[refactor] eval: drop dead batch_unit and trial_forward across metrics

The Evaluator's calibrate() is gone, so batch_unit class attrs and
trial_forward() method overrides have no callers — pure cruft from the
old auto-calibration design. Mechanical removal across 15 metric files.

Internal time-dim chunking (the part of batching that actually does
work) lives on as metric-owned _chunk_size constants set in __init__,
unaffected by this change.

Five files (color/multiple_objects/object_class/spatial_relationship/
optical_flow metrics) skipped — they have dataloader's in-flight
modifications uncommitted; left for that branch's cleanup pass.

Parity verified on fs-mbz-gpu-085: results byte-identical to v3
baseline.

fastvideo/eval/metrics/common/lpips/metric.py

@@ -14,7 +14,6 @@ class LPIPSMetric(BaseMetric):
     higher_is_better = False
     needs_gpu = True
     dependencies = ["lpips"]
-    batch_unit = "frame"
 
     def __init__(self, net: str = "alex") -> None:
         super().__init__()
@@ -34,11 +33,6 @@ class LPIPSMetric(BaseMetric):
         self._model = lpips_lib.LPIPS(net=self.net).to(self.device)
         self._model.eval()
 
-    def trial_forward(self, batch_size, *, height, width, num_frames):
-        dummy = torch.randn(batch_size, 3, height, width, device=self.device) * 2 - 1
-        with torch.no_grad():
-            self._model(dummy, dummy)
-
     def compute(self, sample: dict) -> list[MetricResult]:
         if self._model is None:
             self.setup()

fastvideo/eval/metrics/common/ssim/metric.py

@@ -52,16 +52,11 @@ class SSIMMetric(BaseMetric):
     requires_reference = True
     higher_is_better = True
     needs_gpu = False
-    batch_unit = "frame"
 
     def __init__(self, window_size: int = 11) -> None:
         super().__init__()
         self.window_size = window_size
 
-    def trial_forward(self, batch_size, *, height, width, num_frames):
-        dummy = torch.randn(batch_size, 3, height, width, device=self.device)
-        _ssim_per_frame(dummy, dummy, self.window_size)
-
     def compute(self, sample: dict) -> list[MetricResult]:
         gen = sample["video"].float()       # (B, T, C, H, W)
         ref = sample["reference"].float()   # (B, T, C, H, W)

fastvideo/eval/metrics/physics_iq/metric.py

@@ -22,7 +22,6 @@ class PhysicsIQMetric(BaseMetric):
     name = "physics_iq"
     requires_reference = True
     higher_is_better = True
-    batch_unit = "video"
 
     def __init__(
         self,
@@ -48,15 +47,6 @@ class PhysicsIQMetric(BaseMetric):
         self._spatial_iou = SpatialIoUMetric(**self._prep_kwargs)
         self._weighted_spatial_iou = WeightedSpatialIoUMetric(**self._prep_kwargs)
 
-    def trial_forward(self, batch_size: int, *, height: int, width: int, num_frames: int) -> None:
-        sample = {
-            "video": torch.rand(batch_size, num_frames, 3, height, width, device=self.device),
-            "reference": torch.rand(batch_size, num_frames, 3, height, width, device=self.device),
-            "reference_take2": torch.rand(batch_size, num_frames, 3, height, width, device=self.device),
-        }
-        with torch.no_grad():
-            self.compute(sample)
-
     @staticmethod
     def _extract_payload(result: MetricResult | Mapping[str, Any]) -> Mapping[str, Any]:
         if isinstance(result, MetricResult):

fastvideo/eval/metrics/vbench/aesthetic_quality/metric.py

@@ -38,7 +38,6 @@ class AestheticQualityMetric(BaseMetric):
     requires_reference = False
     higher_is_better = True
     needs_gpu = True
-    batch_unit = "frame"
     dependencies = ["clip"]
     backbone = "clip_vit_l14"
 
@@ -80,13 +79,6 @@ class AestheticQualityMetric(BaseMetric):
         self._aesthetic_head.to(self.device)
         self._aesthetic_head.eval()
 
-    def trial_forward(self, batch_size, *, height, width, num_frames):
-        dummy = torch.randn(batch_size, 3, 224, 224, device=self.device)
-        with torch.no_grad():
-            feats = self._clip_model.encode_image(dummy).float()
-            feats = F.normalize(feats, dim=-1, p=2)
-            self._aesthetic_head(feats)
-
     @torch.no_grad()
     def compute(self, sample: dict) -> list[MetricResult]:
         video = sample["video"]  # (B, T, C, H, W)

fastvideo/eval/metrics/vbench/appearance_style/metric.py

@@ -34,7 +34,6 @@ class AppearanceStyleMetric(BaseMetric):
     requires_reference = False
     higher_is_better = True
     needs_gpu = True
-    batch_unit = "frame"
     dependencies = ["clip"]
     backbone = "clip_vit_b32"
 
