Code

Dev#74

Code

Merged

YdrMaster merged 64 commits intomasterInfiniTensor/RefactorGraph:masterfrom devInfiniTensor/RefactorGraph:devCopy head branch name to clipboardFeb 19, 2024

+5,037-584

Commits

Commits on Jan 4, 2024

• feat(hardware): 实现 mlu 硬件相关的函数；改变编译方式按照以硬件名称命名的目录名区分是否需要编译

  kilinchange

  committed
  bd6ad6b

  Copy full SHA for bd6ad6b

• feat: 在 cmake 中增加 mlu 的相关环境编译

  kilinchange

  committed
  02939c9

  Copy full SHA for 02939c9

Commits on Jan 5, 2024

• fix: 去掉 cncl 编译

  kilinchange

  committed
  7f82d74

  Copy full SHA for 7f82d74

Commits on Jan 17, 2024

• fix (frontend): 修改get输入输出，内存大小，增加compiler设置输入接口

  PanZezhong1725

  committed
  eb4fc4c

  Copy full SHA for eb4fc4c

• Merge pull request #77 from InfiniTensor/front_fix

  Show description for 54ffa1a
  YdrMaster

  authored
  54ffa1a

  Copy full SHA for 54ffa1a

• add head file

  

  zhangyue207

  authored and

  YdrMaster

  committed
  e76c1fc

  Copy full SHA for e76c1fc

• Merge pull request #79 from InfiniTensor/add-headfile

  Show description for 8b6333d
  YdrMaster

  authored
  8b6333d

  Copy full SHA for 8b6333d

• feat: add HardSigmoid cpu/cuda kernel

  bitzyz

  committed
  a7ed032

  Copy full SHA for a7ed032

• feat: add hardswish cpu/cuda kernel

  bitzyz

  committed
  3998833

  Copy full SHA for 3998833

• fix: fix mod kernel

  

  kilinchange

  authored and

  bitzyz

  committed
  1ce6d81

  Copy full SHA for 1ce6d81

Commits on Jan 18, 2024

• Merge pull request #78 from InfiniTensor/dev-hardsigmoid

  Show description for 9e4789c
  YdrMaster

  authored
  9e4789c

  Copy full SHA for 9e4789c

• style: 整理 hardswish 计算

  Show description for afdfe93
  YdrMaster

  committed
  afdfe93

  Copy full SHA for afdfe93

• Merge pull request #73 from InfiniTensor/dev-hardswish

  Show description for a163eee
  YdrMaster

  authored
  a163eee

  Copy full SHA for a163eee

• feat(python_ffi): 添加一个字符型的 trace 格式

  Show description for ec39fd7
  YdrMaster

  committed
  ec39fd7

  Copy full SHA for ec39fd7

• fix(kernel): 双目运算不能交换

  Show description for 20788a3
  YdrMaster

  committed
  20788a3

  Copy full SHA for 20788a3

Commits on Jan 23, 2024

• refactor(kernel): 添加一个显存的全局缓存，避免反复的 h2d 拷贝

  Show description for 41036a0
  YdrMaster

  committed
  41036a0

  Copy full SHA for 41036a0

• refactor(python_ffi): 前端控制算子类型是否添加 onnx:: 前缀以便自定义算子

  Show description for 1375be3
  YdrMaster

  committed
  1375be3

  Copy full SHA for 1375be3

• build: 添加 llm 子项目用于前端大模型自定义算子

  Show description for 88e0712
  YdrMaster

  committed
  88e0712

  Copy full SHA for 88e0712

• docs: 添加 rms normalization 算子定义

  Show description for 39538e7
  YdrMaster

  committed
  39538e7

  Copy full SHA for 39538e7

Commits on Jan 24, 2024

• docs: 初步添加 attention 算子定义

  Show description for 594e06b
  YdrMaster

  committed
  594e06b

  Copy full SHA for 594e06b

Commits on Jan 25, 2024

• fead(kernel): add max cpu kernel

  kilinchange

  committed
  4f9d09b

  Copy full SHA for 4f9d09b

• feat(kernel): add test for max/min cpu kernel

  kilinchange

  committed
  5e13361

  Copy full SHA for 5e13361

• refactor(frontend): 修改 Attributes 为一个类，并将广播机制等代码提升至 frontend 方便不同算子库公用

  Show description for 6877516
  YdrMaster

  committed
  6877516

  Copy full SHA for 6877516

• feat(llm): 添加一个既支持 transpose 又支持广播的 MatMul

  Show description for 9dce4b3
  YdrMaster

  committed
  9dce4b3

  Copy full SHA for 9dce4b3

Commits on Jan 26, 2024

• feat(llm): 添加 rms normalization 的前端算子、计算图算子和 cpu 核函数

  Show description for eca56d8
  YdrMaster

  committed
  eca56d8

  Copy full SHA for eca56d8

• feat(llm): 添加 rms normalization 的 cuda 核函数

  Show description for e15f28f
  YdrMaster

  committed
  e15f28f

  Copy full SHA for e15f28f

• todo(llm): How can I use cub in NVRTC?

