Update app.py

app.py

@@ -512,6 +512,21 @@ def load_models() -> str | None:
                 raw = {k: handle.get_tensor(k) for k in handle.keys()}
             converted = _convert_full_checkpoint(raw, base_shapes)
             custom_transformer.load_state_dict(converted, strict=True, assign=True)
+
+            # `rope.inv_freq` is a *non-persistent* buffer (`persistent=False` in `MiniMaxH3RotaryPosEmbed`) —
+            # excluded from `state_dict()` entirely, which is exactly why `strict=True` above never complained
+            # about its absence from the checkpoint (it was correctly dropped by `_convert_full_checkpoint` too).
+            # But that also means `load_state_dict(assign=True)` never touches it: it's still sitting on
+            # `torch.device("meta")` from construction, with nothing to move — which is what the later
+            # `pipe.transformer.to("cuda")` call was hitting ("Cannot copy out of meta tensor; no data!").
+            # Recomputed here from its own documented formula rather than moved, since a meta buffer has no data
+            # to move in the first place.
+            rope_theta = float(config["rope_theta"])
+            rope_freq_dim = int(config["rope_freq_dim"])
+            custom_transformer.rope.inv_freq = 1.0 / (
+                rope_theta ** (torch.arange(0, 2 * rope_freq_dim, 2, dtype=torch.float32) / (2 * rope_freq_dim))
+            )
+
             pipe.update_components(transformer=custom_transformer)
             print(f"[gen] transformer replaced with {CUSTOM_TRANSFORMER_REPO}/{CUSTOM_TRANSFORMER_FILE}", flush=True)