【PIR API adaptor No.190、191】 Migrate paddle.scatter，paddle.scatter_nd_add into pir

[enkilee]

PR types

Others

PR changes

APIs

Description

• 新IR Python API适配升级 #58067

PIR API 推全升级

将 paddle.scatter 迁移升级至 pir，并更新单测 单测覆盖率：26/27
将 paddle.scatter_nd_add 迁移升级至 pir，并更新单测 单测覆盖率：11/11

scatter的static_test报错：显存不足

2023-11-03 17:31:19 ----------------------------------------------------------------------
2023-11-03 17:31:19 Traceback (most recent call last):
2023-11-03 17:31:19   File "/paddle/build/python/paddle/pir_utils.py", line 119, in impl
2023-11-03 17:31:19     func(*args, **kwargs)
2023-11-03 17:31:19   File "/mnt/paddle/build/test/legacy_test/test_scatter_op.py", line 711, in test_static_graph
2023-11-03 17:31:19     res = gpu_exe.run(
2023-11-03 17:31:19   File "/paddle/build/python/paddle/base/executor.py", line 1637, in run
2023-11-03 17:31:19     res = self._run_pir_impl(
2023-11-03 17:31:19   File "/paddle/build/python/paddle/base/executor.py", line 1950, in _run_pir_impl
2023-11-03 17:31:19     self._pir_feed_data(program, feed, scope)
2023-11-03 17:31:19   File "/paddle/build/python/paddle/base/executor.py", line 1269, in _pir_feed_data
2023-11-03 17:31:19     cur_feed = _as_lodtensor(cur_feed, self.place, var_type)
2023-11-03 17:31:19   File "/paddle/build/python/paddle/base/executor.py", line 721, in _as_lodtensor
2023-11-03 17:31:19     tensor.set(data, place)
2023-11-03 17:31:19 MemoryError: 
2023-11-03 17:31:19 
2023-11-03 17:31:19 --------------------------------------
2023-11-03 17:31:19 C++ Traceback (most recent call last):
2023-11-03 17:31:19 --------------------------------------
2023-11-03 17:31:19 0   float* phi::DenseTensor::mutable_data<float>(phi::Place const&, unsigned long)
2023-11-03 17:31:19 1   phi::DenseTensor::mutable_data(phi::Place const&, phi::DataType, unsigned long)
2023-11-03 17:31:19 2   phi::memory_utils::AllocShared(phi::Place const&, unsigned long)
2023-11-03 17:31:19 3   paddle::memory::AllocShared(phi::Place const&, unsigned long)
2023-11-03 17:31:19 4   paddle::memory::allocation::AllocatorFacade::AllocShared(phi::Place const&, unsigned long)
2023-11-03 17:31:19 5   paddle::memory::allocation::AllocatorFacade::Alloc(phi::Place const&, unsigned long)
2023-11-03 17:31:19 6   paddle::memory::allocation::StatAllocator::AllocateImpl(unsigned long)
2023-11-03 17:31:19 7   paddle::memory::allocation::Allocator::Allocate(unsigned long)
2023-11-03 17:31:19 8   paddle::memory::allocation::Allocator::Allocate(unsigned long)
2023-11-03 17:31:19 9   paddle::memory::allocation::Allocator::Allocate(unsigned long)
2023-11-03 17:31:19 10  paddle::memory::allocation::CUDAAllocator::AllocateImpl(unsigned long)
2023-11-03 17:31:19 11  std::string phi::enforce::GetCompleteTraceBackString<std::string >(std::string&&, char const*, int)
2023-11-03 17:31:19 12  phi::enforce::GetCurrentTraceBackString[abi:cxx11](bool)
2023-11-03 17:31:19 
2023-11-03 17:31:19 ----------------------
2023-11-03 17:31:19 Error Message Summary:
2023-11-03 17:31:19 ----------------------
2023-11-03 17:31:19 ResourceExhaustedError: 
2023-11-03 17:31:19 
2023-11-03 17:31:19 Out of memory error on GPU 0. Cannot allocate 10.260804GB memory on GPU 0, 11.500977GB memory has been allocated and available memory is only 4.271362GB.
2023-11-03 17:31:19 
2023-11-03 17:31:19 Please check whether there is any other process using GPU 0.
2023-11-03 17:31:19 1. If yes, please stop them, or start PaddlePaddle on another GPU.
2023-11-03 17:31:19 2. If no, please decrease the batch size of your model. 
2023-11-03 17:31:19  (at /paddle/paddle/fluid/memory/allocation/cuda_allocator.cc:86)

[YuanRisheng]

这个报错单测先取消在pir模式下执行，Paddle同学看一下这个问题

[MarioLulab]

@enkilee 这个报错单测先取消在pir模式下执行吧，然后这个 pr 先合入。Paddle 同学解决一下 ci oom 的问题~

[enkilee]

收到

[MarioLulab]

LGTM

[Aurelius84]

approved