mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-08 07:39:33 +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)
59 lines
1.9 KiB
C++
59 lines
1.9 KiB
C++
#pragma once
|
|
|
|
#include <c10/macros/Export.h>
|
|
#include <torch/csrc/jit/ir/ir.h>
|
|
#include <torch/csrc/jit/runtime/profiling_record.h>
|
|
|
|
#include <torch/csrc/jit/codegen/cuda/fusion.h>
|
|
|
|
/*
|
|
* This file handles Parsing PyTorch jit ir;
|
|
*
|
|
* It is used in two places:
|
|
* 1. When partitioning PyTorch jit ir to create prim::CudaFusionGroup, each
|
|
* node is queried by `isNodeParsible` to determine whether the node could
|
|
* be handled by the fuser (whether a given PyTorch jit operator should be
|
|
* merged);
|
|
* 2. lowering PyTorch jit ir to CUDA codegen ir.
|
|
* creates a `Fusion` by traversing a PyTorch jit graph.
|
|
*
|
|
* TODO: we could consider exposing API to allow custom registration of parsing
|
|
* rules for a given PyTorch jit operator.
|
|
*/
|
|
|
|
namespace torch {
|
|
namespace jit {
|
|
namespace fuser {
|
|
namespace cuda {
|
|
|
|
constexpr int kPwThreadX = 128;
|
|
constexpr int kFcdReductionThreadX = 128;
|
|
constexpr int kNonFcdReductionThreadX = 32;
|
|
constexpr int kNonFcdReductionThreadY = 32;
|
|
|
|
TORCH_CUDA_CU_API bool hasReductionNode(const Block* block);
|
|
TORCH_CUDA_CU_API bool isReductionToSizeNode(const Node* node);
|
|
TORCH_CUDA_CU_API bool isReductionNode(const Node* node);
|
|
|
|
TORCH_CUDA_CU_API bool hasNormalizationNode(const Block* block);
|
|
TORCH_CUDA_CU_API bool isNormalizationNode(const Node* node);
|
|
|
|
TORCH_CUDA_CU_API bool isElementWiseNode(const Node* node);
|
|
|
|
// returns whether or not a parsing function exists for the given node type.
|
|
TORCH_CUDA_CU_API bool isNodeParsible(const Node* node);
|
|
TORCH_CUDA_CU_API bool shouldProfileNode(const Node* node);
|
|
|
|
TORCH_CUDA_CU_API bool skipNodeKind(const std::string& symbol_str, bool flip);
|
|
|
|
void InsertProfileNodes(ProfilingRecord* pr);
|
|
|
|
// lowers PyTorch jit graph to `Fusion`.
|
|
TORCH_CUDA_CU_API std::unique_ptr<Fusion> parseJitIR(
|
|
const std::shared_ptr<Graph>& graph);
|
|
|
|
} // namespace cuda
|
|
} // namespace fuser
|
|
} // namespace jit
|
|
} // namespace torch
|