fix train / val step mismatch

examples/train/configs/self_forcing_wangame.yaml

@@ -20,6 +20,7 @@ method:
   generator_update_interval: 5
   real_score_guidance_scale: 1.0
   dmd_denoising_steps: [1000, 750, 500, 250]
+  warp_denoising_step: true
 
   chunk_size: 3
   student_sample_type: sde
@@ -88,6 +89,7 @@ callbacks:
     max_grad_norm: 1.0
   validation:
     pipeline_target: fastvideo.pipelines.basic.wan.wangame_causal_dmd_pipeline.WangameCausalSdeDMDPipeline
+    scheduler_target: fastvideo.models.schedulers.scheduling_self_forcing_flow_match.SelfForcingFlowMatchScheduler
     dataset_file: examples/training/finetune/WanGame2.1_1.3b_i2v/validation_4.json
     every_steps: 5
     sampling_steps: [4]

fastvideo/train/callbacks/validation.py

@@ -145,6 +145,7 @@ class ValidationCallback(Callback):
         self._last_scalar_metrics: dict[str, float] = {}
         self._last_mf_angle_err_mean: float = float("inf")
         self._last_eval_step: int = -1
+        self._logged_sampling_step_configs: set[int] = set()
 
     # ----------------------------------------------------------
     # Callback hooks
@@ -626,6 +627,8 @@ class ValidationCallback(Callback):
                 for s in pk["dmd_denoising_steps"]
             ]
 
+        self._logged_sampling_step_configs.add(int(num_inference_steps))
+
     # ----------------------------------------------------------
     # Pipeline management
     # ----------------------------------------------------------
@@ -696,7 +699,19 @@ class ValidationCallback(Callback):
         SchedulerCls = resolve_target(
             self.scheduler_target
         )
-        return SchedulerCls(shift=float(flow_shift))
+        scheduler = SchedulerCls(shift=float(flow_shift))
+        if SchedulerCls.__name__ == "SelfForcingFlowMatchScheduler":
+            if hasattr(scheduler, "sigma_min"):
+                scheduler.sigma_min = 0.0
+            if hasattr(scheduler, "extra_one_step"):
+                scheduler.extra_one_step = True
+            if hasattr(scheduler, "set_timesteps"):
+                scheduler.set_timesteps(
+                    num_inference_steps=1000,
+                    training=True,
+                )
+
+        return scheduler
 
     # ----------------------------------------------------------
     # Batch preparation