Spaces:
Running
on
Zero
Running
on
Zero
Change interface
Browse files
app.py
CHANGED
|
@@ -73,6 +73,10 @@ def noisify_answer(input_ids, answer_start, threshold=1.0, eot_weight=1.0):
|
|
| 73 |
noised[idx] = val
|
| 74 |
return noised
|
| 75 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 76 |
def generate_diffusion_text(input_ids, answer_start):
|
| 77 |
with torch.no_grad():
|
| 78 |
input_tensor = torch.tensor([input_ids], dtype=torch.long).to(model.device)
|
|
@@ -82,34 +86,20 @@ def generate_diffusion_text(input_ids, answer_start):
|
|
| 82 |
sampled = torch.multinomial(probs, num_samples=1).squeeze().tolist()
|
| 83 |
return input_ids[:answer_start] + sampled[answer_start:]
|
| 84 |
|
| 85 |
-
# ---
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
# --- Gradio Interface ---
|
| 90 |
-
|
| 91 |
-
print("Loading model...")
|
| 92 |
-
model = load_model()
|
| 93 |
-
print("✅ Model loaded.")
|
| 94 |
-
|
| 95 |
-
# --- Generation logic ---
|
| 96 |
@spaces.GPU
|
| 97 |
-
def
|
| 98 |
-
|
| 99 |
-
if question.strip() == "":
|
| 100 |
-
question = placeholder
|
| 101 |
-
|
| 102 |
-
prompt = f"User: {question}\nAssistant:"
|
| 103 |
input_ids = tokenizer.encode(prompt, add_special_tokens=False)
|
| 104 |
answer_start = find_answer_start(input_ids, assistant_marker_ids)
|
| 105 |
if answer_start is None:
|
| 106 |
-
|
|
|
|
| 107 |
|
| 108 |
input_ids = (input_ids + [pad_token] * (256 - len(input_ids)))[:256]
|
| 109 |
current_tokens = noisify_answer(input_ids, answer_start, threshold=1.0, eot_weight=eot_weight)
|
| 110 |
prev_decoded_tokens = []
|
| 111 |
last_tokens = []
|
| 112 |
-
history = ["**User:** " + question]
|
| 113 |
|
| 114 |
for i in range(max_it):
|
| 115 |
generated_tokens = generate_diffusion_text(current_tokens, answer_start)
|
|
@@ -129,8 +119,12 @@ def run_diffusion_loop(question, eot_weight, max_it, sharpness):
|
|
| 129 |
highlighted.append(text)
|
| 130 |
|
| 131 |
prev_decoded_tokens = decoded_tokens
|
|
|
|
|
|
|
|
|
|
| 132 |
last_tokens.append(generated_tokens)
|
| 133 |
if len(last_tokens) == 3 and all(t == last_tokens[0] for t in last_tokens):
|
|
|
|
| 134 |
break
|
| 135 |
|
| 136 |
threshold = get_noising_schedule(i, max_it, sharpness=sharpness)
|
|
@@ -140,31 +134,19 @@ def run_diffusion_loop(question, eot_weight, max_it, sharpness):
|
|
| 140 |
final_tokens = tokenizer.convert_ids_to_tokens(current_tokens[answer_start:])
|
| 141 |
final_tokens = [tok for tok in final_tokens if tokenizer.convert_tokens_to_ids(tok) != eot_token_id]
|
| 142 |
final_output = tokenizer.convert_tokens_to_string(final_tokens)
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
def handle_submit(question, eot, max_it, sharp):
|
| 161 |
-
history, _ = run_diffusion_loop(question, eot, max_it, sharp)
|
| 162 |
-
return history
|
| 163 |
-
|
| 164 |
-
send_btn.click(
|
| 165 |
-
fn=handle_submit,
|
| 166 |
-
inputs=[question_input, eot_weight, max_iters, sharpness],
|
| 167 |
-
outputs=[chatbox]
|
| 168 |
-
)
|
| 169 |
-
|
| 170 |
-
demo.queue().launch()
|
|
|
|
| 73 |
noised[idx] = val
|
| 74 |
return noised
|
| 75 |
|
| 76 |
+
print("Loading model...")
|
| 77 |
+
model = load_model()
|
| 78 |
+
print("✅ Model loaded.")
