* add cmake bits about NCCL
* move example to examples/NNmodel
* impl NCCL communicator
* add comm related function to Runtime
* export runtime interface
* add launch.py
* use unique name to distingush the the NCCL ID file
* add timeout to communicator init
* expose communicator obj from runtime obj, add unit test for nccl communicator
* reformat files
* Add allReduce operator and cuda nccl allReduce kernel
* impl model parallel for resnet
* add allGather nccl kernel and operator
* Add allreduce allgather operator tests, change allgather kernel to output list of tensor, fix shape infer, handle nullptr output
* fix format of onnx.py
* use concat following AllGather
* get tensor parallel for resnet
* fix format of graph_handler.cc
* change BUILD_DIST default to OFF
* polish code of communicator
* update .gitignore
* export min/max to python
* fix MatMul
* modify launch.py to run opt
* hack to treat ReduceSum as AllReduceSum
* throw exception in cuda error
* fix parallel_opt.py
* improve the error prompt and cuda error check
* fix GatherObj::GatherObj member init
* fix size calculation for scalar (rank = 0) tensor
* MatMul supports bias
* fix add bias for row parallel gemm
* add --gen_std to launch.py
* fix AllReduceNCCL
* update launch.py
* less log
* update parallel_opt
* update launch.py
* add __eq__ for Placement sub-classes
* less benchmark run
* fix placement infer for matmul
* fix vacabuary size
* fix Exception
* Add shard tensor with group to support gpt2
* Add find successor function to find split op at different depth
* recover CommunicatorObj
* improve error mesasge
* optimize parallel_opt.py
* optimize launch.py
* recover docs for all_reduce and all_gather
* - support concat for kvcache
* - modify allocator
* - add tensorType
- modify allocator to support memory allocation based on tensorType
* - fix allocator init
* - support kvcache by running 2 stub distributively
* - fix name
* - remove unused flag
* - fix wrong pb name
* - fix as constroy suggessed
* - fix launch.py format
---------
Co-authored-by: constroy <constroy.li@gmail.com>
Co-authored-by: panzezhong <panzezhong@qiyuanlab.com>
* add cmake bits about NCCL
* move example to examples/NNmodel
* impl NCCL communicator
* add comm related function to Runtime
* export runtime interface
* add launch.py
* use unique name to distingush the the NCCL ID file
* add timeout to communicator init
* expose communicator obj from runtime obj, add unit test for nccl communicator
* reformat files
* Add allReduce operator and cuda nccl allReduce kernel
* impl model parallel for resnet
* add allGather nccl kernel and operator
* Add allreduce allgather operator tests, change allgather kernel to output list of tensor, fix shape infer, handle nullptr output
* fix format of onnx.py
* use concat following AllGather
* get tensor parallel for resnet
* fix format of graph_handler.cc
* change BUILD_DIST default to OFF
* polish code of communicator
* update .gitignore
* export min/max to python
* fix MatMul
* modify launch.py to run opt
* hack to treat ReduceSum as AllReduceSum
* throw exception in cuda error
* fix parallel_opt.py
* improve the error prompt and cuda error check
* fix GatherObj::GatherObj member init
* fix size calculation for scalar (rank = 0) tensor
* MatMul supports bias
* fix add bias for row parallel gemm
* add --gen_std to launch.py
* fix AllReduceNCCL
* update launch.py
* less log
* update parallel_opt
* update launch.py
* add __eq__ for Placement sub-classes
* less benchmark run
* fix placement infer for matmul
* fix vacabuary size
* fix Exception
* Add shard tensor with group to support gpt2
* Add find successor function to find split op at different depth
* recover CommunicatorObj
* improve error mesasge
* optimize parallel_opt.py
* optimize launch.py
* recover docs for all_reduce and all_gather
* Fix API
* fix format
---------
Co-authored-by: panzezhong <panzezhong@qiyuanlab.com>
Co-authored-by: Haojie Wang <haojie0429@gmail.com>
* add cmake bits about NCCL
* move example to examples/NNmodel
* impl NCCL communicator
* add comm related function to Runtime
* export runtime interface
* add launch.py
* use unique name to distingush the the NCCL ID file
* add timeout to communicator init
* expose communicator obj from runtime obj, add unit test for nccl communicator
* reformat files
* Add allReduce operator and cuda nccl allReduce kernel
* impl model parallel for resnet
* add allGather nccl kernel and operator
* Add allreduce allgather operator tests, change allgather kernel to output list of tensor, fix shape infer, handle nullptr output
* fix format of onnx.py
* use concat following AllGather
* get tensor parallel for resnet
* fix format of graph_handler.cc
* change BUILD_DIST default to OFF
* polish code of communicator
* update .gitignore
* Add broadcast operator and cuda kernel
* Add comments for operators
* remove const of class member
* move communicator to CudaRuntimeObj
* Add an empty line at EOF.
