import time import gradio import numpy as np import torch from transformers import LogitsProcessor from modules import html_generator, shared params = { 'active': True, 'color_by_perplexity': False, 'color_by_probability': False, 'ppl_scale': 15.0, # No slider for this right now, because I don't think it really needs to be changed. Very large perplexity scores don't show up often. 'probability_dropdown': False, 'verbose': False # For debugging mostly } class PerplexityLogits(LogitsProcessor): def __init__(self, verbose=False): self.generated_token_ids = [] self.selected_probs = [] self.top_token_ids_list = [] self.top_probs_list = [] self.perplexities_list = [] self.last_probs = None self.verbose = verbose def __call__(self, input_ids, scores): # t0 = time.time() probs = torch.softmax(scores, dim=-1, dtype=torch.float) log_probs = torch.nan_to_num(torch.log(probs)) # Note: This is to convert log(0) nan to 0, but probs*log_probs makes this 0 not affect the perplexity. entropy = -torch.sum(probs * log_probs) entropy = entropy.cpu().numpy() perplexity = round(float(np.exp(entropy)), 4) self.perplexities_list.append(perplexity) last_token_id = int(input_ids[0][-1].cpu().numpy().item()) # Store the generated tokens (not sure why this isn't accessible in the output endpoint!) self.generated_token_ids.append(last_token_id) # Get last probability, and add to the list if it wasn't there if len(self.selected_probs) > 0: # Is the selected token in the top tokens? if self.verbose: print('Probs: Token after', shared.tokenizer.decode(last_token_id)) print('Probs:', [shared.tokenizer.decode(token_id) for token_id in self.top_token_ids_list[-1][0]]) print('Probs:', [round(float(prob), 4) for prob in self.top_probs_list[-1][0]]) if last_token_id in self.top_token_ids_list[-1][0]: idx = self.top_token_ids_list[-1][0].index(last_token_id) self.selected_probs.append(self.top_probs_list[-1][0][idx]) else: self.top_token_ids_list[-1][0].append(last_token_id) last_prob = round(float(self.last_probs[last_token_id]), 4) self.top_probs_list[-1][0].append(last_prob) self.selected_probs.append(last_prob) else: self.selected_probs.append(1.0) # Placeholder for the last token of the prompt if self.verbose: pplbar = "-" if not np.isnan(perplexity): pplbar = "*" * round(perplexity) print(f"PPL: Token after {shared.tokenizer.decode(last_token_id)}\t{perplexity:.2f}\t{pplbar}") # Get top 5 probabilities top_tokens_and_probs = torch.topk(probs, 5) top_probs = top_tokens_and_probs.values.cpu().numpy().astype(float).tolist() top_token_ids = top_tokens_and_probs.indices.cpu().numpy().astype(int).tolist() self.top_token_ids_list.append(top_token_ids) self.top_probs_list.append(top_probs) probs = probs.cpu().numpy().flatten() self.last_probs = probs # Need to keep this as a reference for top probs # t1 = time.time() # print(f"PPL Processor: {(t1-t0):.3f} s") # About 1 ms, though occasionally up to around 100 ms, not sure why... # Doesn't actually modify the logits! return scores # Stores the perplexity and top probabilities ppl_logits_processor = None def logits_processor_modifier(logits_processor_list, input_ids): global ppl_logits_processor if params['active']: ppl_logits_processor = PerplexityLogits(verbose=params['verbose']) logits_processor_list.append(ppl_logits_processor) def output_modifier(text): global ppl_logits_processor # t0 = time.time() if not params['active']: return text # TODO: It's probably more efficient to do this above rather than modifying all these lists # Remove last element of perplexities_list, top_token_ids_list, top_tokens_list, top_probs_list since everything is off by one because this extension runs before generation perplexities = ppl_logits_processor.perplexities_list[:-1] top_token_ids_list = ppl_logits_processor.top_token_ids_list[:-1] top_tokens_list = [[shared.tokenizer.decode(token_id) for token_id in top_token_ids[0]] for top_token_ids in top_token_ids_list] top_probs_list = ppl_logits_processor.top_probs_list[:-1] # Remove first element of generated_token_ids, generated_tokens, selected_probs because they are for the last token of the prompt gen_token_ids = ppl_logits_processor.generated_token_ids[1:] gen_tokens = [shared.tokenizer.decode(token_id) for token_id in gen_token_ids] sel_probs = ppl_logits_processor.selected_probs[1:] end_part = '' if params['probability_dropdown'] else '' # Helps with finding the index after replacing part of the text. i = 0 for token, prob, ppl, top_tokens, top_probs in zip(gen_tokens, sel_probs, perplexities, top_tokens_list, top_probs_list): color = 'ffffff' if params['color_by_probability'] and params['color_by_perplexity']: color = probability_perplexity_color_scale(prob, ppl) elif params['color_by_perplexity']: color = perplexity_color_scale(ppl) elif params['color_by_probability']: color = probability_color_scale(prob) if token in text[i:]: if params['probability_dropdown']: text = text[:i] + text[i:].replace(token, add_dropdown_html(token, color, top_tokens, top_probs[0], ppl), 1) else: text = text[:i] + text[i:].replace(token, add_color_html(token, color), 