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

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Andreas Karatzas
2025-12-23 22:42:30 -08:00
committed by GitHub
parent 8ee90c83f8
commit 0247a91e00
26 changed files with 432 additions and 116 deletions
@@ -15,11 +15,6 @@ from tests.models.language.pooling_mteb_test.mteb_utils import (
from tests.utils import RemoteOpenAIServer
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
)
os.environ["VLLM_LOGGING_LEVEL"] = "WARNING"
MODEL_NAME = "cross-encoder/ms-marco-MiniLM-L-6-v2"
@@ -30,6 +25,10 @@ st_main_score = 0.33457
def server():
args = ["--runner", "pooling", "--enforce-eager", "--disable-uvicorn-access-log"]
# ROCm: Use Flex Attention to support encoder-only self-attention.
if current_platform.is_rocm():
args.extend(["--attention-backend", "FLEX_ATTENTION"])
with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
yield remote_server
@@ -11,16 +11,17 @@ 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 = "tomaarsen/Qwen3-Reranker-0.6B-seq-cls"
@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(
@@ -30,6 +31,7 @@ def llm():
gpu_memory_utilization=0.75,
enforce_eager=True,
seed=0,
attention_config=attention_config,
)
yield weakref.proxy(llm)
@@ -11,11 +11,6 @@ from vllm.entrypoints.pooling.pooling.protocol import PoolingResponse
from vllm.entrypoints.pooling.score.protocol import RerankResponse
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 = "BAAI/bge-reranker-base"
DTYPE = "bfloat16"
@@ -24,6 +19,10 @@ DTYPE = "bfloat16"
def server():
args = ["--enforce-eager", "--max-model-len", "100", "--dtype", DTYPE]
# ROCm: Use Flex Attention to support encoder-only self-attention.
if current_platform.is_rocm():
args.extend(["--attention-backend", "FLEX_ATTENTION"])
with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
yield remote_server
@@ -12,11 +12,6 @@ from tests.utils import RemoteOpenAIServer
from vllm.entrypoints.pooling.score.protocol import ScoreResponse
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 = [
{"name": "BAAI/bge-reranker-v2-m3", "is_cross_encoder": True},
{"name": "BAAI/bge-base-en-v1.5", "is_cross_encoder": False},
@@ -44,6 +39,10 @@ def model(request):
def server(model: dict[str, Any]):
args = ["--enforce-eager", "--max-model-len", "100", "--dtype", DTYPE]
# ROCm: Use Flex Attention to support encoder-only self-attention.
if current_platform.is_rocm():
args.extend(["--attention-backend", "FLEX_ATTENTION"])
with RemoteOpenAIServer(model["name"], args) as remote_server:
yield remote_server