Don't require do_sample, just listen to temp, top_p, top_k

src/gen.py

@@ -1516,9 +1516,9 @@ def main(
         max_total_input_tokens = max_total_input_tokens_public if max_total_input_tokens is None else max_total_input_tokens
         allow_upload_to_user_data = False
         input_lines = 1  # ensure set, for ease of use
-        temperature = 0.2 if temperature is None else temperature
+        temperature = 0.3 if temperature is None else temperature
         top_p = 1.0 if top_p is None else top_p
-        top_k = 70 if top_k is None else top_k
+        top_k = 1 if top_k is None else top_k
         penalty_alpha = 0.0 if penalty_alpha is None else penalty_alpha
         if is_hf:
             do_sample = True if do_sample is None else do_sample
@@ -3783,8 +3783,6 @@ def evaluate(
         extract_frames = extract_frames0
     if seed is None:
         seed = 0
-    if seed == 0 and do_sample:
-        seed = randint(0, 32000)
 
     if isinstance(langchain_agents, str):
         if langchain_agents.strip().startswith('['):
@@ -3940,11 +3938,15 @@ def evaluate(
     top_p = min(max(1e-3, top_p), 1.0)
     top_k = min(max(1, int(top_k)), 100)
     penalty_alpha = min(2.0, max(0.0, penalty_alpha))
+    if temperature > 0.0 or top_p < 1.0 or top_k > 1:
+        do_sample = True
     if not do_sample:
         temperature = 0
         top_p = 1.0
         top_k = 1
         seed = 1
+    if seed == 0 and do_sample:
+        seed = randint(0, 32000)
     # Note: Could do below, but for now gradio way can control do_sample directly
     # elif temperature >= 0.01:
     #     do_sample = True
@@ -5244,19 +5246,19 @@ def get_generate_params(model_lower,
     else:
         prompt_type = prompt_type or ''
     if use_defaults:
-        temperature = 1.0 if temperature is None else temperature
+        temperature = 0.0 if temperature is None else temperature
         top_p = 1.0 if top_p is None else top_p
-        top_k = 40 if top_k is None else top_k
+        top_k = 1 if top_k is None else top_k
         penalty_alpha = 0 if penalty_alpha is None else penalty_alpha
         num_beams = num_beams or 1
         max_new_tokens = max_new_tokens or 512
         repetition_penalty = repetition_penalty or 1.07
         num_return_sequences = min(num_beams, num_return_sequences or 1)
         do_sample = False if do_sample is None else do_sample
     else:
-        temperature = 0.1 if temperature is None else temperature
+        temperature = 0.0 if temperature is None else temperature
         top_p = 1.0 if top_p is None else top_p
-        top_k = 40 if top_k is None else top_k
+        top_k = 1 if top_k is None else top_k
         penalty_alpha = 0 if penalty_alpha is None else penalty_alpha
         num_beams = num_beams or 1
         max_new_tokens = max_new_tokens or 1024

src/gradio_runner.py

@@ -1703,7 +1703,8 @@ def show_llava(x):
                                                                value=kwargs['stream_output'])
                         do_sample = gr.Checkbox(label="Sample",
                                                 info="Enable sampler (required for use of temperature, top_p, top_k).  If temperature=0 is set, this is forced to False.",
-                                                value=kwargs['do_sample'])
+                                                value=kwargs['do_sample'],
+                                                visible=False)
                         seed = gr.Number(value=0,
                                          minimum=0,
                                          step=1,
@@ -1716,14 +1717,14 @@ def show_llava(x):
                         temperature = gr.Slider(minimum=0, maximum=2,
                                                 value=kwargs['temperature'],
                                                 label="Temperature",
-                                                info="Lower is deterministic, higher more creative")
+                                                info="Lower is deterministic, higher more creative (e.g. 0.3 to 0.75)")
                         top_p = gr.Slider(minimum=1e-3, maximum=1.0,
                                           value=kwargs['top_p'], label="Top p",
-                                          info="Cumulative probability of tokens to sample from")
+                                          info="Cumulative probability of tokens to sample from (e.g. 0.7)")
                         top_k = gr.Slider(
                             minimum=1, maximum=100, step=1,
                             value=kwargs['top_k'], label="Top k",
-                            info='Num. tokens to sample from'
+                            info='Num. tokens to sample from (e.g. 5 to 70)'
                         )
                         penalty_alpha = gr.Slider(
                             minimum=0.0, maximum=2.0, step=0.01,