Spaces:
Running
on
Zero
Running
on
Zero
Patches
Browse files
app.py
CHANGED
@@ -1,6 +1,7 @@
|
|
1 |
import os
|
2 |
import gradio as gr
|
3 |
import torch
|
|
|
4 |
|
5 |
from bytelatent.data.file_util import get_fs
|
6 |
from bytelatent.generate_patcher import patcher_nocache
|
@@ -12,16 +13,78 @@ from download_blt_weights import main as ensure_present
|
|
12 |
# --- Global Setup (Consider loading models outside if necessary) ---
|
13 |
# Kept inside the function for simplicity as before.
|
14 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
def process_text(prompt: str, model_name: str = "blt-1b"):
|
16 |
"""
|
17 |
-
Processes the input prompt using the ByteLatent model and returns
|
|
|
18 |
|
19 |
Args:
|
20 |
prompt: The input text string from the Gradio interface.
|
21 |
model_name: The name of the model to use.
|
22 |
|
23 |
Returns:
|
24 |
-
A
|
|
|
|
|
|
|
25 |
"""
|
26 |
try:
|
27 |
# --- Model and Tokenizer Loading ---
|
@@ -63,55 +126,69 @@ def process_text(prompt: str, model_name: str = "blt-1b"):
|
|
63 |
|
64 |
if not results:
|
65 |
print("Processing returned no results.")
|
66 |
-
return "Processing completed, but no results were generated."
|
67 |
|
68 |
batch_patch_lengths, batch_scores, batch_tokens = results
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
)
|
80 |
-
|
81 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
82 |
# --- End Processing ---
|
83 |
|
84 |
-
return fig
|
85 |
|
86 |
except FileNotFoundError as e:
|
87 |
print(f"Error: {e}")
|
88 |
-
|
89 |
-
return f"Error: {str(e)}" # Return error as text output
|
90 |
except Exception as e:
|
91 |
print(f"An unexpected error occurred: {e}")
|
92 |
import traceback
|
93 |
traceback.print_exc()
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
),
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
allow_flagging="never",
|
108 |
-
)
|
109 |
|
110 |
with gr.Blocks() as iface:
|
111 |
gr.Markdown("# ByteLatent Entropy Visualizer") # Title
|
112 |
gr.Markdown(
|
113 |
"Process any prompt (limited to 512 bytes) with the 100M entropy patcher model "
|
114 |
-
"and visualize the token entropies plot below.<br><br>" # Updated description
|
115 |
"NOTE: this implementation differs slightly by excluding local attention so we limit "
|
116 |
"the characters limit to 512 to avoid any deviation.",
|
117 |
line_breaks=True
|
@@ -121,20 +198,33 @@ with gr.Blocks() as iface:
|
|
121 |
prompt_input = gr.Textbox(
|
122 |
label="Input Prompt",
|
123 |
value="Daenerys Targaryen is in Game of Thrones, a fantasy epic by George R.R. Martin.",
|
124 |
-
placeholder="
|
125 |
-
max_length=512
|
|
|
126 |
)
|
127 |
-
submit_button = gr.Button("Generate
|
128 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
129 |
|
130 |
-
# Define the action
|
131 |
submit_button.click(
|
132 |
fn=process_text,
|
133 |
-
inputs=prompt_input,
|
134 |
-
outputs=plot_output
|
135 |
)
|
136 |
|
137 |
# --- Launch the Gradio App ---
|
138 |
if __name__ == "__main__":
|
139 |
-
ensure_present(["blt-1b"])
|
140 |
iface.launch()
|
|
|
1 |
import os
|
2 |
import gradio as gr
|
3 |
import torch
|
4 |
+
import itertools # Import itertools for color cycling
|
5 |
|
6 |
from bytelatent.data.file_util import get_fs
|
7 |
from bytelatent.generate_patcher import patcher_nocache
|
|
|
13 |
# --- Global Setup (Consider loading models outside if necessary) ---
|
14 |
# Kept inside the function for simplicity as before.
|
15 |
|
16 |
+
# Define colors for patches (similar to the image style)
|
17 |
+
# Using colors from a qualitative colormap (e.g., Colorbrewer Set3 or Paired)
|
18 |
+
PATCH_COLORS = [
|
19 |
+
"#a6cee3", "#1f78b4", "#b2df8a", "#33a02c", "#fb9a99", "#e31a1c",
|
20 |
+
"#fdbf6f", "#ff7f00", "#cab2d6", "#6a3d9a", "#ffff99", "#b15928"
|
21 |
+
] # Add more if you expect many patches
|
22 |
+
|
23 |
+
|
24 |
+
def create_highlighted_text_data(tokenizer, patch_lengths_tensor, tokens_tensor, colors):
|
25 |
+
"""
|
26 |
+
Generates the data structure needed for gr.HighlightedText based on patches.
