[CI] Add Blackwell AsyncTP correctness test (#35871)

Signed-off-by: Stefano Castagnetta <scastagnetta@nvidia.com>

.buildkite/test_areas/compile.yaml

@@ -36,6 +36,16 @@ steps:
   - export VLLM_TEST_CLEAN_GPU_MEMORY=1
   - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py
 
+- label: AsyncTP Correctness Tests (B200)
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/"
+  device: b200
+  optional: true
+  num_devices: 2
+  commands:
+  - export VLLM_TEST_CLEAN_GPU_MEMORY=1
+  - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py
+
 - label: Distributed Compile Unit Tests (2xH100)
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/"

tests/compile/correctness_e2e/test_async_tp.py

@@ -31,7 +31,12 @@ def test_async_tp_pass_correctness(
     distributed_backend: str,
     eager_mode: bool,
     num_gpus_available: int,
+    monkeypatch,
 ):
+    # Disable FlashInfer FP8 scaled_mm kernel as it is incompatible with
+    # async TP patterns. No-op on H100 (kernel requires CC >= 100).
+    monkeypatch.setenv("VLLM_DISABLED_KERNELS", "FlashInferFP8ScaledMMLinearKernel")
+
     model_info = HF_EXAMPLE_MODELS.find_hf_info(model_id)
     model_info.check_transformers_version(on_fail="skip")
     model_info.check_available_online(on_fail="skip")