diff --git a/src/solve.jl b/src/solve.jl
index fd120bde4..dcf8afbd4 100644
--- a/src/solve.jl
+++ b/src/solve.jl
@@ -805,12 +805,79 @@ function promote_f(f::F, ::Val{specialize}, u0, p, t, ::Val{true}) where {F, spe
                       zeros(uElType, n, n)
             f = @set f.jac = wrapfun_jac_iip(f.jac, (J_proto, u0, p, t))
         end
+        # Wrap the initialization data functions if present, so the
+        # NonlinearSolve dispatch also benefits from compilation caching.
+        if f.initialization_data isa SciMLBase.OverrideInitData
+            f = _wrap_initialization_data(f)
+        end
         return unwrapped_f(f, wrapfun_iip(f.f, (u0, u0, p, t)))
     else
         return f
     end
 end
 
+# Wrap functions inside OverrideInitData so their types are erased:
+# - NonlinearFunction.f in initializeprob (IIP, same pattern as jac wrapping)
+# - initializeprobmap (OOP, 1-arg)
+# - initializeprobpmap (OOP, 2-arg)
+function _wrap_initialization_data(f)
+    initdata = f.initialization_data
+    initprob = initdata.initializeprob
+    nlf = initprob.f
+
+    # Wrap NLF's f if IIP and not already wrapped
+    initprob2 = initprob
+    if SciMLBase.isinplace(nlf) && !(nlf.f isa FunctionWrappersWrappers.FunctionWrappersWrapper)
+        init_u0 = SciMLBase.state_values(initprob)
+        init_p = SciMLBase.parameter_values(initprob)
+        if init_u0 !== nothing
+            wrapped_nlf_f = wrapfun_jac_iip(nlf.f, (similar(init_u0), init_u0, init_p))
+            nlf2 = SciMLBase.NonlinearFunction{
+                SciMLBase.isinplace(nlf), SciMLBase.specialization(nlf),
+                typeof(wrapped_nlf_f),
+                typeof(nlf.mass_matrix), Nothing, Nothing,
+                Nothing, Nothing, Nothing, typeof(nlf.jac_prototype),
+                typeof(nlf.sparsity), Nothing, Nothing,
+                Nothing,
+                Any, typeof(nlf.colorvec),
+                typeof(nlf.sys), typeof(nlf.resid_prototype), Nothing,
+            }(
+                wrapped_nlf_f, nlf.mass_matrix,
+                nlf.analytic, nlf.tgrad, nlf.jac,
+                nlf.jvp, nlf.vjp, nlf.jac_prototype,
+                nlf.sparsity, nlf.Wfact,
+                nlf.Wfact_t, nlf.paramjac,
+                nlf.observed, nlf.colorvec, nlf.sys,
+                nlf.resid_prototype, nlf.initialization_data
+            )
+            initprob2 = @set initprob.f = nlf2
+        end
+    end
+
+    # Wrap initializeprobmap(nlsol) -> u0 (OOP, 1-arg)
+    wrapped_map = initdata.initializeprobmap
+    if wrapped_map !== nothing && !(wrapped_map isa FunctionWrappersWrappers.FunctionWrappersWrapper)
+        wrapped_map = FunctionWrappersWrappers.FunctionWrappersWrapper(
+            wrapped_map, (Tuple{Any},), (Any,)
+        )
+    end
+
+    # Wrap initializeprobpmap(valp, nlsol) -> p (OOP, 2-arg)
+    wrapped_pmap = initdata.initializeprobpmap
+    if wrapped_pmap !== nothing && !(wrapped_pmap isa FunctionWrappersWrappers.FunctionWrappersWrapper)
+        wrapped_pmap = FunctionWrappersWrappers.FunctionWrappersWrapper(
+            wrapped_pmap, (Tuple{Any, Any},), (Any,)
+        )
+    end
+
+    f = @set f.initialization_data = SciMLBase.OverrideInitData(
+        initprob2, initdata.update_initializeprob!,
+        wrapped_map, wrapped_pmap,
+        initdata.metadata, initdata.is_update_oop
+    )
+    return f
+end
+
 # Simple path for algorithms that do NOT use ForwardDiff internally (e.g. Tsit5, Verner).
 # Avoids calling hasdualpromote/wrapfun_iip which have extension overrides in
 # DiffEqBaseForwardDiffExt that would create invalidating method table backedges.