Using Torch Autograd ctx to Optimize Memory Leaking Issue #202
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
See Issue #201
This PR create a potential solution to solve the memory leadking issue when using MS-AMP custom GeMM.
Currently the custom GeMM function use
ctxobject to save input tensor x and weight tensor W. In backward gradient computing, x and W are needed.ctx.input_fp8means directly saving this attribute. However,input_fp8is forclass ScalingTensor. In practice, this saving method does not fully leverage the advantage of FP8 tensors!Instead, I suggest using
ctx.save_for_backward(). This method is specially designed for better memory management. Change saved context fromScalingTensortotorch.Tensor+ScalingMeta. This is proved to be efficient in memory saving!Effect for deit-base (86M) model training, batch size 256:
Effect for deit 570M model training, batch size 256: