[ROCm][Perf][MLA] Add AITER FlashAttention MLA prefill backend (ROCM_AITER_FA) (#45033)

Signed-off-by: Xavier Aguilar <xavier.aguilarfruto@amd.com>
Signed-off-by: Xavier Aguilar <Xavier.AguilarFruto@amd.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
This commit is contained in:
xaguilar-amd
2026-06-28 07:52:00 -07:00
committed by GitHub
co-authored by TJian
parent 6eb63a1da6
commit 5ecae3266c
5 changed files with 269 additions and 0 deletions
@@ -133,3 +133,20 @@ def test_clear_override():
def test_unknown_backend_name_raises():
with pytest.raises(ValueError, match="Unknown MLA prefill backend"):
MLAPrefillBackendEnum["NONEXISTENT"]
def test_rocm_aiter_fa_registered():
"""ROCM_AITER_FA is a known backend pointing at the AITER FA class."""
assert "ROCM_AITER_FA" in MLAPrefillBackendEnum.__members__
path = MLAPrefillBackendEnum.ROCM_AITER_FA.get_path()
assert path == (
"vllm.v1.attention.backends.mla.prefill.aiter_flash_attn."
"AiterFlashAttnPrefillBackend"
)
backend_cls = MLAPrefillBackendEnum.ROCM_AITER_FA.get_class()
assert backend_cls.get_name() == "ROCM_AITER_FA"
# The AITER FA path is the fp16/bf16 generic-varlen prefill path.
assert backend_cls.supports_dtype(torch.bfloat16)
assert backend_cls.supports_dtype(torch.float16)
@@ -14,6 +14,7 @@ from vllm.v1.attention.backends.mla.prefill.registry import MLAPrefillBackendEnu
from vllm.v1.attention.backends.mla.prefill.selector import (
MLAPrefillSelectorConfig,
_auto_select_mla_prefill_backend,
_get_mla_prefill_backend_priorities,
get_mla_prefill_backend,
)
@@ -166,6 +167,7 @@ class TestAutoSelectMLAPrefillBackend:
return
with (
patch("vllm.platforms.current_platform") as mock_platform,
patch.object(
MLAPrefillBackendEnum.FLASH_ATTN,
"get_class",
@@ -173,6 +175,8 @@ class TestAutoSelectMLAPrefillBackend:
),
patch.object(trtllm_cls, "validate_configuration", return_value=[]),
):
# Force the non-ROCm priority on the Blackwell.
mock_platform.is_rocm.return_value = False
backend = _auto_select_mla_prefill_backend(
capability,
selector_config,
@@ -272,6 +276,121 @@ class TestBackendValidation:
assert invalid_reasons == []
class TestROCmAiterFAPrefillSelection:
"""Tests for the ROCm AITER FlashAttention MLA prefill backend."""
def test_rocm_priorities_prefer_aiter_fa(self):
"""On ROCm, ROCM_AITER_FA is tried first, FLASH_ATTN as fallback."""
with patch("vllm.platforms.current_platform") as mock_platform:
mock_platform.is_rocm.return_value = True
priorities = _get_mla_prefill_backend_priorities(
DeviceCapability(major=9, minor=5)
)
assert priorities == [
MLAPrefillBackendEnum.ROCM_AITER_FA,
MLAPrefillBackendEnum.FLASH_ATTN,
]
def test_supported_dtypes_are_fp16_bf16_only(self):
from vllm.v1.attention.backends.mla.prefill.aiter_flash_attn import (
AiterFlashAttnPrefillBackend,
)
assert AiterFlashAttnPrefillBackend.supports_dtype(torch.bfloat16)
assert AiterFlashAttnPrefillBackend.supports_dtype(torch.float16)
# FP8 is served by the separate AITER ASM backend, not this one.
assert not AiterFlashAttnPrefillBackend.supports_dtype(torch.float8_e4m3fn)
def test_supports_compute_capability_on_rocm(self):
from vllm.v1.attention.backends.mla.prefill import aiter_flash_attn as mod
# Gating is decided by on_mi3xx(), not by capability
capability = MagicMock()
with patch.object(mod.current_platform, "is_rocm", return_value=False):
assert not mod.AiterFlashAttnPrefillBackend.supports_compute_capability(
capability
)
with (
patch.object(mod.current_platform, "is_rocm", return_value=True),
patch("vllm.platforms.rocm.on_mi3xx", return_value=False),
):
assert not mod.AiterFlashAttnPrefillBackend.supports_compute_capability(
capability
)
with (
patch.object(mod.current_platform, "is_rocm", return_value=True),
patch("vllm.platforms.rocm.on_mi3xx", return_value=True),
):
assert mod.AiterFlashAttnPrefillBackend.supports_compute_capability(
capability
)
def test_is_available_delegates_to_rocm_aiter_ops(self):
from vllm._aiter_ops import rocm_aiter_ops
from vllm.v1.attention.backends.mla.prefill import aiter_flash_attn as mod
with patch.object(rocm_aiter_ops, "is_enabled", return_value=False):
assert not mod.AiterFlashAttnPrefillBackend.is_available()
with patch.object(rocm_aiter_ops, "is_enabled", return_value=True):
assert mod.AiterFlashAttnPrefillBackend.is_available()
def test_auto_select_prefers_aiter_fa_on_rocm(self):
from vllm.v1.attention.backends.mla.prefill.aiter_flash_attn import (
AiterFlashAttnPrefillBackend,
)
# gfx gating is simulated via the mocked validate_configuration,
# not the capability.
capability = MagicMock()
selector_config = MLAPrefillSelectorConfig(dtype=torch.bfloat16)
with (
patch("vllm.platforms.current_platform") as mock_platform,
patch.object(
AiterFlashAttnPrefillBackend,
"validate_configuration",
return_value=[],
),
):
mock_platform.is_rocm.return_value = True
backend = _auto_select_mla_prefill_backend(capability, selector_config)
assert backend.get_name() == "ROCM_AITER_FA"
def test_auto_select_falls_back_to_flash_attn_when_aiter_invalid(self):
from vllm.v1.attention.backends.mla.prefill.aiter_flash_attn import (
AiterFlashAttnPrefillBackend,
)
try:
flash_attn_cls = MLAPrefillBackendEnum.FLASH_ATTN.get_class()
except ImportError:
pytest.skip("FLASH_ATTN backend not available")
return
# the fallback is forced by the mocked validate_configuration,
# not the capability.
capability = MagicMock()
selector_config = MLAPrefillSelectorConfig(dtype=torch.bfloat16)
with (
patch("vllm.platforms.current_platform") as mock_platform,
patch.object(
AiterFlashAttnPrefillBackend,
"validate_configuration",
return_value=["compute capability not supported"],
),
patch.object(flash_attn_cls, "validate_configuration", return_value=[]),
):
mock_platform.is_rocm.return_value = True
backend = _auto_select_mla_prefill_backend(capability, selector_config)
assert backend.get_name() == "FLASH_ATTN"
class TestMLAPrefillBackendParsing:
"""Tests for string-based mla_prefill_backend parsing from CLI args."""