error bars

app.py

@@ -1,3 +1,4 @@
+from math import e
 import pandas as pd
 import gradio as gr
 import matplotlib.pyplot as plt
@@ -43,6 +44,26 @@ def create_matplotlib_bar_charts():
     ttft_values = data['ttft']
     tpot_values = data['tpot']
 
+    # Calculate error bars (standard deviation) for each configuration
+    import numpy as np
+    ttft_errors = []
+    tpot_errors = []
+
+    # Get raw data to calculate standard deviations
+    raw_data = DATA.data
+    for label in labels:
+        if label in raw_data:
+            # Calculate std dev for TTFT
+            ttft_raw = [d['wall_time'] for d in raw_data[label]['ttft']]
+            ttft_errors.append(float(np.std(ttft_raw)))
+
+            # Calculate std dev for TPOT
+            tpot_raw = [d['wall_time'] for d in raw_data[label]['tpot']]
+            tpot_errors.append(float(np.std(tpot_raw)))
+        else:
+            ttft_errors.append(0)
+            tpot_errors.append(0)
+
     # Define color mapping based on configuration keywords
     def get_color_for_config(label):
         is_eager = 'eager' in label.lower()
@@ -66,10 +87,17 @@ def create_matplotlib_bar_charts():
 
     # TTFT Plot (left)
     ax1.set_facecolor('#000000')
-    bars1 = ax1.bar(range(len(labels)), ttft_values,
+    _ = ax1.bar(range(len(labels)), ttft_values,
                     color=colors, width=1.0, edgecolor='white', linewidth=1,
                     label=[label[:15] + '...' if len(label) > 15 else label for label in labels])
 
+    # Add error bars for TTFT
+    ax1.errorbar(
+        range(len(labels)), ttft_values, yerr=ttft_errors,
+        fmt='none', ecolor='white', alpha=0.8,
+        elinewidth=1.5, capthick=1.5, capsize=4,
+    )
+
     ax1.set_ylabel('TTFT (seconds)', color='white', fontsize=14)
     ax1.set_title('Time to first token (lower is better)', color='white', fontsize=16, pad=20)
 
@@ -79,9 +107,16 @@ def create_matplotlib_bar_charts():
 
     # TPOT Plot (right)
     ax2.set_facecolor('#000000')
-    bars2 = ax2.bar(range(len(labels)), tpot_values,
+    _ = ax2.bar(range(len(labels)), tpot_values,
                     color=colors, width=1.0, edgecolor='white', linewidth=1)
 
+    # Add error bars for TPOT
+    ax2.errorbar(
+        range(len(labels)), tpot_values, yerr=tpot_errors,
+        fmt='none', ecolor='white', alpha=0.8,
+        elinewidth=1.5, capthick=1.5, capsize=4,
+    )
+
     ax2.set_ylabel('TPOT (seconds)', color='white', fontsize=14)
     ax2.set_title('Time per output token (lower is better)', color='white', fontsize=16, pad=20)