[ROCm][Spec Decode] Fix probabilistic draft probs test attention backend (#45706)

Signed-off-by: Stefan Koncarevic <[email protected]>

.buildkite/test_areas/misc.yaml

@@ -21,6 +21,12 @@ steps:
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     # TODO: create another `optional` test group for slow tests
     - pytest -v -s -m 'not slow_test' v1/spec_decode
+  mirror:
+    amd:
+      device: mi300_1
+      timeout_in_minutes: 65
+      depends_on:
+      - image-build-amd
 
 - label: V1 Sample + Logits
   key: v1-sample-logits

tests/v1/spec_decode/test_eagle.py

@@ -1002,7 +1002,11 @@ def test_propose(method, attn_backend, num_speculative_tokens, monkeypatch):
     assert torch.equal(result, expected_tokens)
 
 
-def test_propose_stores_probabilistic_draft_probs(monkeypatch):
+@pytest.mark.parametrize(
+    "attn_backend",
+    ["ROCM_ATTN", "TRITON_ATTN"] if current_platform.is_rocm() else ["FLASH_ATTN"],
+)
+def test_propose_stores_probabilistic_draft_probs(attn_backend, monkeypatch):
     device = torch.device(DEVICE_TYPE)
     batch_size = 2
     seq_lens = [5, 3]
@@ -1053,7 +1057,7 @@ def test_propose_stores_probabilistic_draft_probs(monkeypatch):
     )
 
     attn_metadata_builder_cls, _ = try_get_attention_backend(
-        AttentionBackendEnum.FLASH_ATTN
+        AttentionBackendEnum[attn_backend]
     )
     attn_metadata_builder = attn_metadata_builder_cls(
         kv_cache_spec=create_standard_kv_cache_spec(proposer.vllm_config),