|
27 |
+
|
28 |
+
Args:
|
29 |
+
tokenizer: The BltTokenizer instance.
|
30 |
+
patch_lengths_tensor: Tensor containing the length of each patch (in tokens).
|
31 |
+
tokens_tensor: Tensor containing the token IDs for the entire sequence.
|
32 |
+
colors: A list of color hex codes to cycle through.
|
33 |
+
|
34 |
+
Returns:
|
35 |
+
A list of tuples for gr.HighlightedText, e.g., [(text, label), ...].
|
36 |
+
Returns None if input tensors are invalid.
|
37 |
+
"""
|
38 |
+
if patch_lengths_tensor is None or tokens_tensor is None or patch_lengths_tensor.numel() == 0:
|
39 |
+
return None
|
40 |
+
|
41 |
+
patch_lengths = patch_lengths_tensor.tolist()
|
42 |
+
all_tokens = tokens_tensor.tolist()
|
43 |
+
highlighted_data = []
|
44 |
+
current_token_index = 0
|
45 |
+
color_cycler = itertools.cycle(colors) # Use itertools to cycle through colors
|
46 |
+
|
47 |
+
for i, length in enumerate(patch_lengths):
|
48 |
+
if length <= 0: # Skip empty patches if they somehow occur
|
49 |
+
continue
|
50 |
+
patch_token_ids = all_tokens[current_token_index : current_token_index + length]
|
51 |
+
if not patch_token_ids: # Should not happen if length > 0, but good practice
|
52 |
+
continue
|
53 |
+
|
54 |
+
patch_text = tokenizer.decode(patch_token_ids)
|
55 |
+
patch_label = f"Patch {i+1}" # Unique label for each patch
|
56 |
+
patch_color = next(color_cycler) # Get the next color
|
57 |
+
|
58 |
+
# Add to highlighted_data: (text, label_for_coloring)
|
59 |
+
highlighted_data.append((patch_text, patch_label))
|
60 |
+
current_token_index += length
|
61 |
+
|
62 |
+
# Check if all tokens were consumed (optional sanity check)
|
63 |
+
if current_token_index != len(all_tokens):
|
64 |
+
print(f"Warning: Token mismatch. Consumed {current_token_index}, total {len(all_tokens)}")
|
65 |
+
# Decode any remaining tokens if necessary, though this indicates a logic issue
|
66 |
+
remaining_tokens = all_tokens[current_token_index:]
|
67 |
+
if remaining_tokens:
|
68 |
+
remaining_text = tokenizer.decode(remaining_tokens)
|
69 |
+
highlighted_data.append((remaining_text, "Remainder")) # Assign a generic label
|
70 |
+
|
71 |
+
return highlighted_data
|
72 |
+
|
73 |
+
|
74 |
def process_text(prompt: str, model_name: str = "blt-1b"):
|
75 |
"""
|
76 |
+
Processes the input prompt using the ByteLatent model and returns
|
77 |
+
an entropy plot and color-coded text data.
|
78 |
|
79 |
Args:
|
80 |
prompt: The input text string from the Gradio interface.
|
81 |
model_name: The name of the model to use.
|
82 |
|
83 |
Returns:
|
84 |
+
A tuple containing:
|
85 |
+
- Matplotlib Figure for the entropy plot (or None on error).
|
86 |
+
- List of tuples for gr.HighlightedText (or None on error/no results).
|
87 |
+
- Error message string (or None if successful).
|
88 |
"""
|
89 |
try:
|
90 |
# --- Model and Tokenizer Loading ---
|
|
|
126 |
|
127 |
if not results:
|
128 |
print("Processing returned no results.")
|
129 |
+
return None, None, "Processing completed, but no results were generated."
|
130 |
|
131 |
batch_patch_lengths, batch_scores, batch_tokens = results
|
132 |
+
|
133 |
+
# Process the first (and only) result in the batch
|
134 |
+
patch_lengths = batch_patch_lengths[0]
|
135 |
+
scores = batch_scores[0]
|
136 |
+
tokens = batch_tokens[0]
