1. 12 2月, 2022 3 次提交
  2. 26 1月, 2022 1 次提交
  3. 20 1月, 2022 1 次提交
  4. 17 1月, 2022 1 次提交
  5. 16 1月, 2022 3 次提交
  6. 14 1月, 2022 2 次提交
  7. 11 1月, 2022 2 次提交
  8. 29 12月, 2021 1 次提交
    • P
      Updating retracing examples to handle n-ary operators · 54da13e0
      Priyam Parashar 提交于
      Summary:
      Replaced Proxy(object) creations with Proxy(object, GraphAppendingTracer)
      which ensures graceful handling of n-ary operators.
      
      By default Proxy creates a new tracer when no tracer arg
      is provided, these examples encourage use of Proxy creation
      with the GraphAppendingTracer when dealing with n-ary operators.
      
      Test Plan:
      ```bash
      python fx/inline_function.py
      python fx/proxy_based_graph_creation.py
      python fx/primitive_library.py
      ```
      
      No errors on the above and output of `fx/primitive_library` confirmed
      to be the same as inline commented expected output. Also verified  via
      diff of OG output and updated output.
      
      FYI: `fx/nnc_compile.py` might need this fix too, but I was not able to run
      the script at all. Threw the following error:
      `AttributeError: module 'torch._C._te' has no attribute 'KernelScope'`
      
      Reviewers:
      jamesreed
      qianxuqx
      
      Subscribers:
      
      Tasks:
      T105739531
      
      Tags:
      54da13e0
  9. 27 12月, 2021 1 次提交
  10. 25 11月, 2021 1 次提交
  11. 23 11月, 2021 1 次提交
  12. 27 10月, 2021 1 次提交
  13. 17 9月, 2021 2 次提交
  14. 29 7月, 2021 1 次提交
  15. 28 7月, 2021 1 次提交
  16. 12 6月, 2021 2 次提交
  17. 28 5月, 2021 4 次提交
  18. 10 5月, 2021 1 次提交
  19. 08 5月, 2021 1 次提交
    • K
      Fix Transformer init_weights · 3d63fe18
      kyox24 提交于
      Weights of decoders are initialized twice. I think we need to call the zeros_() to initialize bias.
      3d63fe18
  20. 07 5月, 2021 1 次提交
  21. 04 5月, 2021 1 次提交
  22. 03 5月, 2021 1 次提交
  23. 26 4月, 2021 1 次提交
  24. 16 4月, 2021 2 次提交
  25. 03 4月, 2021 1 次提交
  26. 02 4月, 2021 2 次提交
  27. 23 3月, 2021 1 次提交