[ROCm][CI] Move ROCm AITER quantization tests (#40938)
Signed-off-by: Andreas Karatzas <akaratza@amd.com> Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com> Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Andreas Karatzas
co-authored by
OpenAI Codex
OpenAI Codex
parent
d785eb51ce
commit
5fa8ca971a
@@ -1544,11 +1544,13 @@ steps:
|
||||
- vllm/_aiter_ops.py
|
||||
- tests/kernels/attention/test_rocm_aiter_mla_decode.py
|
||||
- tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
|
||||
- tests/kernels/attention/test_rocm_aiter_mla_fp8_support.py
|
||||
- tests/kernels/attention/test_rocm_aiter_mla_op_registration.py
|
||||
commands:
|
||||
- rocm-smi
|
||||
- pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
|
||||
- pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_decode.py
|
||||
- pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_fp8_support.py
|
||||
- pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_op_registration.py
|
||||
|
||||
- label: Kernels Attention Test %N # TBD
|
||||
@@ -1664,6 +1666,7 @@ steps:
|
||||
- tests/kernels/quant_utils.py
|
||||
- tests/kernels/utils.py
|
||||
- vllm/_aiter_ops.py
|
||||
- vllm/kernels/aiter_ops.py
|
||||
- vllm/_custom_ops.py
|
||||
- vllm/envs.py
|
||||
- vllm/platforms/rocm.py
|
||||
@@ -2452,24 +2455,6 @@ steps:
|
||||
- uv pip install --system conch-triton-kernels
|
||||
- VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
|
||||
|
||||
#----------------------------------------------------------- mi300 · rocm ------------------------------------------------------------#
|
||||
|
||||
- label: ROCm AITER Ops Test # TBD
|
||||
timeout_in_minutes: 180
|
||||
mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
|
||||
dind: false
|
||||
agent_pool: mi300_1
|
||||
working_dir: "/vllm-workspace/tests"
|
||||
source_file_dependencies:
|
||||
- vllm/_aiter_ops.py
|
||||
- vllm/envs.py
|
||||
- vllm/platforms/rocm.py
|
||||
- tests/rocm/aiter/
|
||||
- vllm/v1/attention/backends/mla/rocm_aiter_mla.py
|
||||
- vllm/v1/attention/selector.py
|
||||
commands:
|
||||
- pytest -v -s rocm/aiter/
|
||||
|
||||
#--------------------------------------------------------- mi300 · samplers ----------------------------------------------------------#
|
||||
|
||||
- label: Samplers Test # TBD
|
||||
@@ -3461,11 +3446,13 @@ steps:
|
||||
- vllm/_aiter_ops.py
|
||||
- tests/kernels/attention/test_rocm_aiter_mla_decode.py
|
||||
- tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
|
||||
- tests/kernels/attention/test_rocm_aiter_mla_fp8_support.py
|
||||
- tests/kernels/attention/test_rocm_aiter_mla_op_registration.py
|
||||
commands:
|
||||
- rocm-smi
|
||||
- pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
|
||||
- pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_decode.py
|
||||
- pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_fp8_support.py
|
||||
- pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_op_registration.py
|
||||
|
||||
- label: Kernels Attention Test %N # TBD
|
||||
@@ -3525,6 +3512,7 @@ steps:
|
||||
- tests/kernels/quant_utils.py
|
||||
- tests/kernels/utils.py
|
||||
- vllm/_aiter_ops.py
|
||||
- vllm/kernels/aiter_ops.py
|
||||
- vllm/_custom_ops.py
|
||||
- vllm/envs.py
|
||||
- vllm/platforms/rocm.py
|
||||
|
||||
@@ -183,6 +183,7 @@ steps:
|
||||
- tests/kernels/quantization
|
||||
- tests/kernels/quantization/test_rocm_skinny_gemms.py
|
||||
- vllm/_aiter_ops.py
|
||||
- vllm/kernels/aiter_ops.py
|
||||
- vllm/platforms/rocm.py
|
||||
- vllm/model_executor/kernels/
|
||||
depends_on:
|
||||
|
||||
@@ -0,0 +1,137 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Tests for ROCm AITER MLA FP8 support detection."""
