diff --git a/deepmd/pt/model/model/sezm_model.py b/deepmd/pt/model/model/sezm_model.py
index 91ff826a3e..7b2595a663 100644
--- a/deepmd/pt/model/model/sezm_model.py
+++ b/deepmd/pt/model/model/sezm_model.py
@@ -2370,11 +2370,13 @@ def compute_fn(  # type: ignore[misc]
             def _inductor_inference_compiler(
                 fx_gm: torch.fx.GraphModule, fx_inputs: list[Any]
             ) -> Any:
+                # This callback lowers the forward-only AOTAutograd inference
+                # graph (the eval branch below owns it; training compiles
+                # through the separate torch.compile branch above).
                 # ``compile_options`` already includes ``triton.max_tiles=1``
-                # from ``build_inductor_compile_options``, so the 1D launch-grid
-                # constraint applies to both the training and evaluation graphs.
+                # from ``build_inductor_compile_options``.
                 with _ind_cfg.patch(compile_options):
-                    return compile_fx_inner(fx_gm, fx_inputs)
+                    return compile_fx_inner(fx_gm, fx_inputs, is_inference=True)
 
             # select_decomp_table keeps the decomposition set aligned with
             # Inductor's fallback set (a mismatch raises an aten._to_copy