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 8, 07:30 UTC
[Bugfix] Fix ROCm UVA CPU weight offloading broken by #32993 (#34543)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Feb 13, 2026 at 22:01 UTC
a0638d052db74ba28eada4768b9bbf98720b44a4
1 file changed
+1
-1
vllm/utils/torch_utils.py
+1
-1
@@ -683,7 +683,7 @@ def get_accelerator_view_from_cpu_tensor(cpu_tensor: torch.Tensor) -> torch.Tens
683
if current_platform.is_xpu():
684
assert cpu_tensor.is_pinned(), "CPU tensor must be pinned"
685
return torch.ops._C.get_xpu_view_from_cpu_tensor(cpu_tensor)
686
- elif current_platform.is_cuda():
686
+ elif current_platform.is_cuda() or current_platform.is_rocm():
687
return torch.ops._C.get_cuda_view_from_cpu_tensor(cpu_tensor)
688
else:
689
raise ValueError(