|
137 |
+
|
138 |
+
# Decode the full output once for the plot labels (if needed by plot_entropies)
|
139 |
+
# Note: BltTokenizer might decode directly to bytes, then utf-8. Ensure it handles errors.
|
140 |
+
try:
|
141 |
+
# Using the raw tokens tensor for decoding consistency
|
142 |
+
decoded_output_for_plot = tokenizer.decode(tokens.tolist())
|
143 |
+
except Exception as decode_err:
|
144 |
+
print(f"Warning: Error decoding full sequence for plot: {decode_err}")
|
145 |
+
# Fallback: attempt to decode the original prompt if possible, or use generic labels
|
146 |
+
decoded_output_for_plot = prompt # Use original prompt as fallback
|
147 |
+
|
148 |
+
# Generate the plot
|
149 |
+
fig = plot_entropies(
|
150 |
+
patch_lengths,
|
151 |
+
scores,
|
152 |
+
decoded_output_for_plot, # Pass the decoded string for plot labels
|
153 |
+
threshold=patcher.threshold
|
154 |
+
)
|
155 |
+
|
156 |
+
# Generate data for HighlightedText
|
157 |
+
highlighted_data = create_highlighted_text_data(
|
158 |
+
tokenizer, patch_lengths, tokens, PATCH_COLORS
|
159 |
+
)
|
160 |
+
|
161 |
+
print("Processing and visualization data generation complete.")
|
162 |
# --- End Processing ---
|
163 |
|
164 |
+
return fig, highlighted_data, None # Return plot, highlighted text data, no error
|
165 |
|
166 |
except FileNotFoundError as e:
|
167 |
print(f"Error: {e}")
|
168 |
+
return None, None, f"Error: {str(e)}" # Return None for plot/text, error message
|
|
|
169 |
except Exception as e:
|
170 |
print(f"An unexpected error occurred: {e}")
|
171 |
import traceback
|
172 |
traceback.print_exc()
|
173 |
+
return None, None, f"An unexpected error occurred: {e}" # Return None for plot/text, error message
|
174 |
+
|
175 |
+
# --- Gradio Interface ---
|
176 |
+
|
177 |
+
# Create the color map for HighlightedText dynamically
|
178 |
+
# Generate enough patch labels and map them to the cycled colors
|
179 |
+
MAX_EXPECTED_PATCHES = 50 # Estimate a reasonable maximum
|
180 |
+
color_map = {
|
181 |
+
f"Patch {i+1}": color
|
182 |
+
for i, color in zip(range(MAX_EXPECTED_PATCHES), itertools.cycle(PATCH_COLORS))
|
183 |
+
}
|
184 |
+
# Add a color for the potential 'Remainder' label from create_highlighted_text_data
|
185 |
+
color_map["Remainder"] = "#808080" # Grey for any leftovers
|
|
|
|
|
186 |
|
187 |
with gr.Blocks() as iface:
|
188 |
gr.Markdown("# ByteLatent Entropy Visualizer") # Title
|
189 |
gr.Markdown(
|
190 |
"Process any prompt (limited to 512 bytes) with the 100M entropy patcher model "
|
191 |
+
"and visualize the token entropies plot and color-coded patches below.<br><br>" # Updated description
|
192 |
"NOTE: this implementation differs slightly by excluding local attention so we limit "
|
193 |
"the characters limit to 512 to avoid any deviation.",
|
194 |
line_breaks=True
|
|
|
198 |
prompt_input = gr.Textbox(
|
199 |
label="Input Prompt",
|
200 |
value="Daenerys Targaryen is in Game of Thrones, a fantasy epic by George R.R. Martin.",
|
201 |
+
placeholder="Enter text here...",
|
202 |
+
max_length=512,
|
203 |
+
lines=3
|
204 |
)
|
205 |
+
submit_button = gr.Button("Generate Visualization") # Update button text
|
206 |
+
|
207 |
+
# Output for error messages or status
|
208 |
+
status_output = gr.Textbox(label="Status", interactive=False)
|
209 |
+
|
210 |
+
# Output component for the color-coded text
|
211 |
+
highlighted_output = gr.HighlightedText(
|
212 |
+
label="Patched Text Visualization",
|
213 |
+
color_map=color_map,
|
214 |
+
show_legend=False # Show the patch labels and colors
|
215 |
+
)
|
216 |
+
|
217 |
+
# Output component for the plot
|
218 |
+
plot_output = gr.Plot(label="Entropy vs. Token Index (with Patch Threshold)")
|
219 |
|
220 |
+
# Define the action for the button click
|
221 |
submit_button.click(
|
222 |
fn=process_text,
|
223 |
+
inputs=prompt_input,
|
224 |
+
outputs=[plot_output, highlighted_output, status_output] # Order matters!
|
225 |
)
|
226 |
|
227 |
# --- Launch the Gradio App ---
|
228 |
if __name__ == "__main__":
|
229 |
+
ensure_present(["blt-1b"]) # Ensure model is present before launching
|
230 |
iface.launch()
|