diff --git a/.buildkite/test_areas/cuda.yaml b/.buildkite/test_areas/cuda.yaml
index f0b1f4af2b..230781cd5d 100644
--- a/.buildkite/test_areas/cuda.yaml
+++ b/.buildkite/test_areas/cuda.yaml
@@ -30,8 +30,8 @@ steps:
       - vllm/
       - tests/cuda
 
-- label: ":nvidia: (H200 MIG 35GB) CUDAGraph"
-  device: h200_35gb
+- label: ":nvidia: (H200 MIG 18GB) CUDAGraph"
+  device: h200_18gb
   key: cudagraph
   timeout_in_minutes: 30
   source_file_dependencies:
diff --git a/.buildkite/test_areas/engine.yaml b/.buildkite/test_areas/engine.yaml
index a6b98bf931..bb848cd9e5 100644
--- a/.buildkite/test_areas/engine.yaml
+++ b/.buildkite/test_areas/engine.yaml
@@ -74,18 +74,20 @@ steps:
       depends_on:
       - image-build-amd
 
-- label: ":nvidia: (H200 MIG 35GB) E2E Core"
-  device: h200_35gb
+- label: ":nvidia: (H200 MIG 18GB) E2E Core"
+  device: h200_18gb
   key: e2e-core-1-gpu
   timeout_in_minutes: 50
   source_file_dependencies:
     - vllm/v1/
     - tests/v1/e2e/general/
   commands:
-    - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
+    - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py --ignore v1/e2e/general/test_kv_sharing_fast_prefill.py -k "not test_mamba_prefix_cache_mrv2"
   mirror:
     amd:
       label: ":amd: (MI250) E2E Core"
+      commands:
+      - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
       device: mi250_1
       timeout_in_minutes: 50
       depends_on:
@@ -95,6 +97,18 @@ steps:
       - tests/v1/e2e/general/
       - vllm/platforms/rocm.py
 
+# Keep the large Gemma-3n and Mamba MRV2 warmup allocations on 35GB slices.
+- label: ":nvidia: (H200 MIG 35GB) E2E Core Large Memory"
+  device: h200_35gb
+  key: e2e-core-large-memory
+  timeout_in_minutes: 50
+  source_file_dependencies:
+    - vllm/v1/
+    - tests/v1/e2e/general/
+  commands:
+    - pytest -v -s v1/e2e/general/test_kv_sharing_fast_prefill.py
+    - pytest -v -s v1/e2e/general/test_mamba_prefix_cache.py::test_mamba_prefix_cache_mrv2 v1/e2e/general/test_mamba_prefix_cache.py::test_mamba_prefix_cache_mrv2_async
+
 - label: ":nvidia: (L4) V1 E2E"
   key: v1-e2e-2-gpus
   timeout_in_minutes: 25 # TODO: Fix timeout after we have more confidence in the test stability
diff --git a/.buildkite/test_areas/misc.yaml b/.buildkite/test_areas/misc.yaml
index d24ec5b57f..b45b26b2f2 100644
--- a/.buildkite/test_areas/misc.yaml
+++ b/.buildkite/test_areas/misc.yaml
@@ -103,8 +103,8 @@ steps:
       depends_on:
       - image-build-amd
 
-- label: ":nvidia: (H200 MIG 35GB) V1 Executor + Worker"
-  device: h200_35gb
+- label: ":nvidia: (H200 MIG 18GB) V1 Executor + Worker"
+  device: h200_18gb
   key: v1-executor-worker
   timeout_in_minutes: 45
   source_file_dependencies: *v1-core-kv-metrics-deps
diff --git a/.buildkite/test_areas/samplers.yaml b/.buildkite/test_areas/samplers.yaml
index dae7e96468..18f857f65a 100644
--- a/.buildkite/test_areas/samplers.yaml
+++ b/.buildkite/test_areas/samplers.yaml
@@ -2,8 +2,8 @@ group: Samplers
 depends_on: 
   - image-build
 steps:
-- label: ":nvidia: (H200 MIG 35GB) Samplers"
-  device: h200_35gb
+- label: ":nvidia: (H200 MIG 18GB) Samplers"
+  device: h200_18gb
   key: samplers-test
   timeout_in_minutes: 60
   source_file_dependencies:
@@ -15,8 +15,8 @@ steps:
   commands:
     # VLLM_USE_FLASHINFER_SAMPLER defaults to 1 now, so we need to pin both
     # values explicitly to still cover the PyTorch-native (Triton) path.
-    - VLLM_USE_FLASHINFER_SAMPLER=0 pytest -v -s samplers
-    - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
+    - VLLM_USE_FLASHINFER_SAMPLER=0 pytest -v -s samplers -k "not test_beam_search_passes_multimodal_data"
+    - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers -k "not test_beam_search_passes_multimodal_data"
   mirror:
     amd:
       label: ":amd: (MI250) Samplers"
@@ -35,3 +35,18 @@ steps:
       - vllm/platforms/rocm.py
       commands:
       - pytest -v -s samplers
+
+# Qwen2-Audio-7B leaves no room for KV cache on a 16 GiB slice.
+- label: ":nvidia: (H200 MIG 35GB) Samplers Multimodal Beam Search"
+  device: h200_35gb
+  key: samplers-multimodal-beam-search
+  timeout_in_minutes: 30
+  source_file_dependencies:
+  - vllm/model_executor/layers
+  - vllm/sampling_metadata.py
+  - tests/samplers
+  - tests/conftest.py
+  - vllm/entrypoints/generate/beam_search
+  commands:
+    - VLLM_USE_FLASHINFER_SAMPLER=0 pytest -v -s samplers/test_beam_search.py::test_beam_search_passes_multimodal_data
+    - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers/test_beam_search.py::test_beam_search_passes_multimodal_data