[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:
Andreas Karatzas
2026-08-11 14:19:16 -04:00
committed by GitHub
co-authored by OpenAI Codex
parent 4f2f31b82e
commit 1ab2801dde
56 changed files with 500 additions and 558 deletions
@@ -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)