[CI] Migrate non-Megatron GPU CI to run on new inference codepath

[SumanthRH]

What does this PR do?

Migrate the new inference codepath to run only on the new inference codepath.

This is the first part in a series of PRs to migrate completely to the new inference codepath.

I will wait for R3 to land before merging changes from this PR: #1428

Fixes

1. Fix port collisions for prometheus port for vLLM router with multiple concurrent runs of SkyRL
2. Fix world size calculation for new inference codepath with DP > 1: Previously, we incorrectly calculated offsets per server url (count of num_engines * data_parallel_size) - we should really be calculating offsets per deployment (i.e for the count of num_engines). This PR fixes it by including data parallel size in the offset calculation.
3. Fix sleep/ wake-up for tests/backends/skyrl_train/gpu/gpu_ci/test_lora.py : Old codepath performed a sleep + wake up by default -> this lead to some memory savings in temporary buffers etc. New codepath ooms because engines are on GPU by default. Added proper sleep and wake up calls at inference training boundaries as the fix.
4. Migrates test_pause_and_continue_generation.py to the new inference codepath.
5. Removes tests meant solely for legacy inference codepath in test_engine_generation.py

Test Plan

I ran GPU CI E2E with the new changes and all tests pass.

TODO:

• Run GPU CI again and ensure tests pass

Future work

Not all tests are fully migrated to the new codepath. There are two major items pending

1. Megatron migration worker tests skip colocated tests for new inference: We should be able to run these after [feat] chunked ipc support for new inference #1512 lands.
2. Gloo backend support for weight syncing: We should probably just get rid of these for now, Should be implementable on top of SkyRL

[gemini-code-assist]

Code Review

This pull request migrates the CI to the new inference pathway, adds support for data parallelism in the RemoteInferenceClient, and cleans up legacy tests. Key modifications include updating world size and rank offset logic to account for data_parallel_size and implementing port reservation in VLLMRouter to avoid race conditions. Feedback identifies potential ZeroDivisionError issues in RemoteInferenceClient and BroadcastInitInfo if data_parallel_size is zero or if the server count is not a multiple of the parallel size.

[gemini-code-assist]

The calculation of num_deployments assumes data_parallel_size is at least 1 and that len(self.server_urls) is an exact multiple of it. If data_parallel_size is 0, this will raise a ZeroDivisionError. If the division is not exact, it will silently truncate the number of deployments, leading to an incorrect total_world_size. It is recommended to validate these invariants.

Suggested change

	num_deployments = len(self.server_urls) // self.data_parallel_size
	self._world_size = (per_server[0] * num_deployments, per_server[0])
	assert self.data_parallel_size > 0, "data_parallel_size must be at least 1"
	num_deployments, remainder = divmod(len(self.server_urls), self.data_parallel_size)
	assert remainder == 0, "Number of server URLs must be a multiple of data_parallel_size"
	self._world_size = (per_server[0] * num_deployments, per_server[0])

[gemini-code-assist]

Potential ZeroDivisionError at line 97 if dp_size is 0. Although it has a default value of 1 in the method signature, it is passed dynamically from the client's configuration. Adding a validation check at the start of the method would improve robustness.

Suggested change

	result: List[BroadcastInitInfo] = []
	assert dp_size > 0, "dp_size must be at least 1"
	result: List[BroadcastInitInfo] = []

[SumanthRH]

non-Megatron GPU CI is passing: https://github.com/NovaSky-AI/SkyRL/actions/runs/24435811449/job/71389704452
Megatron GPU CI is also passing: https://github.com/NovaSky-AI/SkyRL/actions/runs/24435811479