[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
@@ -1,18 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Pytest configuration for vLLM pooling tests."""
import pytest
from vllm.platforms import current_platform
@pytest.fixture
def siglip_attention_config():
"""Return attention config for SigLIP tests on ROCm.
On ROCm, SigLIP tests require FLEX_ATTENTION backend.
"""
if current_platform.is_rocm():
return {"backend": "FLEX_ATTENTION"}
return None
@@ -39,7 +39,6 @@ def _run_test(
*,
dtype: str,
tokenization_kwargs: dict[str, Any] | None = None,
attention_config: dict[str, Any] | None = None,
) -> None:
if tokenization_kwargs is None:
tokenization_kwargs = {}
@@ -51,7 +50,6 @@ def _run_test(
enforce_eager=True,
max_model_len=64,
gpu_memory_utilization=0.7,
attention_config=attention_config,
) as vllm_model:
vllm_outputs = vllm_model.embed(
input_texts, images=input_images, tokenization_kwargs=tokenization_kwargs
@@ -96,7 +94,6 @@ def test_models_text(
hf_runner,
vllm_runner,
image_assets,
siglip_attention_config,
model: str,
dtype: str,
) -> None:
@@ -115,7 +112,6 @@ def test_models_text(
"padding": "max_length",
"max_length": 64,
}, # siglip2 was trained with this padding setting.
attention_config=siglip_attention_config,
)
@@ -125,7 +121,6 @@ def test_models_image(
hf_runner,
vllm_runner,
image_assets,
siglip_attention_config,
model: str,
dtype: str,
) -> None:
@@ -142,7 +137,6 @@ def test_models_image(
input_images,
model,
dtype=dtype,
attention_config=siglip_attention_config,
)
@@ -151,7 +145,6 @@ def test_models_image(
def test_models_text_image_no_crash(
vllm_runner,
image_assets,
siglip_attention_config,
model: str,
dtype: str,
) -> None:
@@ -165,7 +158,6 @@ def test_models_text_image_no_crash(
enforce_eager=True,
max_model_len=64,
gpu_memory_utilization=0.7,
attention_config=siglip_attention_config,
) as vllm_model:
with pytest.raises(ValueError, match="not both"):
vllm_model.embed(texts, images=images)