lucalp's picture
adding patch counts and cleaning up
545bc06
raw
history blame
21.3 kB
import os
import gradio as gr
import torch
import itertools # For color cycling
import tiktoken # For GPT-4 tokenizer
from transformers import AutoTokenizer # For Llama3 tokenizer - AutoModel usually not needed just for tokenizer
# Bytelatent imports (assuming they are in the python path)
try:
from bytelatent.data.file_util import get_fs
from bytelatent.generate_patcher import patcher_nocache
from bytelatent.tokenizers.blt_tokenizer import BltTokenizer
from bytelatent.plotting.entropy_figure_via_matplot_lib import plot_entropies
from bytelatent.args import TrainArgs
from download_blt_weights import main as ensure_present
BLT_AVAILABLE = True
except ImportError as e:
print(f"Warning: Bytelatent libraries not found. Bytelatent functionality will be disabled. Error: {e}")
BLT_AVAILABLE = False
# Define dummy classes/functions if BLT is not available to avoid NameErrors later
class BltTokenizer: pass
class TrainArgs: pass
def patcher_nocache(*args, **kwargs): return None
def plot_entropies(*args, **kwargs): return None
def ensure_present(*args, **kwargs): pass
# --- Global Setup ---
# Define colors for patches/tokens
VIZ_COLORS = [
"#a6cee3", "#1f78b4", "#b2df8a", "#33a02c", "#fb9a99", "#e31a1c",
"#fdbf6f", "#ff7f00", "#cab2d6", "#6a3d9a", "#ffff99", "#b15928"
] # Add more if you expect many segments
LLAMA3_MODEL_NAME = "meta-llama/Meta-Llama-3-8B" # Or choose another variant like Instruct
# --- Helper Functions ---
def create_bytelatent_highlight_data(tokenizer, patch_lengths_tensor, tokens_tensor, colors):
"""Generates data for gr.HighlightedText based on bytelatent patches."""
if not BLT_AVAILABLE:
return [("Bytelatent library not available.", "Error")]
if patch_lengths_tensor is None or tokens_tensor is None or patch_lengths_tensor.numel() == 0:
return None
patch_lengths = patch_lengths_tensor.tolist()
all_tokens = tokens_tensor.tolist()
highlighted_data = []
current_token_index = 0
patch_count = 0 # Initialize patch count
# color_cycler = itertools.cycle(colors) # Moved inside loop if needed per-patch
for i, length in enumerate(patch_lengths):
if length <= 0: continue
patch_token_ids = all_tokens[current_token_index : current_token_index + length]
if not patch_token_ids: continue
try: patch_text = tokenizer.decode(patch_token_ids)
except Exception as decode_err:
print(f"Warning: Bytelatent patch decoding failed: {decode_err}")
patch_text = f"[Decode Error: {len(patch_token_ids)} tokens]"
patch_label = f"BL Patch {i+1}"
highlighted_data.append((patch_text, patch_label))
patch_count += 1 # Increment count for each valid patch added
current_token_index += length
# Handle remainder separately, don't count it as a 'patch'
if current_token_index != len(all_tokens):
print(f"Warning: Bytelatent token mismatch. Consumed {current_token_index}, total {len(all_tokens)}")
remaining_tokens = all_tokens[current_token_index:]
if remaining_tokens:
try: remaining_text = tokenizer.decode(remaining_tokens)
except Exception: remaining_text = f"[Decode Error: {len(remaining_tokens)} remaining tokens]"
highlighted_data.append((remaining_text, "BL Remainder"))
# Return both highlighted data and the calculated patch count
return highlighted_data, patch_count
def create_tiktoken_highlight_data(prompt, colors):
"""Generates data for gr.HighlightedText based on tiktoken (gpt-4) tokens."""
