[XPU] Enable FLASH_ATTN on XPU (#1332)

Signed-off-by: Yan Ma <yan.ma@intel.com>
This commit is contained in:
Yan Ma authored and GitHub committed 2026-02-26 20:28:09 +08:00
1 parent 7b75c3993a
commit cf1f4c6338
3 files changed
+115 -46

No files matched your search

+9 -6
View File
@@ -15,6 +15,9 @@ import torch
from vllm_omni.diffusion.attention.backends.abstract import AttentionMetadata
from vllm_omni.diffusion.attention.backends.flash_attn import FlashAttentionImpl
from vllm_omni.diffusion.attention.backends.sdpa import SDPAImpl
from vllm_omni.platforms import current_omni_platform
is_gpu = current_omni_platform.is_cuda_alike() or current_omni_platform.is_xpu()
def create_attention_mask(batch_size: int, seq_len: int, valid_len: int, device: torch.device) -> torch.Tensor:
@@ -56,7 +59,7 @@ def pad_tensor(tensor: torch.Tensor, target_seq_len: int, pad_value: float = 0.0
return torch.cat([tensor, padding], dim=1)
@pytest.mark.skipif(not torch.cuda.is_available(), reason="FlashAttention requires CUDA")
@pytest.mark.skipif(not is_gpu, reason="FlashAttention requires CUDA or XPU")
def test_padding_equivalence():
"""
Case 1: Test that padded and unpadded inputs produce similar outputs.
@@ -68,7 +71,7 @@ def test_padding_equivalence():
Expected: Output A and Output B should be very close.
"""
device = torch.device("cuda")
device = torch.device(current_omni_platform.device_type)
dtype = torch.bfloat16
# Configuration
@@ -149,7 +152,7 @@ def test_padding_equivalence():
print("✓ Case 1 PASSED: Padded and unpadded outputs are very close!")
@pytest.mark.skipif(not torch.cuda.is_available(), reason="FlashAttention requires CUDA")
@pytest.mark.skipif(not is_gpu, reason="FlashAttention requires CUDA or XPU")
def test_fa_vs_sdpa():
"""
Case 2: Compare FlashAttention and SDPA backends with padding.
@@ -162,7 +165,7 @@ def test_fa_vs_sdpa():
Expected: FA and SDPA outputs should be very close.
"""
device = torch.device("cuda")
device = torch.device(current_omni_platform.device_type)
dtype = torch.bfloat16
# Configuration
@@ -265,8 +268,8 @@ if __name__ == "__main__":
print("Running FlashAttention Padding Tests...")
print("=" * 60)
# Try to run CUDA tests
if torch.cuda.is_available():
# Try to run tests
if is_gpu:
try:
print("\n[Running Case 1: Padding Equivalence for FA]")
test_padding_equivalence()