compiled_model.cc

// Copyright 2024 Google LLC.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//      http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#include "litert/runtime/compiled_model.h"

#include <algorithm>
#include <array>
#include <cstdarg>
#include <cstddef>
#include <cstdint>
#include <cstring>
#include <functional>
#include <iterator>
#include <limits>
#include <memory>
#include <optional>
#include <string>
#include <utility>
#include <vector>

#if !defined(LITERT_WINDOWS_OS)
#include <unistd.h>
#endif  // !defined(LITERT_WINDOWS_OS)

#include "absl/functional/any_invocable.h"  // from @com_google_absl
#include "absl/status/status.h"  // from @com_google_absl
#include "litert/c/litert_layout.h"
#include "litert/cc/internal/litert_consts.h"

#if defined(__ANDROID__)
#include <android/hardware_buffer.h>
#endif

#include "absl/cleanup/cleanup.h"  // from @com_google_absl
#include "absl/log/absl_check.h"  // from @com_google_absl
#include "absl/strings/match.h"  // from @com_google_absl
#include "absl/strings/str_cat.h"  // from @com_google_absl
#include "absl/strings/str_format.h"  // from @com_google_absl
#include "absl/strings/string_view.h"  // from @com_google_absl
#include "absl/types/span.h"  // from @com_google_absl
#include "litert/c/internal/litert_accelerator.h"
#include "litert/c/internal/litert_delegate_wrapper.h"
#include "litert/c/internal/litert_logging.h"
#include "litert/c/internal/litert_runtime_context.h"
#include "litert/c/internal/litert_scheduling_info.h"
#include "litert/c/litert_any.h"
#include "litert/c/litert_common.h"
#include "litert/c/litert_environment_options.h"
#include "litert/c/litert_opaque_options.h"
#include "litert/c/litert_options.h"
#include "litert/c/litert_profiler_event.h"
#include "litert/c/litert_tensor_buffer.h"
#include "litert/c/litert_tensor_buffer_requirements.h"
#include "litert/c/litert_tensor_buffer_types.h"
#include "litert/cc/internal/litert_handle.h"
#include "litert/cc/internal/litert_tensor_buffer_utils.h"
#include "litert/cc/litert_buffer_ref.h"
#include "litert/cc/litert_expected.h"
#include "litert/cc/litert_macros.h"
#include "litert/cc/litert_opaque_options.h"
#include "litert/core/buffer_error_reporter.h"
#include "litert/core/build_stamp.h"
#include "litert/core/error_reporter.h"
#include "litert/core/model/model.h"
#include "tflite/model_builder.h"
#if !defined(LITERT_DISABLE_NPU)
#include "litert/compiler/plugin/compiler_plugin.h"
#include "litert/core/cache/compilation_cache.h"
#include "litert/core/model/model_serialize.h"
#endif  // !defined(LITERT_DISABLE_NPU)
#include "litert/core/options.h"
#include "litert/core/util/flatbuffer_tools.h"
#include "litert/runtime/accelerator.h"
#include "litert/runtime/custom_op_dispatcher.h"
#include "litert/runtime/dispatch/dispatch_opaque_options.h"
#include "litert/runtime/external_litert_buffer_context.h"
#if !defined(LITERT_DISABLE_CPU)
#include "litert/runtime/litert_cpu_options.h"
#endif  // !defined(LITERT_DISABLE_CPU)
#include "litert/runtime/litert_runtime_options.h"
#include "litert/runtime/magic_number_utils.h"
#include "litert/runtime/metrics.h"
#include "litert/runtime/tensor_buffer.h"
#include "litert/runtime/tensor_buffer_requirements.h"
#include "litert/runtime/tensor_identifier.h"
#include "litert/runtime/tfl_utils.h"
#include "weight_loader/external_weight_loader_litert.h"
#include "tflite/converter/allocation.h"
#include "tflite/builtin_ops.h"
#include "tflite/core/api/profiler.h"
#include "tflite/core/interpreter_builder.h"
#include "tflite/interpreter.h"
#include "tflite/interpreter_options.h"
#if !defined(LITERT_NO_BUILTIN_OPS)
#include "tflite/kernels/register.h"
#endif  // LITERT_NO_BUILTIN_OPS

#if defined(LITERT_NO_BUILTIN_OPS)
#include "litert/runtime/stub_op_resolver.h"
#endif  // LITERT_NO_BUILTIN_OPS

using litert::Error;
using litert::Expected;
using litert::Unexpected;
using litert::internal::DispatchDelegateOptions;
using litert::internal::GetTensorIdentifier;
#if !defined(LITERT_DISABLE_NPU)
using litert::internal::SerializeModel;
#endif  // !defined(LITERT_DISABLE_NPU)
using litert::internal::ParseLiteRtRuntimeOptions;
using litert::internal::TfLiteTensorIdentifier;

namespace {
std::optional<std::string> ExtractDirectory(absl::string_view path) {
  const size_t last_sep = path.find_last_of("/\\");
  if (last_sep == absl::string_view::npos) {
    return std::nullopt;
  }
  return std::string(path.substr(0, last_sep));
}

void* StubOpInit([[maybe_unused]] TfLiteContext* context,
                 [[maybe_unused]] const char* buffer,
                 [[maybe_unused]] size_t length) {
  return nullptr;
}

void StubOpFree([[maybe_unused]] TfLiteContext* context,
                [[maybe_unused]] void* buffer) {}

