Experimental: automatic whitelist/blacklist for fp16 training #2
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Merged by command line. |
Closed
rohithkrn
pushed a commit
to rohithkrn/apex
that referenced
this pull request
May 8, 2020
Open
thorjohnsen
pushed a commit
that referenced
this pull request
Jul 21, 2020
Fixing mask multiplication with grad tensors
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
This PR contains v0.1 of an experimental tool to automatically enable fp16 in PyTorch with a whitelist/blacklist model for handling automatic type conversion. The included README describes both the interface and implementation in more detail.
The primary apex-wide change is a couple hacks to enable compiling the custom fused scaling + overflow checking kernel that amp uses for it's loss scaling. Carilli and I are planning to move all the kernel implementations to the new (as of 0.4) cpp-extension interface, so my hope is these hacks are short-lived.
The strangest hack is including the
__init__.pyfiles underapex/amp/_C/in the source tree. They are synthesized by cffi, but when it's run for packaging (ie,python setup.py install), it doesn't synthesize them -- so the only way to get them correctly copied over to the install directory is for them to already exist. This will all go away soon.