[ROCm] Remove stale SDPA and skinny GEMM workarounds (#50907)
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com> Signed-off-by: Andreas Karatzas <akaratza@amd.com> Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
co-authored by
OpenAI Codex
parent
4f2f31b82e
commit
1ab2801dde
@@ -9,7 +9,7 @@ import pytest_asyncio
|
||||
from transformers import AutoProcessor
|
||||
|
||||
from tests.entrypoints.multimodal.conftest import TEST_IMAGE_ASSETS
|
||||
from tests.utils import ROCM_ENV_OVERRIDES, ROCM_EXTRA_ARGS, RemoteOpenAIServer
|
||||
from tests.utils import ROCM_EXTRA_ARGS, RemoteOpenAIServer
|
||||
from vllm.multimodal.media import MediaWithBytes
|
||||
from vllm.multimodal.utils import encode_image_url, fetch_image
|
||||
from vllm.platforms import current_platform
|
||||
@@ -92,7 +92,6 @@ def server():
|
||||
# ROCm: Increase timeouts to handle potential network delays and slower
|
||||
# video processing when downloading multiple videos from external sources
|
||||
env_overrides = {
|
||||
**ROCM_ENV_OVERRIDES,
|
||||
**(
|
||||
{
|
||||
"VLLM_VIDEO_FETCH_TIMEOUT": "120",
|
||||
|
||||
@@ -10,7 +10,7 @@ import pytest
|
||||
import pytest_asyncio
|
||||
|
||||
# downloading lora to test lora requests
|
||||
from tests.utils import ROCM_ENV_OVERRIDES, ROCM_EXTRA_ARGS, RemoteOpenAIServer
|
||||
from tests.utils import ROCM_EXTRA_ARGS, RemoteOpenAIServer
|
||||
|
||||
# any model with a chat template should work here
|
||||
MODEL_NAME = "Qwen/Qwen3-0.6B"
|
||||
@@ -144,9 +144,7 @@ def server():
|
||||
"--enforce-eager",
|
||||
] + ROCM_EXTRA_ARGS
|
||||
|
||||
with RemoteOpenAIServer(
|
||||
MODEL_NAME, args, env_dict=ROCM_ENV_OVERRIDES
|
||||
) as remote_server:
|
||||
with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
|
||||
yield remote_server
|
||||
|
||||
|
||||
|
||||
@@ -7,7 +7,6 @@ import pytest
|
||||
|
||||
from vllm import LLM, PoolingParams
|
||||
from vllm.distributed import cleanup_dist_env_and_memory
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
MODEL_NAME = "intfloat/multilingual-e5-small"
|
||||
|
||||
@@ -30,12 +29,6 @@ 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(
|
||||
@@ -45,7 +38,6 @@ def llm():
|
||||
gpu_memory_utilization=0.75,
|
||||
enforce_eager=True,
|
||||
seed=0,
|
||||
attention_config=attention_config,
|
||||
)
|
||||
|
||||
yield weakref.proxy(llm)
|
||||
|
||||
@@ -7,7 +7,6 @@ import pytest
|
||||
import pytest_asyncio
|
||||
|
||||
from tests.utils import RemoteOpenAIServer
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
MODEL_NAME = "sentence-transformers/all-MiniLM-L12-v2"
|
||||
max_model_len = 128
|
||||
@@ -39,10 +38,6 @@ 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
|
||||
|
||||
|
||||
@@ -1,28 +0,0 @@
|
||||
# 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,
|
||||
)
|
||||
@@ -11,7 +11,6 @@ from tests.models.language.pooling_mteb_test.mteb_embed_utils import (
|
||||
run_mteb_embed_task,
|
||||
)
|
||||
from tests.utils import RemoteOpenAIServer
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
os.environ["VLLM_LOGGING_LEVEL"] = "WARNING"
|
||||
|
||||
@@ -23,10 +22,6 @@ 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
|
||||
|
||||
|
||||
@@ -8,7 +8,6 @@ import torch.nn.functional as F
|
||||
|
||||
from vllm import LLM, EmbeddingRequestOutput, PoolingParams
|
||||
from vllm.distributed import cleanup_dist_env_and_memory
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.tasks import PoolingTask
|
||||
|
||||
MODEL_NAME = "intfloat/multilingual-e5-small"
|
||||
@@ -20,12 +19,6 @@ embedding_size = 384
|
||||
|
||||
@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(
