Fixed issue for T5 base model quantization issue with IPEX smoothquant (#946)

Author: PenghuiCheng

examples/.config/pytorch_optimize.json

@@ -2068,7 +2068,6 @@
       "cmd": "bash run_tuning.sh",
       "params": {
         "topology": "flan-t5-large",
-        "task": "generation",
         "approach": "static",
         "backend": "ipex",
         "output_model": "saved_results"
@@ -2078,7 +2077,7 @@
       "cmd": "bash run_benchmark.sh",
       "params": {
         "topology": "flan-t5-large",
-        "task": "generation",
+        "lm_eval_tasks": "cnn_dailymail",
         "approach": "static",
         "backend": "ipex",
         "mode": "benchmark",
@@ -2095,7 +2094,6 @@
       "cmd": "bash run_tuning.sh",
       "params": {
         "topology": "t5-base-tag",
-        "task": "generation",
         "approach": "static",
         "backend": "ipex",
         "output_model": "saved_results"
@@ -2105,7 +2103,7 @@
       "cmd": "bash run_benchmark.sh",
       "params": {
         "topology": "t5-base-tag",
-        "task": "generation",
+        "lm_eval_tasks": "cnn_dailymail",
         "approach": "static",
         "backend": "ipex",
         "mode": "benchmark",

examples/huggingface/pytorch/text2text-generation/requirements.txt

@@ -7,5 +7,7 @@ onnx
 rouge-score
 nltk
 neural-compressor
-optimum-intel
-git+https://github.com/intel/intel-extension-for-pytorch.git
+optimum-intel > 1.12.0
+onnxruntime
+intel-extension-for-pytorch
+git+https://github.com/EleutherAI/lm-evaluation-harness.git@cc9778fbe4fa1a709be2abed9deb6180fd40e7e2

examples/huggingface/pytorch/text2text-generation/run_benchmark.sh

@@ -98,7 +98,7 @@ function run_benchmark {
         elif [[ ${int8} == "true" ]]; then
             model_name_or_path=${tuned_checkpoint}
         else
-            model_name_or_path="fabiochiu/t5-base-tag-generation"
+            model_name_or_path="t5-base"
         fi
         if [ "${backend}" = "ipex" ]; then
             extra_cmd=$extra_cmd" --ipex"

examples/huggingface/pytorch/text2text-generation/run_seq2seq_generation.py

@@ -72,10 +72,10 @@ def set_seed(args):
 
 def get_example_inputs(model):
     onnx_config_class = TasksManager.get_exporter_config_constructor(model_type=model.config.model_type, exporter="onnx", task="text2text-generation")
-    onnx_config = onnx_config_class(model.config, use_past=model.config.use_cache)
+    onnx_config = onnx_config_class(model.config, use_past=model.config.use_cache, use_past_in_inputs=model.config.use_cache)
     encoder_onnx_config = onnx_config.with_behavior("encoder")
     decoder_onnx_config = onnx_config.with_behavior("decoder", use_past=False)
-    decoder_with_past_onnx_config = onnx_config.with_behavior("decoder", use_past=True)
+    decoder_with_past_onnx_config = onnx_config.with_behavior("decoder", use_past=True, use_past_in_inputs=model.config.use_cache)
     encoder_dummy_inputs = encoder_onnx_config.generate_dummy_inputs(framework="pt")
     decoder_dummy_inputs = decoder_onnx_config.generate_dummy_inputs(framework="pt")
     decoder_dummy_inputs["encoder_outputs"] = tuple(decoder_dummy_inputs["encoder_outputs"][0:1])

examples/huggingface/pytorch/text2text-generation/run_tuning.sh

@@ -73,7 +73,7 @@ function run_tuning {
         if [ $input_model ];then
             model_name_or_path=${input_model}
         else
-            model_name_or_path="fabiochiu/t5-base-tag-generation"
+            model_name_or_path="t5-base"
         fi
         if [ "${backend}" = "ipex" ]; then
             extra_cmd=$extra_cmd" --ipex"