try:
enc = tiktoken.get_encoding("cl100k_base")
tiktoken_ids = enc.encode(prompt)
highlighted_data = []
# color_cycler = itertools.cycle(colors) # Moved inside loop if needed per-token
for i, token_id in enumerate(tiktoken_ids):
try: token_text = enc.decode([token_id])
except UnicodeDecodeError:
try:
token_bytes = enc.decode_single_token_bytes(token_id)
token_text = f"[Bytes: {token_bytes.hex()}]"
except Exception: token_text = "[Decode Error]"
except Exception as e:
print(f"Unexpected tiktoken decode error: {e}")
token_text = "[Decode Error]"
token_label = f"GPT4 Tk {i+1}"
highlighted_data.append((token_text, token_label))
token_count = len(tiktoken_ids)
print(f"Tiktoken processing complete. Found {token_count} tokens.")
return highlighted_data, token_count
except ImportError:
print("Error: tiktoken library not found. Please install it: pip install tiktoken")
return [("tiktoken library not installed.", "Error")], 0
except Exception as tiktoken_err:
print(f"Error during tiktoken processing: {tiktoken_err}")
return [(f"Error processing with tiktoken: {str(tiktoken_err)}", "Error")], 0
def create_llama3_highlight_data(prompt, colors, model_name=LLAMA3_MODEL_NAME):
"""Generates data for gr.HighlightedText based on Llama 3 tokenizer."""
try:
# Load Llama 3 tokenizer from Hugging Face Hub
print(f"Loading Llama 3 tokenizer: {model_name}")
# Use trust_remote_code=True if required by the specific model revision
tokenizer = AutoTokenizer.from_pretrained(model_name) #, trust_remote_code=True)
print("Llama 3 tokenizer loaded.")
# Encode the prompt
llama_token_ids = tokenizer.encode(prompt)
highlighted_data = []
# color_cycler = itertools.cycle(colors) # Moved inside loop if needed per-token
for i, token_id in enumerate(llama_token_ids):
try:
# Decode individual token.
token_text = tokenizer.decode([token_id])
except Exception as e:
print(f"Unexpected Llama 3 decode error for token {token_id}: {e}")
token_text = "[Decode Error]"
token_label = f"Llama3 Tk {i+1}" # Clearer label prefix
highlighted_data.append((token_text, token_label))
token_count = len(llama_token_ids)
print(f"Llama 3 processing complete. Found {token_count} tokens.")
return highlighted_data, token_count
except ImportError:
print("Error: transformers or sentencepiece library not found. Please install them: pip install transformers sentencepiece")
return [("transformers/sentencepiece library not installed.", "Error")], 0
except OSError as e:
# Handle errors like model not found, network issues, authentication needed
print(f"Error loading Llama 3 tokenizer '{model_name}': {e}")
error_msg = f"Could not load Llama 3 tokenizer '{model_name}'. Check model name and network."
if "authentication" in str(e).lower():
error_msg = f"Authentication required for Llama 3 tokenizer '{model_name}'. Use `huggingface-cli login`."
return [(f"{error_msg} Error: {e}", "Error")], 0
except Exception as llama_err:
print(f"Error during Llama 3 processing: {llama_err}")
import traceback
traceback.print_exc() # Print full traceback for debugging
return [(f"Error processing with Llama 3: {str(llama_err)}", "Error")], 0
# --- Main Processing Function ---
def process_text(prompt: str, model_name: str = "blt-1b"):
"""
Processes the input prompt using ByteLatent, Tiktoken, and Llama 3,
returning visualizations, counts, and status.
Args:
prompt: The input text string from the Gradio interface.
model_name: The name of the bytelatent model to use.
Returns:
A tuple containing:
- Matplotlib Figure for the entropy plot (or None).
- List of tuples for bytelatent gr.HighlightedText (or None).
- Integer count of bytelatent patches.
- List of tuples for tiktoken gr.HighlightedText (or None).
- Integer count of tiktoken tokens.
- List of tuples for Llama 3 gr.HighlightedText (or None).
- Integer count of Llama 3 tokens.
- Status/Error message string.
"""
fig = None
bl_highlighted_data = None
tk_highlighted_data = None
llama_highlighted_data = None
bl_count = 0
tk_count = 0
llama_count = 0
status_message = "Starting processing..."
# --- 1. Tiktoken Processing (Independent) ---
status_message += "\nProcessing with Tiktoken (gpt-4)..."
tk_highlighted_data, tk_count_calc = create_tiktoken_highlight_data(prompt, VIZ_COLORS)
if tk_highlighted_data and tk_highlighted_data[0][1] == "Error":
status_message += f"\nTiktoken Error: {tk_highlighted_data[0][0]}"
tk_count = 0 # Ensure count is 0 on error
else:
tk_count = tk_count_calc # Assign calculated count
status_message += f"\nTiktoken processing successful ({tk_count} tokens)."
# --- 2. Llama 3 Processing (Independent) ---
status_message += "\nProcessing with Llama 3 tokenizer..."
llama_highlighted_data, llama_count_calc = create_llama3_highlight_data(prompt, VIZ_COLORS)
if llama_highlighted_data and llama_highlighted_data[0][1] == "Error":
status_message += f"\nLlama 3 Error: {llama_highlighted_data[0][0]}"
llama_count = 0 # Ensure count is 0 on error
else:
llama_count = llama_count_calc # Assign calculated count
status_message += f"\nLlama 3 processing successful ({llama_count} tokens)."
# --- 3. Bytelatent Processing ---
if BLT_AVAILABLE:
try:
status_message += f"\nLoading Bytelatent entropy model for '{model_name}'..."
# (Bytelatent loading code remains the same)
consolidated_path = os.path.join("hf-weights", model_name)
train_args_path = os.path.join(consolidated_path, "params.json")
if not os.path.exists(train_args_path): raise FileNotFoundError(f"BLT training args not found at {train_args_path}.")
fs = get_fs(train_args_path); train_args = TrainArgs.model_validate_json(fs.read_text(train_args_path))
bl_tokenizer = train_args.data.tokenizer_args.build(); assert isinstance(bl_tokenizer, BltTokenizer)
patcher_args = train_args.data.patcher_args.model_copy(deep=True); patcher_args.realtime_patching = True
device = "cuda" if torch.cuda.is_available() else "cpu"; print(f"Using BLT device: {device}")
patcher_args.patching_device = device; patcher_args.device = device
entropy_model_dir = os.path.join(consolidated_path, "entropy_model")
if not os.path.exists(entropy_model_dir): raise FileNotFoundError(f"Entropy model directory not found at {entropy_model_dir}.")
patcher_args.entropy_model_checkpoint_dir = entropy_model_dir; bl_patcher = patcher_args.build()
status_message += "\nBytelatent entropy model loaded."
# --- Processing ---
status_message += "\nRunning Bytelatent entropy model patching..."
print(f"Processing prompt with entropy model: '{prompt}'")
prompt_bytes = prompt.encode('utf-8')
max_bytes = 512 # Define max bytes
if len(prompt_bytes) > max_bytes:
print(f"Warning: Prompt exceeds {max_bytes} bytes ({len(prompt_bytes)}). Truncating for entropy model.")
# Find the byte position that corresponds to the last full character within the limit
# This avoids splitting a multi-byte character
try:
last_char_pos = prompt_bytes[:max_bytes].rfind(b' ') # Simple whitespace split point find, might not be perfect
if last_char_pos == -1: # If no space, truncate hard (less ideal)
prompt_bl = prompt_bytes[:max_bytes].decode('utf-8', errors='ignore')
else:
prompt_bl = prompt_bytes[:last_char_pos].decode('utf-8', errors='ignore')
except Exception: # Fallback to simple truncation on decode errors
prompt_bl = prompt_bytes[:max_bytes].decode('utf-8', errors='ignore')
status_message += f"\nWarning: Prompt truncated to approx {len(prompt_bl.encode('utf-8'))} bytes for Bytelatent entropy model."
else:
prompt_bl = prompt
results = patcher_nocache([prompt_bl], tokenizer=bl_tokenizer, patcher=bl_patcher)
if not results:
print("Bytelatent entropy processing returned no results.")
status_message += "\nBytelatent entropy model warning: Processing completed, but no results were generated."
bl_highlighted_data = [("No patches generated.", "Info")]
bl_count = 0
else:
batch_patch_lengths, batch_scores, batch_tokens = results
patch_lengths, scores, tokens = batch_patch_lengths[0], batch_scores[0], batch_tokens[0]
# --- Visualization Data Generation ---
try: decoded_output_for_plot = bl_tokenizer.decode(tokens.tolist())
except Exception as decode_err:
print(f"Warning: Error decoding full sequence for plot: {decode_err}")
decoded_output_for_plot = prompt_bl # Use truncated prompt for plot if decode fails
fig = plot_entropies(patch_lengths, scores, decoded_output_for_plot, threshold=bl_patcher.threshold)
bl_highlighted_data, bl_count_calc = create_bytelatent_highlight_data(bl_tokenizer, patch_lengths, tokens, VIZ_COLORS)
bl_count = bl_count_calc # Assign calculated count
status_message += f"\nBytelatent entropy model processing and visualization successful ({bl_count} patches)."
print("Bytelatent Entropy model processing and decoding complete.")
except FileNotFoundError as e:
print(f"Bytelatent Error: {e}")
status_message += f"\nBytelatent FileNotFoundError: {str(e)}"
bl_highlighted_data = [(f"Bytelatent Error: {e}", "Error")]
bl_count = 0
except Exception as e:
print(f"An unexpected Bytelatent error occurred: {e}")
import traceback
traceback.print_exc()
status_message += f"\nBytelatent Unexpected Error: {str(e)}"
bl_highlighted_data = [(f"Bytelatent Error: {e}", "Error")]
bl_count = 0
else:
status_message += "\nBytelatent processing skipped (library not found)."
bl_highlighted_data = [("Bytelatent library not available.", "Error")]
bl_count = 0
fig = None # Ensure fig is None if BLT is skipped
# Return all generated data and the final status message
return fig, bl_highlighted_data, bl_count, tk_highlighted_data, tk_count, llama_highlighted_data, llama_count, status_message
# --- Gradio Interface ---
# Create color maps for HighlightedText dynamically
MAX_EXPECTED_SEGMENTS = 2000 # Increased max segments further just in case
common_error_map = {"Error": "#FF0000", "Info": "#808080"} # Red for errors, Gray for info
bytelatent_color_map = {f"BL Patch {i+1}": color for i, color in zip(range(MAX_EXPECTED_SEGMENTS), itertools.cycle(VIZ_COLORS))}
bytelatent_color_map["BL Remainder"] = "#AAAAAA"; bytelatent_color_map.update(common_error_map)
tiktoken_color_map = {f"GPT4 Tk {i+1}": color for i, color in zip(range(MAX_EXPECTED_SEGMENTS), itertools.cycle(VIZ_COLORS))}
tiktoken_color_map.update(common_error_map)
llama3_color_map = {f"Llama3 Tk {i+1}": color for i, color in zip(range(MAX_EXPECTED_SEGMENTS), itertools.cycle(VIZ_COLORS))}
llama3_color_map.update(common_error_map)
with gr.Blocks(theme=gr.themes.Origin()) as iface:
gr.Markdown("# BLT's Entropy-based Patcher vs. Tokenizer Visualisation")
gr.Markdown(
"Enter text to visualize its segmentation according to different methods:\n"
"1. **Byte Latent Transformer (BLT):** Entropy-based patching plot and patched text (_for this space ONLY_ - limited to ~512 bytes).\n"
"2. **Tiktoken (GPT-4):** Text segmented by `cl100k_base` tokens.\n"
f"3. **Llama 3:** Text segmented by the `{LLAMA3_MODEL_NAME}` tokenizer."
