Added throughput

bar_plot.py

@@ -90,11 +90,11 @@ def create_matplotlib_bar_plot() -> None:
             )
             return None
 
-    # TTFT Plot (left)
+    # TTFT Plot (top)
     ttft_bars, ttft_errors, x_ticks = make_bar_kwargs(per_device_data, "ttft")
     draw_bar_plot(axs[0], ttft_bars, ttft_errors, "TTFT (seconds)", x_ticks)
 
-    # # ITL Plot (right)
+    # # ITL Plot (bottom)
     itl_bars, itl_errors, x_ticks = make_bar_kwargs(per_device_data, "itl")
     draw_bar_plot(axs[1], itl_bars, itl_errors, "ITL (seconds)", x_ticks)
 

data.py

@@ -21,6 +21,12 @@ class ModelBenchmarkData:
             for dts in measures["dt_tokens"]
         ]
 
+    def compute_throughput(self, measures: dict, batch_size: int) -> list[float]:
+        return [
+            (batch_size * len(dts) / e2e) if e2e > 0 else 0
+            for e2e, dts in zip(measures["e2e_latency"], measures["dt_tokens"])
+        ]
+
     def compute_e2e_latency(self, measures: dict) -> list[float]:
         return measures["e2e_latency"][:]