Skip to content

Commit b5c5331

Browse files
committed
fix max calib recipe
Signed-off-by: Jennifer Chen <jennifchen@nvidia.com>
1 parent b5f4b56 commit b5c5331

1 file changed

Lines changed: 3 additions & 3 deletions

File tree

modelopt_recipes/models/Nemotron-3-Super-120B-A12B/super-nvfp4-max-calib.yaml

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -42,14 +42,14 @@ quantize:
4242
enable: false
4343

4444
# MoE routed experts -> NVFP4 W4A4, block_size 16, e4m3 scale.
45-
# Weight uses static block scales (chosen by MSE); activations stay dynamic.
45+
# Max/amax calibration uses dynamic block scales for both weight and activation.
4646
# HF/export names: backbone.layers.*.mixer.experts.*.{up,down}_proj.
4747
- quantizer_name: '*mixer.experts.*weight_quantizer'
4848
enable: true
4949
cfg:
5050
block_sizes:
5151
-1: 16
52-
type: static
52+
type: dynamic
5353
scale_bits: e4m3
5454
num_bits: e2m1
5555
- quantizer_name: '*mixer.experts.*input_quantizer'
@@ -66,7 +66,7 @@ quantize:
6666
cfg:
6767
block_sizes:
6868
-1: 16
69-
type: static
69+
type: dynamic
7070
scale_bits: e4m3
7171
num_bits: e2m1
7272
- quantizer_name: '*mlp.experts*input_quantizer'

0 commit comments

Comments
 (0)