Upcast gradually when computing variance #4283
Merged
Conversation
This file contains hidden or 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
Going all the way to f64 is undesirable, especially for low-precision tensors in bf16 or f8 variants. Upcast only to the next type, e.g., bf16->f32 or f8->bf16. This is consistent with what Pytorch seems to be doing internally. Signed-off-by: Alex Zinenko <git@ozinenko.com>
ftynse
force-pushed
the
users/ftynse/dont-upscale-variance
branch
from
b661b9c to
5b0479c
Compare
ftynse
changed the title
qedawkins
approved these changes
Jul 31, 2025
Comment on lines
+9325
to
9334
| // Upcasting the input tensor to a double-bitwidth dtype for higher precision | ||
| // during the computation of the result. | ||
| unsigned bitwidth = inputTensorTy.getDtype().getIntOrFloatBitWidth(); | ||
| if (bitwidth != 64) { | ||
| Type targetTy = rewriter.getF64Type(); | ||
| if (bitwidth == 8) | ||
| targetTy = rewriter.getBF16Type(); | ||
| else if (bitwidth == 16) | ||
| targetTy = rewriter.getF32Type(); | ||
| self = convertTensorToDtype(rewriter, loc, self, rewriter.getF64Type()); |
There was a problem hiding this comment.
Don't you need to replace rewriter.getF64Type() with the targetTy here?
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Outdated suggestions cannot be applied.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.
Going all the way to f64 is undesirable, especially for low-precision tensors in bf16 or f8 variants. Upcast only to the next type, e.g., bf16->f32 or f8->bf16. This is consistent with what Pytorch seems to be doing internally.