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 12, 17:15 UTC
Simplify `from_blob` usage in `get_cuda_view_from_cpu_tensor` (#29027)
Signed-off-by: Jane Xu <janeyx@meta.com>
Jane (Yuan) Xu committed
Nov 22, 2025 at 05:35 UTC
e6309acdba3a26e803d1ea7f66804f4ad30c2b9a
1 file changed
+3
-8
csrc/cuda_view.cu
+3
-8
@@ -22,15 +22,10 @@ torch::Tensor get_cuda_view_from_cpu_tensor(torch::Tensor& cpu_tensor) {
22
auto strides = cpu_tensor.strides();
23
auto options = cpu_tensor.options().device(torch::kCUDA);
24
25
- // from_blob signature: from_blob(void *data, IntArrayRef sizes, ..., Deleter,
26
- // const TensorOptions &) Provide a no-op deleter. The CPU tensor holds the
27
- // memory, so we don't free it here.
28
- auto deleter = [](void*) {
29
- // no-op, since the memory is owned by the original CPU tensor
30
- };
31
-
25
+ // use default no-op deleter, since the memory is owned by the original CPU
26
+ // tensor
27
torch::Tensor cuda_tensor =
33
- torch::from_blob(device_ptr, sizes, strides, deleter, options);
28
+ torch::from_blob(device_ptr, sizes, strides, options);
29
30
TORCH_CHECK(cuda_tensor.device().is_cuda(),
31
"Resulting tensor is not on CUDA device");