TfLiteStatus StubOpPrepare([[maybe_unused]] TfLiteContext* context,
                           [[maybe_unused]] TfLiteNode* node) {
  return kTfLiteOk;
}

TfLiteStatus StubOpEval(TfLiteContext* context,
                        [[maybe_unused]] TfLiteNode* node) {
  // This should never be called as accelerators will handle the operations
  context->ReportError(
      context, "Stub operation invoked. This function should not be called.");
  return kTfLiteError;
}

TfLiteRegistration sStubRegistration = {
    .init = StubOpInit,
    .free = StubOpFree,
    .prepare = StubOpPrepare,
    .invoke = StubOpEval,
};

#if !defined(LITERT_DISABLE_NPU)
LiteRtLogSeverity GetLogSeverityForJitCompilationFailure(
    LiteRtHwAcceleratorSet hw_accelerators) {
  return (hw_accelerators & kLiteRtHwAcceleratorNpu) ? LITERT_WARNING
                                                     : LITERT_VERBOSE;
}
#endif  // !defined(LITERT_DISABLE_NPU)

constexpr uint32_t kAllSchedulingInfoFields =
    kLiteRtSchedulingInfoFieldOriginalUid |
    kLiteRtSchedulingInfoFieldDebugFeatureId |
    kLiteRtSchedulingInfoFieldJobPriority | kLiteRtSchedulingInfoFieldGroupId;

bool IsValidSchedulingPriority(int32_t priority) {
  return priority >= kLiteRtSchedulingInfoJobPriorityHighest &&
         priority <= kLiteRtSchedulingInfoJobPriorityLowest;
}

int32_t GetDefaultOriginalUid() {
#if !defined(LITERT_WINDOWS_OS)
  return static_cast<int32_t>(getuid());
#else
  return -1;
#endif  // !defined(LITERT_WINDOWS_OS)
}

LiteRtSchedulingInfo GetDefaultSchedulingInfo() {
  LiteRtSchedulingInfo info{};
  info.fields_mask = kAllSchedulingInfoFields;
  info.original_uid = GetDefaultOriginalUid();
  info.debug_feature_id = "";
  info.job_priority = kLiteRtSchedulingInfoJobPriorityHighest;
  memset(info.group_id, 0, sizeof(info.group_id));
  return info;
}

Expected<void> ValidateSchedulingInfo(
    const LiteRtSchedulingInfo& scheduling_info) {
  if ((scheduling_info.fields_mask &
       kLiteRtSchedulingInfoFieldDebugFeatureId) &&
      scheduling_info.debug_feature_id == nullptr) {
    return Unexpected(kLiteRtStatusErrorInvalidArgument,
                      "Scheduling info debug_feature_id cannot be null when "
                      "kLiteRtSchedulingInfoFieldDebugFeatureId is set");
  }
  if ((scheduling_info.fields_mask & kLiteRtSchedulingInfoFieldJobPriority) &&
      !IsValidSchedulingPriority(scheduling_info.job_priority)) {
    return Unexpected(
        kLiteRtStatusErrorInvalidArgument,
        absl::StrFormat(
            "Scheduling info job_priority=%d is out of range [%d, %d]",
            scheduling_info.job_priority,
            kLiteRtSchedulingInfoJobPriorityHighest,
            kLiteRtSchedulingInfoJobPriorityLowest));
  }
  return {};
}

void ApplySchedulingInfoOverrides(const LiteRtSchedulingInfo& overrides,
                                  LiteRtSchedulingInfo& target) {
  if (overrides.fields_mask & kLiteRtSchedulingInfoFieldOriginalUid) {
    target.original_uid = overrides.original_uid;
  }
  if (overrides.fields_mask & kLiteRtSchedulingInfoFieldDebugFeatureId) {
    target.debug_feature_id =
        overrides.debug_feature_id != nullptr ? overrides.debug_feature_id : "";
  }
  if (overrides.fields_mask & kLiteRtSchedulingInfoFieldJobPriority) {
    target.job_priority = overrides.job_priority;
  }
  if (overrides.fields_mask & kLiteRtSchedulingInfoFieldGroupId) {
    memcpy(target.group_id, overrides.group_id, sizeof(target.group_id));
  }
}

}  // namespace

Expected<void> LiteRtCompiledModelT::InitializeRuntime(
    LiteRtEnvironmentT* env, LiteRtHwAcceleratorSet hardware_accelerators,
    LiteRtOptions jit_compilation_options) {
#ifdef LITERT_NO_BUILTIN_OPS
  // Use StubOpResolver which provides minimal stub implementations for all
  // builtin ops. These stubs allow the model to pass validation, but the
  // actual operations will be handled by LiteRT's accelerator system
  // (NPU > GPU > CPU) through their respective delegates.
  litert::internal::StubOpResolver resolver;
#else
  tflite::ops::builtin::BuiltinOpResolverWithoutDefaultDelegates resolver;
#endif  // LITERT_NO_BUILTIN_OPS

