[bugfix]: keep _rms_norm_dispatch output in the input dtype under autocast

2026-07-07 02:08:20 -07:00
parent e2f4d1a7b5
commit 23bcf88026
1 changed files with 5 additions and 1 deletions
@@ -72,8 +72,12 @@ def _rms_norm_dispatch(
    """Use QuACK RMSNorm only for LTX-2 refine stage, else Torch RMSNorm."""
    if _is_ltx2_refine_stage():
        return _quack_rmsnorm(x, weight=weight, eps=eps)
+    # torch 2.12 added rms_norm to autocast's float32 cast policy, so under
+    # autocast(bfloat16) this norm now returns float32 where torch <= 2.11
+    # returned bfloat16. Restore the input dtype so autocast-blind consumers
+    # downstream don't receive upcast activations.
    return torch.nn.functional.rms_norm(
-        x, (x.shape[-1], ), weight=weight, eps=eps)
+        x, (x.shape[-1], ), weight=weight, eps=eps).to(x.dtype)


 class StageAwareRMSNorm(nn.RMSNorm):