[ROCm][CI] Fix entrypoints tests and Python-only installation test on ROCm (#28979)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>

tests/entrypoints/pooling/embed/test_online_dimensions.py

@@ -14,11 +14,6 @@ from tests.utils import RemoteOpenAIServer
 from vllm.entrypoints.pooling.embed.protocol import EmbeddingResponse
 from vllm.platforms import current_platform
 
-if current_platform.is_rocm():
-    pytest.skip(
-        "Encoder self-attention is not implemented on ROCm.", allow_module_level=True
-    )
-
 MODELS = [
     EmbedModelInfo("intfloat/multilingual-e5-small", is_matryoshka=False),
     EmbedModelInfo(
@@ -62,6 +57,10 @@ def server(model_info, dtype: str):
             ["--trust_remote_code", "--hf_overrides", '{"matryoshka_dimensions":[256]}']
         )
 
+    # ROCm: Use Flex Attention to support encoder-only self-attention.
+    if current_platform.is_rocm():
+        args.extend(["--attention-backend", "FLEX_ATTENTION"])
+
     with RemoteOpenAIServer(model_info.name, args) as remote_server:
         yield remote_server