|
||||
@@ -35,7 +28,6 @@ def llm():
|
||||
gpu_memory_utilization=0.75,
|
||||
enforce_eager=True,
|
||||
seed=0,
|
||||
attention_config=attention_config,
|
||||
)
|
||||
assert embedding_size == llm.model_config.embedding_size
|
||||
|
||||
|
||||
@@ -50,14 +50,6 @@ input_tokens = [
|
||||
2,
|
||||
]
|
||||
|
||||
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)
|
||||
|
||||
# On ROCm, floating-point reductions in attention and GEMM kernels are
|
||||
# non-associative and sensitive to batch geometry. Force LLM instances
|
||||
# into an identical, deterministic execution mode:
|
||||
@@ -507,11 +499,17 @@ async def test_base64_embedding(hf_model, client: openai.AsyncOpenAI, model_name
|
||||
"The best thing about vLLM is that it supports many different models",
|
||||
]
|
||||
|
||||
def check_outputs(outputs: list[list[float]]) -> None:
|
||||
# The embeddings endpoint executes list elements as separate scheduler
|
||||
# requests, so use the same per-request geometry for the HF reference.
|
||||
for input_text, output in zip(input_texts, outputs, strict=True):
|
||||
run_embedding_correctness_test(hf_model, [input_text], [output])
|
||||
|
||||
responses_float = await client.embeddings.create(
|
||||
input=input_texts, model=model_name, encoding_format="float"
|
||||
)
|
||||
float_data = [d.embedding for d in responses_float.data]
|
||||
run_embedding_correctness_test(hf_model, input_texts, float_data)
|
||||
check_outputs(float_data)
|
||||
|
||||
responses_base64 = await client.embeddings.create(
|
||||
input=input_texts, model=model_name, encoding_format="base64"
|
||||
@@ -522,14 +520,14 @@ async def test_base64_embedding(hf_model, client: openai.AsyncOpenAI, model_name
|
||||
np.frombuffer(base64.b64decode(data.embedding), dtype="float32").tolist()
|
||||
)
|
||||
|
||||
run_embedding_correctness_test(hf_model, input_texts, base64_data)
|
||||
check_outputs(base64_data)
|
||||
|
||||
# Default response is float32 decoded from base64 by OpenAI Client
|
||||
responses_default = await client.embeddings.create(
|
||||
input=input_texts, model=model_name
|
||||
)
|
||||
default_data = [d.embedding for d in responses_default.data]
|
||||
run_embedding_correctness_test(hf_model, input_texts, default_data)
|
||||
check_outputs(default_data)
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
|
||||
@@ -12,7 +12,6 @@ from tests.models.language.pooling.embed_utils import run_embedding_correctness_
|
||||
from tests.models.utils import EmbedModelInfo
|
||||
from tests.utils import ROCM_EXTRA_ARGS, RemoteOpenAIServer
|
||||
from vllm.entrypoints.pooling.embed.protocol import EmbeddingResponse
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
MODELS = [
|
||||
EmbedModelInfo("intfloat/multilingual-e5-small", is_matryoshka=False),
|
||||
@@ -57,10 +56,6 @@ 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
|
||||
|
||||
|
||||
@@ -16,7 +16,6 @@ import pytest_asyncio
|
||||
|
||||
from tests.utils import RemoteOpenAIServer
|
||||
from vllm.entrypoints.pooling.embed.protocol import EmbeddingResponse
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
|
||||
def _generate_random_text(word_count: int) -> str:
|
||||
@@ -225,10 +224,6 @@ 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
|
||||
|
||||
|
||||
@@ -8,7 +8,6 @@ import pytest
|
||||
from tests.entrypoints.pooling.scoring.util import EncoderScoringHfRunner
|
||||
from vllm import LLM
|
||||
from vllm.distributed import cleanup_dist_env_and_memory
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
MODEL_NAME = "intfloat/multilingual-e5-small"
|
||||
PROMPT = "The chef prepared a delicious meal."