  Show description for 39785bd
  YdrMaster

  committed
  39785bd

  Copy full SHA for 39785bd

• fix(kernel): 修正 SliceInfo 优化计算

  Show description for 5f518a3
  YdrMaster

  committed
  5f518a3

  Copy full SHA for 5f518a3

• fix(kernel): 引入最新版本的 cccl 以支持 nvrtc 与 cub 配合使用

  Show description for c462d7d
  YdrMaster

  committed
  c462d7d

  Copy full SHA for c462d7d

• build: update backward-cpp to its master branch

  Show description for 605c4c0
  YdrMaster

  committed
  605c4c0

  Copy full SHA for 605c4c0

• fix: format CMakeLists.txt

  

  Chamberlain0w0

  authored and

  kilinchange

  committed
  0ebf34c

  Copy full SHA for 0ebf34c

• feat(hardware): add cnnl

  kilinchange

  committed
  cc19556

  Copy full SHA for cc19556

• Merge branch 'dev' into dev-mlu-runtime

  kilinchange

  committed
  d8a2dea

  Copy full SHA for d8a2dea

• feat: support pad operator | cpu/cuda kernel

  bitzyz

  committed
  f344ae8

  Copy full SHA for f344ae8

• feat: 优化Pad算子

  bitzyz

  committed
  777d9c8

  Copy full SHA for 777d9c8

• fix(hardware): fix file encoding bug and include bug

  kilinchange

  committed
  b3b7d09

  Copy full SHA for b3b7d09

• Merge pull request #87 from InfiniTensor/dev-mlu-runtime

  Show description for d91d89c
  YdrMaster

  authored
  d91d89c

  Copy full SHA for d91d89c

• fix(kernel): 修改pad/slice的diminfo; 删除部分注释

  bitzyz

  committed
  85a8004

  Copy full SHA for 85a8004

• Merge pull request #86 from InfiniTensor/dev-pad

  Show description for 0e3c71d
  YdrMaster

  authored
  0e3c71d

  Copy full SHA for 0e3c71d

• fix: fix for PRs

  Show description for 817cc8f
  YdrMaster

  committed
  817cc8f

  Copy full SHA for 817cc8f

• fix(kernel): 改正 rms norm 以支持任意长的 BlockReduce

  Show description for 5f54011
  YdrMaster

  committed
  5f54011

  Copy full SHA for 5f54011

Commits on Jan 29, 2024

• feat(llm): 注册 rms normalization

  Show description for add61cb
  YdrMaster

  committed
  add61cb

  Copy full SHA for add61cb

• feat: 搭建 Attention 在各层的基本结构

  Show description for daeac28
  YdrMaster

  committed
  daeac28

  Copy full SHA for daeac28

• feat(llm): 实现 attention 形状推导

  Show description for 722e6d9
  YdrMaster

  committed
  722e6d9

  Copy full SHA for 722e6d9

• fix(hardware): typo

  Show description for 12e264a
  YdrMaster

  committed
  12e264a

  Copy full SHA for 12e264a

• refactor(kernel): 重构 transpose info 计算

  Show description for 8595be4
  YdrMaster

  committed
  8595be4

  Copy full SHA for 8595be4

• refactor(kernel): 重构 transpose info 和 kernel 的构造

  Show description for f4d063a
  YdrMaster

  committed
  f4d063a

  Copy full SHA for f4d063a

• refactor(kernel): transpose 支持合并连续访存

  Show description for bbed31e
  YdrMaster

  committed
  bbed31e

  Copy full SHA for bbed31e

• feat: 添加 attention cuda

  Show description for f1faf3e
  YdrMaster

  committed
  f1faf3e

  Copy full SHA for f1faf3e

Commits on Jan 30, 2024

• build(kernel): 添加 cublasLt

  Show description for 9495516
  YdrMaster

  committed
  9495516

  Copy full SHA for 9495516

• fix(kernel): 为 Gather 支持负的 indices 值

  Show description for 6630866
  YdrMaster

  committed
  6630866

  Copy full SHA for 6630866

Commits on Jan 31, 2024

• add max/min cuda kernel and test

  kilinchange

  committed
  e237349

  Copy full SHA for e237349

• Merge pull request #84 from InfiniTensor/add_max_min_kernel

  Show description for 35dc6c8
  YdrMaster

  authored
  35dc6c8

  Copy full SHA for 35dc6c8

• fix(frontend): 填充边信息时不能接受输入边不存在

  Show description for 27a8ad6
  YdrMaster

  committed
  27a8ad6

  Copy full SHA for 27a8ad6

• fix(fronted): 为适配torchvision模型，修改前端从而接调用cpu kernel

  bitzyz

  committed
  eea2249

  Copy full SHA for eea2249

• fix(fronted): 修复reduce算子对reducesum的opset特判bug

  bitzyz

  committed
  3f5b0ee

  Copy full SHA for 3f5b0ee

• fix(kernel):修复transpose info存在的bug

  bitzyz

  committed
  7db54c1

  Copy full SHA for 7db54c1

• fix(kernel): 为 TransposeInfo 应对更多 coner case 并增加一个单元测试

  Show description for b3e89a6
  YdrMaster

  committed
  b3e89a6

  Copy full SHA for b3e89a6

Commits on Feb 2, 2024

• Merge pull request #88 from InfiniTensor/fix-torchvision-models

  Show description for d076c20
  YdrMaster

  authored
  d076c20

  Copy full SHA for d076c20

Commits on Feb 4, 2024

• feat(kernel): 添加exp算子

  bitzyz

  committed
  3199eb5

  Copy full SHA for 3199eb5

Commits on Feb 19, 2024

• fix(kernel): 没有引用的输出张量也要分配，但计算完后立即释放

  Show description for e3febd9
  YdrMaster

  committed
  e3febd9

  Copy full SHA for e3febd9

• feat (dist): nccl通信库接入，allreduce算子

  PanZezhong1725

  committed
  16870d6

  Copy full SHA for 16870d6

• Merge pull request #92 from InfiniTensor/dist-merge

  Show description for 34ed834
  YdrMaster

  authored
  34ed834

  Copy full SHA for 34ed834

• Merge pull request #91 from InfiniTensor/dev-exp

  Show description for 3cb6f5d
  YdrMaster

  authored
  3cb6f5d

  Copy full SHA for 3cb6f5d