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 13, 20:45 UTC
[Bugfix][plugin] fla crash on plugin (#27322)
Hank_ committed
Nov 4, 2025 at 05:27 UTC
ccd3e55e51d44bf3a17b2203a304c9609aa5dfe2
1 file changed
+3
-2
vllm/model_executor/layers/fla/ops/utils.py
+3
-2
@@ -17,6 +17,7 @@ from typing import Any, Literal
17
18
import torch
19
20
+from vllm.platforms import current_platform
21
from vllm.triton_utils import triton
22
23
logger = logging.getLogger(__name__)
@@ -137,8 +138,8 @@ def _check_platform() -> Literal["nvidia", "amd", "intel", "musa"]:
138
# For AMD GPUs, the triton backend is 'hip', while for Nvidia GPUs, the triton backend is 'cuda'.
139
# However, the torch backend is 'cuda' for both Nvidia and AMD GPUs.
140
# Therefore, we need to check the triton backend to determine the actual GPU vendor.
140
-device = get_available_device() if get_available_device() != "hip" else "cuda"
141
-device_torch_lib = getattr(torch, device)
141
+device = "cuda" if current_platform.is_cuda_alike() else get_available_device()
142
+device_torch_lib = getattr(torch, device, None)
143
device_platform = _check_platform()
144
145
is_amd = device_platform == "amd"