add functionnal sliders for hyperparameters

app_dialogue.py

@@ -282,7 +282,15 @@ with gr.Blocks(title="IDEFICS", theme=gr.themes.Base()) as demo:
                     interactive=True,
                     label="Top P",
                 )
-                max_output_tokens = gr.Slider(
+                top_k = gr.Slider(
+                    minimum=0.0,
+                    maximum=100.0,
+                    value=50.0,
+                    step=1.0,
+                    interactive=True,
+                    label="Top K",
+                )
+                max_new_tokens = gr.Slider(
                     minimum=0,
                     maximum=1024,
                     value=512,
@@ -290,6 +298,46 @@ with gr.Blocks(title="IDEFICS", theme=gr.themes.Base()) as demo:
                     interactive=True,
                     label="Max output tokens",
                 )
+                repetition_penalty = gr.Slider(
+                    minimum=0.0,
+                    maximum=10.0,
+                    value=1.0,
+                    step=0.1,
+                    interactive=True,
+                    label="Repetition penalty",
+                )
+                min_length = gr.Slider(
+                    minimum=0.0,
+                    maximum=50.0,
+                    value=0.0,
+                    step=1.0,
+                    interactive=True,
+                    label="No repeat ngram size",
+                )
+                length_penalty = gr.Slider(
+                    minimum=0.0,
+                    maximum=10.0,
+                    value=1.0,
+                    step=0.1,
+                    interactive=True,
+                    label="Length penalty",
+                )
+                no_repeat_ngram_size = gr.Slider(
+                    minimum=0.0,
+                    maximum=10.0,
+                    value=0.0,
+                    step=1.0,
+                    interactive=True,
+                    label="No repeat ngram size",
+                )
+                penalty_alpha = gr.Slider(
+                    minimum=0.0,
+                    maximum=10.0,
+                    value=0.95,
+                    step=1.0,
+                    interactive=True,
+                    label="Penalty alpha",
+                )
 
         with gr.Column(scale=6):
             chatbot = gr.Chatbot(
@@ -357,22 +405,30 @@ with gr.Blocks(title="IDEFICS", theme=gr.themes.Base()) as demo:
     def model_inference(
         user_prompt,
         chat_history,
+        temperature = 1.0,
+        no_repeat_ngram_size = 0,
+        max_new_tokens = 512,
+        min_length = 16,
+        repetition_penalty = 1.0,
+        length_penalty = 1.0,
+        top_k = 50,
+        top_p = 0.95,
+        penalty_alpha = 0.95,
     ):
         global processor, model, tokenizer
-
-        temperature = 1.0
-        no_repeat_ngram_size = 0
-        max_new_tokens = 512
-        min_length = 16
+        # temperature = 1.0
+        # no_repeat_ngram_size = 0
+        # max_new_tokens = 512
+        # min_length = 16
         force_words = ""
-        repetition_penalty = 1.0
+        # repetition_penalty = 1.0
         hide_special_tokens = False
         decoding_strategy = "greedy"
         num_beams = 3
-        length_penalty = 1.0
-        top_k = 50
-        top_p = 0.95
-        penalty_alpha = 0.95
+        # length_penalty = 1.0
+        # top_k = 50
+        # top_p = 0.95
+        # penalty_alpha = 0.95
 
         formated_prompt = format_prompt_with_history_and_system_conditioning(
             current_user_prompt=user_prompt.strip(),
@@ -406,13 +462,13 @@ with gr.Blocks(title="IDEFICS", theme=gr.themes.Base()) as demo:
 
     textbox.submit(
         fn=model_inference,
-        inputs=[textbox, chatbot],
+        inputs=[textbox, chatbot, temperature, ],
         outputs=[textbox, chatbot],
     )
     submit_btn.click(
         fn=model_inference,
-        inputs=[textbox, chatbot],
-        outputs=[textbox, chatbot],
+        inputs=[textbox, chatbot, temperature, no_repeat_ngram_size, max_new_tokens, min_length, repetition_penalty, length_penalty, top_k, top_p, penalty_alpha],
+        outputs=[textbox, chatbot, temperature, no_repeat_ngram_size, max_new_tokens, min_length, repetition_penalty, length_penalty, top_k, top_p, penalty_alpha],
     )
 
 demo.queue()