[ROCm][CI] Fix ROCm attention backend validation for head sizes, block sizes, and compute capability checks (#36292)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
@@ -29,11 +29,18 @@ def mock_vllm_config():
|
||||
|
||||
@pytest.fixture
|
||||
def mock_on_gfx9():
|
||||
"""Mock the on_gfx9 function to return True."""
|
||||
"""Mock gfx9 arch detection to return True."""
|
||||
with patch("vllm.platforms.rocm.on_gfx9", return_value=True):
|
||||
yield
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def mock_on_mi3xx():
|
||||
"""Mock mi3xx arch detection to return True."""
|
||||
with patch("vllm.platforms.rocm.on_mi3xx", return_value=True):
|
||||
yield
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"env_vars, selected_backend, expected_backend_path",
|
||||
[
|
||||
@@ -122,6 +129,7 @@ def test_standard_attention_backend_selection(
|
||||
expected_backend_path,
|
||||
mock_vllm_config,
|
||||
mock_on_gfx9,
|
||||
mock_on_mi3xx,
|
||||
monkeypatch,
|
||||
):
|
||||
"""Test standard attention backend selection with various configurations."""
|
||||
@@ -313,16 +321,16 @@ def test_mla_backend_selection(
|
||||
assert backend_path == expected_backend_path
|
||||
|
||||
|
||||
def test_aiter_fa_requires_gfx9(mock_vllm_config):
|
||||
"""Test that ROCM_AITER_FA requires gfx9 architecture."""
|
||||
def test_aiter_fa_requires_mi3xx(mock_vllm_config):
|
||||
"""Test that ROCM_AITER_FA requires mi3xx architecture."""
|
||||
from vllm.platforms.rocm import RocmPlatform
|
||||
|
||||
# Mock on_gfx9 to return False
|
||||
# Mock on_mi3xx to return False (used by supports_compute_capability)
|
||||
with (
|
||||
patch("vllm.platforms.rocm.on_gfx9", return_value=False),
|
||||
patch("vllm.platforms.rocm.on_mi3xx", return_value=False),
|
||||
pytest.raises(
|
||||
ValueError,
|
||||
match="only supported on gfx9",
|
||||
match="compute capability not supported",
|
||||
),
|
||||
):
|
||||
attn_selector_config = AttentionSelectorConfig(
|
||||
@@ -342,11 +350,12 @@ def test_aiter_fa_requires_gfx9(mock_vllm_config):
|
||||
|
||||
|
||||
def test_sparse_not_supported(mock_vllm_config):
|
||||
"""Test that sparse attention is not supported on ROCm."""
|
||||
"""Test that sparse MLA without use_mla flag raises an error."""
|
||||
from vllm.platforms.rocm import RocmPlatform
|
||||
|
||||
with pytest.raises(
|
||||
AssertionError, match="Sparse MLA backend on ROCm only supports block size 1"
|
||||
ValueError,
|
||||
match="No valid attention backend found",
|
||||
):
|
||||
attn_selector_config = AttentionSelectorConfig(
|
||||
head_size=128,
|
||||
|
||||
Reference in New Issue
Block a user