[ROCm][CI] Add ROCm misc ops and env tests (#53279)
Signed-off-by: Divakar Verma <divakar.verma@amd.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
co-authored by
Andreas Karatzas
parent
6c58595c2d
commit
c0adee9231
@@ -0,0 +1,226 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""ROCm platform and extension smoke tests.
|
||||
|
||||
These tests catch ROCm-specific regressions early:
|
||||
|
||||
1. **Extension import smoke tests**: Verify that ROCm native extensions load
|
||||
without crashing. Build/linking issues surface here before runtime.
|
||||
|
||||
2. **GCN arch parsing tests**: Verify _capability_from_gcn_arch correctly
|
||||
parses all known GPU architectures. Wrong parsing causes silent feature
|
||||
gate failures or crashes from selecting unsupported kernels.
|
||||
|
||||
3. **AITER ops availability tests**: Verify rocm_aiter_ops accessor methods
|
||||
work without crashing, regardless of hardware support.
|
||||
|
||||
4. **Env var propagation tests**: Verify ROCm-specific env vars correctly
|
||||
gate their respective features (e.g., shuffle KV cache, skinny GEMM).
|
||||
|
||||
Related coverage:
|
||||
- AITER kernel numerics: ``tests/kernels/core/test_rocm_aiter_ops.py``
|
||||
- Attention backend selection: ``tests/v1/attention/test_rocm_attention_backends_selection.py``
|
||||
- GEMM dispatch: ``tests/model_executor/layers/test_rocm_unquantized_gemm.py``
|
||||
""" # noqa: E501
|
||||
|
||||
import importlib
|
||||
|
||||
import pytest
|
||||
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
pytestmark = pytest.mark.skipif(
|
||||
not current_platform.is_rocm(), reason="ROCm-specific tests"
|
||||
)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Extension import smoke tests
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def test_rocm_extension_imports():
|
||||
"""ROCm native extension loads without error.
|
||||
|
||||
Catches build/linking issues before runtime. If this fails, the ROCm
|
||||
build is broken and nothing else will work.
|
||||
"""
|
||||
import vllm._rocm_C # noqa: F401
|
||||
|
||||
|
||||
def test_aiter_ops_module_imports():
|
||||
"""AITER ops module imports without error.
|
||||
|
||||
The _aiter_ops module registers custom ops and initializes AITER state.
|
||||
Import failures here indicate missing dependencies or incompatible
|
||||
AITER versions.
|
||||
"""
|
||||
import vllm._aiter_ops # noqa: F401
|
||||
|
||||
|
||||
def test_rocm_platform_module_imports():
|
||||
"""ROCm platform module imports without error.
|
||||
|
||||
The platform module queries hardware via amdsmi and sets up arch flags.
|
||||
Import failures indicate missing amdsmi or HIP runtime issues.
|
||||
"""
|
||||
import vllm.platforms.rocm # noqa: F401
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# GCN arch parsing tests
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
("gcn_arch", "expected_major", "expected_minor"),
|
||||
[
|
||||
pytest.param("gfx90a", 9, 0, id="MI200-gfx90a"),
|
||||
pytest.param("gfx942", 9, 4, id="MI300X-gfx942"),
|
||||
pytest.param("gfx950", 9, 5, id="MI350-gfx950"),
|
||||
pytest.param("gfx1100", 11, 0, id="RX7900-gfx1100"),
|
||||
pytest.param("gfx1101", 11, 0, id="RX7800-gfx1101"),
|
||||
pytest.param("gfx1151", 11, 5, id="StrixHalo-gfx1151"),
|
||||
pytest.param("gfx1200", 12, 0, id="RDNA4-gfx1200"),
|
||||
pytest.param("gfx1201", 12, 0, id="RX9070-gfx1201"),
|
||||
pytest.param("gfx1250", 12, 5, id="CDNA5-gfx1250"),
|
||||
],
|
||||
)
|
||||
def test_capability_from_gcn_arch_known_gpus(gcn_arch, expected_major, expected_minor):
|
||||
"""_capability_from_gcn_arch parses known GPU architectures correctly.
|
||||
|
||||
This is critical for feature gating - wrong parsing causes:
|
||||
- FP8 kernels selected on hardware that doesn't support them
|
||||
- MX format support incorrectly enabled/disabled
|
||||
- Custom allreduce used on unsupported hardware
|
||||
"""
|
||||
from vllm.platforms.rocm import _capability_from_gcn_arch
|
||||
|
||||
result = _capability_from_gcn_arch(gcn_arch)
|
||||
assert result is not None, f"Failed to parse {gcn_arch}"
|
||||
assert result == (expected_major, expected_minor), (
|
||||
f"Wrong capability for {gcn_arch}: got {result}, "
|
||||
f"expected ({expected_major}, {expected_minor})"
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"gcn_arch",
|
||||
[
|
||||
pytest.param("not_a_gfx_string", id="non-gfx-string"),
|
||||
pytest.param("cuda_arch", id="cuda-style"),
|
||||
pytest.param("", id="empty-string"),
|
||||
],
|
||||
)
|
||||
def test_capability_from_gcn_arch_returns_none_for_non_gfx(gcn_arch):
|
||||
"""_capability_from_gcn_arch returns None for non-GFX strings.
