gelu_approx(gate) * up, consuming both inputs. compiled is the owner's
construction-time choice: the compiled closure (the oracle's @mx.compile'd
geglu, one kernel set; the default the Gemma graphs compose), or the
spelled-out operation (same L1
parity, slower). Inside a compiled-decode trace the closure records the
spelled-out ops.
gelu_approx(gate) * up, consuming both inputs.
compiledis the owner's construction-time choice: the compiled closure (the oracle's @mx.compile'd geglu, one kernel set; the default the Gemma graphs compose), or the spelled-out operation (same L1 parity, slower). Inside a compiled-decode trace the closure records the spelled-out ops.