|
||||
@@ -29,12 +28,6 @@ DTYPE = "half"
|
||||
|
||||
@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(
|
||||
@@ -44,7 +37,6 @@ def llm():
|
||||
gpu_memory_utilization=0.75,
|
||||
enforce_eager=True,
|
||||
seed=0,
|
||||
attention_config=attention_config,
|
||||
)
|
||||
|
||||
yield weakref.proxy(llm)
|
||||
|
||||
@@ -8,7 +8,6 @@ from tests.entrypoints.pooling.scoring.util import EncoderScoringHfRunner
|
||||
from tests.utils import RemoteOpenAIServer
|
||||
from vllm.entrypoints.pooling.pooling.protocol import PoolingResponse
|
||||
from vllm.entrypoints.pooling.scoring.protocol import RerankResponse, ScoreResponse
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
MODEL_NAME = "BAAI/bge-base-en-v1.5"
|
||||
input_text = "This product was excellent and exceeded my expectations"
|
||||
@@ -30,10 +29,6 @@ TEXTS_2 = [
|
||||
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
|
||||
|
||||
|
||||
@@ -13,7 +13,6 @@ from tests.models.language.pooling_mteb_test.mteb_score_utils import (
|
||||
run_mteb_rerank,
|
||||
)
|
||||
from tests.utils import RemoteOpenAIServer
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
os.environ["VLLM_LOGGING_LEVEL"] = "WARNING"
|
||||
|
||||
@@ -25,10 +24,6 @@ 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
|
||||
|
||||
|
||||
@@ -9,7 +9,6 @@ import torch
|
||||
from tests.models.utils import softmax
|
||||
from vllm import LLM, PoolingParams
|
||||
from vllm.distributed import cleanup_dist_env_and_memory
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
MODEL_NAME = "tomaarsen/Qwen3-Reranker-0.6B-seq-cls"
|
||||
PROMPT = "The chef prepared a delicious meal."
|
||||
@@ -26,12 +25,6 @@ TEXTS_2 = [
|
||||
|
||||
@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(
|
||||
@@ -41,7 +34,6 @@ def llm():
|
||||
gpu_memory_utilization=0.75,
|
||||
enforce_eager=True,
|
||||
seed=0,
|
||||
attention_config=attention_config,
|
||||
)
|
||||
|
||||
yield weakref.proxy(llm)
|
||||
|
||||
@@ -9,7 +9,6 @@ import torch.nn.functional as F
|
||||
from tests.utils import RemoteOpenAIServer
|
||||
from vllm.entrypoints.pooling.pooling.protocol import PoolingResponse
|
||||
from vllm.entrypoints.pooling.scoring.protocol import RerankResponse, ScoreResponse
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
MODEL_NAME = "BAAI/bge-reranker-base"
|
||||
DTYPE = "half"
|
||||
@@ -32,10 +31,6 @@ TEXTS_2 = [
|
||||
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
|
||||
|
||||
|
||||
@@ -53,12 +53,6 @@ BACKEND_ABS_TOL: dict[str, float] = {
|
||||
"FLEX_ATTENTION": 0.006,
|
||||
}
|
||||
|
||||
# ROCm: disable skinny GEMM to avoid non-deterministic results from
|
||||
# atomic reductions in wvSplitKrc kernel.
|
||||
# See: https://github.com/vllm-project/vllm/pull/33493#issuecomment-3906083975
|
||||
ROCM_ENV_OVERRIDES = (
|
||||
{"VLLM_ROCM_USE_SKINNY_GEMM": "0"} if current_platform.is_rocm() else {}
|
||||
)
