Create loading-plot.py

91c8e18 verified 9 months ago

2.39 kB

import matplotlib.pyplot as plt
import numpy as np

# Set up the figure and axis
fig, ax = plt.subplots(figsize=(12, 7))

# Define categories and data
categories = ['v4.57.6\ndevice_map=auto\nThreadpool',
              'v4.57.6\ndevice_map=auto\nNormal',
              'v4.57.6\nTP',
              'v5\ndevice_map=auto\nAsync',
              'v5\ndevice_map=auto\nSync',
              'v5\nTP\nAsync',
              'v5\nTP\nSync']

times = [66.24, 67.29, np.nan, 20.71, 45.3, 10.1, 19.28]
colors = ['#3498db', '#2980b9', '#e74c3c', '#2ecc71', '#27ae60', '#f39c12', '#e67e22']

# Create bar positions
x_pos = np.arange(len(categories))

# Create bars
bars = ax.bar(x_pos, times, color=colors, alpha=0.8, edgecolor='black', linewidth=1.2)

# Add value labels on top of bars
for i, (bar, time) in enumerate(zip(bars, times)):
    if np.isnan(time):
        # Mark OOM with text
        ax.text(bar.get_x() + bar.get_width()/2, 5, 'OOM',
                ha='center', va='bottom', fontsize=12, fontweight='bold', color='red')
    else:
        ax.text(bar.get_x() + bar.get_width()/2, time + 1.5, f'{time}s',
                ha='center', va='bottom', fontsize=10, fontweight='bold')

# Customize the plot
ax.set_xlabel('Configuration', fontsize=12, fontweight='bold')
ax.set_ylabel('Loading Time (seconds)', fontsize=12, fontweight='bold')
ax.set_title('Model Loading Benchmark: Qwen/Qwen1.5-110B-Chat\nGPU: 1x A100 (80 GB)',
             fontsize=14, fontweight='bold', pad=20)

ax.set_xticks(x_pos)
ax.set_xticklabels(categories, fontsize=9, ha='center')
ax.set_ylim(0, max([t for t in times if not np.isnan(t)]) * 1.15)

# Add grid for better readability
ax.yaxis.grid(True, linestyle='--', alpha=0.3)
ax.set_axisbelow(True)

# Add a note about the versions
# note_text = ('Transformers v4.57.6:\n'
#              '  • Threadpool: HF_ENABLE_PARALLEL_LOADING\n'
#              '  • Normal: no threadpool\n'
#              '\n'
#              'Transformers v5:\n'
#              '  • Async: default behavior\n'
#              '  • Sync: HF_DEACTIVATE_ASYNC_LOAD')

# ax.text(1.02, 0.5, note_text, transform=ax.transAxes,
#         fontsize=9, verticalalignment='center',
#         bbox=dict(boxstyle='round', facecolor='wheat', alpha=0.3))

plt.tight_layout()
plt.savefig('loading_benchmark.png', dpi=300, bbox_inches='tight')
plt.show()

print("Plot saved as 'loading_benchmark.png'")

	import matplotlib.pyplot as plt
	import numpy as np

	# Set up the figure and axis
	fig, ax = plt.subplots(figsize=(12, 7))

	# Define categories and data
	categories = ['v4.57.6\ndevice_map=auto\nThreadpool',
	'v4.57.6\ndevice_map=auto\nNormal',
	'v4.57.6\nTP',
	'v5\ndevice_map=auto\nAsync',
	'v5\ndevice_map=auto\nSync',
	'v5\nTP\nAsync',
	'v5\nTP\nSync']

	times = [66.24, 67.29, np.nan, 20.71, 45.3, 10.1, 19.28]
	colors = ['#3498db', '#2980b9', '#e74c3c', '#2ecc71', '#27ae60', '#f39c12', '#e67e22']

	# Create bar positions
	x_pos = np.arange(len(categories))

	# Create bars
	bars = ax.bar(x_pos, times, color=colors, alpha=0.8, edgecolor='black', linewidth=1.2)

	# Add value labels on top of bars
	for i, (bar, time) in enumerate(zip(bars, times)):
	if np.isnan(time):
	# Mark OOM with text
	ax.text(bar.get_x() + bar.get_width()/2, 5, 'OOM',
	ha='center', va='bottom', fontsize=12, fontweight='bold', color='red')
	else:
	ax.text(bar.get_x() + bar.get_width()/2, time + 1.5, f'{time}s',
	ha='center', va='bottom', fontsize=10, fontweight='bold')

	# Customize the plot
	ax.set_xlabel('Configuration', fontsize=12, fontweight='bold')
	ax.set_ylabel('Loading Time (seconds)', fontsize=12, fontweight='bold')
	ax.set_title('Model Loading Benchmark: Qwen/Qwen1.5-110B-Chat\nGPU: 1x A100 (80 GB)',
	fontsize=14, fontweight='bold', pad=20)

	ax.set_xticks(x_pos)
	ax.set_xticklabels(categories, fontsize=9, ha='center')
	ax.set_ylim(0, max([t for t in times if not np.isnan(t)]) * 1.15)

	# Add grid for better readability
	ax.yaxis.grid(True, linestyle='--', alpha=0.3)
	ax.set_axisbelow(True)

	# Add a note about the versions
	# note_text = ('Transformers v4.57.6:\n'
	# ' • Threadpool: HF_ENABLE_PARALLEL_LOADING\n'
	# ' • Normal: no threadpool\n'
	# '\n'
	# 'Transformers v5:\n'
	# ' • Async: default behavior\n'
	# ' • Sync: HF_DEACTIVATE_ASYNC_LOAD')

	# ax.text(1.02, 0.5, note_text, transform=ax.transAxes,
	# fontsize=9, verticalalignment='center',
	# bbox=dict(boxstyle='round', facecolor='wheat', alpha=0.3))

	plt.tight_layout()
	plt.savefig('loading_benchmark.png', dpi=300, bbox_inches='tight')
	plt.show()

	print("Plot saved as 'loading_benchmark.png'")