diff --git a/python/sglang/multimodal_gen/test/unit/test_ipc_a2a_lifecycle.py b/python/sglang/multimodal_gen/test/unit/test_ipc_a2a_lifecycle.py index 72edb91252..65dae67f65 100644 --- a/python/sglang/multimodal_gen/test/unit/test_ipc_a2a_lifecycle.py +++ b/python/sglang/multimodal_gen/test/unit/test_ipc_a2a_lifecycle.py @@ -9,10 +9,18 @@ from sglang.multimodal_gen.runtime.distributed.device_communicators.ipc_a2a impo ipc_a2a_ready, ipc_shareable_zeros, ) +from sglang.multimodal_gen.runtime.platforms import current_platform _IPC = "sglang.multimodal_gen.runtime.distributed.device_communicators.ipc_a2a" +# ipc_a2a_ready() only sets up the transport on CUDA. Elsewhere the platform may +# lack is_full_nvlink and torch may lack the accelerator allocator-settings API. +requires_cuda = pytest.mark.skipif( + not current_platform.is_cuda(), reason="the IPC all-to-all transport is CUDA-only" +) + +@requires_cuda @pytest.mark.parametrize("rank", [0, 1]) @pytest.mark.parametrize("nvlink", [False, True]) def test_ipc_requires_nvlink_before_mapping_peer_memory(monkeypatch, rank, nvlink): @@ -140,6 +148,7 @@ def test_drop_a2a_staging_buffers_clears_the_ulysses_cache(): assert usp._A2A_STAGING_BUFFERS == {} +@requires_cuda @pytest.mark.parametrize("expandable", [False, True]) def test_ipc_buffers_are_allocated_outside_expandable_segments(expandable): settings = "expandable_segments:True" if expandable else ""