fix CI

Signed-off-by: Mengni Wang <[email protected]>

examples/nlp/huggingface_model/text_to_image/stable_diffusion/quantization/ptq_static/main.py → examples/nlp/huggingface_model/text_to_image/stable_diffusion_v1_5/quantization/ptq_static/main.py

@@ -16,18 +16,20 @@
 # under the License.
 # pylint:disable=redefined-outer-name,logging-format-interpolation
 import argparse
+import inspect
 import logging
 import os
 import time
-import inspect
+from typing import List
+
 import numpy as np
 import onnx
 import onnxruntime as ort
 import torch
-from typing import List
+from diffusers import OnnxRuntimeModel, OnnxStableDiffusionPipeline
+
 from onnx_neural_compressor import data_reader
-from onnx_neural_compressor.quantization import config, quantize, QuantType
-from diffusers import OnnxStableDiffusionPipeline, OnnxRuntimeModel
+from onnx_neural_compressor.quantization import QuantType, config, quantize
 
 logging.basicConfig(
     format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.WARN
@@ -70,17 +72,13 @@
     "tensor(double)": np.float64,
 }
 
+
 def benchmark(model):
     generator = None if args.seed is None else np.random.RandomState(args.seed)
 
-    pipe = OnnxStableDiffusionPipeline.from_pretrained(
-        args.model_path,
-        provider=args.provider
-    )
+    pipe = OnnxStableDiffusionPipeline.from_pretrained(args.model_path, provider=args.provider)
     if args.quantized_unet_path is not None:
-        unet = OnnxRuntimeModel(
-            model=ort.InferenceSession(args.quantized_unet_path, providers=[args.provider])
-        )
+        unet = OnnxRuntimeModel(model=ort.InferenceSession(args.quantized_unet_path, providers=[args.provider]))
         pipe.unet = unet
 
     image = None
@@ -249,15 +247,15 @@ def rewind(self):
             data_reader,
             weight_type=QuantType.QInt8,
             activation_type=QuantType.QUInt8,
-            op_types_to_quantize=["MatMul","Gemm"],
+            op_types_to_quantize=["MatMul", "Gemm"],
             per_channel=True,
             extra_options={
                 "SmoothQuant": True,
                 "SmoothQuantAlpha": args.alpha,
                 "WeightSymmetric": True,
                 "ActivationSymmetric": False,
                 "OpTypesToExcludeOutputQuantization": ["MatMul", "Gemm"],
-                }
-            )
+            },
+        )
         input_path = os.path.join(args.model_path, "unet/model.onnx")
         quantize(input_path, args.output_model, cfg, optimization_level=ort.GraphOptimizationLevel.ORT_ENABLE_EXTENDED)

examples/nlp/huggingface_model/text_to_image/stable_diffusion/quantization/ptq_static/run_benchmark.sh → examples/nlp/huggingface_model/text_to_image/stable_diffusion_v1_5/quantization/ptq_static/run_benchmark.sh

@@ -14,19 +14,19 @@ function init_params {
   do
     case $var in
       --input_model=*)
-          input_model=$(echo $var |cut -f2 -d=)
+          input_model=$(echo "$var" |cut -f2 -d=)
       ;;
       --quantized_unet_path=*)
-          quantized_unet_path=$(echo $var |cut -f2 -d=)
+          quantized_unet_path=$(echo "$var" |cut -f2 -d=)
       ;;
       --batch_size=*)
-          batch_size=$(echo $var |cut -f2 -d=)
+          batch_size=$(echo "$var" |cut -f2 -d=)
       ;;
       --prompt=*)
-	  prompt=$(echo $var |cut -f2 -d=)
+	  prompt=$(echo "$var" |cut -f2 -d=)
       ;;
       --image_path=*)
-	  image_path=$(echo $var |cut -f2 -d=)
+	  image_path=$(echo "$var" |cut -f2 -d=)
       ;;
     esac
   done
@@ -44,21 +44,21 @@ function run_benchmark {
 
     extra_cmd=""
 
-    if [ ! -z $quantized_unet_path ]; then
-        extra_cmd=$extra_cmd"--quantized_unet_path ${quantized_unet_path} "
+    if [ "$quantized_unet_path" ]; then
+        extra_cmd=$extra_cmd"--quantized_unet_path=${quantized_unet_path} "
     fi
 
-    if [ ! -z $prompt ]; then
-	extra_cmd=$extra_cmd"--prompt ${prompt} "
+    if [ "$prompt" ]; then
+	extra_cmd=$extra_cmd"--prompt=${prompt} "
     fi
 
-    if [ ! -z $image_path ]; then
-	extra_cmd=$extra_cmd"--image_path ${image_path} "
+    if [ "$image_path" ]; then
+	extra_cmd=$extra_cmd"--image_path=${image_path} "
     fi
 
     python main.py \
-            --model_path ${input_model} \
-            --batch_size=${batch_size-1} \
+            --model_path="${input_model}" \
+            --batch_size="${batch_size-1}" \
             --benchmark \
 	    ${extra_cmd}
 

examples/nlp/huggingface_model/text_to_image/stable_diffusion/quantization/ptq_static/run_quant.sh → examples/nlp/huggingface_model/text_to_image/stable_diffusion_v1_5/quantization/ptq_static/run_quant.sh

@@ -12,13 +12,13 @@ function init_params {
   do
     case $var in
       --input_model=*)
-          input_model=$(echo $var |cut -f2 -d=)
+          input_model=$(echo "$var" |cut -f2 -d=)
       ;;
       --output_model=*)
-          output_model=$(echo $var |cut -f2 -d=)
+          output_model=$(echo "$var" |cut -f2 -d=)
       ;;
       --alpha=*)
-          alpha=$(echo $var |cut -f2 -d=)
+          alpha=$(echo "$var" |cut -f2 -d=)
       ;;
     esac
   done
@@ -42,9 +42,9 @@ function run_tuning {
     fi
 
     python main.py \
-            --model_path ${input_model} \
-            --output_model ${output_model} \
-	    --alpha ${alpha-0.7} \
+            --model_path "${input_model}" \
+            --output_model "${output_model}" \
+	    --alpha "${alpha-0.7}" \
             --tune
 }
 

onnx_neural_compressor/algorithms/smoother/core.py

@@ -276,7 +276,9 @@ def mul(node, scale):  # pragma: no cover
                     # Mul operator has broadcast mechanism
                     self.model.remove_initializer(tensor)
                     self.model.add_initializer(
-                        onnx.helper.make_tensor(inp, tensor.data_type, list(new_tensor.shape), new_tensor.flatten().tolist())
+                        onnx.helper.make_tensor(
+                            inp, tensor.data_type, list(new_tensor.shape), new_tensor.flatten().tolist()
+                        )
                     )
                     self.tensor_scales_info[key] = (
                         1.0 / scale