Update example to cover some common issues

extensions_built_in/diffusion_models/example_model/example_model.py

@@ -330,6 +330,15 @@ class ExampleModel(BaseModel):
              batch max is deferred to get_noise_prediction / the pipeline,
              which keeps caches small and lets any prompts share a batch.
 
+             Each per-prompt tensor MUST be 2D ``(L, D)`` -- BaseModel infers the
+             text batch size from the list and only treats it as one-per-prompt
+             when the tensors are 2D; a 3D per-prompt tensor is misread as an
+             already-batched ``(B, L, D)`` and training fails with a latents-vs-
+             text batch-size mismatch. If your conditioning has an extra axis
+             (e.g. N stacked encoder layers -> ``(L, N, D)``), flatten it here
+             (``(L, N*D)``) and restore it (``reshape(B, Lt, N, D)``) at the
+             model call.
+
              You can store any number of keys (pooled embeds, image features,
              ...). If a key must keep its dtype when everything else is cast
              (masks, token ids), list it in ``embeds.frozen_dtype_keys``.