chore(deps): update Ray to 2.57.0

[praateekmahajan]

Summary

• Bump Ray/Ray Serve 2.56.1 → 2.57.0 and Dynamo 1.3.0.post1 → 1.3.1.
• Remove obsolete Serve client-cache, HAProxy, protobuf, and ray_dist.jar workarounds.
• Retarget the opt-in Ray Data diagnostics shim for Ray 2.57 and add low-overhead decision-state attribution.

Relevant upstream behavior

• Ray 2.57 lowers downstream-capacity backpressure from 10× to 2× (ray-project/ray#64352).
  • Ray now lets a fast stage queue about 2×, rather than 10×, what the next stage can process. This reduces object-store memory, but can create pipeline bubbles when downstream consumption is uneven or throttling causes upstream actors to scale down.
• Ray 2.57 optimizes ExecutionResources and reservation/budget scheduling loops without intentionally changing allocation semantics.
• Ray 2.57 fetches block metadata on a background thread by default, preventing per-block metadata ray.get calls from blocking the streaming scheduler. This is an expected optimization, not an attribution for the observed regressions.

Dynamo 1.3.1 has no relevant production-source change from 1.3.0.post1.

Verified

• Dependency sync, affected tests, lint, and live Ray Serve deploy/reconnect/shutdown passed.
• The final Ray 2.57 run emitted 46,072 valid Curator diagnostic events with no missing fields, invalid values, or installation errors.
• Diagnostics now report blocked durations and internal-vs-output object-store memory for admission/autoscaling decisions. These additions run only when a decision state changes; no per-block or metadata-fetch hot path is instrumented. A real Ray 2.57 pipeline test validates these fields in ray-data.log under Docker.
• Benchmark runs:
  • All 41 EOS benchmarks passed in the nightly-to-PR comparison.
  • The final-SHA full EOS run passed 39/41 requirements. Both ArXiv backends completed successfully with exit code 0; their count checks were stale because the shared input grew from 45 to 100 tar files.
  • The final diagnostics added no systematic runtime or storage regression: runtime changes did not correlate with diagnostic-event volume, and the full result grew only 0.11% versus the prior Ray 2.57 run.
  • The targeted regression rerun passed and reproduced the meaningful timing signals.
  • Image curation was 8–17s slower with the new 2× ratio. A Ray 2.57 ratio-10 control recovered baseline runtime (318.39s vs 321.07s), eliminated 154–170 downstream-capacity blocks and related actor-scaling decisions, and restored GPU utilization near baseline.
• Video captioning/embedding also encountered 2× backpressure but showed no meaningful, reproducible end-to-end regression. ALM, transcoding, modifier, and FLEURS were not attributed to this policy.

Follow-up

• Preserve Ray 2.57's upstream 2× default in this dependency update.
• Update #2295 to Ray 2.57 and document the measured 2×-versus-10× behavior, threaded metadata fetching, their scheduling/autoscaling interactions, and the limits of the available diagnostics.

Closes NMCUR-312

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[praateekmahajan]

/claude review
@greptileai review

[praateekmahajan]

/ok to test 3470da2

[github-actions]

🌿 Preview your docs: https://nvidia-preview-praateek-ray-2-57.docs.buildwithfern.com/nemo/curator

Here are the markdown pages you've updated:

• nemo/curator/curate-text/synthetic/inference-server
• nemo/curator/reference/infra/container-environments

[greptile-apps]

Greptile Summary

The PR updates Ray and Ray Serve to 2.57.0 and Dynamo to 1.3.1 while removing workarounds made obsolete by the upgraded dependencies.

• Retargets opt-in Ray Data scheduler diagnostics and adds decision-state duration and object-store attribution.
• Updates Ray Serve HAProxy discovery and shutdown behavior.
• Synchronizes dependency constraints, lock data, container behavior, tests, and documentation.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
nemo_curator/backends/ray_data/diagnostics.py	Retargets private scheduler hooks to Ray 2.57 and enriches change-oriented diagnostic events with blocked durations and object-store memory attribution.
nemo_curator/core/serve/ray_serve/backend.py	Removes the obsolete Serve client-cache workaround and relies on Ray 2.57's packaged HAProxy support.
nemo_curator/core/utils.py	Updates local-cluster HAProxy discovery and configures dedicated metrics and statistics ports when a supported binary source exists.
pyproject.toml	Raises Ray, Ray Serve, and Dynamo minimum versions and relaxes the obsolete protobuf upper bound.
docker/Dockerfile	Removes explicit ray-haproxy setup and the obsolete ray_dist.jar security workaround after the upstream dependency fix.
tests/backends/ray_data/test_diagnostics.py	Extends diagnostics validation to cover the newly emitted duration and object-store attribution fields.
uv.lock	Synchronizes the resolved environment with Ray 2.57.0, Dynamo 1.3.1, protobuf 6.33.5, and Ray's transitive HAProxy package.

_{Reviews (5): Last reviewed commit: "test(ray-data): exercise diagnostics wit..." | Re-trigger Greptile}

[praateekmahajan]

/claude review
@greptileai review

[praateekmahajan]

/ok to test ce29c6b

[praateekmahajan]

/claude review
@greptileai review

[praateekmahajan]

/ok to test 3f82a7a

[praateekmahajan]

/claude review
@greptileai review

[praateekmahajan]

/ok to test 985bafb