|
||||
# ROCm: disable prefix caching and eliminate batch variance to reduce
|
||||
# test flakiness.
|
||||
ROCM_EXTRA_ARGS = (
|
||||
@@ -132,7 +126,6 @@ def server(request):
|
||||
args += ["--attention-config", json.dumps({"backend": backend})]
|
||||
args += ROCM_EXTRA_ARGS
|
||||
|
||||
env = dict(ROCM_ENV_OVERRIDES)
|
||||
if backend != "ROCM_AITER_FA":
|
||||
env["VLLM_ROCM_USE_AITER"] = "0"
|
||||
|
||||
|
||||
@@ -7,7 +7,6 @@ import pytest
|
||||
|
||||
from vllm import LLM
|
||||
from vllm.distributed import cleanup_dist_env_and_memory
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
from .util import ColBERTScoringHfRunner
|
||||
|
||||
@@ -32,12 +31,6 @@ DTYPE = "half"
|
||||
|
||||
@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(
|
||||
@@ -47,7 +40,6 @@ def llm():
|
||||
gpu_memory_utilization=0.75,
|
||||
enforce_eager=True,
|
||||
seed=0,
|
||||
attention_config=attention_config,
|
||||
)
|
||||
|
||||
yield weakref.proxy(llm)
|
||||
|
||||
@@ -7,7 +7,6 @@ import pytest
|
||||
|
||||
from vllm import LLM
|
||||
from vllm.distributed import cleanup_dist_env_and_memory
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
from .util import make_base64_image, make_image_mm_param
|
||||
|
||||
@@ -16,12 +15,6 @@ MODEL_NAME = "vidore/colpali-v1.3-hf"
|
||||
|
||||
@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(
|
||||
@@ -31,7 +24,6 @@ def llm():
|
||||
gpu_memory_utilization=0.75,
|
||||
enforce_eager=True,
|
||||
seed=0,
|
||||
attention_config=attention_config,
|
||||
)
|
||||
|
||||
yield weakref.proxy(llm)
|
||||
|
||||
@@ -7,7 +7,6 @@ import pytest
|
||||
from vllm import LLM, PoolingRequestOutput
|
||||
from vllm.config import PoolerConfig
|
||||
from vllm.distributed import cleanup_dist_env_and_memory
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.tasks import PoolingTask
|
||||
|
||||
MODEL_NAME = "intfloat/multilingual-e5-small"
|
||||
@@ -19,12 +18,6 @@ embedding_size = 384
|
||||
|
||||
@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(
|
||||
@@ -35,7 +28,6 @@ def llm():
|
||||
gpu_memory_utilization=0.75,
|
||||
enforce_eager=True,
|
||||
seed=0,
|
||||
attention_config=attention_config,
|
||||
)
|
||||
assert embedding_size == llm.model_config.embedding_size
|
||||
|
||||
|
||||
@@ -9,8 +9,6 @@ processor in the pytest parent process to keep os.fork() in sibling
|
||||
tests (e.g. test_weight_transfer_llm.py) deadlock-free.
|
||||
"""
|
||||
|
||||
import os
|
||||
|
||||
import httpx
|
||||
import pytest
|
||||
import pytest_asyncio
|
||||
@@ -41,10 +39,7 @@ def server():
|
||||
"--no-enable-prefix-caching",
|
||||
]
|
||||
|
||||
envs = os.environ.copy()
|
||||
envs["VLLM_ROCM_USE_SKINNY_GEMM"] = "0"
|
||||
|
||||
with RemoteOpenAIServer(MODEL_NAME, args, env_dict=envs) as remote_server:
|
||||
with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
|
||||
yield remote_server
|
||||
|
||||
|
||||
|
||||
@@ -2,7 +2,6 @@
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
import json
|
||||
import os
|
||||
|
||||
import httpx
|
||||
import pytest
|
||||
@@ -79,11 +78,7 @@ def server(request):
|
||||
else [str(extra_args)]
|
||||
)
|
||||
|
||||
envs = os.environ.copy()
|
||||
# See: https://github.com/vllm-project/vllm/pull/33493#issuecomment-3888060787
|
||||
envs["VLLM_ROCM_USE_SKINNY_GEMM"] = "0"
|
||||
|
||||
with RemoteOpenAIServer(MODEL_NAME, args, env_dict=envs) as remote_server:
|
||||
with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
|
||||
yield remote_server
|
||||
|
||||
|
||||
|
||||
@@ -16,19 +16,6 @@ def add_attention_backend(server_args, attention_config):
|
||||
server_args.extend(["--attention-backend", attention_config["backend"]])
|
||||
|
||||
|
||||
@pytest.fixture(scope="module")
|
||||
def rocm_aiter_fa_attention():
|
||||
"""Return attention config for transcription/translation tests on ROCm.
