mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 12:21:27 +01:00
e4e19d5beb
Summary:
added python API to disable nvfuser on certain opkind.
```
"_jit_set_nvfuser_skip_node_kind",
[](const std::string& op_name, bool flip = true) {
return fuser::cuda::skipNode(op_name, flip);
})
```
Args:
`op_name`: Symbol of op;
`flip`: flag indicating whether to flip the given op in the skip list.
Returns:
a bool flag indicating if `op_name` was already in the skip list.
The python example that disables the fusion of `aten::add` afterwards.
`torch._C._jit_set_nvfuser_skip_node_kind("aten::add", True) # returns False, as no op is in skip list by default`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/74520
Reviewed By: saketh-are
Differential Revision: D35046110
Pulled By: davidberard98
fbshipit-source-id: 689f5286513dbab206768823a852467b9f6b49b6
(cherry picked from commit 9a31129f7591ba2d393ab057b1cd137a6a25e7e8)
58 lines
1.8 KiB
C++
58 lines
1.8 KiB
C++
#pragma once
|
|
|
|
#include <c10/macros/Export.h>
|
|
#include <torch/csrc/jit/ir/ir.h>
|
|
#include <torch/csrc/jit/runtime/profiling_record.h>
|
|
|
|
/*
|
|
* This file contains APIs for cuda fuser;
|
|
*
|
|
* We use an empty static struct to hold the function pointers, which are
|
|
* registered separately. This is to support cpu-only compilation.
|
|
* Registration is done in torch/csrc/jit/codegen/cuda/register_interface.cpp
|
|
*/
|
|
|
|
namespace torch {
|
|
namespace jit {
|
|
namespace fuser {
|
|
namespace cuda {
|
|
|
|
TORCH_API std::atomic<bool>& getCudaFusionGuardMode();
|
|
|
|
TORCH_API bool getSingletonFusion();
|
|
TORCH_API bool setSingletonFusion(bool value);
|
|
TORCH_API bool getHorizontalFusion();
|
|
TORCH_API bool setHorizontalFusion(bool value);
|
|
|
|
// dummy struct to allow API registration
|
|
struct CudaFuserInterface {
|
|
void (*fn_compile_n)(Node*) = nullptr;
|
|
void (*fn_run_n_s)(const Node*, Stack&) = nullptr;
|
|
void (*fn_fuse_graph)(std::shared_ptr<Graph>&) = nullptr;
|
|
bool (*fn_can_fuse_n)(const Node*) = nullptr;
|
|
void (*fn_insert_profile_inodes)(ProfilingRecord* pr) = nullptr;
|
|
bool (*fn_profile_n)(const Node*) = nullptr;
|
|
bool (*fn_skip_n)(const std::string&, bool flip) = nullptr;
|
|
};
|
|
|
|
// Get interface, this is used by registration and user facing API internally
|
|
TORCH_API CudaFuserInterface* getFuserInterface();
|
|
|
|
TORCH_API void compileFusionGroup(Node* fusion_node);
|
|
TORCH_API void runFusionGroup(const Node* fusion_node, Stack& stack);
|
|
TORCH_API void fuseGraph(std::shared_ptr<Graph>&);
|
|
TORCH_API bool canFuseNode(const Node* node);
|
|
TORCH_API void InsertProfileNodesForCUDAFuser(ProfilingRecord* pr);
|
|
TORCH_API bool profileNode(const Node* node);
|
|
|
|
TORCH_API bool skipNode(const std::string& symbol_str, bool flip = true);
|
|
|
|
TORCH_API bool complyWith(
|
|
const at::Tensor& tensor,
|
|
const c10::TensorTypePtr& guard_tensor_type);
|
|
|
|
} // namespace cuda
|
|
} // namespace fuser
|
|
} // namespace jit
|
|
} // namespace torch
|