[PTQ] Store FP32 global scaling factors (absmax or scale_inv) for all quantized activations and weights. - #3296

Open

cspades wants to merge 16 commits into

NVIDIA:mainfrom

cspades:cye/scale_factor_buffering

Commits on Sep 13, 2026

Add optional tooling to accumulate quantization scaling factors for inference.
cspades
committed
dd5b59d
Add tests.
cspades
committed
9074933
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
cspades
committed
745658d
Fix minor bugs.
cspades
committed
25c6864
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
cspades
committed
aa4292a
Weed out NaN.
cspades
committed
f53509b
Rebase fixes.
cspades
committed
c87e8db
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
cspades
committed
a4915ef
Merge calibration tooling into the calibrate() API.
cspades
committed
f2943b8
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
cspades
committed
39da280
Tweaks.
cspades
committed
b00f1f8
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
cspades
committed
0a64a73
Fix activation recompute.
cspades
committed
19cf8bb
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
cspades
committed
1842f94
Minor fixes, still build quantizers when calibrating BF16/FP32, some calibration parity issues with RHT, and torch compile support.
cspades
committed
41923e4
nit: fix comment for bf16/fp32 calibration
cspades
committed
dea6c30