[quantization] Increase precision#704
Merged
Merged
Conversation
This PR increases precision for GPTQ quantization: 1. replaces moving average in Hessian with just a sum 2. increases precision to double in Cholesky and its inversion to improve reproducibility. TICO-DCO-1.0-Signed-off-by: s.malakhov <s.malakhov@partner.samsung.com>
mhs4670go
reviewed
May 13, 2026
| diag = torch.arange(self.columns, device=self.dev) | ||
| H[diag, diag] += damp | ||
| H = torch.linalg.cholesky(H) | ||
| H = torch.linalg.cholesky(H.double()).float() |
Contributor
There was a problem hiding this comment.
IIUC, currently the code casts each intermediate result back to float.
Since the goal of this PR is to increase precision in Cholesky and its inversion, it would be better to keep H as fp64 across the full linear algebra path and cast back to fp32 only once before entering the quantization loop.
H = H.double()
damp = percdamp * torch.mean(torch.diag(H))
diag = torch.arange(self.columns, device=self.dev)
H[diag, diag] += damp
H = torch.linalg.cholesky(H)
H = torch.cholesky_inverse(H)
Hinv = torch.linalg.cholesky(H, upper=True).float()This preserves the intended double precision during Cholesky -> inverse -> Cholesky, while still keeping the existing fp32 quantization loop mostly unchanged.
Contributor
Author
There was a problem hiding this comment.
@mhs4670go
Fixed.
Please take another look.
mhs4670go
reviewed
May 13, 2026
Co-authored-by: seongwoo chae <mhs4670go@naver.com>
TICO-DCO-1.0-Signed-off-by: s.malakhov <s.malakhov@partner.samsung.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 increases precision for GPTQ quantization:
Draft: #670
Related: #656
GPTQ quantization for
HuggingFaceTB/SmolLM2-135M-Instructprovides the sameppl as in #702 (comment)
TICO-DCO-1.0-Signed-off-by: s.malakhov s.malakhov@partner.samsung.com