update

app.py

@@ -81,7 +81,7 @@ def get_preprocess_examples(inst, input):
 def update_base_llm_dropdown_along_examples(dummy_text):
     candidates = CANDIDATE_EXAMPLES[dummy_text]
     ex_llm_outputs = {f"LLM-{i+1}": candidates[i]['text'] for i in range(len(candidates))}
-    return ex_llm_outputs
+    return ex_llm_outputs, "", ""
     
 def check_save_ranker_inputs(inst, input, llm_outputs, blender_config):
     if not inst and not input:
@@ -97,7 +97,11 @@ def check_save_ranker_inputs(inst, input, llm_outputs, blender_config):
     }
 
 def check_fuser_inputs(blender_state, blender_config, ranks):
-    pass
+    if not (blender_state.get("inst", None) or blender_state.get("input", None)):
+        raise gr.Error("Please enter instruction or input context")
+    if "candidates" not in blender_state or len(ranks)==0:
+        raise gr.Error("Please rank LLM outputs first")
+    return 
 
 def llms_rank(inst, input, llm_outputs, blender_config):
     candidates = list(llm_outputs.values())
@@ -115,9 +119,11 @@ def llms_fuse(blender_state, blender_config, ranks):
     candidates = blender_state['candidates']
     top_k_for_fuser = blender_config['top_k_for_fuser']
     fuse_params = blender_config.copy()
-    del fuse_params["top_k_for_fuser"]
+    fuse_params.pop("top_k_for_fuser")
+    fuse_params.pop("source_max_length")
+    fuse_params['no_repeat_ngram_size'] = 3
     top_k_candidates = get_topk_candidates_from_ranks([ranks], [candidates], top_k=top_k_for_fuser)[0]
-    fuser_outputs = blender.fuse(instructions=[inst], inputs=[input], candidates=[top_k_candidates], **fuse_params)[0]
+    fuser_outputs = blender.fuse(instructions=[inst], inputs=[input], candidates=[top_k_candidates], **fuse_params, batch_size=1)[0]
     return [fuser_outputs, fuser_outputs]
 
 def display_fuser_output(fuser_output):
@@ -157,7 +163,7 @@ with gr.Blocks(theme='ParityError/Anime') as demo:
     saved_fuse_outputs = gr.State(value=[])
     gr.Markdown("## Blender Outputs")
     with gr.Group():
-        rank_outputs = gr.Textbox(lines=1, label="Ranking outputs", placeholder="Ranking outputs", show_label=True)
+        rank_outputs = gr.Textbox(lines=1, label="Ranks of each LLM's output", placeholder="Ranking outputs", show_label=True)
         fuser_outputs = gr.Textbox(lines=4, label="Fusing outputs", placeholder="Fusing outputs", show_label=True)
     with gr.Row():
         rank_button = gr.Button('Rank LLM Outputs', variant='primary')
@@ -173,6 +179,13 @@ with gr.Blocks(theme='ParityError/Anime') as demo:
     })
         
     with gr.Accordion(label='Advanced options', open=False):
+        top_k_for_fuser = gr.Slider(
+            label='Top-k ranked candidates to fuse',
+            minimum=1,
+            maximum=3,
+            step=1,
+            value=3,
+        )
         source_max_length = gr.Slider(
             label='Max length of Instruction + Input',
             minimum=1,
@@ -187,13 +200,6 @@ with gr.Blocks(theme='ParityError/Anime') as demo:
             step=1,
             value=DEFAULT_CANDIDATE_MAX_LENGTH,
         )
-        top_k_for_fuser = gr.Slider(
-            label='Top-k ranked candidates to fuse',
-            minimum=1,
-            maximum=3,
-            step=1,
-            value=3,
-        )
         max_new_tokens = gr.Slider(
             label='Max new tokens fuser can generate',
             minimum=1,
@@ -201,19 +207,26 @@ with gr.Blocks(theme='ParityError/Anime') as demo:
             step=1,
             value=DEFAULT_FUSER_MAX_NEW_TOKENS,
         )
-        temperature = gr.Slider(
-            label='Temperature of fuser generation',
-            minimum=0.1,
-            maximum=2.0,
-            step=0.1,
-            value=0.7,
-        )
-        top_p = gr.Slider(
-            label='Top-p of fuser generation',
-            minimum=0.05,
-            maximum=1.0,
-            step=0.05,
-            value=1.0,
+        # temperature = gr.Slider(
+        #     label='Temperature of fuser generation',
+        #     minimum=0.1,
+        #     maximum=2.0,
+        #     step=0.1,
+        #     value=0.7,
+        # )
+        # top_p = gr.Slider(
+        #     label='Top-p of fuser generation',
+        #     minimum=0.05,
+        #     maximum=1.0,
+        #     step=0.05,
+        #     value=1.0,
+        # )
+        beam_size = gr.Slider(
+            label='Beam size of fuser generation',
+            minimum=1,
+            maximum=10,
+            step=1,
+            value=4,
         )
     
     examples_dummy_textbox = gr.Textbox(lines=1, label="", placeholder="", show_label=False, visible=False)     
@@ -235,7 +248,7 @@ with gr.Blocks(theme='ParityError/Anime') as demo:
     examples_dummy_textbox.change(
         fn=update_base_llm_dropdown_along_examples,
         inputs=[examples_dummy_textbox],
-        outputs=saved_llm_outputs,
+        outputs=[saved_llm_outputs, rank_outputs, fuser_outputs],
     ).then(
         fn=display_llm_output,
         inputs=[saved_llm_outputs, selected_base_llm_name_dropdown],
@@ -278,7 +291,7 @@ with gr.Blocks(theme='ParityError/Anime') as demo:
     fuse_button.click(
         fn=check_fuser_inputs,
         inputs=[blender_state, blender_config, saved_rank_outputs],
-        outputs=[],
+        outputs=fuser_outputs,
     ).success(
         fn=llms_fuse,
         inputs=[blender_state, blender_config, saved_rank_outputs],
@@ -312,14 +325,19 @@ with gr.Blocks(theme='ParityError/Anime') as demo:
         inputs=[max_new_tokens, blender_config],
         outputs=blender_config,
     )
-    temperature.change(
-        fn=lambda x, y: y.update({"temperature": x}) or y,
-        inputs=[temperature, blender_config],
-        outputs=blender_config,
-    )
-    top_p.change(
-        fn=lambda x, y: y.update({"top_p": x}) or y,
-        inputs=[top_p, blender_config],
+    # temperature.change(
+    #     fn=lambda x, y: y.update({"temperature": x}) or y,
+    #     inputs=[temperature, blender_config],
+    #     outputs=blender_config,
+    # )
+    # top_p.change(
+    #     fn=lambda x, y: y.update({"top_p": x}) or y,
+    #     inputs=[top_p, blender_config],
+    #     outputs=blender_config,
+    # )
+    beam_size.change(
+        fn=lambda x, y: y.update({"num_beams": x}) or y,
+        inputs=[beam_size, blender_config],
         outputs=blender_config,
     )