[fix] Fix placement group bundle ordering for inference engines

[SumanthRH]

Fixes #1306

Ray placement groups don't guarantee bundle order - bundles on the same node may not have consecutive indices.

This PR introduces a ResolvedPlacementGroup wrapper that computes GPU-aware reordered bundle indices (sorted by node_id, gpu_id) as a cached property, and use them for all bundle indexing in both inference engine and trainer placement.

---

[gemini-code-assist]

Code Review

This pull request introduces a SkyRLPlacementGroup wrapper class to manage Ray placement groups, ensuring GPU-aware ordering of resource bundles. The change addresses the issue that Ray's native placement groups do not guarantee bundle order, which is crucial for consistent GPU resource allocation in distributed training and inference. The code updates replace direct PlacementGroup usage with SkyRLPlacementGroup across various modules, including inference engines, trainers, and server groups. This involves extracting the raw Ray placement group and its reordered bundle indices from the wrapper, and modifying scheduling strategies and bundle index assignments to utilize these reordered indices. Additionally, server actor initialization now accepts a list of bundle_indices instead of a starting index, providing more precise control over resource allocation.

_{Note: Security Review did not run due to the size of the PR.}

[erictang000]

@SumanthRH this pretty much lgtm, but let's merge #1300 first since there's some overlap/merge conflicts to resolve?

[kouroshHakha]

Can we call this OrderedPlacementGroup?

[SumanthRH]

Hm I don't think that name would be right given that the underlying pg object is still "unordered" and the wrapper class here is simply resolving the physical ordering of the bundles. OrderedPlacementGroup makes it sound like we have additional guarantees on the pg object

I've renamed to ResolvedPlacementGroup - this should be better.

[gemini-code-assist]

Code Review

This pull request introduces a ResolvedPlacementGroup wrapper to address incorrect bundle ordering in Ray placement groups, which is a critical fix for ensuring correct GPU-aware resource allocation. The new wrapper cleverly caches bundle placement information, improving both correctness and efficiency. The changes are consistently applied across the codebase, including inference engines, trainers, and tests. The overall implementation is robust and well-designed. I have one minor suggestion to improve code clarity in server_group.py.

[gemini-code-assist]

This block can be simplified to improve clarity and remove redundancy. The assignment self._external_pg = placement_group on line 90 is immediately overwritten within the if block, and the else block repeats the default None assignment. Initializing these attributes to None before the if block would make the logic more straightforward.

Suggested change

	self._external_pg = placement_group
	# Extract the raw PG, reordered indices, and GPU IDs from ResolvedPlacementGroup.
	if placement_group is not None:
	self._external_pg = placement_group.pg
	self._reordered_bundle_indices = placement_group.reordered_bundle_indices
	self._bundle_gpu_ids = placement_group.bundle_gpu_ids
	else:
	self._external_pg = None
	self._reordered_bundle_indices = None
	self._bundle_gpu_ids = None
	self._external_pg = None
	self._reordered_bundle_indices = None
	self._bundle_gpu_ids = None
	if placement_group is not None:
	self._external_pg = placement_group.pg
	self._reordered_bundle_indices = placement_group.reordered_bundle_indices
	self._bundle_gpu_ids = placement_group.bundle_gpu_ids