From f932e78f1b49360ae34fe9404bccf059b2fb95b6 Mon Sep 17 00:00:00 2001 From: "wenxin.zhang" Date: Wed, 30 Sep 2026 17:10:35 +0000 Subject: [PATCH 1/4] test(xpu): skip cases affected by missing paged decode kernel Signed-off-by: wenxin.zhang --- tests/mm_router/test_vllm_mm_router_e2e.py | 1 - tests/serve/multimodal_profiles/vllm_xpu.py | 38 ++++++++++++++++++--- tests/serve/test_vllm_omni.py | 1 - 3 files changed, 33 insertions(+), 7 deletions(-) diff --git a/tests/mm_router/test_vllm_mm_router_e2e.py b/tests/mm_router/test_vllm_mm_router_e2e.py index dc71834d073b..e8862480a10d 100644 --- a/tests/mm_router/test_vllm_mm_router_e2e.py +++ b/tests/mm_router/test_vllm_mm_router_e2e.py @@ -46,7 +46,6 @@ pytest.mark.vllm, pytest.mark.multimodal, pytest.mark.gpu_1, - pytest.mark.xpu_1, pytest.mark.model(VLLM_MM_MODEL), ] diff --git a/tests/serve/multimodal_profiles/vllm_xpu.py b/tests/serve/multimodal_profiles/vllm_xpu.py index 65dc7047d4db..48ee98861825 100644 --- a/tests/serve/multimodal_profiles/vllm_xpu.py +++ b/tests/serve/multimodal_profiles/vllm_xpu.py @@ -61,7 +61,13 @@ gpu_marker="xpu_1", topologies={ "agg": TopologyConfig( - marks=[pytest.mark.post_merge, pytest.mark.xpu_1], + marks=[ + pytest.mark.skip( + reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false" + ), + pytest.mark.post_merge, + pytest.mark.xpu_1, + ], # TODO: re-enable XPU-parallel scheduling with # profiled_vram_gib=9.6 once this has a bounded --kv-bytes profile. timeout_s=300, @@ -135,7 +141,13 @@ # single-GPU CI environment. "agg_router_chat_processor": TopologyConfig( health_check_workers=True, - marks=[pytest.mark.post_merge, pytest.mark.xpu_2], + marks=[ + pytest.mark.skip( + reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false" + ), + pytest.mark.post_merge, + pytest.mark.xpu_2, + ], gpu_marker="xpu_2", timeout_s=400, profiled_vram_gib=18.7, @@ -153,7 +165,13 @@ # CI; the content-hash correctness assertion lives in # tests/mm_router/test_router_rust_mm_frontend_decode_e2e.py. "agg_router_frontend_decode": TopologyConfig( - marks=[pytest.mark.post_merge, pytest.mark.xpu_2], + marks=[ + pytest.mark.skip( + reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false" + ), + pytest.mark.post_merge, + pytest.mark.xpu_2, + ], gpu_marker="xpu_2", timeout_s=400, profiled_vram_gib=18.7, @@ -180,7 +198,12 @@ topologies={ "agg_router": TopologyConfig( health_check_workers=True, - marks=[pytest.mark.post_merge], + marks=[ + pytest.mark.skip( + reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false" + ), + pytest.mark.post_merge, + ], timeout_s=500, gpu_marker="xpu_2", profiled_vram_gib=19.0, @@ -196,7 +219,12 @@ topologies={ "agg_router": TopologyConfig( health_check_workers=True, - marks=[pytest.mark.post_merge], + marks=[ + pytest.mark.skip( + reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false" + ), + pytest.mark.post_merge, + ], timeout_s=500, gpu_marker="xpu_2", profiled_vram_gib=16.0, diff --git a/tests/serve/test_vllm_omni.py b/tests/serve/test_vllm_omni.py index a2dfec803d45..cce8a94d3911 100644 --- a/tests/serve/test_vllm_omni.py +++ b/tests/serve/test_vllm_omni.py @@ -177,7 +177,6 @@ class VLLMOmniConfig(EngineConfig): script_name="agg_omni_audio.sh", marks=[ pytest.mark.gpu_1, - pytest.mark.xpu_1, pytest.mark.pre_merge, pytest.mark.timeout(1200), ], From c5c8fbaa269c9a2510b94ab1f9b39d372edd8538 Mon Sep 17 00:00:00 2001 From: "wenxin.zhang" Date: Wed, 30 Sep 2026 18:23:13 +0000 Subject: [PATCH 2/4] test(xpu): skip Qwen3-VL router with missing paged decode kernel Signed-off-by: wenxin.zhang --- tests/serve/multimodal_profiles/vllm_xpu.py | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/tests/serve/multimodal_profiles/vllm_xpu.py b/tests/serve/multimodal_profiles/vllm_xpu.py index 48ee98861825..bd60c04dba71 100644 --- a/tests/serve/multimodal_profiles/vllm_xpu.py +++ b/tests/serve/multimodal_profiles/vllm_xpu.py @@ -122,7 +122,13 @@ # (post_merge). "agg_router": TopologyConfig( health_check_workers=True, - marks=[pytest.mark.pre_merge, pytest.mark.xpu_2], + marks=[ + pytest.mark.skip( + reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false" + ), + pytest.mark.pre_merge, + pytest.mark.xpu_2, + ], gpu_marker="xpu_2", timeout_s=400, profiled_vram_gib=18.7, From ede3f6fa4bb0e7411348180741e534b961ba3d4d Mon Sep 17 00:00:00 2001 From: "wenxin.zhang" Date: Wed, 30 Sep 2026 22:45:09 +0000 Subject: [PATCH 3/4] test(xpu): track skipped multimodal coverage in #15417 Signed-off-by: wenxin.zhang --- tests/mm_router/test_vllm_mm_router_e2e.py | 1 + tests/serve/multimodal_profiles/vllm_xpu.py | 12 ++++++------ tests/serve/test_vllm_omni.py | 1 + 3 files changed, 8 insertions(+), 6 deletions(-) diff --git a/tests/mm_router/test_vllm_mm_router_e2e.py b/tests/mm_router/test_vllm_mm_router_e2e.py index e8862480a10d..0df70dc2d5ef 100644 --- a/tests/mm_router/test_vllm_mm_router_e2e.py +++ b/tests/mm_router/test_vllm_mm_router_e2e.py @@ -45,6 +45,7 @@ pytest.mark.e2e, pytest.mark.vllm, pytest.mark.multimodal, + # Restore xpu_1 after the paged decode kernel is available; see #15417. pytest.mark.gpu_1, pytest.mark.model(VLLM_MM_MODEL), ] diff --git a/tests/serve/multimodal_profiles/vllm_xpu.py b/tests/serve/multimodal_profiles/vllm_xpu.py index bd60c04dba71..4c5955c97f83 100644 --- a/tests/serve/multimodal_profiles/vllm_xpu.py +++ b/tests/serve/multimodal_profiles/vllm_xpu.py @@ -63,7 +63,7 @@ "agg": TopologyConfig( marks=[ pytest.mark.skip( - reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false" + reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false; see #15417" ), pytest.mark.post_merge, pytest.mark.xpu_1, @@ -124,7 +124,7 @@ health_check_workers=True, marks=[ pytest.mark.skip( - reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false" + reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false; see #15417" ), pytest.mark.pre_merge, pytest.mark.xpu_2, @@ -149,7 +149,7 @@ health_check_workers=True, marks=[ pytest.mark.skip( - reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false" + reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false; see #15417" ), pytest.mark.post_merge, pytest.mark.xpu_2, @@ -173,7 +173,7 @@ "agg_router_frontend_decode": TopologyConfig( marks=[ pytest.mark.skip( - reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false" + reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false; see #15417" ), pytest.mark.post_merge, pytest.mark.xpu_2, @@ -206,7 +206,7 @@ health_check_workers=True, marks=[ pytest.mark.skip( - reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false" + reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false; see #15417" ), pytest.mark.post_merge, ], @@ -227,7 +227,7 @@ health_check_workers=True, marks=[ pytest.mark.skip( - reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false" + reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false; see #15417" ), pytest.mark.post_merge, ], diff --git a/tests/serve/test_vllm_omni.py b/tests/serve/test_vllm_omni.py index cce8a94d3911..1ef48d206f14 100644 --- a/tests/serve/test_vllm_omni.py +++ b/tests/serve/test_vllm_omni.py @@ -176,6 +176,7 @@ class VLLMOmniConfig(EngineConfig): directory=vllm_dir, script_name="agg_omni_audio.sh", marks=[ + # Restore XPU with a two-card case after startup is fixed; see #15417. pytest.mark.gpu_1, pytest.mark.pre_merge, pytest.mark.timeout(1200), From c66a26d0cea78d41b711cc8570e2f668aac58b09 Mon Sep 17 00:00:00 2001 From: "wenxin.zhang" Date: Tue, 13 Oct 2026 18:03:01 +0000 Subject: [PATCH 4/4] test(xpu): restore vLLM Omni audio coverage Signed-off-by: wenxin.zhang --- tests/serve/test_vllm_omni.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/serve/test_vllm_omni.py b/tests/serve/test_vllm_omni.py index 8ad61aa4e210..9b337af6c72d 100644 --- a/tests/serve/test_vllm_omni.py +++ b/tests/serve/test_vllm_omni.py @@ -180,8 +180,8 @@ class VLLMOmniConfig(EngineConfig): directory=vllm_dir, script_name="agg_omni_audio.sh", marks=[ - # Restore XPU with a two-card case after startup is fixed; see #15417. pytest.mark.gpu_1, + pytest.mark.xpu_1, pytest.mark.pre_merge, pytest.mark.timeout(1200), ],