microsoft/microxcaling

Issues

• Can this library support flash attention?

  #34 opened 2 months ago by ryusaeba
  1

• Incompatible requirements?

  #31 opened 3 months ago by awf
  0

• mx.matmul overhead

  #29 opened 4 months ago by tsengalb99
  1

• How to set gradient and activation to different formats???

  #30 opened 4 months ago by rensushan
  1

• How to specify the MX format for Gradients in BP pass ??

  #25 opened 4 months ago by rensushan
  2

• Could You Offer a More Comprehensive Environment Setup Guide?

  #16 opened 4 months ago by Mr-Philo
  1

• examples ffn_mx.py - Module not found error - no module named mx (needs path.append('..'))

  #9 opened 4 months ago by lessw2020
  3

• How is the matmul for MX format implemented?

  #23 opened 4 months ago by xijiu9
  1

• BFloat16 compatibility contributions

  #21 opened 4 months ago by hmellor
  1

• Large accuracy loss on MobileVit model

  #27 opened 5 months ago by zks878
  0

• Consider a PyPI package?

  #22 opened 7 months ago by awf
  1

• Inference Error with OPT models

  #18 opened 10 months ago by ruisizhang123
  0

• TypeError: gelu() got an unexpected keyword argument 'approximate'

  #17 opened 10 months ago by Arthur-Ling
  4

• Can MX formats emulate Block Floating Point (BFP)?

  #15 opened a year ago by JeevanBhoot
  1

• Error from test_conv.py and convolution.py to the latest version

  #14 opened a year ago by zks878
  1

• Support for LSTMCell

  #12 opened a year ago by rjfnobre
  2

• Custom CUDA code vs. Pytorch CPU/GPU

  #11 opened a year ago by rjfnobre
  1

• Activations unit tests (gelu, tanh) hang on both A10 and A100 gpus...

  #8 opened a year ago by lessw2020
  23

• Example ffn hangs on multi-gpu scenario (works on single gpu)

  #10 opened a year ago by lessw2020
  1

• example or docs for getting started and converting an existing model to MX dtypes ala FP6?

  #5 opened a year ago by lessw2020
  4

• using fp32 data format to simulate mx data format is not equivalent？

  #6 opened a year ago by jackiechen19940708
  1

• 957 is quantized as 896

  #4 opened a year ago by zhuango
  1

• MXINT8 elements: MX spec: 2's complement vs. README: sign-magnitude

  #3 opened a year ago by grayresearch
  1

• Action required: migrate or opt-out of migration to GitHub inside Microsoft

  #2 opened a year ago by microsoft-github-policy-service
  3