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

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

tests/entrypoints/pooling/basic/test_encode.py

@@ -9,11 +9,6 @@ from vllm import LLM, PoolingParams
 from vllm.distributed import cleanup_dist_env_and_memory
 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
-    )
-
 MODEL_NAME = "intfloat/multilingual-e5-small"
 
 PROMPTS = [
@@ -35,6 +30,12 @@ TOKEN_IDS = [
 
 @pytest.fixture(scope="module")
 def llm():
+    # ROCm: Use FLEX_ATTENTION backend as it's the only attention backend
+    # that supports encoder-only models on ROCm.
+    attention_config = None
+    if current_platform.is_rocm():
+        attention_config = {"backend": "FLEX_ATTENTION"}
+
     # pytest caches the fixture so we use weakref.proxy to
     # enable garbage collection
     llm = LLM(
@@ -44,6 +45,7 @@ def llm():
         gpu_memory_utilization=0.75,
         enforce_eager=True,
         seed=0,
+        attention_config=attention_config,
     )
 
     yield weakref.proxy(llm)