@@ -60,11 +59,6 @@ class AppearanceStyleMetric(BaseMetric):
         )
         self._model.eval()
 
-    def trial_forward(self, batch_size, *, height, width, num_frames):
-        dummy = torch.randn(batch_size, 3, 224, 224, device=self.device)
-        with torch.no_grad():
-            self._model.encode_image(dummy)
-
     @torch.no_grad()
     def compute(self, sample: dict) -> list[MetricResult]:
         import clip

fastvideo/eval/metrics/vbench/background_consistency/metric.py

@@ -36,7 +36,6 @@ class BackgroundConsistencyMetric(BaseMetric):
     requires_reference = False
     higher_is_better = True
     needs_gpu = True
-    batch_unit = "frame"
     dependencies = ["clip"]
     backbone = "clip_vit_b32"
 
@@ -63,11 +62,6 @@ class BackgroundConsistencyMetric(BaseMetric):
         model.eval()
         self._model = model
 
-    def trial_forward(self, batch_size, *, height, width, num_frames):
-        dummy = torch.randn(batch_size, 3, 224, 224, device=self.device)
-        with torch.no_grad():
-            self._model.encode_image(dummy)
-
     @torch.no_grad()
     def compute(self, sample: dict) -> list[MetricResult]:
         video = sample["video"]  # (B, T, C, H, W)

fastvideo/eval/metrics/vbench/dynamic_degree/metric.py

@@ -25,7 +25,6 @@ class DynamicDegreeMetric(BaseMetric):
     higher_is_better = True
     needs_gpu = True
     dependencies = ["easydict"]
-    batch_unit = "frame_pair"
 
     def __init__(self) -> None:
         super().__init__()
@@ -59,18 +58,6 @@ class DynamicDegreeMetric(BaseMetric):
         model.eval()
         self._model = model
 
-    def trial_forward(self, batch_size, *, height, width, num_frames):
-        from vbench.third_party.RAFT.core.utils_core.utils import InputPadder
-        # RAFT's CorrBlock creates a (B*H*W, 1, H, W) correlation volume.
-        # Cap resolution to avoid integer overflow at large batch sizes.
-        h, w = min(height, 256), min(width, 256)
-        img1 = torch.randn(batch_size, 3, h, w, device=self.device)
-        img2 = torch.randn(batch_size, 3, h, w, device=self.device)
-        padder = InputPadder(img1.shape)
-        img1, img2 = padder.pad(img1, img2)
-        with torch.no_grad():
-            self._model(img1, img2, iters=20, test_mode=True)
-
     def _get_score(self, flow: torch.Tensor) -> float:
         """Top-5% mean flow magnitude (matching VBench dynamic_degree.get_score)."""
         flo = flow.permute(1, 2, 0).cpu().numpy()

fastvideo/eval/metrics/vbench/human_action/metric.py

@@ -48,7 +48,6 @@ class HumanActionMetric(BaseMetric):
     requires_reference = False
     higher_is_better = True
     needs_gpu = True
-    batch_unit = "video"
     dependencies = ["timm"]
 
     def __init__(self) -> None:
@@ -97,11 +96,6 @@ class HumanActionMetric(BaseMetric):
         self._model.to(self.device)
         self._model.eval()
 
-    def trial_forward(self, batch_size, *, height, width, num_frames):
-        dummy = torch.randn(batch_size, 3, 16, 224, 224, device=self.device)
-        with torch.no_grad():
-            self._model(dummy)
-
     @torch.no_grad()
     def compute(self, sample: dict) -> list[MetricResult]:
         video = sample["video"]  # (B, T, C, H, W) float [0, 1]

fastvideo/eval/metrics/vbench/imaging_quality/metric.py

@@ -21,7 +21,6 @@ class ImagingQualityMetric(BaseMetric):
     requires_reference = False
     higher_is_better = True
     needs_gpu = True
-    batch_unit = "frame"
     dependencies = ["pyiqa"]
 
     def __init__(self) -> None:
@@ -41,15 +40,6 @@ class ImagingQualityMetric(BaseMetric):
         self._model = pyiqa.create_metric("musiq-spaq", device=self.device)
         self._model.eval()
 
