Update app.py

app.py

@@ -427,6 +427,19 @@ def load_models() -> str | None:
         print(f"[gen] loading {[c.name for c in blocks.expected_components]} from {MODEL_REPO} ...", flush=True)
         pipe = blocks.init_pipeline(MODEL_REPO, components_manager=manager, collection="h3")
         pipe.load_components(dtype=torch.bfloat16)
+
+        # 10Eros-Max beta_5: an in-place weight replacement on the already-loaded official transformer, not a
+        # meta-device swap — the reference file is *partial* (it never carries `time_embedder`, reusing the
+        # base model's own), so a `strict=True` load would fail on those missing keys regardless. Because this
+        # never replaces the parameter objects themselves, only their data, the dtype each parameter already
+        # has — correctly set by `load_components` above, `_keep_in_fp32_modules` included — is preserved
+        # automatically. None of the meta-device pitfalls from the full-checkpoint path apply here.
+        import h3_eros
+
+        eros_status = h3_eros.apply_eros(pipe.transformer)
+        if eros_status:
+            print(f"[gen] {eros_status}", flush=True)
+
         pipe.transformer.set_attention_backend(ATTENTION)
 
         # --- Diagnostic: dump the LoRA files' key names/shapes and the transformer's own shapes to the Space