|
||||
|
||||
Non-GFX strings should return None so the caller can fall back to
|
||||
torch.cuda.get_device_capability.
|
||||
"""
|
||||
from vllm.platforms.rocm import _capability_from_gcn_arch
|
||||
|
||||
result = _capability_from_gcn_arch(gcn_arch)
|
||||
assert result is None
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"gcn_arch",
|
||||
[
|
||||
pytest.param("gfx8", id="too-few-digits"),
|
||||
pytest.param("gfx12345", id="too-many-digits"),
|
||||
pytest.param("gfx700", id="unsupported-major-7"),
|
||||
pytest.param("gfx1500", id="future-major-15"),
|
||||
],
|
||||
)
|
||||
def test_capability_from_gcn_arch_raises_for_malformed(gcn_arch):
|
||||
"""_capability_from_gcn_arch raises ValueError for malformed GFX strings.
|
||||
|
||||
Malformed strings that look like GFX but can't be parsed should raise
|
||||
so the user gets a clear error message asking them to file an issue.
|
||||
"""
|
||||
from vllm.platforms.rocm import _capability_from_gcn_arch
|
||||
|
||||
with pytest.raises(ValueError, match="GCN arch"):
|
||||
_capability_from_gcn_arch(gcn_arch)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# AITER ops availability tests
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def test_rocm_aiter_ops_accessor_methods_dont_crash():
|
||||
"""rocm_aiter_ops accessor methods work without crashing.
|
||||
|
||||
These methods gate kernel selection. Even if AITER isn't available,
|
||||
they should return False gracefully rather than crashing.
|
||||
"""
|
||||
from vllm._aiter_ops import rocm_aiter_ops
|
||||
|
||||
# These should all return bool without crashing
|
||||
assert isinstance(rocm_aiter_ops.is_enabled(), (bool, type(None)))
|
||||
assert isinstance(rocm_aiter_ops.is_linear_enabled(), (bool, type(None)))
|
||||
assert isinstance(rocm_aiter_ops.is_fused_moe_enabled(), (bool, type(None)))
|
||||
assert isinstance(rocm_aiter_ops.is_mla_enabled(), (bool, type(None)))
|
||||
assert isinstance(rocm_aiter_ops.is_mha_enabled(), (bool, type(None)))
|
||||
|
||||
|
||||
def test_rocm_aiter_ops_is_aiter_found():
|
||||
"""is_aiter_found returns consistent results."""
|
||||
from vllm._aiter_ops import IS_AITER_FOUND, is_aiter_found
|
||||
|
||||
# The cached global should match the function result
|
||||
assert is_aiter_found() == IS_AITER_FOUND
|
||||
|
||||
|
||||
def test_rocm_aiter_ops_refresh_env_variables_doesnt_crash():
|
||||
"""refresh_env_variables works without crashing.
|
||||
|
||||
This is called after monkeypatching env vars in tests. It should
|
||||
never crash, even with unusual env var combinations.
|
||||
"""
|
||||
from vllm._aiter_ops import rocm_aiter_ops
|
||||
|
||||
# Should not raise
|
||||
rocm_aiter_ops.refresh_env_variables()
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Env var propagation tests
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
@pytest.mark.parametrize("enabled", [True, False])
|
||||
def test_shuffle_kv_cache_env_propagates_to_rocm_aiter_ops(enabled, monkeypatch):
|
||||
"""VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT propagates to AITER state."""
|
||||
import vllm.envs as envs
|
||||
from vllm._aiter_ops import rocm_aiter_ops
|
||||
|
||||
monkeypatch.setenv("VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT", "1" if enabled else "0")
|
||||
importlib.reload(envs)
|
||||
rocm_aiter_ops.refresh_env_variables()
|
||||
|
||||
assert envs.VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT is enabled
|
||||
assert rocm_aiter_ops.is_shuffle_kv_cache_enabled() is enabled
|
||||
|
||||
|
||||
@pytest.mark.parametrize("enabled", [True, False])
|
||||
def test_skinny_gemm_env_controls_rocm_fp8_scaled_mm_support(enabled, monkeypatch):
|
||||
"""VLLM_ROCM_USE_SKINNY_GEMM gates ROCm FP8 scaled-mm kernel eligibility."""
|
||||
import vllm.envs as envs
|
||||
from vllm.model_executor.kernels.linear.scaled_mm.rocm import (
|
||||
ROCmFP8ScaledMMLinearKernel,
|
||||
)
|
||||
|
||||
monkeypatch.setenv("VLLM_ROCM_USE_SKINNY_GEMM", "1" if enabled else "0")
|
||||
importlib.reload(envs)
|
||||
monkeypatch.setattr("vllm.platforms.rocm.on_mi3xx", lambda: True)
|
||||
|
||||
supported, reason = ROCmFP8ScaledMMLinearKernel.is_supported()
|
||||
|
||||
assert supported is enabled
|
||||
if enabled:
|
||||
assert reason is None
|
||||
else:
|
||||
assert reason == "requires VLLM_ROCM_USE_SKINNY_GEMM to be enabled."
|
||||
Reference in New Issue
Block a user