refactor fmoe tuner profile (ROCm#1614)

* refactor fmoe profile data, and log

* Update fused_moe.py

Author: yzhou103

aiter/fused_moe.py

@@ -615,6 +615,9 @@ def MainFunc():
         )
 
     def FinalFunc():
+        logger.info(
+            f"[Hint] tuned configs are saved in {tune_file}, you can set AITER_CONFIG_FMOE to this file to use tuned configs"
+        )
         logger.info("\033[0m")
 
     # cfg = cfg_2stages.get(keys, None)

aiter/utility/base_tuner.py

@@ -58,6 +58,7 @@ def __init__(self, name, key, resultList, description=None):
         self.failed = pd.DataFrame(columns=self.columns)
 
         self.remain_untuned = pd.DataFrame(columns=self.keys)
+        self.sort_keys = key
         self.start_time = 0
 
     def get_arg_defaults(self):
@@ -305,7 +306,14 @@ def post_process(self, rets, args, topk=-1, fast_mode=False):
         """post process, post process all results to return topk results"""
         rets = list(rets)
         if args.profile_file != "":
+            if args.verbose:
+                logger.info(f"saving profile to {args.profile_file}")
             profiledf = self.result_to_df(sorted(rets, key=itemgetter(0)))
+            if os.path.exists(args.profile_file):
+                old_df = pd.read_csv(args.profile_file)
+            else:
+                old_df = pd.DataFrame(columns=self.columns)
+            profiledf = pd.concat([old_df, profiledf], ignore_index=True)
             profiledf.to_csv(args.profile_file, index=False, na_rep="Null")
 
         if fast_mode or topk == -1:
@@ -394,7 +402,7 @@ def run(self, args, fast_mode=False):
             logger.info(f"args: {args}")
         if len(self.untunedf) == 0:
             # self.update_tflops_bw(args.tune_file)
-            self.sortResults(output_file, args.sort, self.keys)
+            self.sortResults(output_file, args.sort, self.sort_keys)
             logger.info(
                 f"no shapes to be tuned, skip tuning, tuned file is {args.tune_file}"
             )
@@ -426,7 +434,7 @@ def run(self, args, fast_mode=False):
                     logger.info(
                         f"tune result is none or all shape is tuned in {args.tune_file}!"
                     )
-            self.sortResults(output_file, args.sort, self.keys)
+            self.sortResults(output_file, args.sort, self.sort_keys)
         except KeyboardInterrupt:
             tuning_status = "Interrupted"
             logger.error(
@@ -460,6 +468,14 @@ def __init__(
         description=None,
     ):
         super().__init__(name, key, resultList, description)
+        # Swap M and N positions to ensure N comes before M
+        self.sort_keys = list(key)
+        m_idx = self.sort_keys.index("M")
+        n_idx = self.sort_keys.index("N")
+        self.sort_keys[m_idx], self.sort_keys[n_idx] = (
+            self.sort_keys[n_idx],
+            self.sort_keys[m_idx],
+        )
 
     def pre_process(self, args):
         if args.all:

csrc/cktile_gemm_a8w8_bpreshuffle/gen_instances.py

@@ -240,7 +240,7 @@ def get_tune_dict(tune_dict_csv):
             N = tune_df.loc[i, "N"]
             K = tune_df.loc[i, "K"]
             kid = tune_df.loc[i, "kernelId"]
-            if kid < 0 or kid > len(kernels_list):
+            if kid < 0 or kid >= len(kernels_list):
                 print(f"[Warning]: kernelId {kid} is out of range, skip it")
                 continue
             tune_dict[(M, N, K)] = kernels_list[kid]

hsa/gfx942/fmoe_2stages/tune.py

@@ -49,7 +49,7 @@ class FmoeTuner(TunerCommon):
         "untune_file": "aiter/configs/untuned_fmoe.csv",
         "errRatio": 0.5,
         "batch": 100,
-        "profile_file": "aiter/configs/profile_fmoe.csv",  # for all results
+        "profile_file": "",  # for all results
     }
 
     def _setup_specific_arguments(self):
@@ -1958,6 +1958,13 @@ def post_process(self, results, args, topk=-1, fast_mode=False):
             profileDF.drop(["tflops1", "tflops2", "bw1", "bw2"], axis=1, inplace=True)
             profileDF["err1"] = profileDF["err1"].apply(lambda x: f"{x:.1%}")
             profileDF["err2"] = profileDF["err2"].apply(lambda x: f"{x:.1%}")
+            if args.profile_file != "":
+                if os.path.exists(args.profile_file):
+                    old_df = pd.read_csv(args.profile_file)
+                else:
+                    old_df = pd.DataFrame(columns=self.columns)
+                tmpprofileDF = pd.concat([old_df, profileDF], ignore_index=True)
+                tmpprofileDF.to_csv(args.profile_file, index=False)
             best_one = profileDF.loc[profileDF["us"].idxmin()].copy()
             print(
                 f"Tuning result for {key} is {best_one['block_m'] ,best_one['kernelName1'], best_one['kernelName2'], best_one['err1'], best_one['err2'],  best_one['run_1stage']} {best_one['us']} us, {best_one['tflops']} TFLOPS, {best_one['bw']} GB/s"
@@ -1971,12 +1978,12 @@ def post_process(self, results, args, topk=-1, fast_mode=False):
         if len(prorfiles) > 0:
             profile_result = pd.concat(prorfiles)
             profile_result["err"] = profile_result["err"].apply(lambda x: f"{x:.1%}")
+            profile_file = f"aiter/configs/profile_fmoe.csv"
             old_profile = self.get_tuned_gemm_list(
-                args.profile_file, profile_result.columns.tolist()
+                profile_file, profile_result.columns.tolist()
             )
-
             profile_result = pd.concat([old_profile, profile_result])
-            profile_result.to_csv(args.profile_file, index=False)
+            profile_result.to_csv(profile_file, index=False)
         if len(bests) > 0:
             return pd.concat(bests, axis=1).T
         else: