diff --git a/backends/cuda/runtime/cuda_backend.cpp b/backends/cuda/runtime/cuda_backend.cpp
index 1f989ab1b5b..921c15ae7af 100644
--- a/backends/cuda/runtime/cuda_backend.cpp
+++ b/backends/cuda/runtime/cuda_backend.cpp
@@ -288,6 +288,11 @@ class ET_EXPERIMENTAL CudaBackend final
         get_constant_dtype, AOTInductorModelContainerGetConstantDtype);
     LOAD_OPTIONAL_SYMBOL(
         get_constant_data_size, AOTInductorModelContainerGetConstantDataSize);
+    LOAD_OPTIONAL_SYMBOL(
+        get_constant_type, AOTInductorModelContainerGetConstantType);
+    LOAD_OPTIONAL_SYMBOL(
+        get_constant_from_folded,
+        AOTInductorModelContainerGetConstantFromFolded);
     LOAD_OPTIONAL_SYMBOL(
         extract_constants_map, AOTInductorModelContainerExtractConstantsMap);
     LOAD_OPTIONAL_SYMBOL(
diff --git a/backends/cuda/runtime/cuda_delegate_handle.h b/backends/cuda/runtime/cuda_delegate_handle.h
index 93c11eaeb52..f42f0f2382f 100644
--- a/backends/cuda/runtime/cuda_delegate_handle.h
+++ b/backends/cuda/runtime/cuda_delegate_handle.h
@@ -56,6 +56,15 @@ using AOTInductorModelContainerGetConstantDataSizeFunc =
         aoti::AOTInductorModelContainerHandle container_handle,
         size_t idx,
         size_t* data_size);
+using AOTInductorModelContainerGetConstantTypeFunc = aoti::AOTIRuntimeError (*)(
+    aoti::AOTInductorModelContainerHandle container_handle,
+    size_t idx,
+    int32_t* type);
+using AOTInductorModelContainerGetConstantFromFoldedFunc =
+    aoti::AOTIRuntimeError (*)(
+        aoti::AOTInductorModelContainerHandle container_handle,
+        size_t idx,
+        bool* from_folded);
 struct CudaWeightStorage {
   void* data{nullptr};
   size_t nbytes{0};
@@ -286,6 +295,11 @@ struct CudaDelegateHandle : public aoti::AOTIDelegateHandle {
   // storage against it before binding.
   AOTInductorModelContainerGetConstantDataSizeFunc get_constant_data_size{
       nullptr};
+  // With these, FQN weight binding checks completeness itself instead of
+  // AOTI's full update, which warns once per off-graph KV buffer on every load.
+  AOTInductorModelContainerGetConstantTypeFunc get_constant_type{nullptr};
+  AOTInductorModelContainerGetConstantFromFoldedFunc get_constant_from_folded{
+      nullptr};
 
   // The per-thread stream. Nothing owns it: the value is a fixed sentinel the
   // driver resolves to a different stream on each host thread, so releasing the
diff --git a/backends/cuda/runtime/cuda_weight_cache.cpp b/backends/cuda/runtime/cuda_weight_cache.cpp
index 017629f4760..104b4461318 100644
--- a/backends/cuda/runtime/cuda_weight_cache.cpp
+++ b/backends/cuda/runtime/cuda_weight_cache.cpp
@@ -140,6 +140,59 @@ bool is_offgraph_kv_fqn(const std::string& fqn) {
   return fqn.rfind(kOffGraphKVFqnPrefix, 0) == 0;
 }
 
+// torch::aot_inductor::ConstantType.
+constexpr int32_t kAotiParameter = 1;
+constexpr int32_t kAotiBuffer = 2;
+constexpr int32_t kAotiTensorConstant = 3;
+
+bool can_check_constants_bound(const CudaDelegateHandle& handle) {
+  return handle.get_constant_type && handle.get_constant_from_folded &&
+      handle.get_constant_data_size;
+}
+
+// The same rule as AOTI's validate_full_update, without its warning for each
+// absent buffer: folded constants, buffers, tensor constants and empty
+// parameters may be absent, every other constant must be bound.
+Error check_constants_bound(
+    const CudaDelegateHandle& handle,
+    size_t num_constants,
+    const std::unordered_set<std::string>& bound) {
+  for (size_t index = 0; index < num_constants; ++index) {
+    bool from_folded = false;
+    int32_t type = 0;
+    size_t data_size = 0;
+    ET_CHECK_OK_OR_RETURN_ERROR(
+        handle.get_constant_from_folded(
+            handle.container_handle, index, &from_folded),
+        "Failed to read CUDA AOTI constant folding at index %zu",
+        index);
+    ET_CHECK_OK_OR_RETURN_ERROR(
+        handle.get_constant_type(handle.container_handle, index, &type),
+        "Failed to read CUDA AOTI constant type at index %zu",
+        index);
+    ET_CHECK_OK_OR_RETURN_ERROR(
+        handle.get_constant_data_size(
+            handle.container_handle, index, &data_size),
+        "Failed to read CUDA AOTI constant size at index %zu",
+        index);
+    if (from_folded || type == kAotiBuffer || type == kAotiTensorConstant ||
+        (type == kAotiParameter && data_size == 0)) {
+      continue;
+    }
+    const char* name = nullptr;
+    ET_CHECK_OK_OR_RETURN_ERROR(
+        handle.get_constant_name(handle.container_handle, index, &name),
+        "Failed to read CUDA AOTI constant name at index %zu",
+        index);
+    ET_CHECK_OR_RETURN_ERROR(
+        name != nullptr && bound.count(name) != 0,
+        InvalidProgram,
+        "CUDA AOTI constant '%s' has no bound FQN weight",
+        name == nullptr ? "<unnamed>" : name);
+  }
+  return Error::Ok;
+}
+
 } // namespace
 
 bool CudaWeightCache::is_serialized(const void* data, size_t size) {
@@ -650,16 +703,25 @@ Error CudaWeightCache::load(
   }
 
   if (!pairs.empty()) {
-    // Full validation stays on with off-graph KV present: those constants are
-    // AOTI buffers, which the check permits to be absent, while every weight
-    // must still be supplied.
+    // Off-graph KV constants are AOTI buffers bound later by the KV pool, so
+    // they are allowed to be absent here, while every weight must be supplied.
+    const bool check_here = can_check_constants_bound(*handle);
+    if (check_here) {
+      std::unordered_set<std::string> bound_names;
+      for (const auto& pair : pairs) {
+        bound_names.emplace(pair.name);
+      }
+      ET_CHECK_OK_OR_RETURN_ERROR(
+          check_constants_bound(*handle, num_constants, bound_names),
+          "CUDA FQN weights do not cover the AOTI library");
+    }
     ET_CHECK_OK_OR_RETURN_ERROR(
         handle->update_user_managed_constant_buffer_pairs(
             handle->container_handle,
             pairs.data(),
             pairs.size(),
             /*use_inactive=*/false,
-            /*validate_full_update=*/true),
+            /*validate_full_update=*/!check_here),
         "Failed to bind CUDA FQN weights");
   }
   ET_LOG(