sync from github

src/leaderboard/read_evals.py

@@ -116,9 +116,9 @@ class EvalResult:
                         multiplier = 1.0
                     if "time" in metric:
                         multiplier = 1.0
-                    if "throughput" in metric:
+                    if "throughput" in metric or "mfu" in metric or "mbu" in metric:
                         multiplier = 1.0
-                    if "batch_" in metric or "Mem" in metric or "Util" in metric or "mfu" in metric or "mbu" in metric:
+                    if "batch_" in metric or "Mem" in metric or "Util" in metric:
                         multiplier = 1
 
 

src/populate.py

@@ -75,7 +75,7 @@ def get_leaderboard_df(
             df[col] = np.nan
 
     if not df.empty:
-        df = df.round(decimals=2)
+        df = df.round(decimals=4)
 
         # filter out if any of the benchmarks have not been produced
         # df = df[has_no_nan_values(df, benchmark_cols)]