A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 23:45 UTC
[XPU] Conditionally import CUDA-specific passes to avoid import errors on xpu platform (#21036)
Signed-off-by: chzhang <chaojun.zhang@intel.com>
Chaojun Zhang committed
Jul 24, 2025 at 23:23 UTC
d9f9a3fd96b8461680495dee8f004037eef61d04
1 file changed
+6
-3
vllm/compilation/pass_manager.py
+6
-3
@@ -5,12 +5,15 @@ from torch import fx as fx
5
6
from vllm.config import VllmConfig
7
from vllm.logger import init_logger
8
+from vllm.platforms import current_platform
9
+
10
+if current_platform.is_cuda_alike():
11
+ from .fusion import FusionPass
12
+ from .collective_fusion import AllReduceFusionPass, AsyncTPPass
13
+ from .fusion_attn import AttnFusionPass
14
15
from .activation_quant_fusion import ActivationQuantFusionPass
10
-from .collective_fusion import AllReduceFusionPass, AsyncTPPass
16
from .fix_functionalization import FixFunctionalizationPass
12
-from .fusion import FusionPass
13
-from .fusion_attn import AttnFusionPass
17
from .inductor_pass import CustomGraphPass, InductorPass, get_pass_context
18
from .noop_elimination import NoOpEliminationPass
19
from .sequence_parallelism import SequenceParallelismPass