* fix Slice
* change default rounds of timeit to 10 to reduce time
* fix slice with large ends
* Reshape support Int64
* support position_ids as input
* skip last MatMul in Llama
* skip infer_shapes to parse large model
* update launch.py
* fix split_concat_kernel
* print more message in launch.py
* Reshape supports both Int32 and Int64
* try infer_shapes and warn about failure
* fix format
---------
Co-authored-by: whjthu <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
* 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>
* Function tune and corresponding testcase.
*Add: Tune function in /src/kernel/cuda/conv.cc and corresponding testcase in test_conv.
*Fix: A little bug of perfRecord using in /src/core/runtime.cc.
* Tune part debug
*Add: recover the code, fixed the commit error.
*Add: some anotations in tune function
* clang formmat test
* Fix: mem leak in CUDA Runtime and Conv
* Fix: sync in conv and default sync in timeit
* Change the way to tune operator conv.
Timeit function cudNNUnfused -> Timeit function cudnnConvolutionForward.
* Change: merge the common part of cudnnunfused&tune into cudnndescriptoraccess
* clang test
* clang-format
* clang-format bash.
* Chore: remove print and blank lines
Co-authored-by: wcz112 <wcz19@mails.tsinghua.edu.cn>
Co-authored-by: Liyan Zheng <liyan-zheng@outlook.com>
Class "Cuda Runtime" fulfills function "tune" and adds corresponding testcase.
*Add: convCudnn::tune, convCudnn::cuDNNdescriptorAccess.
*Add: testcase tune.
*Fix: a brief bug in CPU Runtime.
* Fix: add warm-up and repetition in timing
* Add: CUDA runtime and float support
* Refactor: Cuda and Cpu runtimes inherit Runtime
* Add: environment script for Lotus
* Add: Lotus build instructions
* Update README.md
Co-authored-by: Liyan Zheng <liyan-zheng@outlook.com>
* Refactor: operator hash and inferShape
* Add: hash without shape
* Add: inferShape interface for given input tensors
* Add: construct outputs in op ctor
* Add: comments for matmul
* Add: opType in AttrVector and WorkloadVector
* Chore: _graph -> graph in Op ctor
* Chore: change the "Node" suffix to "Obj"
Co-authored-by: Liyan Zheng <liyan-zheng@outlook.com>
zhangyue