Remove per-call torch.compile JIT from gated_deltanet helpers (ZeroGPU warmup)

kairos/third_party/fla/layers/gated_deltanet.py

@@ -21,12 +21,16 @@ if TYPE_CHECKING:
     from fla.models.utils import Cache
 
 
-@torch.compile
+# NOTE: These are intentionally left as eager functions (no @torch.compile).
+# On ZeroGPU each @spaces.GPU call runs in a fresh forked worker where
+# torch.compile (JIT) is not supported / would re-compile on every call,
+# re-triggering the per-call warmup tax. They are tiny elementwise ops, so
+# eager execution costs nothing at steady state while avoiding the first-use
+# compilation on every fresh worker process.
 def elu_p1(x):
     return (F.elu(x, 1., False) + 1.).to(x)
 
 
-@torch.compile
 def sum_norm(x):
     return (x / x.sum(-1, keepdim=True)).to(x)