add vae_encoder_pre_latent duration metrics (#381)

Co-authored-by: yihuiwen <yihuiwen@sensetime.com>

lightx2v/models/runners/qwen_image/qwen_image_runner.py

@@ -125,7 +125,7 @@ class QwenImageRunner(DefaultRunner):
             text_encoder_output["image_info"] = image_info
         return text_encoder_output
 
-    @ProfilingContext4DebugL1("Run VAE Encoder", recorder_mode=GET_RECORDER_MODE(), metrics_func=monitor_cli.lightx2v_run_vae_encode_duration, metrics_labels=["QwenImageRunner"])
+    @ProfilingContext4DebugL1("Run VAE Encoder", recorder_mode=GET_RECORDER_MODE(), metrics_func=monitor_cli.lightx2v_run_vae_encoder_image_duration, metrics_labels=["QwenImageRunner"])
     def run_vae_encoder(self, image):
         image_latents = self.vae.encode_vae_image(image)
         return {"image_latents": image_latents}

lightx2v/models/runners/wan/wan_animate_runner.py

@@ -154,7 +154,7 @@ class WanAnimateRunner(WanRunner):
     @ProfilingContext4DebugL1(
         "Run VAE Encoder",
         recorder_mode=GET_RECORDER_MODE(),
-        metrics_func=monitor_cli.lightx2v_run_vae_encode_duration,
+        metrics_func=monitor_cli.lightx2v_run_vae_encoder_image_duration,
         metrics_labels=["WanAnimateRunner"],
     )
     def run_vae_encoder(

lightx2v/models/runners/wan/wan_audio_runner.py

@@ -470,7 +470,7 @@ class WanAudioRunner(WanRunner):  # type:ignore
     @ProfilingContext4DebugL1(
         "Run VAE Encoder",
         recorder_mode=GET_RECORDER_MODE(),
-        metrics_func=monitor_cli.lightx2v_run_vae_encode_duration,
+        metrics_func=monitor_cli.lightx2v_run_vae_encoder_image_duration,
         metrics_labels=["WanAudioRunner"],
     )
     def run_vae_encoder(self, img):
@@ -533,7 +533,12 @@ class WanAudioRunner(WanRunner):  # type:ignore
             self.vae_encoder = self.load_vae_encoder()
 
         _, nframe, height, width = self.model.scheduler.latents.shape
-        with ProfilingContext4DebugL1("vae_encoder in init run segment"):
+        with ProfilingContext4DebugL1(
+            "vae_encoder in init run segment",
+            recorder_mode=GET_RECORDER_MODE(),
+            metrics_func=monitor_cli.lightx2v_run_vae_encoder_pre_latent_duration,
+            metrics_labels=["WanAudioRunner"],
+        ):
             if self.config["model_cls"] == "wan2.2_audio":
                 if prev_video is not None:
                     prev_latents = self.vae_encoder.encode(prev_frames.to(dtype))

lightx2v/models/runners/wan/wan_runner.py

@@ -273,7 +273,7 @@ class WanRunner(DefaultRunner):
     @ProfilingContext4DebugL1(
         "Run VAE Encoder",
         recorder_mode=GET_RECORDER_MODE(),
-        metrics_func=monitor_cli.lightx2v_run_vae_encode_duration,
+        metrics_func=monitor_cli.lightx2v_run_vae_encoder_image_duration,
         metrics_labels=["WanRunner"],
     )
     def run_vae_encoder(self, first_frame, last_frame=None):
@@ -509,7 +509,7 @@ class Wan22DenseRunner(WanRunner):
     @ProfilingContext4DebugL1(
         "Run VAE Encoder",
         recorder_mode=GET_RECORDER_MODE(),
-        metrics_func=monitor_cli.lightx2v_run_vae_encode_duration,
+        metrics_func=monitor_cli.lightx2v_run_vae_encoder_image_duration,
         metrics_labels=["Wan22DenseRunner"],
     )
     def run_vae_encoder(self, img):

lightx2v/models/runners/wan/wan_vace_runner.py

@@ -92,7 +92,7 @@ class WanVaceRunner(WanRunner):
     @ProfilingContext4DebugL1(
         "Run VAE Encoder",
         recorder_mode=GET_RECORDER_MODE(),
-        metrics_func=monitor_cli.lightx2v_run_vae_encode_duration,
+        metrics_func=monitor_cli.lightx2v_run_vae_encoder_image_duration,
         metrics_labels=["WanVaceRunner"],
     )
     def run_vae_encoder(self, frames, ref_images, masks):

lightx2v/server/metrics/metrics.py

@@ -254,13 +254,20 @@ METRICS_INFO = {
         labels=["model_cls"],
         buckets=HYBRID_10_50MS_BUCKETS,
     ),
-    "lightx2v_run_vae_encode_duration": MetricsConfig(
-        name="lightx2v_run_vae_encode_duration",
-        desc="Duration of run vae encode (s)",
+    "lightx2v_run_vae_encoder_image_duration": MetricsConfig(
+        name="lightx2v_run_vae_encoder_image_duration",
+        desc="Duration of run vae encode for image (s)",
         type_="histogram",
         labels=["model_cls"],
         buckets=HYBRID_60_120MS_BUCKETS,
     ),
+    "lightx2v_run_vae_encoder_pre_latent_duration": MetricsConfig(
+        name="lightx2v_run_vae_encoder_pre_latent_duration",
+        desc="Duration of run vae encode for pre latents (s)",
+        type_="histogram",
+        labels=["model_cls"],
+        buckets=HYBRID_1_30S_BUCKETS,
+    ),
     "lightx2v_run_vae_decode_duration": MetricsConfig(
         name="lightx2v_run_vae_decode_duration",
         desc="Duration of run vae decode (s)",