-    def trial_forward(self, batch_size, *, height, width, num_frames):
-        h, w = min(height, 512), min(width, 512)
-        if max(height, width) > 512:
-            scale = 512.0 / max(height, width)
-            h, w = int(height * scale), int(width * scale)
-        dummy = torch.randn(batch_size, 3, h, w, device=self.device)
-        with torch.no_grad():
-            self._model(dummy)
-
     @torch.no_grad()
     def compute(self, sample: dict) -> list[MetricResult]:
         video = sample["video"]  # (B, T, C, H, W)

fastvideo/eval/metrics/vbench/motion_smoothness/metric.py

@@ -25,7 +25,6 @@ class MotionSmoothnessMetric(BaseMetric):
     requires_reference = False
     higher_is_better = True
     needs_gpu = True
-    batch_unit = "frame_pair"
     dependencies = ["omegaconf"]
 
     def __init__(self) -> None:
@@ -87,14 +86,6 @@ class MotionSmoothnessMetric(BaseMetric):
         scale = 1 / np.floor(1 / np.sqrt(scale) * 16) * 16
         return float(scale)
 
-    def trial_forward(self, batch_size, *, height, width, num_frames):
-        h, w = min(height, 512), min(width, 512)
-        img0 = torch.randn(batch_size, 3, h, w, device=self.device)
-        img1 = torch.randn(batch_size, 3, h, w, device=self.device)
-        embt = self._embt.expand(batch_size, -1, -1, -1)
-        with torch.no_grad():
-            self._model(img0, img1, embt, scale_factor=1.0, eval=True)
-
     @torch.no_grad()
     def compute(self, sample: dict) -> list[MetricResult]:
         from vbench.third_party.amt.utils.utils import (

fastvideo/eval/metrics/vbench/overall_consistency/metric.py

@@ -34,7 +34,6 @@ class OverallConsistencyMetric(BaseMetric):
     requires_reference = False
     higher_is_better = True
     needs_gpu = True
-    batch_unit = "video"
     dependencies = ["timm", "einops", "clip"]
     backbone = "viclip"
 
@@ -73,11 +72,6 @@ class OverallConsistencyMetric(BaseMetric):
         self._model.to(self.device)
         self._model.eval()
 
-    def trial_forward(self, batch_size, *, height, width, num_frames):
-        dummy = torch.randn(batch_size, 8, 3, 224, 224, device=self.device)
-        with torch.no_grad():
-            self._model.encode_vision(dummy, test=True)
-
     @torch.no_grad()
     def compute(self, sample: dict) -> list[MetricResult]:
         video = sample["video"]  # (B, T, C, H, W)

fastvideo/eval/metrics/vbench/scene/metric.py

@@ -37,7 +37,6 @@ class SceneMetric(BaseMetric):
     requires_reference = False
     higher_is_better = True
     needs_gpu = True
-    batch_unit = "video"
     dependencies = ["transformers", "qwen_omni_utils"]
     backbone = "avocado"
 

fastvideo/eval/metrics/vbench/subject_consistency/metric.py

@@ -28,7 +28,6 @@ class SubjectConsistencyMetric(BaseMetric):
     requires_reference = False
     higher_is_better = True
     needs_gpu = True
-    batch_unit = "frame"
     backbone = "dino_vitb16"
 
     def __init__(self) -> None:
@@ -49,11 +48,6 @@ class SubjectConsistencyMetric(BaseMetric):
         model.eval()
         self._model = model
 
-    def trial_forward(self, batch_size, *, height, width, num_frames):
-        dummy = torch.randn(batch_size, 3, 224, 224, device=self.device)
-        with torch.no_grad():
-            self._model(dummy)
-
     @torch.no_grad()
     def compute(self, sample: dict) -> list[MetricResult]:
         video = sample["video"]  # (B, T, C, H, W)

fastvideo/eval/metrics/vbench/temporal_flickering/metric.py

@@ -21,7 +21,6 @@ class TemporalFlickeringMetric(BaseMetric):
     requires_reference = False
     higher_is_better = True
     needs_gpu = False
-    batch_unit = "video"
 
     @torch.no_grad()
     def compute(self, sample: dict) -> list[MetricResult]:

fastvideo/eval/metrics/videoscore2/metric.py

@@ -148,7 +148,6 @@ class VideoScore2Metric(BaseMetric):
     higher_is_better = True
     needs_gpu = True
     dependencies = ["transformers", "qwen_vl_utils"]
-    batch_unit = "video"
 
     def __init__(
         self,