Optimize MONA: fp16, remove conv7x7, bottleneck 64→32

- Remove forced fp32 cast in MONA forward (runs in AMP fp16 now)
- Remove conv7x7 from MonaOp (keep 3x3 + 5x5 only)
- Reduce default bottleneck from 64 to 32
- MONA params: 3.5M (was 7.0M, -50%)
- Total trainable: 7.0M (was 10.5M)
- Peak VRAM at bs=24: 18.6 GB (was 20.3 GB before fp16)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

src/models/adapters.py

@@ -29,18 +29,17 @@ coloredlogs.install(level="INFO", logger=LOGGER, fmt="%(asctime)s %(name)s %(lev
 # ---------------------------------------------------------------------------
 
 class MonaOp(nn.Module):
-    """Multi-cognitive visual filter: parallel depthwise convs (3×3, 5×5, 7×7)."""
+    """Multi-cognitive visual filter: parallel depthwise convs (3×3, 5×5)."""
 
     def __init__(self, channels: int) -> None:
         super().__init__()
         self.conv3 = nn.Conv2d(channels, channels, kernel_size=3, padding=1, groups=channels)
         self.conv5 = nn.Conv2d(channels, channels, kernel_size=5, padding=2, groups=channels)
-        self.conv7 = nn.Conv2d(channels, channels, kernel_size=7, padding=3, groups=channels)
         self.projector = nn.Conv2d(channels, channels, kernel_size=1)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         identity = x
-        x = (self.conv3(x) + self.conv5(x) + self.conv7(x)) / 3.0 + identity
+        x = (self.conv3(x) + self.conv5(x)) / 2.0 + identity
         return x + self.projector(x)
 
 
@@ -56,7 +55,7 @@ class MonaAdapter(nn.Module):
         dropout: Dropout rate.
     """
 
-    def __init__(self, dim: int = 1024, bottleneck: int = 64, dropout: float = 0.1) -> None:
+    def __init__(self, dim: int = 1024, bottleneck: int = 32, dropout: float = 0.1) -> None:
         super().__init__()
         self.down = nn.Linear(dim, bottleneck)
         self.up = nn.Linear(bottleneck, dim)
@@ -68,7 +67,7 @@ class MonaAdapter(nn.Module):
         self.gammax = nn.Parameter(torch.ones(dim))
 
     def forward(self, x: torch.Tensor, hw: tuple[int, int]) -> torch.Tensor:
-        """Apply MONA adapter (runs in fp32 to avoid AMP overflow).
+        """Apply MONA adapter in AMP-native fp16.
 
         Args:
             x: Token features [B, N, D] where N includes CLS + register + patch tokens.
@@ -77,10 +76,6 @@ class MonaAdapter(nn.Module):
         Returns:
             Adapted features [B, N, D] (same shape, residual connection).
         """
-        with torch.amp.autocast("cuda", enabled=False):
-            return self._forward_fp32(x.float(), hw)
-
-    def _forward_fp32(self, x: torch.Tensor, hw: tuple[int, int]) -> torch.Tensor:
         identity = x
         # Scaled LayerNorm.
         x = self.norm(x) * self.gamma + x * self.gammax
@@ -151,7 +146,7 @@ class LoRALinear(nn.Module):
 
 def inject_mona_into_dinov3(
     model: nn.Module,
-    bottleneck: int = 64,
+    bottleneck: int = 32,
     dropout: float = 0.1,
 ) -> int:
     """Inject MONA adapters into a frozen DINOv3ViT model.

src/models/asymmetric_encoder.py

@@ -298,7 +298,7 @@ class AsymmetricEncoder(nn.Module):
         init_gate: float = 0.7,
         baseline_mode: bool = False,
         shared_encoder: bool = True,
-        mona_bottleneck: int = 64,
+        mona_bottleneck: int = 32,
         lora_rank: int = 4,
         device: str = "cuda",
     ) -> None: