Skip to content

Commit f7c458e

Browse files
committed
Remove INT8 DQ
Signed-off-by: Andrea Fasoli <andrea.fasoli@ibm.com>
1 parent dfd6758 commit f7c458e

File tree

2 files changed

+1
-272
lines changed

2 files changed

+1
-272
lines changed

aiu_fms_testing_utils/utils/direct_quantization.py

Lines changed: 0 additions & 260 deletions
This file was deleted.

scripts/run_encoder.py

Lines changed: 1 addition & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -37,17 +37,6 @@
3737
# Main model setup
3838
default_dtype, device, dist_strat = setup_model(args)
3939

40-
model_path = args.model_path
41-
if args.int8_direct_quantization:
42-
save_path = None
43-
44-
# !!! insert DQ for encoders here
45-
# pass default_dtype to DQ function
46-
47-
# if DQ is used, args.model_path represent FP16 ckpt but we need to load the
48-
# newly-created INT8 ckpt. Without DQ, args.model_path is the INT8 ckpt already.
49-
model_path = save_path
50-
5140
# Retrieve linear configuration (quantized or not) to instantiate FMS model
5241
linear_config = get_linear_config(args)
5342

@@ -64,7 +53,7 @@
6453
model = get_model(
6554
args.architecture,
6655
args.variant,
67-
model_path=model_path,
56+
model_path=args.model_path,
6857
device_type="cpu" if args.is_aiu_backend else args.device_type,
6958
data_type=default_dtype,
7059
source=args.model_source,

0 commit comments

Comments
 (0)