[ROCm] [PyTorch] Move to stable abi since ROCm upgraded to torch 2.11 (#47128)
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
@@ -2,7 +2,6 @@
|
||||
// cache.h, which is no longer included here after cache ops moved to
|
||||
// _C_stable_libtorch).
|
||||
#include <torch/all.h>
|
||||
#include "cuda_utils.h"
|
||||
#include "ops.h"
|
||||
#include "core/registration.h"
|
||||
#include <torch/library.h>
|
||||
@@ -18,19 +17,6 @@
|
||||
// https://docs.google.com/document/d/1_W62p8WJOQQUzPsJYa7s701JXt0qf2OfLub2sbkHOaU/edit#heading=h.ptttacy8y1u9
|
||||
// https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/README.md#annotations
|
||||
|
||||
TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
|
||||
// vLLM custom ops
|
||||
|
||||
#ifdef USE_ROCM
|
||||
// TODO: Remove this once we upgrade to torch 2.11.
|
||||
// ROCm still uses torch 2.10,
|
||||
// So we still need to use unstable torch ABI for now.
|
||||
ops.def("get_cuda_view_from_cpu_tensor(Tensor cpu_tensor) -> Tensor");
|
||||
ops.impl("get_cuda_view_from_cpu_tensor", torch::kCPU,
|
||||
&get_cuda_view_from_cpu_tensor);
|
||||
#endif
|
||||
}
|
||||
|
||||
#ifdef USE_ROCM
|
||||
TORCH_LIBRARY_FRAGMENT(CONCAT(TORCH_EXTENSION_NAME, _custom_ar), custom_ar) {
|
||||
// Quick Reduce all-reduce kernels (ROCm-only; stays on legacy _C).
|
||||
@@ -48,20 +34,6 @@ TORCH_LIBRARY_FRAGMENT(CONCAT(TORCH_EXTENSION_NAME, _custom_ar), custom_ar) {
|
||||
|
||||
custom_ar.def("qr_max_size", &qr_max_size);
|
||||
}
|
||||
|
||||
// TODO: Remove this once ROCm upgrade to torch 2.11.
|
||||
TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cuda_utils), cuda_utils) {
|
||||
// Cuda utils
|
||||
// Gets the specified device attribute.
|
||||
cuda_utils.def("get_device_attribute(int attribute, int device_id) -> int");
|
||||
cuda_utils.impl("get_device_attribute", &get_device_attribute);
|
||||
|
||||
// Gets the maximum shared memory per block device attribute.
|
||||
cuda_utils.def(
|
||||
"get_max_shared_memory_per_block_device_attribute(int device_id) -> int");
|
||||
cuda_utils.impl("get_max_shared_memory_per_block_device_attribute",
|
||||
&get_max_shared_memory_per_block_device_attribute);
|
||||
}
|
||||
#endif
|
||||
|
||||
REGISTER_EXTENSION(TORCH_EXTENSION_NAME)
|
||||
|
||||
Reference in New Issue
Block a user