perf: avoid unused SVD reconstruction and select the CUDA driver in SVDQuant #2356
perf: avoid unused SVD reconstruction in SVDQuant
ef46bad8
changwangss
marked this pull request as ready for review 27 days ago
perf: reuse the selected SVD driver during quantization
f59dae70
fix: initialize the SVD driver once on the quantization device
bc5527b5
refactor: keep SVD driver initialization private to the transform
d9a6122a
refactor: pass SVD driver and device explicitly
44e30b8f
refactor: reuse the configured device for SVDQuant
da9cde2a
refactor: run SVDQuant on the weight device
4de495e3
changwangss
changed the title perf: avoid unused SVD reconstruction in SVDQuant perf: avoid unused SVD reconstruction and reuse the CUDA driver in SVDQuant 27 days ago
changwangss
changed the title perf: avoid unused SVD reconstruction and reuse the CUDA driver in SVDQuant perf: avoid unused SVD reconstruction and select the CUDA driver in SVDQuant 27 days ago
yiliu30
approved these changes
on 2026-09-16
chensuyue
merged
903ba0ce
into main 25 days ago
chensuyue
deleted the wangchang/svdquant-svd-factors branch 25 days ago
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub