add sampling decoding strategies

app_dialogue.py

@@ -283,6 +283,14 @@ with gr.Blocks(title="IDEFICS", theme=gr.themes.Base()) as demo:
             )
 
             with gr.Accordion("Parameters", open=False, visible=True) as parameter_row:
+                decoding_strategy = gr.Radio(
+                    [
+                        "greedy",
+                        "sampling_top_k",
+                        "sampling_top_p",
+                    ],
+                    label="Decoding strategy",
+                )
                 temperature = gr.Slider(
                     minimum=0.0,
                     maximum=1.0,
@@ -426,6 +434,7 @@ with gr.Blocks(title="IDEFICS", theme=gr.themes.Base()) as demo:
     def model_inference(
         user_prompt,
         chat_history,
+        decoding_strategy="greedy",
         temperature=1.0,
         no_repeat_ngram_size=0,
         max_new_tokens=512,
@@ -444,7 +453,7 @@ with gr.Blocks(title="IDEFICS", theme=gr.themes.Base()) as demo:
         force_words = ""
         # repetition_penalty = 1.0
         hide_special_tokens = False
-        decoding_strategy = "greedy"
+        # decoding_strategy = "greedy"
         num_beams = 3
         # length_penalty = 1.0
         # top_k = 50
@@ -486,6 +495,7 @@ with gr.Blocks(title="IDEFICS", theme=gr.themes.Base()) as demo:
         inputs=[
             textbox,
             chatbot,
+            decoding_strategy,
             temperature,
             no_repeat_ngram_size,
             max_new_tokens,
@@ -503,6 +513,7 @@ with gr.Blocks(title="IDEFICS", theme=gr.themes.Base()) as demo:
         inputs=[
             textbox,
             chatbot,
+            decoding_strategy,
             temperature,
             no_repeat_ngram_size,
             max_new_tokens,