Skip to content

chore(deps): update Ray to 2.58.0 - #2368

Draft
praateekmahajan wants to merge 3 commits into
NVIDIA-NeMo:mainfrom
praateekmahajan:codex/ray-2.58
Draft

chore(deps): update Ray to 2.58.0#2368
praateekmahajan wants to merge 3 commits into
NVIDIA-NeMo:mainfrom
praateekmahajan:codex/ray-2.58

Conversation

@praateekmahajan

@praateekmahajan praateekmahajan commented Sep 3, 2026

Copy link
Copy Markdown
Contributor

Summary

  • require and lock Ray 2.58.0 for Ray Data and Ray Serve
  • update Curator's CUDA 12.9 vLLM stack to 0.26.0, matching Ray 2.58's llm extra
  • align ai-dynamo 1.4.2, Transformers 5.5.3, TorchCodec 0.14, and PyNvVideoCodec 2.0.4
  • preserve Qwen-ASR compatibility with Transformers 5 and remove the obsolete Qwen-Omni multimodal workaround
  • update the Ray Data diagnostics shim for 2.58 output-pressure backpressure semantics
  • document Ray 2.58 scheduler, no-progress timeout, and actor-init behavior
  • remove a stale Ray cluster TODO whose assertion is already implemented

Validation

Using the latest nvcr.io/nvidian/nemo-curator:nightly with this worktree mounted at /opt/Curator:

  • uv lock --check
  • uv sync --all-extras --all-groups --dry-run
  • installed the resolved Ray 2.58.0 / vLLM 0.26.0+cu129 environment and directly imported Ray, vLLM, Transformers, and Qwen-ASR
  • pytest -q -m "not gpu" tests/models/asr/test_qwen_omni.py tests/models/asr/test_qwen_asr.py tests/config/test_run.py tests/core/serve/dynamo/test_vllm.py tests/backends/ray_data/test_diagnostics.py tests/backends/ray_data/test_adapter.py tests/backends/ray_data/test_utils.py (148 passed, 2 deselected)
  • scoped pre-commit hooks

Signed-off-by: Praateek <praateekm@gmail.com>
@github-actions

github-actions Bot commented Sep 3, 2026

Copy link
Copy Markdown
Contributor

@praateekmahajan
praateekmahajan marked this pull request as draft September 3, 2026 18:01
@copy-pr-bot

copy-pr-bot Bot commented Sep 3, 2026

Copy link
Copy Markdown

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.

@greptile-apps

greptile-apps Bot commented Sep 3, 2026

Copy link
Copy Markdown
Contributor

Greptile Summary

This PR raises the Ray Data and Ray Serve baseline to 2.58.0 and updates the lockfile, diagnostic shim, tests, and operational guidance to match Ray’s output-pressure semantics.

  • Replaces queue-ratio diagnostic fields with output-pressure and output-size fields.
  • Documents Ray 2.58 backpressure, no-progress timeout, and actor-initialization behavior.
  • Updates dependency constraints, locked Ray wheels, and inference-server prerequisites.
  • Removes a stale test TODO while preserving its implemented assertion.

Confidence Score: 5/5

The PR appears safe to merge, with no concrete changed-code failure established.

The dependency constraints and lockfile consistently select Ray 2.58.0, and the diagnostic implementation and integration test were updated together for the revised output-pressure fields.

Important Files Changed

Filename Overview
nemo_curator/backends/ray_data/diagnostics.py Retargets the opt-in diagnostic monkeypatch to Ray 2.58 and aligns downstream admission logging with output-pressure semantics; no actionable defect was established.
pyproject.toml Raises both core Ray Data/default and inference-server Ray Serve minimum versions to 2.58.0.
uv.lock Locks Ray 2.58.0 with wheels covering the repository’s declared Python 3.11–3.13 targets and supported architectures.
tests/backends/ray_data/test_diagnostics.py Extends the runtime diagnostics test to require the new output-size and output-pressure event fields.
nemo_curator/backends/ray_data/AGENTS.md Updates scheduler tuning guidance for Ray 2.58 output pressure, execution timeouts, and actor initialization behavior.
fern/versions/main/pages/curate-text/synthetic/inference-server.mdx Updates the documented Ray Serve prerequisite to 2.58.0 or later.
tests/backends/test_utils.py Removes a stale cluster-related TODO from an assertion already checking one setup log per Ray node.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    Config[Ray 2.58 dependency constraints] --> Runtime[Ray Data 2.58 runtime]
    Env[NEMO_CURATOR_RAY_DATA_DIAGNOSTICS] --> Installer[Diagnostics installer]
    Runtime --> Installer
    Installer --> Policy[Downstream capacity policy shim]
    Policy --> Pressure[Compute output pressure]
    Pressure --> Decision[Allow or block producer]
    Decision --> Logs[Structured Ray Data diagnostics]
    Logs --> Operator[Pipeline diagnosis and tuning]
Loading

Reviews (1): Last reviewed commit: "chore(deps): update Ray to 2.58.0" | Re-trigger Greptile

Signed-off-by: Praateek <praateekm@gmail.com>
Signed-off-by: Praateek <praateekm@gmail.com>
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