[Metrics] move prompt_tokens_total to main process metrics (#5118)

* [update] move prompt_tokens_total to main process metrics

* [update] move again from zmq thread to scheduler

fastdeploy/engine/common_engine.py

@@ -562,6 +562,8 @@ class EngineService:
                 else:
                     continue
 
+                for task in tasks:
+                    main_process_metrics.inc_value("prompt_tokens_total", task.prompt_token_ids_len)
                 main_process_metrics.dec_value("num_requests_waiting", len(tasks))
                 main_process_metrics.inc_value("num_requests_running", len(tasks))
             except Exception as e:
@@ -597,6 +599,7 @@ class EngineService:
                 )
 
                 for task in tasks:
+                    main_process_metrics.inc_value("prompt_tokens_total", task.prompt_token_ids_len)
                     task.schedule_start_time = time.time()
 
                 self.llm_logger.debug(f"get tasks from scheduler: {tasks}")

fastdeploy/entrypoints/engine_client.py

@@ -159,7 +159,7 @@ class EngineClient:
                 del task["messages"]
             api_server_logger.info(f"task['max_tokens']:{task['max_tokens']}")
             work_process_metrics.obs_value("request_params_max_tokens", task["max_tokens"])
-            work_process_metrics.inc_value("prompt_tokens_total", input_ids_len)
+            # work_process_metrics.inc_value("prompt_tokens_total", input_ids_len)
             work_process_metrics.obs_value("request_prompt_tokens", input_ids_len)
         except Exception as e:
             api_server_logger.error(f"add_requests error: {e}, {str(traceback.format_exc())}")

fastdeploy/metrics/metrics.py

@@ -148,6 +148,7 @@ class MetricsManager(MetricsManagerInterface):
     request_queue_time: "Histogram"
     gpu_cache_usage_perc: "Gauge"
     generation_tokens_total: "Counter"
+    prompt_tokens_total: "Counter"
     request_prefill_time: "Histogram"
     request_decode_time: "Histogram"
     request_generation_tokens: "Histogram"
@@ -252,6 +253,12 @@ class MetricsManager(MetricsManagerInterface):
             "description": "GPU KV-cache usage. 1 means 100 percent usage",
             "kwargs": {},
         },
+        "prompt_tokens_total": {
+            "type": Counter,
+            "name": "fastdeploy:prompt_tokens_total",
+            "description": "Total number of prompt tokens processed",
+            "kwargs": {},
+        },
         "generation_tokens_total": {
             "type": Counter,
             "name": "fastdeploy:generation_tokens_total",
@@ -504,7 +511,7 @@ class MetricsManager(MetricsManagerInterface):
         if workers == 1:
             registry.register(work_process_metrics.e2e_request_latency)
             registry.register(work_process_metrics.request_params_max_tokens)
-            registry.register(work_process_metrics.prompt_tokens_total)
+            # registry.register(work_process_metrics.prompt_tokens_total)
             registry.register(work_process_metrics.request_prompt_tokens)
         if hasattr(main_process_metrics, "spec_decode_draft_acceptance_rate"):
             self.register_speculative_metrics(registry)

fastdeploy/metrics/work_metrics.py

@@ -72,11 +72,11 @@ class WorkMetricsManager(MetricsManagerInterface):
             buckets=build_1_2_5_buckets(33792),
             labelnames=LABEL_NAMES,
         )
-        self.prompt_tokens_total = Counter(
-            name="fastdeploy:prompt_tokens_total",
-            documentation="Total number of prompt tokens processed",
-            labelnames=LABEL_NAMES,
-        )
+        # self.prompt_tokens_total = Counter(
+        #     name="fastdeploy:prompt_tokens_total",
+        #     documentation="Total number of prompt tokens processed",
+        #     labelnames=LABEL_NAMES,
+        # )
         self.request_prompt_tokens = Histogram(
             name="fastdeploy:request_prompt_tokens",
             documentation="Number of prefill tokens processed.",