1) i += text[i:].find(end_part) + len(end_part) # Use full perplexity list for calculating the average here. print('Average perplexity:', round(np.mean(ppl_logits_processor.perplexities_list[:-1]), 4)) # t1 = time.time() # print(f"Modifier: {(t1-t0):.3f} s") # About 50 ms return text def probability_color_scale(prob): ''' Green-yellow-red color scale ''' rv = 0 gv = 0 if prob <= 0.5: rv = 'ff' gv = hex(int(255 * prob * 2))[2:] if len(gv) < 2: gv = '0' * (2 - len(gv)) + gv else: rv = hex(int(255 - 255 * (prob - 0.5) * 2))[2:] gv = 'ff' if len(rv) < 2: rv = '0' * (2 - len(rv)) + rv return rv + gv + '00' def perplexity_color_scale(ppl): ''' Red component only, white for 0 perplexity (sorry if you're not in dark mode) ''' value = hex(max(int(255.0 - params['ppl_scale'] * (float(ppl) - 1.0)), 0))[2:] if len(value) < 2: value = '0' * (2 - len(value)) + value return 'ff' + value + value def probability_perplexity_color_scale(prob, ppl): ''' Green-yellow-red for probability and blue component for perplexity ''' rv = 0 gv = 0 bv = hex(min(max(int(params['ppl_scale'] * (float(ppl) - 1.0)), 0), 255))[2:] if len(bv) < 2: bv = '0' * (2 - len(bv)) + bv if prob <= 0.5: rv = 'ff' gv = hex(int(255 * prob * 2))[2:] if len(gv) < 2: gv = '0' * (2 - len(gv)) + gv else: rv = hex(int(255 - 255 * (prob - 0.5) * 2))[2:] gv = 'ff' if len(rv) < 2: rv = '0' * (2 - len(rv)) + rv return rv + gv + bv def add_color_html(token, color): return f'{token}' # TODO: Major issue: Applying this to too many tokens will cause a permanent slowdown in generation speed until the messages are removed from the history. # I think the issue is from HTML elements taking up space in the visible history, and things like history deepcopy add latency proportional to the size of the history. # Potential solution is maybe to modify the main generation code to send just the internal text and not the visible history, to avoid moving too much around. # I wonder if we can also avoid using deepcopy here. def add_dropdown_html(token, color, top_tokens, top_probs, perplexity=0): html = f'
{token}
' return html # About 750 characters per token... def custom_css(): return """ .dropdown { display: none; position: absolute; z-index: 50; background-color: var(--block-background-fill); box-shadow: 0px 8px 16px 0px rgba(0,0,0,0.2); width: max-content; overflow: visible; padding: 5px; border-radius: 10px; border: 1px solid var(--border-color-primary); } .dropdown-content { border: none; z-index: 50; } .dropdown-content tr.selected { background-color: var(--block-label-background-fill); } .dropdown-content td { color: var(--body-text-color); } .hoverable { color: var(--body-text-color); position: relative; display: inline-block; overflow: visible; font-size: 15px; line-height: 1.75; margin: 0; padding: 0; } .hoverable:hover .dropdown { display: block; } pre { white-space: pre-wrap; } # TODO: This makes the hover menus extend outside the bounds of the chat area, which is good. # However, it also makes the scrollbar disappear, which is bad. # The scroll bar needs to still be present. So for now, we can't see dropdowns that extend past the edge of the chat area. #.chat { # overflow-y: auto; #} """ # Monkeypatch applied to html_generator.py # We simply don't render markdown into HTML. We wrap everything in
 tags to preserve whitespace
# formatting. If you're coloring tokens by perplexity or probability, or especially if you're using
# the probability dropdown, you probably care more about seeing the tokens the model actually outputted
# rather than rendering ```code blocks``` or *italics*.
def convert_to_markdown(string):
    return '
' + string + '
' html_generator.convert_to_markdown = convert_to_markdown def ui(): def update_active_check(x): params.update({'active': x}) def update_color_by_ppl_check(x): params.update({'color_by_perplexity': x}) def update_color_by_prob_check(x): params.update({'color_by_probability': x}) def update_prob_dropdown_check(x): params.update({'probability_dropdown': x}) active_check = gradio.Checkbox(value=True, label="Compute probabilities and perplexity scores", info="Activate this extension. Note that this extension currently does not work with exllama or llama.cpp.") color_by_ppl_check = gradio.Checkbox(value=False, label="Color by perplexity", info="Higher perplexity is more red. If also showing probability, higher perplexity has more blue component.") color_by_prob_check = gradio.Checkbox(value=False, label="Color by probability", info="Green-yellow-red linear scale, with 100% green, 50% yellow, 0% red.") prob_dropdown_check = gradio.Checkbox(value=False, label="Probability dropdown", info="Hover over a token to show a dropdown of top token probabilities. Currently slightly buggy with whitespace between tokens.") active_check.change(update_active_check, active_check, None) color_by_ppl_check.change(update_color_by_ppl_check, color_by_ppl_check, None) color_by_prob_check.change(update_color_by_prob_check, color_by_prob_check, None) prob_dropdown_check.change(update_prob_dropdown_check, prob_dropdown_check, None)