|
||||
|
||||
import sys
|
||||
import types
|
||||
from typing import Any
|
||||
from unittest.mock import patch
|
||||
|
||||
import pytest
|
||||
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
_SKIP_UNSUPPORTED_AITER_HARDWARE = True
|
||||
if current_platform.is_rocm():
|
||||
from vllm.platforms.rocm import get_cdna_version
|
||||
|
||||
_SKIP_UNSUPPORTED_AITER_HARDWARE = get_cdna_version() <= 2
|
||||
|
||||
pytestmark = pytest.mark.skipif(
|
||||
not current_platform.is_rocm(), reason="ROCm-specific tests"
|
||||
)
|
||||
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def reset_aiter_mla_support_cache(monkeypatch: pytest.MonkeyPatch) -> None:
|
||||
import vllm._aiter_ops as aiter_ops
|
||||
|
||||
monkeypatch.setattr(aiter_ops, "_AITER_MLA_SUPPORTS_FP8", None)
|
||||
|
||||
|
||||
def _install_fake_aiter_modules(
|
||||
monkeypatch: pytest.MonkeyPatch, *, supports_fp8: bool
|
||||
) -> None:
|
||||
aiter_mod: Any = types.ModuleType("aiter")
|
||||
mla_mod: Any = types.ModuleType("aiter.mla")
|
||||
|
||||
if supports_fp8:
|
||||
|
||||
def mla_decode_fwd_with_fp8(
|
||||
q,
|
||||
kv_buffer,
|
||||
kv_indptr,
|
||||
kv_indices,
|
||||
o,
|
||||
sm_scale,
|
||||
q_scale=None,
|
||||
kv_scale=None,
|
||||
):
|
||||
return None
|
||||
|
||||
mla_decode_fwd: Any = mla_decode_fwd_with_fp8
|
||||
|
||||
else:
|
||||
|
||||
def mla_decode_fwd_without_fp8(
|
||||
q,
|
||||
kv_buffer,
|
||||
kv_indptr,
|
||||
kv_indices,
|
||||
o,
|
||||
sm_scale,
|
||||
):
|
||||
return None
|
||||
|
||||
mla_decode_fwd = mla_decode_fwd_without_fp8
|
||||
|
||||
mla_mod.mla_decode_fwd = mla_decode_fwd
|
||||
aiter_mod.mla = mla_mod
|
||||
|
||||
monkeypatch.setitem(sys.modules, "aiter", aiter_mod)
|
||||
monkeypatch.setitem(sys.modules, "aiter.mla", mla_mod)
|
||||
|
||||
|
||||
def test_aiter_mla_fp8_support_detects_fp8_signature(monkeypatch):
|
||||
"""The support check should detect q_scale and kv_scale parameters."""
|
||||
from vllm._aiter_ops import _check_aiter_mla_fp8_support
|
||||
|
||||
_install_fake_aiter_modules(monkeypatch, supports_fp8=True)
|
||||
|
||||
assert _check_aiter_mla_fp8_support() is True
|
||||
|
||||
|
||||
def test_aiter_mla_fp8_support_rejects_missing_fp8_signature(monkeypatch):
|
||||
"""The support check should return False when FP8 params are absent."""
|
||||
from vllm._aiter_ops import _check_aiter_mla_fp8_support
|
||||
|
||||
_install_fake_aiter_modules(monkeypatch, supports_fp8=False)
|
||||
|
||||
assert _check_aiter_mla_fp8_support() is False
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
_SKIP_UNSUPPORTED_AITER_HARDWARE,
|
||||
reason="Installed AITER MLA FP8 check requires CDNA 3 or newer",
|
||||
)
|
||||
def test_installed_aiter_mla_supports_fp8():
|
||||
"""Supported ROCm CI must provide AITER with MLA FP8 scaling."""
|
||||
from vllm._aiter_ops import (
|
||||
_check_aiter_mla_fp8_support,
|
||||
is_aiter_found_and_supported,
|
||||
)
|
||||
|
||||
assert is_aiter_found_and_supported(), (
|
||||
"AITER must be installed on supported ROCm hardware"
|
||||
)
|
||||
assert _check_aiter_mla_fp8_support() is True
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"error_type",
|
||||
[ImportError, ModuleNotFoundError, AttributeError, ValueError, TypeError],
|
||||
)
|
||||
def test_aiter_mla_fp8_support_handles_signature_errors(monkeypatch, error_type):
|
||||
"""The support check should fail closed on import or signature problems."""
|
||||
import vllm._aiter_ops as aiter_ops
|
||||
from vllm._aiter_ops import _check_aiter_mla_fp8_support
|
||||
|
||||
_install_fake_aiter_modules(monkeypatch, supports_fp8=True)
|
||||
|
||||
with patch("inspect.signature", side_effect=error_type("boom")):
|
||||
assert _check_aiter_mla_fp8_support() is False
|
||||
assert aiter_ops._AITER_MLA_SUPPORTS_FP8 is False
|
||||
|
||||
|
||||
def test_aiter_mla_fp8_support_result_is_cached(monkeypatch):
|
||||
"""The support check should reuse the cached result on later calls."""
|
||||
import inspect
|
||||
|
||||
from vllm._aiter_ops import _check_aiter_mla_fp8_support
|
||||
|
||||
_install_fake_aiter_modules(monkeypatch, supports_fp8=True)
|
||||
|
||||
with patch("inspect.signature", wraps=inspect.signature) as signature_mock:
|
||||
assert _check_aiter_mla_fp8_support() is True
|
||||
assert _check_aiter_mla_fp8_support() is True
|
||||
assert signature_mock.call_count == 1
|
||||
@@ -1,118 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Unit tests for AITER MLA FP8 support detection.
|
||||
|
||||
These tests verify that the _check_aiter_mla_fp8_support() function
|
||||
correctly handles various error conditions without crashing.
|
||||
"""
|
||||
|
||||
from unittest.mock import patch
|
||||
|
||||
import pytest
|
||||
|
||||
|
||||
class TestAiterMlaFp8SupportCheck:
|
||||
"""Test cases for _check_aiter_mla_fp8_support() function."""