|
||||
|
||||
On ROCm, audio tests require ROCM_AITER_FA attention backend.
|
||||
"""
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
if current_platform.is_rocm():
|
||||
return {"backend": "ROCM_AITER_FA"}
|
||||
return None
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def mary_had_lamb():
|
||||
path = AudioAsset("mary_had_lamb").get_local_path()
|
||||
|
||||
@@ -10,15 +10,13 @@ import pybase64 as base64
|
||||
import pytest
|
||||
import websockets
|
||||
|
||||
from tests.entrypoints.speech_to_text.conftest import add_attention_backend
|
||||
from tests.utils import ROCM_ENV_OVERRIDES, ROCM_EXTRA_ARGS, RemoteOpenAIServer
|
||||
from tests.utils import ROCM_EXTRA_ARGS, RemoteOpenAIServer
|
||||
from vllm.assets.audio import AudioAsset
|
||||
from vllm.multimodal.media.audio import load_audio
|
||||
|
||||
# Increase engine iteration timeout for ROCm where first-use JIT compilation
|
||||
# can exceed the default 60s, causing a silent deadlock in feed_tokens.
|
||||
REALTIME_ENV_OVERRIDES = {
|
||||
**ROCM_ENV_OVERRIDES,
|
||||
"VLLM_ENGINE_ITERATION_TIMEOUT_S": "600",
|
||||
}
|
||||
|
||||
@@ -72,17 +70,13 @@ def mary_had_lamb_audio_chunks() -> list[str]:
|
||||
|
||||
@pytest.mark.asyncio
|
||||
@pytest.mark.parametrize("model_name", [MODEL_NAME])
|
||||
async def test_multi_chunk_streaming(
|
||||
model_name, mary_had_lamb_audio_chunks, rocm_aiter_fa_attention
|
||||
):
|
||||
async def test_multi_chunk_streaming(model_name, mary_had_lamb_audio_chunks):
|
||||
"""Test streaming multiple audio chunks before committing."""
|
||||
server_args = ["--enforce-eager", "--max-model-len", "2048"]
|
||||
|
||||
if model_name.startswith("mistralai"):
|
||||
server_args += MISTRAL_FORMAT_ARGS
|
||||
|
||||
add_attention_backend(server_args, rocm_aiter_fa_attention)
|
||||
|
||||
with RemoteOpenAIServer(
|
||||
model_name, server_args, env_dict=REALTIME_ENV_OVERRIDES
|
||||
) as remote_server:
|
||||
@@ -174,7 +168,7 @@ async def test_multi_chunk_streaming(
|
||||
@pytest.mark.asyncio
|
||||
@pytest.mark.parametrize("model_name", [MODEL_NAME])
|
||||
async def test_empty_commit_does_not_crash_engine(
|
||||
model_name, mary_had_lamb_audio_chunks, rocm_aiter_fa_attention
|
||||
model_name, mary_had_lamb_audio_chunks
|
||||
):
|
||||
"""Test that committing without audio does not crash the engine.
|
||||
|
||||
@@ -189,8 +183,6 @@ async def test_empty_commit_does_not_crash_engine(
|
||||
if model_name.startswith("mistralai"):
|
||||
server_args += MISTRAL_FORMAT_ARGS
|
||||
|
||||
add_attention_backend(server_args, rocm_aiter_fa_attention)
|
||||
|
||||
with RemoteOpenAIServer(
|
||||
model_name, server_args, env_dict=REALTIME_ENV_OVERRIDES
|
||||
) as remote_server:
|
||||
@@ -273,17 +265,13 @@ async def test_empty_commit_does_not_crash_engine(
|
||||
|
||||
@pytest.mark.asyncio
|
||||
@pytest.mark.parametrize("model_name", [MODEL_NAME])
|
||||
async def test_session_update_invalid_model_returns_error(
|
||||
model_name, rocm_aiter_fa_attention
|
||||
):
|
||||
async def test_session_update_invalid_model_returns_error(model_name):
|
||||
"""Test that session.update with an invalid model returns an error."""
