Create PlacementGroup for steps using vLLM

[gabrielmbmb]

Description

This PR updates RayPipeline to create a PlacementGroup when an step is using vLLM. The created placement group, contains as many GPU bundles as tensor_parallel_size specified in the vLLM initialisation and it uses the STRICT_PACK to have the GPU bundles in the same node. It also creates the placement group specifying the _soft_target_node_id, assuring that the _RayStepWrapper actor for the step using vLLM will be created in a specific node. This avoid having vLLM raising the exception Ray does not allocate any GPUs on the driver node. Consider adjusting the Ray placement group or running the driver on a GPU node, as it assures that the driver _StepWrapperRay created resides in the same node as the ray actors created by vLLM for the distributed inference.

[github-actions]

Documentation for this PR has been built. You can view it at: https://distilabel.argilla.io/pr-842/

[codspeed-hq]

CodSpeed Performance Report

Merging #842 will not alter performance

_{Comparing ray-placement-group (f56a9bb) with develop (2aa977f)}

Summary

✅ 1 untouched benchmarks