[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
@@ -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
|
||||
|
||||
|
||||
Reference in New Issue
Block a user