Add files via upload

Author: pass-lin

examples/test-Qwen-generate.py

@@ -0,0 +1,38 @@
+import os
+os.environ["KERAS_BACKEND"] = "torch"
+
+max_len=300
+import keras
+keras.config.set_dtype_policy("bfloat16")
+from transformers import AutoTokenizer
+from bert4keras3.models import build_transformer_model
+
+base_path = 'Qwen1.5-0.5B/'
+config_path = base_path+'config.json'
+weights_path = base_path+'QWen.weights.h5'
+dict_path = base_path+'qwen_tokenizer'
+
+tokenizer = AutoTokenizer.from_pretrained(dict_path)
+
+Qwen = build_transformer_model(
+    config_path,
+    keras_weights_path=weights_path, 
+    model='qwen',
+    with_lm=True,
+    return_keras_model=False,
+)
+
+gemma = Qwen.model
+gemma.summary()
+import numpy as np 
+
+cache_model=Qwen.build_cache_model([max_len],end_token=151643,progress_print=True,search_mode='topp',k=0.7)
+
+test_text = '''楼房大小的血睛鬃毛狮，力大无穷的紫睛金毛猿，毁天灭地的九头蛇皇，携带着毁灭雷电的恐怖雷龙…… '''
+tokens2 = tokenizer(test_text)['input_ids']
+#print(tokens2)
+tokens2=np.expand_dims(tokens2+[0]*(max_len-len(tokens2)),0)
+tokens2 = np.concatenate([tokens2,tokens2],0)
+tokens2 = cache_model.predict(tokens2)
+print(tokenizer.decode(tokens2[0]))
+print(tokenizer.decode(tokens2[1]))

examples/test-llama-generate.py

@@ -0,0 +1,39 @@
+import os
+os.environ["KERAS_BACKEND"] = "torch"
+
+max_len=300
+import keras
+keras.config.set_dtype_policy("bfloat16")
+from transformers import AutoTokenizer
+from bert4keras3.models import build_transformer_model
+
+
+base_path = 'Yi-6B/'
+config_path = base_path+'config.json'
+weights_path = base_path+'model.weights.h5'
+dict_path = base_path+'Yi_tokenizer'
+
+tokenizer = AutoTokenizer.from_pretrained(dict_path)
+
+Qwen = build_transformer_model(
+    config_path,
+    keras_weights_path=weights_path, 
+    model='qwen',
+    with_lm=True,
+    return_keras_model=False,
+)
+
+gemma = Qwen.model
+gemma.summary()
+import numpy as np 
+#yi的end_token是2，llama3是128001，具体自己看hf的config
+cache_model=Qwen.build_cache_model([max_len],end_token=2,progress_print=True,search_mode='topp',k=0.7)
+
+test_text = '''楼房大小的血睛鬃毛狮，力大无穷的紫睛金毛猿，毁天灭地的九头蛇皇，携带着毁灭雷电的恐怖雷龙…… '''
+tokens2 = tokenizer(test_text)['input_ids']
+print(tokens2)
+tokens2=np.expand_dims(tokens2+[0]*(max_len-len(tokens2)),0)
+tokens2 = np.concatenate([tokens2,tokens2],0)
+tokens2 = cache_model.predict(tokens2)
+print(tokenizer.decode(tokens2[0]))
+print(tokenizer.decode(tokens2[1]))