You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
OP move task from ernie-core to framework (#72957)
* init
* insert moe_combine
* init
* update yaml
* update python API
* delete useless header file
* remove supported by DCU
* add expand_modality_expert_id kernel
* reorder the new code and refine OP type
* add unit test
* add cal_aux_loss_op and build_src_rank_and_local_expert_id_op
* moegatedispatch init
* insert moegatedispatch
* remove DCU support
* fix-bugs
fix-bugs
fix-bugs
* fix log2 in windows maybe
* update header file format
* fix-bugs
* delete op test for pass CI
* add cmath header
* tmp
* pass int_bincount
* add moe_dispatch_bwd
* add moe_gate_dispatch
* fix-bugs
* fix optional Tensor
* update cal_aux_loss_kernel
* Finished moe_combine & expand_modality_expert_id integrate and optests.
* add python interface
* nosoftmax forward has finished
* finishi fused_rms_norm fwd
* finish rms_norm bwd
* finish rms norm bwd
* add optional in ops.yaml
* nosoftmax bwd has finished
* update python api
* Verified cal_aux_loss op and bwd.
* Verified build_src_rank_and_local_expert_id
* gate_dispatch_permute has finished
* Verified fused_rms_norm_ext(with bwd) and int_bincount.
* Add stage2 fwd and bwd optests.
* Clean print
* Fix conflict, move some headers.
* sync with dev
* Add incubate port.
* fix miscs
* Fix module issue
* Add missing yamls
* Fix stale package problems
* fix moe_combine bug.
* Fix miscs
* Align with original initializations.
* fix typos and pre-commit warnings
* Fix miscs
* try to pass CI
* format header file
* remove win32 supported
* check OP type
* remove optest for WIN & APPLE
* fix bug for (int32_t and int)
* rename fused_rms_norm op
* select op test env not for Volta
* fix openblas mistake
* CMake code format
* fix bugs in CPU
* CodeStyle format
* fix bugs in CPU
* fix bugs in CPU
* skip some op when CUDA<12.0
* skip op when CUDA<12.0
* fix bugs in CPU
---------
Co-authored-by: pesionzhao <pesionzhao@gmail.com>
Co-authored-by: feixi21 <1802550529@qq.com>
Co-authored-by: zhenghuaijin <zhenghuaijin@baidu.com>
0 commit comments