5min

app.py

@@ -1,6 +1,8 @@
 import subprocess
+
 # Installing flash_attn
-subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
+subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"},
+               shell=True)
 
 from threading import Thread
 import spaces
@@ -27,7 +29,7 @@ class StopOnTokens(StoppingCriteria):
         return False
 
 
-@spaces.GPU(duration=120)
+@spaces.GPU(duration=300)
 def predict(history, prompt, max_length, top_p, temperature):
     stop = StopOnTokens()
     messages = []
@@ -80,10 +82,10 @@ with gr.Blocks() as demo:
             <a href="https://arxiv.org/pdf/2408.07055">📜 arxiv </a>
         </div>
         <div style="text-align: center; font-size: 15px; font-weight: bold; color: red; margin-bottom: 20px;">
-            ⚠️ This is just a basic demo. Due to the scheduling limitations of Huggingface ZERO GPUs, the output
-            length is restricted to under 4K per request (Otherwise may cause timeout ERROR). 
-            If you wish to experience the full capabilities of the model (output exceeding 10K), 
-            please deploy the model yourself. 
+            ⚠️ Due to the limitations of Huggingface ZERO GPUs, in order to output 10K characters in one go, 
+            we need to request a 5-minute quota each time. 
+            This will result in you only being able to use it once every 4 hours. 
+            If you plan to use it long-term, please consider deploying the model yourself.
         </div>
         """
     )
@@ -100,7 +102,8 @@ with gr.Blocks() as demo:
             pBtn = gr.Button("Set Prompt")
         with gr.Column(scale=1):
             emptyBtn = gr.Button("Clear History")
-            max_length = gr.Slider(0, 128000, value=4096, step=1.0, label="Maximum length(Input + Output)", interactive=True)
+            max_length = gr.Slider(0, 128000, value=4096, step=1.0, label="Maximum length(Input + Output)",
+                                   interactive=True)
             top_p = gr.Slider(0, 1, value=0.8, step=0.01, label="Top P", interactive=True)
             temperature = gr.Slider(0.01, 1, value=0.6, step=0.01, label="Temperature", interactive=True)