|
| 79 |
+
|
| 80 |
def generate_diffusion_text(input_ids, answer_start):
|
| 81 |
with torch.no_grad():
|
| 82 |
input_tensor = torch.tensor([input_ids], dtype=torch.long).to(model.device)
|
|
|
|
| 86 |
sampled = torch.multinomial(probs, num_samples=1).squeeze().tolist()
|
| 87 |
return input_ids[:answer_start] + sampled[answer_start:]
|
| 88 |
|
| 89 |
+
# --- Diffusion Chat Function ---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 90 |
@spaces.GPU
|
| 91 |
+
def diffusion_chat(message, system_prompt, eot_weight, max_it, sharpness):
|
| 92 |
+
prompt = f"{system_prompt}\nUser: {message}\nAssistant:"
|
|
|
|
|
|
|
|
|
|
|
|
|
| 93 |
input_ids = tokenizer.encode(prompt, add_special_tokens=False)
|
| 94 |
answer_start = find_answer_start(input_ids, assistant_marker_ids)
|
| 95 |
if answer_start is None:
|
| 96 |
+
yield "<span style='color:red'><b>Error:</b> Could not find Assistant marker in input.</span>"
|
| 97 |
+
return
|
| 98 |
|
| 99 |
input_ids = (input_ids + [pad_token] * (256 - len(input_ids)))[:256]
|
| 100 |
current_tokens = noisify_answer(input_ids, answer_start, threshold=1.0, eot_weight=eot_weight)
|
| 101 |
prev_decoded_tokens = []
|
| 102 |
last_tokens = []
|
|
|
|
| 103 |
|
| 104 |
for i in range(max_it):
|
| 105 |
generated_tokens = generate_diffusion_text(current_tokens, answer_start)
|
|
|
|
| 119 |
highlighted.append(text)
|
| 120 |
|
| 121 |
prev_decoded_tokens = decoded_tokens
|
| 122 |
+
yield ("<div style='padding:0.5em'><b>Iteration {}</b><br>"
|
| 123 |
+
"<div style='background:#f5f5f5;padding:0.5em;border-radius:0.5em'>{}</div></div>").format(i+1, ''.join(highlighted))
|
| 124 |
+
|
| 125 |
last_tokens.append(generated_tokens)
|
| 126 |
if len(last_tokens) == 3 and all(t == last_tokens[0] for t in last_tokens):
|
| 127 |
+
yield f"<div style='color:gray'><i>Stopped early after {i+1} iterations (converged).</i></div>"
|
| 128 |
break
|
| 129 |
|
| 130 |
threshold = get_noising_schedule(i, max_it, sharpness=sharpness)
|
|
|
|
| 134 |
final_tokens = tokenizer.convert_ids_to_tokens(current_tokens[answer_start:])
|
| 135 |
final_tokens = [tok for tok in final_tokens if tokenizer.convert_tokens_to_ids(tok) != eot_token_id]
|
| 136 |
final_output = tokenizer.convert_tokens_to_string(final_tokens)
|
| 137 |
+
yield f"<div style='padding:0.5em'><b>Final Output:</b><br><div style='background:#e0ffe0;padding:0.5em;border-radius:0.5em'>{final_output}</div></div>"
|
| 138 |
+
|
| 139 |
+
# --- Chat Interface ---
|
| 140 |
+
demo = gr.ChatInterface(
|
| 141 |
+
diffusion_chat,
|
| 142 |
+
additional_inputs=[
|
| 143 |
+
gr.Textbox(value="You are a helpful assistant.", label="System message"),
|
| 144 |
+
gr.Slider(0, 1, value=0.4, step=0.05, label="EOT token weight (lower = longer output)"),
|
| 145 |
+
gr.Slider(1, 512, value=64, step=1, label="Max Iterations"),
|
| 146 |
+
gr.Slider(1.0, 20.0, value=5.0, step=0.5, label="Noising sharpness (lower = more noise)")
|
| 147 |
+
],
|
| 148 |
+
title="Diffusion Language Model Chat",
|
| 149 |
+
description="Iterative denoising chat interface using a fine-tuned LLaMA model."
|
| 150 |
+
)
|
| 151 |
+
|
| 152 |
+
demo.launch()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|