[vLLM] MRV2 nondeterministic OOM error on B580

[quinnlp]

See nondeterminism: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/31307857412/attempts/1

• Attempts 1 and 2 fail
• Attempt 3 passes

Error:

ValueError: To serve at least one request with the model's max seq len (40960), (4.38 GiB KV cache is needed, which is larger than the available KV cache memory (3.92 GiB). Based on the available memory, the estimated maximum model length is 36672. Try increasing `gpu_memory_utilization` (which also controls CPU memory on the CPU backend) or decreasing `max_model_len` when initializing the engine. See https://docs.vllm.ai/en/latest/configuration/conserving_memory/ for more details.

Failing test:

tests/v1/kv_connector/unit/test_nixl_connector.py::test_abort_timeout_on_prefiller[ray]

[kwasd]

They were ran on runners with different configurations, and now this is fixed. Please re-check.