Code

Xnnpack backend support#159

Code

Merged

yirongjie merged 69 commits intoUbiquitousLearning:mainUbiquitousLearning/mllm:mainfrom chenghuaWang:mainchenghuaWang/mllm:mainCopy head branch name to clipboardOct 31, 2024

+5,545-88

Commits

Commits on Oct 9, 2024

• feat: XpOps, XpDirect.

  chenghuaWang

  committed
  59cdc15

  Copy full SHA for 59cdc15

• feat: Xnnpack Add Example

  chenghuaWang

  committed
  8fbd7df

  Copy full SHA for 8fbd7df

Commits on Oct 10, 2024

• feat: mllm frontend -> xnn static graph

  chenghuaWang

  committed
  7024eab

  Copy full SHA for 7024eab

• fix: Add Example Done.

  chenghuaWang

  committed
  a3f271c

  Copy full SHA for a3f271c

• feat: xnnpack wrap

  chenghuaWang

  committed
  9c59cd0

  Copy full SHA for 9c59cd0

• fix: include path, update xnnpack to latest

  chenghuaWang

  committed
  c4bad7b

  Copy full SHA for c4bad7b

Commits on Oct 11, 2024

• feat: xnn backend element wise op function

  chenghuaWang

  committed
  0a6f706

  Copy full SHA for 0a6f706

• feat: xnn weight register and linear op

  chenghuaWang

  committed
  f156f35

  Copy full SHA for f156f35

• fix: XpLinear error with NoLoadWeightsDtype

  chenghuaWang

  committed
  efb3fa4

  Copy full SHA for efb3fa4

Commits on Oct 14, 2024

• feat: xnnpack matmul rope

  chenghuaWang

  committed
  5781312

  Copy full SHA for 5781312

Commits on Oct 15, 2024

• feat: fix redefine tensor in xnnpack bug

  chenghuaWang

  committed
  72dfb7a

  Copy full SHA for 72dfb7a

• feat: add relu and rope bug fix

  chenghuaWang

  committed
  4ad8d31

  Copy full SHA for 4ad8d31

Commits on Oct 16, 2024

• feat: xnnpack GELU, Softmax, SiLU impl

  chenghuaWang

  committed
  ad3b148

  Copy full SHA for ad3b148

• feat: rms norm, tranpose

  chenghuaWang

  committed
  886eba7

  Copy full SHA for 886eba7

Commits on Oct 17, 2024

• feat: kvcache, still buggy, rfc.

  chenghuaWang

  committed
  b37ad92

  Copy full SHA for b37ad92

• feat: update 3rd party packages

  chenghuaWang

  committed
  90164da

  Copy full SHA for 90164da

Commits on Oct 18, 2024

• feat: xp kvcache fix.

  chenghuaWang

  committed
  b0883bb

  Copy full SHA for b0883bb

• fix: github action main.yml

  chenghuaWang

  committed
  05f167b

  Copy full SHA for 05f167b

Commits on Oct 20, 2024

• feat: !!!SDPA!!! (Support B, H, S, D) layout

  chenghuaWang

  committed
  69e80f9

  Copy full SHA for 69e80f9

• feat: XpSDPA torch impl for check mllm's correctness

  chenghuaWang

  committed
  a7471be

  Copy full SHA for a7471be

• fix: XpRoPE, add view func

  chenghuaWang

  committed
  4a95180

  Copy full SHA for 4a95180

Commits on Oct 22, 2024

• fix: rope test example

  chenghuaWang

  committed
  a34240f

  Copy full SHA for a34240f

• fix: transpose xnn example bugs find

  chenghuaWang

  committed
  5b4bd71

  Copy full SHA for 5b4bd71

• fix: xnnpack uuid register bug

  chenghuaWang

  committed
  314ca42

  Copy full SHA for 314ca42

• fix: xnnpack uuid register bug

  chenghuaWang

  committed
  bea70c1

  Copy full SHA for bea70c1

Commits on Oct 23, 2024

• fix: rope xnnpack test error

  chenghuaWang

  committed
  063f37f

  Copy full SHA for 063f37f

• feat: matmul xnnpack, failed at stl malloc.

  chenghuaWang

  committed
  3b75a25

  Copy full SHA for 3b75a25

• fix: xnnpack illegal memory r/w by using valgrind.

