GRU units and FC layers - 24kHz sampling frequency requires 40us sampling net runtime - First PyTorch model used a 3,400us sampling net runtime Image from LPCNetExit, Pursued By A Bear - 3400us (baseline) (baseline), 40us (target) - 85x speedup - Uh ohEnter, TVM and model co-design - PyTorch operator overhead makes interpreter infeasible - Reduce FLOPs with block-sparsified weight matrices - not trades off icache/ dcache - also available today in FBGEMMPyTorch and TVM - Lots of opportunity in PyTorch - Graph optimization - Existing fusion infrastructure fairly limited (CUDA-only, injective-only)

wildcard. For example, *py* returns all packages that have "py" in their names, such as "numpy", "pytorch", "python", etc. version (Optional) Package version. May include the * wildcard or a version range(s) objects. Internally, conda will translate your command line requests, like numpy>=1.19, python=3.* or pytorch=1.8.*=*cuda*, into in- stances of this class. This query language has its own syntax and rules, detailed here. The most important fields of a MatchSpec object are: – name: the name of the package (e.g. pytorch); always expected. – version: the version constraints (e.g. 1.8.*); can be empty but if build is

分时切换成功率11/19 计算型任务混部 Part 2：双十一 Kubernetes 实践 其他自研 Operators AI 训练 Kubernetes TensorFlow PyTorch Horovod/mpi XGBoost Spark SparkApplication Spark Operator 在线应用 无状态服务 Node Node Node Node Node

容易更新及维护 • 基于FUSE提供POSIX兼容⽂件接⼝ • 问题 • 相对kernel⽂件系统的实现(ext4, xfs)性能 差异⼤，延迟⾼FUSE⽂件IO读写流程 • 场景1 pytorch example word_language_model • LOOKUP inode 返回 fstat + timeout设置 • OPEN 打开 inode返回ok • GETATTR

Your chip can run any models Your compiler (TVM) supports multiple frontends (e.g., TensorFlow, PyTorch, MXNet) Non Maximum Suppression ResNet-50 Dense Your Chip Your Chip© 2019, Amazon Web Services

for the user to be able to leverage your hot path capabilities written in C++ ○ Numpy / Pandas / PyTorch are nice DSLs that provide “hot path” C++ functionsWhat Functions to expose in the API ● Init some

wildcard. For example, *py* returns all packages that have "py" in their names, such as "numpy", "pytorch", "python", etc. version (Optional) Package version. May include the * wildcard or a version range(s) objects. Internally, conda will translate your command line requests, like numpy>=1.19, python=3.* or pytorch=1.8.*=*cuda*, into in- stances of this class. This query language has its own syntax and rules, detailed here. The most important fields of a MatchSpec object are: – name: the name of the package (e.g. pytorch); always expected. – version: the version constraints (e.g. 1.8.*); can be empty but if build is

wildcard. For example, *py* returns all packages that have "py" in their names, such as "numpy", "pytorch", "python", etc. version (Optional) Package version. May include the * wildcard or a version range(s) chdir in- stead. (#12678) Docs • Change the README example from IPython Notebook and NumPy to PyTorch. (#12579) • Discuss options available to properly configure mirrored channels. (#12583, #12641) objects. Internally, conda will translate your command line requests, like numpy>=1.19, python=3.* or pytorch=1.8.*=*cuda*, into in- stances of this class. This query language has its own syntax and rules, detailed

wildcard. For example, *py* returns all packages that have "py" in their names, such as "numpy", "pytorch", "python", etc. version (Optional) Package version. May include the * wildcard or a version range(s) objects. Internally, conda will translate your command line requests, like numpy>=1.19, python=3.* or pytorch=1.8.*=*cuda*, into in- stances of this class. This query language has its own syntax and rules, detailed here. The most important fields of a MatchSpec object are: – name: the name of the package (e.g. pytorch); always expected. – version: the version constraints (e.g. 1.8.*); can be empty but if build is

wildcard. For example, *py* returns all packages that have "py" in their names, such as "numpy", "pytorch", "python", etc. version (Optional) Package version. May include the * wildcard or a version range(s) chdir in- stead. (#12678) Docs • Change the README example from IPython Notebook and NumPy to PyTorch. (#12579) • Discuss options available to properly configure mirrored channels. (#12583, #12641) objects. Internally, conda will translate your command line requests, like numpy>=1.19, python=3.* or pytorch=1.8.*=*cuda*, into in- stances of this class. This query language has its own syntax and rules, detailed