Merge pull request !3626 from guozhijian/fix_device_occupied_tdt_hung_r0.6

Merge pull request !3629 from xiefangqi/md_fix_numpyslice_r0.6

Merge pull request !3589 from guansongsong/fix_cache_core_for_r0.6

Merge pull request !3579 from meixiaowei/r0.6

Merge pull request !3566 from ms_yan/r0.6_api_format

Merge pull request !3599 from riemann_penn/merge_fix_sparse_doc_to_r0.6

Merge pull request !3595 from jinyaohui/sink_size

Merge pull request !3582 from xiefangqi/r0.6

Merge pull request !3572 from liyong126/r0.6_fix_save_pydoc_log

Merge pull request !3580 from gziyan/fix_allreduce_fusion

Merge pull request !3533 from dinghao/r0.6

Merge pull request !3564 from Xiaoda/15-r0.6-add-new-gpu-costmodel

Merge pull request !3530 from hewei/fix_parameter_bug_r0.6

Merge pull request !3539 from LiHongzhang/f50_t20_r

Merge pull request !3518 from hexia/fix_python_import_r0.6

Merge pull request !3501 from lirongzhen1/r0.6

Merge pull request !3524 from yoonlee666/bertci

When `collect_tensor_freq` is specified as `None`,
the `collect_tensor_freq` would be auto calculated.

The previous behavior is to collect at most 50 steps,
now changing to 20

Merge pull request !3482 from LiHongzhang/limit_summary_r

1. Parameter's init_data() should have no effect if default_input already set;
2. This bug is introduced by 'decouple ParamValue from python';
3. An unit test case is added to ensure the right init_data() behavior.

Merge pull request !3492 from yangyongjie/r0.6

Merge pull request !3509 from ZPaC/add-ps-test-cases

Merge pull request !3493 from Simson/push-to-r06

Merge pull request !3498 from YuJianfeng/r0.6

Merge pull request !3505 from riemann_penn/merge_eager_mode_enable_sparse_to_r0.6

Merge pull request !3499 from ZPaC/r0.6-delete-param-name-hard-code

Merge pull request !3470 from jiangzhiwen/fix_column_name_r_0_6

Merge pull request !3466 from kisnwang/r0.6-fix-cpu-nonop-fp16-error