|
||||
server_args = ["--enforce-eager", "--max-model-len", "2048"]
|
||||
|
||||
if model_name.startswith("mistralai"):
|
||||
server_args += MISTRAL_FORMAT_ARGS
|
||||
|
||||
add_attention_backend(server_args, rocm_aiter_fa_attention)
|
||||
|
||||
with RemoteOpenAIServer(
|
||||
model_name, server_args, env_dict=REALTIME_ENV_OVERRIDES
|
||||
) as remote_server:
|
||||
@@ -305,9 +293,7 @@ async def test_session_update_invalid_model_returns_error(
|
||||
|
||||
@pytest.mark.asyncio
|
||||
@pytest.mark.parametrize("model_name", [MODEL_NAME])
|
||||
async def test_commit_without_session_update_returns_error(
|
||||
model_name, rocm_aiter_fa_attention
|
||||
):
|
||||
async def test_commit_without_session_update_returns_error(model_name):
|
||||
"""Test that committing before validating the model returns an error
|
||||
and does not fall through to processing."""
|
||||
server_args = ["--enforce-eager", "--max-model-len", "2048"]
|
||||
@@ -315,8 +301,6 @@ async def test_commit_without_session_update_returns_error(
|
||||
if model_name.startswith("mistralai"):
|
||||
server_args += MISTRAL_FORMAT_ARGS
|
||||
|
||||
add_attention_backend(server_args, rocm_aiter_fa_attention)
|
||||
|
||||
with RemoteOpenAIServer(
|
||||
model_name, server_args, env_dict=REALTIME_ENV_OVERRIDES
|
||||
) as remote_server:
|
||||
|
||||
@@ -6,8 +6,7 @@ import json
|
||||
|
||||
import pytest
|
||||
|
||||
from tests.entrypoints.speech_to_text.conftest import add_attention_backend
|
||||
from tests.utils import ROCM_ENV_OVERRIDES, ROCM_EXTRA_ARGS, RemoteOpenAIServer
|
||||
from tests.utils import ROCM_EXTRA_ARGS, RemoteOpenAIServer
|
||||
|
||||
MISTRAL_FORMAT_ARGS = [
|
||||
"--tokenizer_mode",
|
||||
@@ -66,18 +65,14 @@ async def transcribe_and_check(
|
||||
@pytest.mark.parametrize(
|
||||
"model_name", ["mistralai/Voxtral-Mini-3B-2507", "Qwen/Qwen3-ASR-0.6B"]
|
||||
)
|
||||
async def test_basic_audio(mary_had_lamb, model_name, rocm_aiter_fa_attention):
|
||||
async def test_basic_audio(mary_had_lamb, model_name):
|
||||
server_args = ["--enforce-eager", *ROCM_EXTRA_ARGS]
|
||||
|
||||
if model_name.startswith("mistralai"):
|
||||
server_args += MISTRAL_FORMAT_ARGS
|
||||
|
||||
add_attention_backend(server_args, rocm_aiter_fa_attention)
|
||||
|
||||
# Based on https://github.com/openai/openai-cookbook/blob/main/examples/Whisper_prompting_guide.ipynb.
|
||||
with RemoteOpenAIServer(
|
||||
model_name, server_args, env_dict=ROCM_ENV_OVERRIDES
|
||||
) as remote_server:
|
||||
with RemoteOpenAIServer(model_name, server_args) as remote_server:
|
||||
client = remote_server.get_async_client()
|
||||
await transcribe_and_check(
|
||||
client,
|
||||
@@ -90,7 +85,7 @@ async def test_basic_audio(mary_had_lamb, model_name, rocm_aiter_fa_attention):
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_basic_audio_with_lora(mary_had_lamb, rocm_aiter_fa_attention):
|
||||
async def test_basic_audio_with_lora(mary_had_lamb):
|
||||
"""Ensure STT (transcribe) requests can pass LoRA through to generate."""
