Fix conv2d bf16 crash on T4: iris/blocks.py

iris/blocks.py

@@ -134,7 +134,10 @@ class UIBFFN(nn.Module):
         h = self.pw_up(x)
         g = F.silu(self.gate(x))
         h_2d = h.view(B, H, W, -1).permute(0, 3, 1, 2)
-        h = self.dw_conv(h_2d).permute(0, 2, 3, 1).reshape(B, N, -1)
+        # Run depthwise conv in float32 — grouped convs lack bf16 cuDNN kernels on T4
+        with torch.amp.autocast(device_type='cuda', enabled=False):
+            h = self.dw_conv(h_2d.float()).permute(0, 2, 3, 1).reshape(B, N, -1)
+        h = h.to(g.dtype)
         return residual + self.pw_down(h * g)