transformers
FP8 backward: chunked expert dequant, skip autograd.Function at inference
#50117
Merged
Merged
FP8 backward: chunked expert dequant, skip autograd.Function at inference #50117
ArthurZucker
merged
11 commits into
main
from
fp8-backward-chunk-guard
Pass the input gradient through FP8 layers, so adapters can train on …
49a2aadb
Same backward for FP8GroupedLinear, with a contiguous gradient for th…
9d66027a
Dequantize in place in the FP8 backward
7e829585
Fit the EP sentinel comment in five lines
93a2d1a0
Merge branch 'main' into fp8-adapter-training
3a4bd9d7
Merge branch 'main' into fp8-adapter-training
f8bd36e9
FP8 backward: chunk expert dequant, broadcast scales, skip autograd.F…
e284869b
Only chunk the FP8 grouped backward when the expert stack exceeds 1 GiB
f239b1a8
Base automatically changed from
fp8-adapter-training
to
main
6 days ago
Merge branch 'main' into fp8-backward-chunk-guard
e1a96509
SunMarc
approved these changes
on 2026-10-29
Free each dequantized expert chunk before the next one
84fcf14a
Update measured peak in comment
c1354887
disabled auto-merge 6 days ago
Manually disabled by user
Manually disabled by user
Login to write a write a comment.
Login via GitHub