pengwa

icon company @Microsoft icon location Beijing icon location Work on LLM training/inference runtime with ONNX

Results 13 comments of pengwa

Trade operator recompute for memory

> Can you explain what config the user needs to set to enable this feature? From the code I gather user needs to set env variable ORTMODULE_ENABLE_MEMORY_ALLEVIATION to "Dropout:0,Gelu:1,Tile:0" is...

Trade operator recompute for memory

Thanks a lot @askhade @baijumeswani for the efforts to review!!!

Export schema in onnxruntime-training if --gen_doc is on

Just as a FYI, I had an old PR (closed) here. https://github.com/microsoft/onnxruntime/pull/10980

Export schema in onnxruntime-training if --gen_doc is on

> You can check the PR description, there are few reasons.

Fix include order build failure training build

> @pengwa do you have a moment to comment on this change? Asking you since this issue occurred as a result of a CPU AdamW implementation commit. sorry I somehow...

Refactor with std::variant (on device training)

As mentioned in some comments, inlined containers are not used in orttraining/orttraining/test/training_api/common/synthetic_data_loader.cc/.h and orttraining/orttraining/test/training_api/trainer/trainer.cc , for two reasons: 1. those tests are independent of ORT internal libs, so we should...

Refactor with std::variant (on device training)

Thank @edgchen1 , @yuslepukhin a lot for the suggestions. Please let me know if you have more comments.

Refactor with std::variant (on device training)

@edgchen1 @yuslepukhin I updated as suggested, let me know if there is more comment. Thanks!

Refactor with std::variant (on device training)

Thanks everyone! @edgchen1 @yuslepukhin @baijumeswani. 💯

Enhance constant folding for Shape node

Will create a new specialized shape optimizer for this, to avoid any backward incompatibility.