Wan LoRA trainer: fix fp32/bf16 dtype mismatch in the train step

torch.rand defaults to fp32, so the rectified-flow interpolation promoted x_t to
fp32 while the patch-embedding Conv3d stays bf16 (bitsandbytes 4-bit quantizes
only Linear layers), raising "Input type (float) and bias type (BFloat16) should
be the same". Compute the interpolation in fp32 then cast x_t/target back to the
model compute dtype, and pass timestep as fp32 (Wan casts it internally).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

codai/api/loras.py

@@ -952,12 +952,16 @@ def _train_wan(req, base_path, images, instance_prompt,
         x0 = latents_list[step % n].to(device, dtype=compute_dtype)
         noise = torch.randn_like(x0)
         # Rectified-flow timestep with Wan resolution shift applied to sigma.
-        u = torch.rand(1, device=device)
+        # Compute the interpolation in fp32 for stability, then cast x_t back to
+        # the model's compute dtype — the patch-embedding Conv3d is NOT quantized
+        # (bnb only touches Linear), so it needs inputs in compute_dtype.
+        u = torch.rand(1, device=device, dtype=torch.float32)
         sigma = (shift * u) / (1.0 + (shift - 1.0) * u)
         s = sigma.view(-1, 1, 1, 1, 1)
-        x_t = (1.0 - s) * x0 + s * noise
-        target = noise - x0                              # flow-matching velocity
-        timestep = (sigma * num_train_t).to(compute_dtype)
+        x0f = x0.float()
+        x_t = ((1.0 - s) * x0f + s * noise.float()).to(compute_dtype)
+        target = (noise.float() - x0f).to(compute_dtype)  # flow-matching velocity
+        timestep = (sigma * num_train_t).to(torch.float32)  # cast internally by Wan
         tr = _pick_expert(float(sigma.item()))
         pred = tr(hidden_states=x_t, timestep=timestep,
                   encoder_hidden_states=encoder_hidden_states.to(device),