|
||||
# ROCm SPECIFIC CONFIGURATION:
|
||||
# To ensure the test passes on ROCm, we modify the max model length to 512.
|
||||
@@ -112,12 +107,8 @@ async def test_basic_audio_with_lora(mary_had_lamb, rocm_aiter_fa_attention):
|
||||
"1",
|
||||
]
|
||||
|
||||
add_attention_backend(server_args, rocm_aiter_fa_attention)
|
||||
|
||||
# Based on https://github.com/openai/openai-cookbook/blob/main/examples/Whisper_prompting_guide.ipynb.
|
||||
with RemoteOpenAIServer(
|
||||
model_name, server_args, env_dict=ROCM_ENV_OVERRIDES
|
||||
) as remote_server:
|
||||
with RemoteOpenAIServer(model_name, server_args) as remote_server:
|
||||
client = remote_server.get_async_client()
|
||||
await transcribe_and_check(
|
||||
client,
|
||||
@@ -133,18 +124,15 @@ async def test_basic_audio_with_lora(mary_had_lamb, rocm_aiter_fa_attention):
|
||||
@pytest.mark.parametrize(
|
||||
"model_name", ["google/gemma-3n-E2B-it", "Qwen/Qwen3-ASR-0.6B"]
|
||||
)
|
||||
async def test_basic_audio_foscolo(foscolo, rocm_aiter_fa_attention, model_name):
|
||||
async def test_basic_audio_foscolo(foscolo, model_name):
|
||||
# Gemma accuracy on some of the audio samples we use is particularly bad,
|
||||
# hence we use a different one here. WER is evaluated separately.
|
||||
server_args = ["--enforce-eager", *ROCM_EXTRA_ARGS]
|
||||
|
||||
add_attention_backend(server_args, rocm_aiter_fa_attention)
|
||||
|
||||
with RemoteOpenAIServer(
|
||||
model_name,
|
||||
server_args,
|
||||
max_wait_seconds=480,
|
||||
env_dict=ROCM_ENV_OVERRIDES,
|
||||
) as remote_server:
|
||||
client = remote_server.get_async_client()
|
||||
await transcribe_and_check(
|
||||
|
||||
@@ -28,27 +28,33 @@ def _get_rocm_attention_config(model_name):
|
||||
|
||||
Whisper uses cross-attention (ENCODER_DECODER) which ROCM_AITER_FA does
|
||||
not support. For Whisper we use ROCM_AITER_UNIFIED_ATTN (or TRITON_ATTN
|
||||
as fallback); other models can use ROCM_AITER_FA.
|
||||
as fallback). Other models use ROCM_AITER_FA only on supported CDNA3+
|
||||
hardware and otherwise use the platform default.
|
||||
"""
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
if not current_platform.is_rocm():
|
||||
return None
|
||||
|
||||
if "whisper" in model_name.lower():
|
||||
try:
|
||||
from vllm.platforms.rocm import get_cdna_version
|
||||
try:
|
||||
from vllm.platforms.rocm import get_cdna_version
|
||||
|
||||
if get_cdna_version() > 2:
|
||||
return {"backend": "ROCM_AITER_UNIFIED_ATTN"}
|
||||
except ImportError:
|
||||
logger.warning(
|
||||
"Could not check cdna version from rocm platform, "
|
||||
"falling back to TRITON_ATTN for Whisper."
|
||||
)
|
||||
cdna_version = get_cdna_version()
|
||||
except ImportError:
|
||||
logger.warning(
|
||||
"Could not check cdna version from rocm platform; "
|
||||
"using a compatible fallback attention backend."
|
||||
)
|
||||
cdna_version = -1
|
||||
|
||||
if "whisper" in model_name.lower():
|
||||
if cdna_version > 2:
|
||||
return {"backend": "ROCM_AITER_UNIFIED_ATTN"}
|
||||
return {"backend": "TRITON_ATTN"}
|
||||
|
||||
return {"backend": "ROCM_AITER_FA"}
|
||||
if cdna_version > 2:
|
||||
return {"backend": "ROCM_AITER_FA"}
|
||||
return None
|
||||
|
||||
|
||||
def _get_server_args(attention_config):
|
||||
|
||||
Reference in New Issue
Block a user