feat: Fix vLLM placement group conflicts in Ray clusters and add local mode…

[ffrujeri]

What does this PR do?

Fixes vLLM placement group conflicts in Ray clusters and adds support for local model paths.

This PR patches vLLM's v1 engine to handle multiple Ray placement groups, preventing crashes when running multiple vLLM instances in the same cluster. It also adds support for using locally stored models instead of always downloading from HuggingFace.

Issues

List issues that this PR closes (syntax):

• Fixes #148

Changes in this PR

• vLLM Placement Group Patch: _patch_vllm_placement_group_filter()

  • Filters out placement group node resource keys (e.g., node:IP_group_N_hash)
  • Allows multiple vLLM instances to coexist in the same Ray cluster
  • Adds unique timestamp-based suffixes to avoid placement group name conflicts
  • Includes comprehensive logging for debugging placement group creation

• Local Model Support: Modified download_model()

  • Checks if model path exists locally before attempting HuggingFace download
  • Skips download step for local models, improving startup time

• Debugging Infrastructure: _cleanup_stale_placement_groups()

  • Logs existing placement groups when debug=True
  • Helps diagnose placement group conflicts and resource allocation issues

• Import Updates

  • Added list_placement_groups from ray.util.state for placement group inspection

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.