|
||||
|
||||
def setup_method(self):
|
||||
"""Reset the global cache before each test."""
|
||||
import vllm._aiter_ops as aiter_ops
|
||||
|
||||
aiter_ops._AITER_MLA_SUPPORTS_FP8 = None
|
||||
|
||||
@patch("vllm._aiter_ops.is_aiter_found_and_supported", return_value=True)
|
||||
def test_import_error_handling(self, mock_supported):
|
||||
"""Test that ImportError is handled gracefully."""
|
||||
import vllm._aiter_ops as aiter_ops
|
||||
from vllm._aiter_ops import _check_aiter_mla_fp8_support
|
||||
|
||||
aiter_ops._AITER_MLA_SUPPORTS_FP8 = None
|
||||
|
||||
# Should return False without raising
|
||||
with patch(
|
||||
"inspect.signature",
|
||||
side_effect=ImportError("No module"),
|
||||
):
|
||||
result = _check_aiter_mla_fp8_support()
|
||||
assert result is False
|
||||
|
||||
@patch("vllm._aiter_ops.is_aiter_found_and_supported", return_value=True)
|
||||
def test_module_not_found_error_handling(self, mock_supported):
|
||||
"""Test that ModuleNotFoundError is handled gracefully."""
|
||||
import vllm._aiter_ops as aiter_ops
|
||||
from vllm._aiter_ops import _check_aiter_mla_fp8_support
|
||||
|
||||
aiter_ops._AITER_MLA_SUPPORTS_FP8 = None
|
||||
|
||||
with patch(
|
||||
"inspect.signature",
|
||||
side_effect=ModuleNotFoundError("Module not found"),
|
||||
):
|
||||
# Should return False without raising
|
||||
assert _check_aiter_mla_fp8_support() is False
|
||||
# Cache should be set to False
|
||||
assert aiter_ops._AITER_MLA_SUPPORTS_FP8 is False
|
||||
|
||||
@patch("vllm._aiter_ops.is_aiter_found_and_supported", return_value=True)
|
||||
def test_attribute_error_handling(self, mock_supported):
|
||||
"""Test that AttributeError is handled gracefully."""
|
||||
import vllm._aiter_ops as aiter_ops
|
||||
from vllm._aiter_ops import _check_aiter_mla_fp8_support
|
||||
|
||||
aiter_ops._AITER_MLA_SUPPORTS_FP8 = None
|
||||
|
||||
with patch(
|
||||
"inspect.signature",
|
||||
side_effect=AttributeError("No attribute"),
|
||||
):
|
||||
assert _check_aiter_mla_fp8_support() is False
|
||||
assert aiter_ops._AITER_MLA_SUPPORTS_FP8 is False
|
||||
|
||||
@patch("vllm._aiter_ops.is_aiter_found_and_supported", return_value=True)
|
||||
def test_value_error_handling(self, mock_supported):
|
||||
"""Test that ValueError is handled gracefully (no signature)."""
|
||||
import vllm._aiter_ops as aiter_ops
|
||||
from vllm._aiter_ops import _check_aiter_mla_fp8_support
|
||||
|
||||
aiter_ops._AITER_MLA_SUPPORTS_FP8 = None
|
||||
|
||||
with patch(
|
||||
"inspect.signature",
|
||||
side_effect=ValueError("No signature"),
|
||||
):
|
||||
assert _check_aiter_mla_fp8_support() is False
|
||||
assert aiter_ops._AITER_MLA_SUPPORTS_FP8 is False
|
||||
|
||||
@patch("vllm._aiter_ops.is_aiter_found_and_supported", return_value=True)
|
||||
def test_type_error_handling(self, mock_supported):
|
||||
"""Test that TypeError is handled gracefully (not callable)."""
|
||||
import vllm._aiter_ops as aiter_ops
|
||||
from vllm._aiter_ops import _check_aiter_mla_fp8_support
|
||||
|
||||
aiter_ops._AITER_MLA_SUPPORTS_FP8 = None
|
||||
|
||||
with patch(
|
||||
"inspect.signature",
|
||||
side_effect=TypeError("Not a callable"),
|
||||
):
|
||||
assert _check_aiter_mla_fp8_support() is False
|
||||
assert aiter_ops._AITER_MLA_SUPPORTS_FP8 is False
|
||||
|
||||
@patch("vllm._aiter_ops.is_aiter_found_and_supported", return_value=True)
|
||||
def test_result_caching(self, mock_supported):
|
||||
"""Test that the result is cached after first check."""
|
||||
import vllm._aiter_ops as aiter_ops
|
||||
|
||||
# Set cache to True
|
||||
aiter_ops._AITER_MLA_SUPPORTS_FP8 = True
|
||||
|
||||
from vllm._aiter_ops import _check_aiter_mla_fp8_support
|
||||
|
||||
# Should return cached value without re-checking
|
||||
result = _check_aiter_mla_fp8_support()
|
||||
assert result is True
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
pytest.main([__file__, "-v"])
|
||||
Reference in New Issue
Block a user