Merge pull request !5190 from anthonyaje/gpu_queue_size

Merge pull request !5629 from zhengjun10/master

Merge pull request !5626 from wangmin0104/master

Merge pull request !5549 from jianghui58/master

Merge pull request !5597 from liuyang/MT

Merge pull request !5616 from hewei/fix_tensor_from_numpy2

fix space_to_batch bug

Merge pull request !5633 from pengyongrong/stack

Merge pull request !5594 from Simson/push-to-opensource

Merge pull request !5624 from heleiwang/gnn_perf

Merge pull request !5628 from guozhijian/fix_tinybert_err

Merge pull request !5622 from baihuawei/0901

Merge pull request !5598 from kisnwang/async-run-graph

Merge pull request !5618 from flywind/fix_pynative_tuple_bug

Merge pull request !5609 from pengyongrong/concat

Merge pull request !5614 from chenzupeng/master-lite

Merge pull request !4922 from amongo/TupleTransform

Merge pull request !5617 from limingqi107/master

Merge pull request !5494 from zhaozhenlong/lite/op/optimize_resize_cache_line

Merge pull request !5596 from gukecai/profiling

add ops supported fp16

add dimension 0-3 for concat

Merge pull request !5577 from chenfei_mindspore/master

Merge pull request !5192 from Peilin/scatternd-type-support

Merge pull request !5461 from laiyongqiang/atomic_clean_every_master

Merge pull request !4958 from 34bunny/GPU-argmaxwithvalue-fix

add support for pynative pass

add testcases

Use buffer_info instead of py::array in TensorDataNumpy,
to avoid requesting buffer every time the data() is called.
This is also a workaround for py::array.request() hangup problem.

Merge pull request !5608 from guozhijian/fix_tinybert_var_error