
💬 Kilocode’s conformance check passed at |
|
🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/42113477993 |
|
The PR is no longer an open ready-to-merge PR targeting main. |
code_gen now verifies every candidate in a provider sandbox. Remove the legacy Ray dispatch, its batch-metric helpers, and the Ray dependency/lock entries; mark the default
app.pyentrypoint as non-Ray.sandbox_app.pyremains a compatibility alias to the same server, with no separate execution branch. This addresses the request to finish the migration from #4607; no separate issue is needed.Migration: existing code_gen configurations must supply a sandbox provider and
sandbox_config.image(the shipped config readsCODE_GEN_SANDBOX_IMAGE). Missing provider/image values fail validation. The README documents provider setup and the migration. Scoring, candidate harness and execution deadlines are unchanged; switching execution environments can change time-limit-sensitive grades.Declare verification stateless for
gym eval reverify: saved responses and test inputs fully specify grading. Both entrypoints import with Ray blocked, and launcher tests verify that neither requires or initializes parent Ray.Validation after merging main
2ac7908c4:pre-commit run --all-filesand diff whitespace checks passed; commits have DCO sign-offs.openai/openai/gpt-5.6-lunagenerated a stdin-sum program, which passed all three test inputs through/verifyand a real OpenSandbox sandbox (reward 1). An incorrect control scored 0. Both requests returned HTTP 200 without masks or failures; Ray was never imported. Inspected the generated code and verifier results. This is an integration smoke, not a benchmark score comparison.Earlier evidence for the unchanged sandbox verifier: a 9/9 model-generated canary; a 1362/1362 cached-answer regrade with exact input/response preservation, no masked rows, and one timeout changing to a pass; and a fresh 1362/1362 benchmark with no missing, duplicate, failed, masked, unscored or truncated rows. Five generations were inspected in each model run. These are earlier-revision results, not fresh full-benchmark results for this head.