* fix xpu kp compilation

* add depends

* lazy creating work queue

* fix dry_run

* Optimize prod's python implementation for dygraph.

* Change key_dim to head_dim.

* Add comment in unittest.

* Disable TF32 in unittest.

* Decoupling ipu sharding and modeling (#665)

* feat(shard): decoupling shard setting with modeling.

* fix(shard): split test cases to avoid failure.

* fix(shard): add function docs and fix typo.

* test(shard): add tests.

* test(shard): more test case.

* fix(): change ipu_index/stage default value to -1.

* fix format
Co-authored-by: Nczr-gc <96037699+czr-gc@users.noreply.github.com>

* set_state_dict not use state_dict hook

* add ut

* refine doc

* use tempfile to place temporary files

* update

* revert test_communicator

* fix test_dist_base

* rename yaml file

* fix merge conflict

* fix infrt

* add some kernel(csr*dense->csr, dense*dense->csr) of SparseTensor matmul

* fix CI

* fix CI

* fix comment

* fix comment

Use int64_t in GetGpuLaunchConfig1D and ElementwiseKernel as index type to support large tensor. (#43506)

* Change some data type from int to int64_t in GetGpuLaunchConfig1D to support large tensor.

* Use int64_t in ElementwiseKernel as index type to support large tensor.

* Refactor port.h

* Remove some unnecessary code

* Fix CI errors

* fix starganv2

* fix starganv2 stop_gradient end error

* fix edvr_starganv2

* fix mul kernel to fix optional ddx

* fix typo

* Support sequential run for standalone executor

* Add UTs

* Fix test_standalone_multiply_write

* Remove unnecessary UTs

* [MLU]: add elementwise_max mlu kernel

* [MLU]: add int32 support for elementwise maxk MLU kernel

* Refine ifelse early return

* update paddle.distributed.launch

* add sample code

* update shell

* fix typo

* fix typo

* update docs

* rm code

* fix doc 2

* fix doc 3

* fix doc 4
Co-authored-by: Nroot <root@sgjur-pod004-1.ipu.graphcore.cn>

* open inference_ut;test=windows_ci_inference

* inference_ut need onnx;test=windows_ci_inference

* disable trt_split_converter_test; use higher parallel level

* too high parallel will cause ut timeout

* change logic for optimize

* modifty

* optimize the backward speed of EinsumOp

* add cache optimizer for einsum op

* EinsumOp: fix new dygraph mode error

* fix bug

* change Cache->InnerCache

* fix code

* fix

* add nan inf utils for einsum op

* add as_extra

* memory optimizer for einsum

* update code

* fix the bug that _DataLoaderIterMultiProcess use time to generate the seed

* use np.random.randint to generate a base seed