fix speedups and savings

app.py

@@ -5,7 +5,7 @@ from optimum_benchmark.task_utils import (
     infer_task_from_model_name_or_path,
 )
 from run import run_benchmark
-from configs import (
+from config_store import (
     get_training_config,
     get_inference_config,
     get_neural_compressor_config,

run.py

@@ -97,41 +97,7 @@ def run_benchmark(kwargs):
         table = pd.read_csv(f"runs/{experiment_name}/{benchmark}_results.csv", index_col=0)
         # concat tables
         table = pd.concat([baseline_table, table], axis=0)
-        table["experiment_name"] = ["baseline", experiment_name]
-        table = table.set_index("experiment_name")
-        table.reset_index(inplace=True)
-        # compute speedups
-        if "forward.latency(s)" in table.columns:
-            table["forward.latency.speedup(%)"] = (
-                table["forward.latency(s)"] / table["forward.latency(s)"].iloc[0] - 1
-            ) * 100
-            table["forward.latency.speedup(%)"] = table["forward.latency.speedup(%)"].round(2)
-        if "forward.throughput(samples/s)" in table.columns:
-            table["forward.throughput.speedup(%)"] = (
-                table["forward.throughput(samples/s)"] / table["forward.throughput(samples/s)"].iloc[0] - 1
-            ) * 100
-            table["forward.throughput.speedup(%)"] = table["forward.throughput.speedup(%)"].round(2)
-        if "forward.peak_memory(MB)" in table.columns:
-            table["forward.peak_memory.savings(%)"] = (
-                table["forward.peak_memory(MB)"] / table["forward.peak_memory(MB)"].iloc[0] - 1
-            ) * 100
-            table["forward.peak_memory.savings(%)"] = table["forward.peak_memory.savings(%)"].round(2)
-        if "generate.latency(s)" in table.columns:
-            table["generate.latency.speedup(%)"] = (
-                table["generate.latency(s)"] / table["generate.latency(s)"].iloc[0] - 1
-            ) * 100
-            table["generate.latency.speedup(%)"] = table["generate.latency.speedup(%)"].round(2)
-        if "generate.throughput(tokens/s)" in table.columns:
-            table["generate.throughput.speedup(%)"] = (
-                table["generate.throughput(tokens/s)"] / table["generate.throughput(tokens/s)"].iloc[0] - 1
-            ) * 100
-            table["generate.throughput.speedup(%)"] = table["generate.throughput.speedup(%)"].round(2)
-        if "generate.peak_memory(MB)" in table.columns:
-            table["generate.peak_memory.savings(%)"] = (
-                table["generate.peak_memory(MB)"] / table["generate.peak_memory(MB)"].iloc[0] - 1
-            ) * 100
-            table["generate.peak_memory.savings(%)"] = table["generate.peak_memory.savings(%)"].round(2)
-
+        table = postprocess_table(table, experiment_name)
     else:
         table = pd.read_csv(f"runs/{experiment_name}/{benchmark}_results.csv", index_col=0)
 
@@ -176,3 +142,46 @@ def run_experiment(args, html_text=""):
         yield gr.update(value=cumul_html_text), gr.update(interactive=False), gr.update(visible=False)
 
     return process.returncode, cumul_html_text
+
+
+def postprocess_table(table, experiment_name):
+    table["experiment_name"] = ["baseline", experiment_name]
+    table = table.set_index("experiment_name")
+    table.reset_index(inplace=True)
+    if "forward.latency(s)" in table.columns:
+        table["forward.latency.speedup(%)"] = (
+            1 - table["forward.latency(s)"] / table["forward.latency(s)"].iloc[0]
+        ) * 100
+        table["forward.latency.speedup(%)"] = table["forward.latency.speedup(%)"].round(2)
+
+    if "forward.throughput(samples/s)" in table.columns:
+        table["forward.throughput.speedup(%)"] = (
+            table["forward.throughput(samples/s)"] / table["forward.throughput(samples/s)"].iloc[0] - 1
+        ) * 100
+        table["forward.throughput.speedup(%)"] = table["forward.throughput.speedup(%)"].round(2)
+
+    if "forward.peak_memory(MB)" in table.columns:
+        table["forward.peak_memory.savings(%)"] = (
+            1 - table["forward.peak_memory(MB)"] / table["forward.peak_memory(MB)"].iloc[0]
+        ) * 100
+        table["forward.peak_memory.savings(%)"] = table["forward.peak_memory.savings(%)"].round(2)
+
+    if "generate.latency(s)" in table.columns:
+        table["generate.latency.speedup(%)"] = (
+            1 - table["generate.latency(s)"] / table["generate.latency(s)"].iloc[0]
+        ) * 100
+        table["generate.latency.speedup(%)"] = table["generate.latency.speedup(%)"].round(2)
+
+    if "generate.throughput(tokens/s)" in table.columns:
+        table["generate.throughput.speedup(%)"] = (
+            table["generate.throughput(tokens/s)"] / table["generate.throughput(tokens/s)"].iloc[0] - 1
+        ) * 100
+        table["generate.throughput.speedup(%)"] = table["generate.throughput.speedup(%)"].round(2)
+
+    if "generate.peak_memory(MB)" in table.columns:
+        table["generate.peak_memory.savings(%)"] = (
+            1 - table["generate.peak_memory(MB)"] / table["generate.peak_memory(MB)"].iloc[0]
+        ) * 100
+        table["generate.peak_memory.savings(%)"] = table["generate.peak_memory.savings(%)"].round(2)
+
+    return table