[data] Clamp rolling utilization averages to zero

[jeffreywang88]

Description

LLM post-merge tests are failing with the following error:

ray/python/ray/data/_internal/cluster_autoscaler/resource_utilization_gauge.py

Lines 129 to 137 in 645d4f1

	def get(self) -> ClusterUtil:
	"""Get the average cluster utilization based on global usage / global limits."""
	return ClusterUtil(
	cpu=self._cluster_cpu_util_calculator.get_average() or 0,
	gpu=self._cluster_gpu_util_calculator.get_average() or 0,
	memory=self._cluster_mem_util_calculator.get_average() or 0,
	object_store_memory=self._cluster_obj_mem_util_calculator.get_average()
	or 0,
	)

get() reads rolling averages from TimeWindowAverageCalculator, which maintains a running _sum updated on report() and reduced in _trim() as values expire. Due to floating-point rounding, _sum can drift slightly negative after all values are trimmed, producing a negative average and violating the ClusterUtil assertion.

Approach

Clamp the rolling average to max(0, ...) in RollingLogicalUtilizationGauge.get(), since utilization ratios are inherently non-negative and small negative values are just floating-point artifacts.

Test

python -m pytest python/ray/llm/tests/batch/gpu/processor/test_vllm_engine_proc.py -vs passes after the fix.

Related issues

• Postmerge failure: https://buildkite.com/ray-project/postmerge/builds/16343/steps/canvas?sid=019cc1c1-e7c2-4a1c-9a86-a44211b14e66&tab=output
• Related PR: [Data] Add ClusterUtil dataclass for ResourceUtilizationGauge #61436

Additional information

Optional: Add implementation details, API changes, usage examples, screenshots, etc.

[gemini-code-assist]

Code Review

The pull request correctly addresses a floating-point precision issue by clamping utilization values to be non-negative, which prevents assertion failures. The change is correct. I have provided a suggestion to refactor the implementation slightly to improve readability and maintainability by reducing code duplication.

[jeffreywang88]

Postmerge test failures are fixed by #61580. This PR is no longer needed.