google/praxis

Issues

• Is there an example of training LLM using praxis?

  #87 opened 10 days ago by knightXun
  0

• Support for loading fp8 checkpoint

  #68 opened 7 months ago by wenscarl
  12

• Add activation checkpoint offloading

  #63 opened 8 months ago by jon-chuang
  0

• TransformerLm docs say `start_time_step` should be `prefix_len` but LanguageModel uses `prefix_len-1`

  #56 opened 8 months ago by DCtheTall
  0

• Cross-layer attention weight sharing fails in different scopes

  #52 opened 9 months ago by mqyqlx
  3

• Incorrect conversion from tf dtype to jax dtype

  #49 opened a year ago by backpropper
  0

• [Feature Request] Need Matmul Attention layer instead of Einsum to support GPU running

  #46 opened a year ago by MoFHeka
  0

• [Feature Request] Need ZeRo-1/2 to cooperate with PP+TP+DP. Which may more faster than FSDP sometimes.

  #45 opened a year ago by MoFHeka
  0

• Support custom FP8 dtype in Pipelined Transformer

  #44 opened a year ago by kaixih
  1

• gpu_fast_attention not passing segment_ids to jax pallas attention mha

  #30 opened a year ago by Cjkkkk
  0

• Praxis layers don't support user-specified collection names

  #22 opened a year ago by kaixih
  2

• Any publicly available document?

  #7 opened 2 years ago by Lekja00160612
  1