)
with gr.Row():
with gr.Column(scale=1): # Input Column
prompt_input = gr.Textbox(
label="Input Prompt",
value="Daenerys Targaryen is in Game of Thrones, a fantasy epic by George R.R. Martin.",
placeholder="Enter text here...",
max_length=512, # Allow even longer input, Bytelatent will truncate
lines=5,
info="For this space ONLY, processing is limited to ~512 bytes."
)
submit_button = gr.Button("Generate Visualizations", variant="primary")
status_output = gr.Textbox(label="Processing Status", interactive=False, lines=7) # Increased lines slightly
with gr.Column(scale=2): # Output Column
# --- Bytelatent Output Area ---
with gr.Row(equal_height=False): # Use Row to place title and count together
gr.Markdown("### BLT Entropy Patcher Output (`blt_main_entropy_100m_512w`)")
bl_count_output = gr.Number(label="Patch Count", value=0, interactive=False, scale=1, step=1) # Added Number output
highlighted_output_bl = gr.HighlightedText(
label="BLT's Entropy-based Patches",
color_map=bytelatent_color_map,
show_legend=False, # Legend can get very long
# show_label=False, # Hide the HighlightedText label as we have the markdown title
show_inline_category=False,
# container=False, # Reduces vertical space slightly
)
plot_output = gr.Plot(label="Entropy vs. Token Index", show_label=True)
# --- Tiktoken Output Area ---
with gr.Row(equal_height=False):
gr.Markdown("### Tiktoken Output (`cl100k_base`)")
tk_count_output = gr.Number(label="Token Count", value=0, interactive=False, scale=1, step=1) # Added Number output
highlighted_output_tk = gr.HighlightedText(
label="Tiktoken Segmented Text",
color_map=tiktoken_color_map,
show_legend=False,
show_inline_category=False,
# show_label=False,
# container=False,
)
# --- Llama 3 Output Area ---
with gr.Row(equal_height=False):
gr.Markdown(f"### Llama 3 Output (`{LLAMA3_MODEL_NAME}`)")
llama_count_output = gr.Number(label="Token Count", value=0, interactive=False, scale=1, step=1) # Added Number output
highlighted_output_llama = gr.HighlightedText(
label="Llama 3 Segmented Text",
color_map=llama3_color_map,
show_legend=False,
show_inline_category=False,
# show_label=False,
# container=False,
)
# Define the action for the button click
submit_button.click(
fn=process_text,
inputs=prompt_input,
# Ensure order matches the 8 return values of process_text
outputs=[
plot_output, # fig
highlighted_output_bl, # bl_highlighted_data
bl_count_output, # bl_count <-- New
highlighted_output_tk, # tk_highlighted_data
tk_count_output, # tk_count <-- New
highlighted_output_llama,# llama_highlighted_data
llama_count_output, # llama_count <-- New
status_output # status_message
]
)
# --- Launch the Gradio App ---
if __name__ == "__main__":
print("Checking required libraries...")
try:
import tiktoken
print("- tiktoken found.")
except ImportError:
print("WARNING: 'tiktoken' not found. GPT-4 visualization will fail. Install with: pip install tiktoken")
try:
import transformers
import sentencepiece
print("- transformers found.")
print("- sentencepiece found.")
except ImportError:
print("WARNING: 'transformers' or 'sentencepiece' not found. Llama 3 visualization will fail. Install with: pip install transformers sentencepiece")
if BLT_AVAILABLE:
print("- Bytelatent libraries found.")
# Ensure bytelatent model is present only if library is available
try:
print("Ensuring Bytelatent model 'blt-1b' weights are present...")
ensure_present(["blt-1b"])
print("Bytelatent model check complete.")
except Exception as blt_dl_err:
print(f"WARNING: Failed to ensure Bytelatent model presence: {blt_dl_err}")
else:
print("INFO: Bytelatent libraries not found, skipping related functionality.")
print(f"Attempting to use Llama 3 Tokenizer: {LLAMA3_MODEL_NAME}. Ensure you have access (e.g., via `huggingface-cli login` if needed).")
print("Launching Gradio interface...")
iface.launch()