  chenghuaWang

  committed
  2945423

  Copy full SHA for 2945423

Commits on Oct 24, 2024

• fix: xnnpack attention impl bug

  chenghuaWang

  committed
  d778974

  Copy full SHA for d778974

Commits on Oct 25, 2024

• feat: XpEmbedding op

  chenghuaWang

  committed
  dbb642b

  Copy full SHA for dbb642b

• feat: xp llama

  chenghuaWang

  committed
  75a2850

  Copy full SHA for 75a2850

Commits on Oct 26, 2024

• fix: use_layername_2_tensorname = false;

  chenghuaWang

  committed
  895e84e

  Copy full SHA for 895e84e

• change: llama xp example -> qwen xp example

  chenghuaWang

  committed
  d0e31e9

  Copy full SHA for d0e31e9

• Merge remote-tracking branch 'upstream/main'

  chenghuaWang

  committed
  19ba8aa

  Copy full SHA for 19ba8aa

• fix: megre confict bugs

  chenghuaWang

  committed
  a4a8c70

  Copy full SHA for a4a8c70

• fix: llm_model_ptr null error

  chenghuaWang

  committed
  bf278ef

  Copy full SHA for bf278ef

• Merge branch 'main' into main

  UbiquitousLearning

  authored
  1c043ee

  Copy full SHA for 1c043ee

Commits on Oct 28, 2024

• fix: xnn backend load bug fix

  chenghuaWang

  committed
  d5982d2

  Copy full SHA for d5982d2

Commits on Oct 29, 2024

• Merge remote-tracking branch 'upstream/main'

  chenghuaWang

  committed
  c689339

  Copy full SHA for c689339

• fix: add OUTPUT_TENSOR to TensorType. Remove this type after xnnpack is all done

  chenghuaWang

  committed
  189fa72

  Copy full SHA for 189fa72

• update: third_party submodule

  chenghuaWang

  committed
  ca7b7f5

  Copy full SHA for ca7b7f5

• fix: freeze xnnpack version

  chenghuaWang

  committed
  f97ee69

  Copy full SHA for f97ee69

• feat: QWen version 1.5 0.5B and 1.8B xnnpack backend.

  chenghuaWang

  committed
  6b185ec

  Copy full SHA for 6b185ec

• Merge branch 'main' into main

  yirongjie

  authored
  1bc39c2

  Copy full SHA for 1bc39c2

• fix: xnnpack backend rope

  chenghuaWang

  committed
  5e9345e

  Copy full SHA for 5e9345e

• feat: reduce memory load time in xnnpack

  chenghuaWang

  committed
  cc27a58

  Copy full SHA for cc27a58

• fix: xnnpack qwen example token backend

  chenghuaWang

  committed
  11cfa6f

  Copy full SHA for 11cfa6f

• Merge branch 'UbiquitousLearning:main' into main

  chenghuaWang

  authored
  8c0c300

  Copy full SHA for 8c0c300

• fix: use_layername_2_tensorname

  yirongjie

  committed
  f77622f

  Copy full SHA for f77622f

• fix: MLLM_BUILD_XNNPACK OFF error and redundant xnnpack exec targets

  chenghuaWang

  committed
  d3d9211

  Copy full SHA for d3d9211

• Merge branch 'main' of https://github.com/chenghuaWang/mllm

  chenghuaWang

  committed
  2cd51cf

  Copy full SHA for 2cd51cf

• Merge branch 'main' into main

  yirongjie

  authored
  172bbd9

  Copy full SHA for 172bbd9

Commits on Oct 30, 2024

• fix: remove memory test due to previous workflow.yaml remove it from MLLM_TEST.

  chenghuaWang

  committed
  757905e

  Copy full SHA for 757905e

• fix: mask init error in xnnpack

  chenghuaWang

  committed
  c5302e0

  Copy full SHA for c5302e0

• Merge branch 'UbiquitousLearning:main' into main

  chenghuaWang

  authored
  7f38bdf

  Copy full SHA for 7f38bdf

• Merge branch 'main' into main

  yirongjie

  authored
  20393cc

  Copy full SHA for 20393cc

• Merge branch 'main' into main

  yirongjie

  authored
  b43180e

  Copy full SHA for b43180e

• fix: merge tokenize error

  yirongjie

  committed
  da3c295

  Copy full SHA for da3c295

• fix: tokenizer apply

  yirongjie

  committed
  9eb944c

  Copy full SHA for 9eb944c

Commits on Oct 31, 2024

• fix: change HardSwish to original Swish function

  chenghuaWang

  committed
  779207b

  Copy full SHA for 779207b

• fix: mask bug in xnnpack

  chenghuaWang

  committed
  458a1e3

  Copy full SHA for 458a1e3

• fix: tokenizer.tokenize

  yirongjie

  committed
  fd1a449

  Copy full SHA for fd1a449

• fix: remove unused

  yirongjie

  committed
  1d62462

  Copy full SHA for 1d62462

• update: move Xp*Test to MLLM_TEST

  chenghuaWang

  committed
  211ad56

  Copy full SHA for 211ad56

• Merge branch 'main' of https://github.com/chenghuaWang/mllm

  chenghuaWang

  committed
  5ceadc3

  Copy full SHA for 5ceadc3

• fix: test bug

  chenghuaWang

  committed
  ddddab3

  Copy full SHA for ddddab3

• fix: xnnpack test setup

  chenghuaWang

  committed
  919520f

  Copy full SHA for 919520f

• fix: XpTest Error

  chenghuaWang

  committed
  99f2ebc

  Copy full SHA for 99f2ebc

• fix: set xnn default threads to 4

  chenghuaWang

  committed
  b829c24

  Copy full SHA for b829c24