[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:
@@ -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)
|
||||
|
||||
@@ -9,11 +9,6 @@ import pytest_asyncio
|
||||
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
|
||||
)
|
||||
|
||||
MODEL_NAME = "sentence-transformers/all-MiniLM-L12-v2"
|
||||
max_model_len = 128
|
||||
|
||||
@@ -44,6 +39,10 @@ def server():
|
||||
str(max_model_len),
|
||||
]
|
||||
|
||||
# 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
|
||||
|
||||
|
||||
@@ -0,0 +1,28 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Pytest configuration for vLLM pooling embed tests."""
|
||||
|
||||
import warnings
|
||||
|
||||
import torch
|
||||
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
|
||||
def pytest_collection_modifyitems(config, items):
|
||||
"""Configure ROCm-specific settings based on collected tests."""
|
||||
if not current_platform.is_rocm():
|
||||
return
|
||||
|
||||
# Disable Flash/MemEfficient SDP on ROCm to avoid HF Transformers
|
||||
# accuracy issues: https://github.com/vllm-project/vllm/issues/30167
|
||||
# TODO: Remove once ROCm SDP accuracy issues are resolved on HuggingFace
|
||||
torch.backends.cuda.enable_flash_sdp(False)
|
||||
torch.backends.cuda.enable_mem_efficient_sdp(False)
|
||||
torch.backends.cuda.enable_math_sdp(True)
|
||||
warnings.warn(
|
||||
"ROCm: Disabled flash_sdp and mem_efficient_sdp, enabled math_sdp "
|
||||
"to avoid HuggingFace Transformers accuracy issues",
|
||||
UserWarning,
|
||||
stacklevel=1,
|
||||
)
|
||||
@@ -13,11 +13,6 @@ 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 = "intfloat/e5-small"
|
||||
@@ -28,6 +23,10 @@ MAIN_SCORE = 0.7422994752439667
|
||||
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
|
||||
|
||||
|
||||
@@ -11,11 +11,6 @@ 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 = ["The chef prepared a delicious meal."]
|
||||
@@ -23,6 +18,12 @@ prompts = ["The chef prepared a delicious meal."]
|
||||
|
||||
@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(
|
||||
@@ -32,6 +33,7 @@ def llm():
|
||||
gpu_memory_utilization=0.75,
|
||||
enforce_eager=True,
|
||||
seed=0,
|
||||
attention_config=attention_config,
|
||||
)
|
||||
|
||||
yield weakref.proxy(llm)
|
||||
|
||||
@@ -28,16 +28,20 @@ from vllm.utils.serial_utils import (
|
||||
decode_pooling_output,
|
||||
)
|
||||
|
||||
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"
|
||||
DUMMY_CHAT_TEMPLATE = """{% for message in messages %}{{message['role'] + ': ' + message['content'] + '\\n'}}{% endfor %}""" # noqa: E501
|
||||
DTYPE = "bfloat16"
|
||||
|
||||
|
||||
if current_platform.is_rocm():
|
||||
# Disable Flash/MemEfficient SDP on ROCm to avoid HF Transformers
|
||||
# accuracy issues: https://github.com/vllm-project/vllm/issues/30167
|
||||
# TODO: Remove once ROCm SDP accuracy issues are resolved on HuggingFace
|
||||
torch.backends.cuda.enable_flash_sdp(False)
|
||||
torch.backends.cuda.enable_mem_efficient_sdp(False)
|
||||
torch.backends.cuda.enable_math_sdp(True)
|
||||
|
||||
|
||||
@pytest.fixture(scope="module")
|
||||
def server():
|
||||
args = [
|
||||
@@ -53,6 +57,10 @@ def server():
|
||||
DUMMY_CHAT_TEMPLATE,
|
||||
]
|
||||
|
||||
# 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
|
||||
|
||||
|
||||
@@ -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
|
||||
|
||||
|
||||
@@ -18,11 +18,6 @@ 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
|
||||
)
|
||||
|
||||
|
||||
def _generate_random_text(word_count: int) -> str:
|
||||
"""Generate random text with approximately the specified word count."""
|
||||
@@ -228,6 +223,10 @@ def server_with_chunked_processing():
|
||||
"0.8",
|
||||
]
|
||||
|
||||
# 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
|
||||
|
||||
|
||||
@@ -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
|
||||
|
||||
|
||||
Reference in New Issue
Block a user