Skip to content

(retriever) Bump torch to 2.10, torchvision to 0.25, and vLLM to 0.17#1762

Merged
charlesbluca merged 11 commits intomainfrom
retriever-torch210
Apr 21, 2026
Merged

(retriever) Bump torch to 2.10, torchvision to 0.25, and vLLM to 0.17#1762
charlesbluca merged 11 commits intomainfrom
retriever-torch210

Conversation

@charlesbluca
Copy link
Copy Markdown
Collaborator

@charlesbluca charlesbluca commented Apr 1, 2026

Description

Bumps core ML dependency versions in the local extra and updates the nemotron-ocr nightly wheel build to match:

  • torch: ~=2.9.1~=2.10.0
  • torchvision: >=0.24,<0.25>=0.25,<0.26
  • accelerate: >=1.1.0==1.12.0
  • vLLM: 0.16.00.17.0 (adds VLM support)
  • huggingface-nightly CI: OCR wheel now built against torch==2.10.0 / torchvision==0.25.0 (CUDA 13.0 unchanged)

Checklist

  • I am familiar with the Contributing Guidelines.
  • New or existing tests cover these changes.
  • The documentation is up to date with these changes.
  • If adjusting docker-compose.yaml environment variables have you ensured those are mimicked in the Helm values.yaml file.

@charlesbluca charlesbluca requested review from a team as code owners April 1, 2026 14:37
@charlesbluca charlesbluca requested a review from jperez999 April 1, 2026 14:37
… conflicts

Keep vllm==0.17.0 from branch with sys_platform=='linux' guard from upstream;
retain new upstream deps (duckdb, neo4j, langchain-nvidia-ai-endpoints).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
@charlesbluca charlesbluca marked this pull request as draft April 14, 2026 18:01
@greptile-apps
Copy link
Copy Markdown
Contributor

greptile-apps Bot commented Apr 14, 2026

Greptile Summary

This PR bumps core ML dependencies in the local extra: torch ~=2.9.1~=2.10.0, torchvision >=0.24,<0.25>=0.25,<0.26, accelerate to ==1.12.0, and vLLM 0.16.00.17.0 with updated CUDA 13.0 wheel URLs. The nightly OCR build workflow is updated to pin torch==2.10.0 / torchvision==0.25.0, which is consistent with the new pyproject.toml range.

Confidence Score: 5/5

This PR is safe to merge — it is a straightforward version bump with consistent updates across both changed files.

All changes are dependency version bumps: pyproject.toml and the nightly workflow are in sync (torch 2.10.0, torchvision 0.25.x). The vLLM wheel URLs are pinned to specific GitHub release artifacts for both x86_64 and aarch64. No logic changes, no security regressions, and no API surface alterations were introduced.

No files require special attention.

Important Files Changed

Filename Overview
nemo_retriever/pyproject.toml Bumps torch to ~=2.10.0, torchvision to >=0.25,<0.26, pins accelerate==1.12.0, and upgrades vllm to 0.17.0 with updated wheel URLs; changes are internally consistent
.github/workflows/huggingface-nightly.yml Updates OCR nightly wheel build to use torch==2.10.0 and torchvision==0.25.0, matching the pyproject.toml version bump

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[nemo-retriever local extra] --> B[torch ~=2.10.0\npytorch-cu130 index]
    A --> C[torchvision >=0.25,<0.26\npytorch-cu130 index]
    A --> D[accelerate ==1.12.0]
    A --> E[vllm ==0.17.0\ndirect wheel URL cu130]
    B --> F[CUDA 13.0]
    C --> F
    E --> F
    G[huggingface-nightly.yml\nOCR wheel build] --> H[torch==2.10.0\ntorchvision==0.25.0\nbuild-time pins]
    H --> F
Loading

Reviews (2): Last reviewed commit: "Build nemotron-ocr wheel against torch 2..." | Re-trigger Greptile

@charlesbluca charlesbluca changed the title (retriever) Bump torch pinning to 2.10 (retriever) Bump torch to 2.10, torchvision to 0.25, and vLLM to 0.17 Apr 21, 2026
@charlesbluca charlesbluca marked this pull request as ready for review April 21, 2026 17:40
@charlesbluca charlesbluca merged commit baf6c88 into main Apr 21, 2026
5 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.

4 participants