---------
Co-authored-by: panzezhong <panzezhong@qiyuanlab.com>
Co-authored-by: Haojie Wang <haojie0429@gmail.com>
* feat: support to sqrt op
* feat: support to erf op
* feat: support to expand op
* feat: support to where op
* fix: gather op index can be int64_t(hard coding)
* fix: some wrong use
* style: fix the format style
* test: add test for change op
* fix: rebase to master
* fix: fix matmul b compute wrong
* add expand and where kernel
* Add int64 support for cuda gather kernel
* add test_where.cc
* add "expand.(cu/cc,test,cuda),modified where.cu"
* Separate initialization of datatypes to avoid compile error
* modify where.(cu/cc/h,test), expand and clip
* Format fix
* Format fix
---------
Co-authored-by: xgqdut2016 <kenan_gewei@163.com>
Co-authored-by: panzezhong <panzezhong@qiyuanlab.com>
Co-authored-by: Haojie Wang <haojie0429@gmail.com>
* impl sqrt on CUDA
fix parser of Gather and ReduceMean
* fix test_gather
* fix test_cuda_gather
* impl sqrt cpu and add sqrt to test_cuda_unary
* cuda_unary supports arbitary shapes
* fix SplitOp with dim=-1
* fix SplitOp with dim=-1
* feat: support mixed dtype
* feat: support cast op
* test: add test for cast op
* feat: support datatype BFloat16
* feat: support data convert fp32 <-> bfp16
* fix: fix all op's infershape func
* fix as review comment
* add conv_half kernel
* Conv Kernel FP16
* dcj:
replace "DataType::Float32" with "op->getDType()" to support more DataType
* feat: support Float16 dtype
* fix: set default clang-format to 14 version
* fix: 按照review意见修改
* fix: add data convert to convfp16 kernel test
* test: add conv_fp16 kernel test
---------
Co-authored-by: zhangyue207 <zhangyue@qiyuanlab.com>
Co-authored-by: kilinchange <kilinchange@163.com>
fix review
change Device::MKL to Device::INTELCPU
fix mkl linkage
fix errors according to merge from master
now can call mkl backend
fix softmax/flatten with axis from onnx.
modify README.md
fix memory refree
add env_lotus_intelcpu.sh
fix compile
merge from branch cpu_backend
fix something add gather
fix something
FIX: directory rename from "mkl" to "intelcpu"
ADD: use oneMKL dpcpp interface to implement matmul kernel.
ADD: add dpcpp as compiler for mkl, and fix warnings for clang compiling.
add dpcpp kernel for pow.
ADD: mkl kernel for pad.
ADD: slice mkl kernel.
ADD: reshape/flatten/identity mkl kernel.
ADD: split mkl kernel.
fix compile error
FIX: fix flattenObj with axis.
ADD reduce_mean mkl kernel.
Add concat mkl kernel.
bathNorm for mkl kernel.
sigmoid mkl kernel.
ADD：add mkl kernel for pooling
add more tests for softmax
Now softmax cuda kernel supports any axises.
mkl kernel for softmax
softmax
add axis to softmax operator
add mkl kernel for abs tanh
ADD: relu kernel for mkl
fix binary mkl primitives.
add mkl kernel for binary operators
fix compiler error
move stream to runtime
clang format
add MemoryFormat for tensorObj.
use post_ops for fused conv/deconv
Distinguish mkl op_timer from cuda op timer.
add act optype to conv and deconv
add operator timer
add mkl kernel for convTransposed
minor fix for group conv
do not use cblas_sgemm_batch
CpuRuntimeObj->NativeCpuRuntimeObj
add matmul op for mkl
kilinchange
Derui Yang