[Feature] support rollout DP+EP for vLLM

[princepride]

Wire vLLM rollout data parallelism into TP sizing and use conservative DP sleep-mode defaults so Qwen3 DP+EP DeepEP validation can run reliably.

[gemini-code-assist]

Code Review

This pull request introduces support for vLLM data parallelism (DP) alongside expert parallelism (EP) and the DeepEP communication kernel. It adds compatibility shims for legacy slime imports, updates parallel size resolution to account for DP, extends engine timeouts for DP configurations, and applies conservative defaults for DP engines in sleep mode. Additionally, comprehensive integration and unit tests are added. Feedback on the changes suggests enforcing positive integer values for pipeline and data parallel sizes to prevent potential division-by-zero errors, and using safe dictionary access for dp_size in server_args to avoid KeyError exceptions.

[gemini-code-assist]

If vllm-pipeline-parallel-size or vllm-data-parallel-size are passed as 0 or negative values, this division will raise a ZeroDivisionError or result in an invalid negative parallel size. We should enforce a minimum value of 1 for both sizes to ensure robustness.

Suggested change

	pp_size = temp_args.vllm_pipeline_parallel_size
	vllm_tp_size = temp_args.rollout_num_gpus_per_engine // pp_size
	dp_size = temp_args.vllm_data_parallel_size
	vllm_tp_size = temp_args.rollout_num_gpus_per_engine // (pp_size * dp_size)
	pp_size = max(1, temp_args.vllm_pipeline_parallel_size)
	dp_size = max(1, temp_args.vllm_data_parallel_size)
	vllm_tp_size = temp_args.rollout_num_gpus_per_engine // (pp_size * dp_size)

[gemini-code-assist]

If denom is 0 or negative due to invalid or zero parallel sizes, this modulo operation will raise a ZeroDivisionError or result in an invalid parallel configuration. Adding a defensive check to ensure denom is positive prevents unexpected crashes.

Suggested change

	denom = pp * dp
	if gpus_per_engine % denom != 0:
	denom = pp * dp
	if denom <= 0:
	raise ValueError(
	f"Invalid parallel configuration: vllm_pipeline_parallel_size ({pp}) "
	f"and vllm_data_parallel_size ({dp}) must be positive integers."
	)
	if gpus_per_engine % denom != 0:

[gemini-code-assist]

Accessing server_args["dp_size"] directly can raise a KeyError if the dictionary is partially populated (e.g., in certain unit tests or custom orchestration flows). Using .get("dp_size", 1) is safer and consistent with how dp_size is accessed in build_vllm_subprocess_env.

Suggested change

	if (
	server_args["dp_size"] > 1
	and getattr(args, "vllm_enable_sleep_mode", False)
	and not _user_overrode(args, "vllm_async_scheduling")
	):
	cmd += ["--no-async-scheduling"]
	if server_args["dp_size"] > 1 and getattr(args, "vllm_enable_sleep_mode", False):
	dp_size = server_args.get("dp_size", 1)
	if (
	dp_size > 1
	and getattr(args, "vllm_enable_sleep_mode", False)
	and not _user_overrode(args, "vllm_async_scheduling")
	):
	cmd += ["--no-async-scheduling"]
	if dp_size > 1 and getattr(args, "vllm_enable_sleep_mode", False):

[aoshen02]

Why disable async-scheduling and enforce eager here?

[princepride]

Closed because #108 has covered.