DeepSpeed
add SD injection policy
#2381
Merged

Commits
  • add initial sd policy
    jeffra committed 3 years ago
  • Merge branch 'master' into jeffra/sd-policy
    jeffra committed 3 years ago
  • formatting
    jeffra committed 3 years ago
  • add attention kernel and enable cuda-graph for SD models
    Reza Yazdani committed 3 years ago
  • add new files
    Reza Yazdani committed 3 years ago
  • formatting and add dtype to unet
    jeffra committed 3 years ago
  • adding more optitmization by enabling ds-encoder with CUDA-Graph
    Reza Yazdani committed 3 years ago
  • Merge branch 'master' into jeffra/sd-policy
    RezaYazdaniAminabadi committed 3 years ago
  • add missing file
    Reza Yazdani committed 3 years ago
  • Merge branch 'jeffra/sd-policy' of github.com:microsoft/DeepSpeed into jeffra/sd-policy
    Reza Yazdani committed 3 years ago
  • adapt the triton kernel to be used in more places
    Reza Yazdani committed 3 years ago
  • add more fusion
    Reza Yazdani committed 3 years ago
  • allocate workspace using the padded hidden_size
    Reza Yazdani committed 3 years ago
  • skip the clip-encoder injection for now
    Reza Yazdani committed 3 years ago
  • Merge branch 'master' into jeffra/sd-policy
    jeffra committed 3 years ago
  • add triton to new extra
    jeffra committed 3 years ago
  • lazy import triton, add sd extra, formatting
    jeffra committed 3 years ago
  • delay import
    jeffra committed 3 years ago
  • fix previous issue i added
    jeffra committed 3 years ago
  • fix bug with adding bias
    Reza Yazdani committed 3 years ago
  • Merge branch 'master' into jeffra/sd-policy
    RezaYazdaniAminabadi committed 3 years ago
  • fixes for triton import and add acks to triton-ops file
    jeffra committed 3 years ago
  • Merge branch 'jeffra/sd-policy' of github.com:microsoft/DeepSpeed into jeffra/sd-policy
    jeffra committed 3 years ago
  • Merge branch 'master' into jeffra/sd-policy
    jeffra committed 3 years ago
  • merge fix & formatting
    Reza Yazdani committed 3 years ago
  • Merge branch 'jeffra/sd-policy' of github.com:microsoft/DeepSpeed into jeffra/sd-policy
    Reza Yazdani committed 3 years ago
  • fix small issue
    Reza Yazdani committed 3 years ago
  • skip cuda-graph for clip-encoder for now (it has issue on larger batch size)
    Reza Yazdani committed 3 years ago
  • Merge branch 'master' into jeffra/sd-policy
    jeffra committed 3 years ago
Loading