glm_api

app.py

@@ -4,7 +4,6 @@ import requests
 import json
 import os
 
-
 APIKEY = os.environ.get("APIKEY")
 APISECRET = os.environ.get("APISECRET")
 
@@ -17,12 +16,12 @@ def predict(text, seed, out_seq_length, min_gen_length, sampling_strategy,
     if text == '':
         return 'Input should not be empty!'
 
-    url = 'https://tianqi.aminer.cn/api/v2/completions_130B'
+    url = 'https://wudao.aminer.cn/os/api/api/v2/completions_130B'
 
     payload = json.dumps({
         "apikey": APIKEY,
         "apisecret": APISECRET ,
-        "model_name": "glm-130b-v1",
+        "language": "zh-CN",
         "prompt": text,
         "length_penalty": length_penalty,
         "temperature": temperature,
@@ -33,7 +32,6 @@ def predict(text, seed, out_seq_length, min_gen_length, sampling_strategy,
         "num_beams": num_beams,
         "max_tokens": out_seq_length,
         "no_repeat_ngram": no_repeat_ngram_size,
-        "quantization": "int4",
         "seed": seed
     })
 
@@ -61,9 +59,9 @@ def predict(text, seed, out_seq_length, min_gen_length, sampling_strategy,
 if __name__ == "__main__":
 
     en_fil = ['The Starry Night is an oil-on-canvas painting by [MASK] in June 1889.']
-    en_gen = ['Question: What\'s the best winter resort city? User: A 10-year professional traveler. Answer: [gMASK]'] #['Eight planets in solar system are [gMASK]']
+    en_gen = ['Eight planets in solar system are [gMASK]']
     ch_fil = ['凯旋门位于意大利米兰市古城堡旁。1807年为纪念[MASK]而建，门高25米，顶上矗立两武士青铜古兵车铸像。']
-    ch_gen = ['三亚位于海南岛的最南端,是[gMASK]']
+    ch_gen = ['三亚位于海南岛的最南端,是中国最南部的热带滨海旅游城市 [gMASK]']
     en_to_ch = ['Pencil in Chinese is [MASK].']
     ch_to_en = ['"我思故我在"的英文是"[MASK]"。']
 
@@ -72,13 +70,8 @@ if __name__ == "__main__":
     with gr.Blocks() as demo:
         gr.Markdown(
             """
-            Dear friends,
-            
-            Nice to meet you here! This is a toy demo of GLM-130B, an open bilingual pre-trained model from Tsinghua Univeristy. GLM-130B uses two different mask tokens: `[MASK]` for short blank filling and `[gMASK]` for left-to-right long text generation. When the input does not contain any MASK token, `[gMASK]` will be automatically appended to the end of the text. We recommend that you use `[MASK]` to try text fill-in-the-blank to reduce wait time (ideally within seconds without queuing).
-            
-            This demo is a raw language model **without** instruction fine-tuning (which is applied to FLAN-* series) and RLHF (which is applied to ChatGPT); its ability is roughly between OpenAI `davinci` and `text-davinci-001`. Thus, it is currently worse than ChatGPT and other instruction fine-tuned models :(
-            
-            However, we are sparing no effort to improve it, and its updated versions will meet you soon! If you find the open-source effort useful, please star our [GitHub repo](https://github.com/THUDM/GLM-130B) to encourage our following development :)
+            An Open Bilingual Pre-Trained Model. [Visit our github repo](https://github.com/THUDM/GLM-130B)
+            GLM-130B uses two different mask tokens: `[MASK]` for short blank filling and `[gMASK]` for left-to-right long text generation. When the input does not contain any MASK token, `[gMASK]` will be automatically appended to the end of the text. We recommend that you use `[MASK]` to try text fill-in-the-blank to reduce wait time (ideally within seconds without queuing).
             """)
 
         with gr.Row():
@@ -100,7 +93,7 @@ if __name__ == "__main__":
                 out_seq_length = gr.Slider(maximum=256, value=128, minimum=32, step=1, label='Output Sequence Length')
             with gr.Column():
                 min_gen_length = gr.Slider(maximum=64, value=0, step=1, label='Min Generate Length')
-                sampling_strategy = gr.Radio(choices=['BeamSearchStrategy', 'BaseStrategy'], value='BaseStrategy', label='Search Strategy')
+                sampling_strategy = gr.Radio(choices=['BeamSearchStrategy', 'BaseStrategy'], value='BeamSearchStrategy', label='Search Strategy')
 
         with gr.Row():
             with gr.Column():
@@ -118,9 +111,9 @@ if __name__ == "__main__":
                     """
                     BaseStrategy
                     """)
-                temperature = gr.Slider(maximum=1, value=1.0, minimum=0, label='Temperature')
-                topk = gr.Slider(maximum=40, value=0, minimum=0, step=1, label='Top K')
-                topp = gr.Slider(maximum=1, value=0.7, minimum=0, label='Top P')
+                temperature = gr.Slider(maximum=1, value=0.7, minimum=0, label='Temperature')
+                topk = gr.Slider(maximum=40, value=1, minimum=0, step=1, label='Top K')
+                topp = gr.Slider(maximum=1, value=0, minimum=0, label='Top P')
             
         inputs = [model_input, seed, out_seq_length, min_gen_length, sampling_strategy, num_beams, length_penalty, no_repeat_ngram_size, temperature, topk, topp]
         gen.click(fn=predict, inputs=inputs, outputs=outputs)