diff --git a/ais_bench/benchmark/calculators/default_perf_metric_calculator.py b/ais_bench/benchmark/calculators/default_perf_metric_calculator.py
index a9a78642dfb52aefd0acc49356cda165069ac8af..5a9288228b27f65ed49e3c2cb500b2ad2eca12d7 100644
--- a/ais_bench/benchmark/calculators/default_perf_metric_calculator.py
+++ b/ais_bench/benchmark/calculators/default_perf_metric_calculator.py
@@ -63,11 +63,14 @@ class DefaultPerfMetricCalculator(BasePerfMetricCalculator):
         self.infer_time[stage_name] = max(result["end_time"]) - min(result["start_time"])
         per_request_avg_decode_time = []
         # Compute the average decode latency per request
-        for i, value in enumerate(result["seq_latency"]):
-            if value:  # Skip empty lists
-                tpot = (value - result["prefill_latency"][i]) / result["generate_tokens_len"][i]
-                per_request_avg_decode_time.append(tpot)
-        result["average_decode_latencies"] = per_request_avg_decode_time[:]
+        if not math.isclose(sum(result["prefill_latency"]), 0):
+            for i, value in enumerate(result["seq_latency"]):
+                if value:  # Skip empty lists
+                    tpot = (value - result["prefill_latency"][i]) / result["generate_tokens_len"][i]
+                    per_request_avg_decode_time.append(tpot)
+            result["average_decode_latencies"] = per_request_avg_decode_time[:]
+        else:
+            result["average_decode_latencies"] = result["prefill_latency"]
         self.logger.info("Converting perf results of stage ...")
         self.result[stage_name] = self.convert_result(copy.deepcopy(result))
 
diff --git a/ais_bench/benchmark/calculators/stable_perf_metric_calculator.py b/ais_bench/benchmark/calculators/stable_perf_metric_calculator.py
index 50ab6089ac0009b5fac4af5a1be5aac05e7d78df..ff91bb56a54a6596ccc21a8b6ae4c7cb13b5d196 100644
--- a/ais_bench/benchmark/calculators/stable_perf_metric_calculator.py
+++ b/ais_bench/benchmark/calculators/stable_perf_metric_calculator.py
@@ -107,10 +107,14 @@ class StablePerfMetricCalculator(BasePerfMetricCalculator):
         self.infer_time[stage_name] = self.stage_section[1] - self.stage_section[0]
         per_request_avg_decode_time = []
         # Compute the average decode latency per request
-        for i, value in enumerate(result["seq_latency"]):
-            if value:  # Skip empty lists
-                tpot = (value - result["prefill_latency"][i]) / result["generate_tokens_len"][i]
-                per_request_avg_decode_time.append(tpot)
+        if not math.isclose(sum(result["prefill_latency"]), 0):
+            for i, value in enumerate(result["seq_latency"]):
+                if value:  # Skip empty lists
+                    tpot = (value - result["prefill_latency"][i]) / result["generate_tokens_len"][i]
+                    per_request_avg_decode_time.append(tpot)
+            result["average_decode_latencies"] = per_request_avg_decode_time[:]
+        else:
+            result["average_decode_latencies"] = result["prefill_latency"]
         result["average_decode_latencies"] = per_request_avg_decode_time[:]
         self.logger.info("Converting perf results of stage ...")
         self.result[stage_name] = self.convert_result(copy.deepcopy(result))
diff --git a/ais_bench/benchmark/clients/openai_chat_text_client.py b/ais_bench/benchmark/clients/openai_chat_text_client.py
index ccd1e17d32e0791992c9de8260251c28b2593c62..08fb01d6a4b45ca22923caff961e76d493a52824 100644
--- a/ais_bench/benchmark/clients/openai_chat_text_client.py
+++ b/ais_bench/benchmark/clients/openai_chat_text_client.py
@@ -25,4 +25,9 @@ class OpenAIChatTextClient(BaseClient, ABC):
         if generated_text:
             inputs.output = generated_text
             inputs.num_generated_chars = len(generated_text)
+            inputs.prefill_latency = res.get("prefill_time", 0)
+            inputs.chunk_time_point_list.append(inputs.chunk_time_point_list[0] + inputs.prefill_latency)
+            inputs.decode_cost = res.get("decode_time_arr", [])
+            inputs.chunk_time_point_list.append(inputs.chunk_time_point_list[1] + sum(inputs.decode_cost))
+
         return generated_text