From cfad212b4e35a1e3fecdef77ec541553bb6829f3 Mon Sep 17 00:00:00 2001
From: "hongliang.yuan" <hongliang.yuan@iluvatar.com>
Date: Tue, 31 Dec 2024 10:31:04 +0800
Subject: [PATCH] fix load int8 dataset from local and pip package version

---
 .../bert_large_squad/igie/Int8QAT/run_qat.py         |  3 ++-
 .../bert_large_squad/igie/Int8QAT/squad_download.py  |  3 ++-
 .../language_model/bert_large_squad/igie/README.md   | 12 +++++++++++-
 .../bert_large_squad/igie/requirements.txt           |  4 ++--
 4 files changed, 17 insertions(+), 5 deletions(-)

diff --git a/models/nlp/language_model/bert_large_squad/igie/Int8QAT/run_qat.py b/models/nlp/language_model/bert_large_squad/igie/Int8QAT/run_qat.py
index 943d04f4..e53b85b3 100644
--- a/models/nlp/language_model/bert_large_squad/igie/Int8QAT/run_qat.py
+++ b/models/nlp/language_model/bert_large_squad/igie/Int8QAT/run_qat.py
@@ -350,6 +350,7 @@ def main():
             os.path.join(cur_dir,'squad_download.py'),
             data_args.dataset_config_name,
             cache_dir=model_args.cache_dir,
+            trust_remote_code=True,
         )
     else:
         data_files = {}
@@ -685,7 +686,7 @@ def main():
         ]
         return EvalPrediction(predictions=formatted_predictions, label_ids=references)
 
-    metric = load_metric(os.path.join(cur_dir,'squad_data.py'))
+    metric = load_metric(os.path.join(cur_dir,'squad_data.py'), trust_remote_code=True)
 
     def compute_metrics(p: EvalPrediction):
         return metric.compute(predictions=p.predictions, references=p.label_ids)
diff --git a/models/nlp/language_model/bert_large_squad/igie/Int8QAT/squad_download.py b/models/nlp/language_model/bert_large_squad/igie/Int8QAT/squad_download.py
index 400b626a..3a4bedf8 100644
--- a/models/nlp/language_model/bert_large_squad/igie/Int8QAT/squad_download.py
+++ b/models/nlp/language_model/bert_large_squad/igie/Int8QAT/squad_download.py
@@ -121,7 +121,8 @@ class Squad(datasets.GeneratorBasedBuilder):
         )
 
     def _split_generators(self, dl_manager):
-        downloaded_files = dl_manager.download_and_extract(_URLS)
+        # downloaded_files = dl_manager.download_and_extract(_URLS)
+        downloaded_files = { "train": "./data/train-v1.1.json", "dev": "./data/dev-v1.1.json" }
 
         return [
             datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"filepath": downloaded_files["train"]}),
diff --git a/models/nlp/language_model/bert_large_squad/igie/README.md b/models/nlp/language_model/bert_large_squad/igie/README.md
index f7bbf87a..7302c3f0 100644
--- a/models/nlp/language_model/bert_large_squad/igie/README.md
+++ b/models/nlp/language_model/bert_large_squad/igie/README.md
@@ -25,11 +25,21 @@ Dataset: <https://rajpurkar.github.io/SQuAD-explorer>
 # Get FP16 Onnx Model
 python3 export.py --output bert-large-uncased-squad-v1.onnx
 
-# Do QAT for INT8 test, will take a long time  
+# Do QAT for INT8 test, will take a long time (16 gpus need 1h)
 cd Int8QAT
+
+# prepare dataset
+mkdir -p data
+cp /path/to/SQuAD/train-v1.1.json /path/to/SQuAD/dev-v1.1.json data/
+
+# prepare model into bert-large-uncased from <https://huggingface.co/google-bert/bert-large-uncased/tree/main>
+mkdir -p bert-large-uncased
+
 bash run_qat.sh
 
+# model: quant_bert_large/pytorch_model.bin or quant_bert_large/model.safetensors
 python3 export_hdf5.py --model quant_bert_large/pytorch_model.bin --model_name large
+
 cd ..
 
 ```
diff --git a/models/nlp/language_model/bert_large_squad/igie/requirements.txt b/models/nlp/language_model/bert_large_squad/igie/requirements.txt
index 21949dbe..2b340d58 100644
--- a/models/nlp/language_model/bert_large_squad/igie/requirements.txt
+++ b/models/nlp/language_model/bert_large_squad/igie/requirements.txt
@@ -1,4 +1,4 @@
 onnx
 tqdm
-numpy==1.23.5
-transformers
+numpy==1.23.2
+transformers==4.37.1
-- 
Gitee