Update src/pipeline.py
Browse files- src/pipeline.py +3 -2
src/pipeline.py
CHANGED
|
@@ -15,7 +15,7 @@ from torchao.quantization.smoothquant import swap_linear_with_smooth_fq_linear,
|
|
| 15 |
from torchao.quantization.quant_api import Int8DynActInt4WeightQuantizer
|
| 16 |
# from torchao.quantization import quantize_, PerRow, float8_dynamic_activation_float8_weight
|
| 17 |
from torchao.quantization import quantize_, float8_dynamic_activation_float8_weight, PerTensor
|
| 18 |
-
|
| 19 |
|
| 20 |
HOME = os.environ["HOME"]
|
| 21 |
Pipeline = None
|
|
@@ -50,7 +50,8 @@ def load_pipeline() -> Pipeline:
|
|
| 50 |
# pipeline.vae = quantizer.quantize(pipeline.vae)
|
| 51 |
|
| 52 |
# quantize_(pipeline.vae, float8_dynamic_activation_float8_weight(granularity=PerRow()))
|
| 53 |
-
quantize_(pipeline.vae, float8_dynamic_activation_float8_weight(granularity=PerTensor()))
|
|
|
|
| 54 |
|
| 55 |
# quant_api.swap_conv2d_1x1_to_linear(pipeline.vae, conv_filter_fn)
|
| 56 |
|
|
|
|
| 15 |
from torchao.quantization.quant_api import Int8DynActInt4WeightQuantizer
|
| 16 |
# from torchao.quantization import quantize_, PerRow, float8_dynamic_activation_float8_weight
|
| 17 |
from torchao.quantization import quantize_, float8_dynamic_activation_float8_weight, PerTensor
|
| 18 |
+
from torchao.quantization import quantize_, float8_weight_only
|
| 19 |
|
| 20 |
HOME = os.environ["HOME"]
|
| 21 |
Pipeline = None
|
|
|
|
| 50 |
# pipeline.vae = quantizer.quantize(pipeline.vae)
|
| 51 |
|
| 52 |
# quantize_(pipeline.vae, float8_dynamic_activation_float8_weight(granularity=PerRow()))
|
| 53 |
+
# quantize_(pipeline.vae, float8_dynamic_activation_float8_weight(granularity=PerTensor()))
|
| 54 |
+
quantize_(model, float8_weight_only())
|
| 55 |
|
| 56 |
# quant_api.swap_conv2d_1x1_to_linear(pipeline.vae, conv_filter_fn)
|
| 57 |
|