diff --git a/prototype_source/gpu_quantization_torchao_tutorial.py b/prototype_source/gpu_quantization_torchao_tutorial.py index 513d54faba..4050a88e56 100644 --- a/prototype_source/gpu_quantization_torchao_tutorial.py +++ b/prototype_source/gpu_quantization_torchao_tutorial.py @@ -35,14 +35,12 @@ # # Segment Anything Model checkpoint setup: # -# 1. Go to the `segment-anything repo `_ and download the ``vit_h`` checkpoint. Alternatively, you can just use ``wget``: `wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth --directory-prefix= +# 1. Go to the `segment-anything repo checkpoint `_ and download the ``vit_h`` checkpoint. Alternatively, you can use ``wget`` (for example, ``wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth --directory-prefix=``). # 2. Pass in that directory by editing the code below to say: # -# .. code-block:: -# -# {sam_checkpoint_base_path}= +# .. code-block:: bash # -# This was run on an A100-PG509-200 power limited to 330.00 W +# {sam_checkpoint_base_path}= # import torch @@ -297,7 +295,7 @@ def get_sam_model(only_one_block=False, batchsize=1): # ----------------- # In this tutorial, we have learned about the quantization and optimization techniques # on the example of the segment anything model. - +# # In the end, we achieved a full-model apples to apples quantization speedup # of about 7.7% on batch size 16 (677.28ms to 729.65ms). We can push this a # bit further by increasing the batch size and optimizing other parts of