  // Apply custom ops.
  if (jit_compilation_options) {
    for (auto& option : jit_compilation_options->custom_op_options) {
      custom_op_dispatchers_.push_back(
          std::make_unique<litert::internal::CustomOpDispatcher>(option));
      auto* tflite_registration =
          custom_op_dispatchers_.back()->GetTfLiteRegistration();
      resolver.AddCustom(option.op_name.c_str(), tflite_registration);
    }
  }

  // Add custom ops that are supported by the CPU / GPU accelerators.
  if (hardware_accelerators & kLiteRtHwAcceleratorGpu) {
    const char* accelerator_supported_custom_ops[] = {
        "Convolution2DTransposeBias", "MaxPoolingWithArgmax2D",
        "MaxUnpooling2D", "Resampler", "custom_call.GroupNorm",
        "custom_call.LayerNorm", "custom_call.RmsNorm",
        "custom_call.PixelShuffle"};
    for (const auto& op_name : accelerator_supported_custom_ops) {
      resolver.AddCustom(op_name, &sStubRegistration);
    }
  } else if (hardware_accelerators & kLiteRtHwAcceleratorCpu) {
    const char* accelerator_supported_custom_ops[] = {
        "Convolution2DTransposeBias", "MaxPoolingWithArgmax2D",
        "MaxUnpooling2D"};
    for (const auto& op_name : accelerator_supported_custom_ops) {
      resolver.AddCustom(op_name, &sStubRegistration);
    }
  }
#ifdef __EMSCRIPTEN__
  else if (hardware_accelerators & kLiteRtHwAcceleratorWebNn) {
    const char* accelerator_supported_custom_ops[] = {
        "Convolution2DTransposeBias"};
    for (const auto& op_name : accelerator_supported_custom_ops) {
      resolver.AddCustom(op_name, &sStubRegistration);
    }
  }
#endif  // __EMSCRIPTEN__

  tflite::InterpreterOptions interpreter_options;
  interpreter_options.SetUseSignatureTensorNames(true);
  int num_threads = 1;
  if (jit_compilation_options) {
    auto opaque_options = litert::OpaqueOptions::WrapCObject(
        jit_compilation_options->options, litert::OwnHandle::kNo);

    if (auto runtime_options_data = litert::FindOpaqueData<const char>(
            opaque_options, LiteRtRuntimeOptionsT::Identifier());
        runtime_options_data) {
      LiteRtRuntimeOptionsT runtime_options;
      absl::string_view data_str(*runtime_options_data);
      if (ParseLiteRtRuntimeOptions(data_str.data(), data_str.size(),
                                    &runtime_options) != kLiteRtStatusOk) {
        LITERT_LOG(LITERT_WARNING, "Failed to parse runtime options");
      } else {
        interpreter_options.SetShloCompositeInlining(true);
        interpreter_options.SetCompressQuantizationZeroPoints(
            runtime_options.compress_quantization_zero_points);
        if (runtime_options.enable_profiling) {
          profiler_ =
              new LiteRtProfilerT(/*max_profiling_buffer_entries=*/2048);
        }

        // Create error reporter based on mode
        switch (runtime_options.error_reporter_mode) {
          case kLiteRtErrorReporterModeNone:
            // No error reporter
            break;
          case kLiteRtErrorReporterModeStderr:
            error_reporter_ = std::make_unique<litert::StderrReporter>();
            break;
          case kLiteRtErrorReporterModeBuffer:
            error_reporter_ = std::make_unique<litert::BufferErrorReporter>();
            break;
        }
      }
    }

#if !defined(LITERT_DISABLE_CPU)
    if (auto cpu_options_data = litert::FindOpaqueData<const char>(
            opaque_options, LiteRtCpuOptionsT::Identifier());
        cpu_options_data) {
      LiteRtCpuOptionsT cpu_options;
      absl::string_view data_str(*cpu_options_data);
      if (litert::internal::ParseLiteRtCpuOptions(
              data_str.data(), data_str.size(), &cpu_options) !=
          kLiteRtStatusOk) {
        LITERT_LOG(LITERT_WARNING, "Failed to parse CPU options");
      } else {
        num_threads = cpu_options.xnn.num_threads;
      }
    }
#endif  // !defined(LITERT_DISABLE_CPU)
  }

  tflite::InterpreterBuilder builder(
      fb_model_->GetModel(), resolver, error_reporter_.get(),
      &interpreter_options, fb_model_->allocation());
  builder(&interp_);
  if (interp_ == nullptr) {
    return Unexpected(kLiteRtStatusErrorRuntimeFailure,
                      "Failed to build TFL interpreter");
  }
  interp_->SetNumThreads(num_threads);

  if (jit_compilation_options) {
    const auto& bindings = jit_compilation_options->external_tensor_bindings;
    for (const auto& binding : bindings) {
      if (litert::internal::SetCustomAllocationForInputTensor(
              interp_.get(), binding) != kTfLiteOk) {
        ReportError("Failed to set custom allocation for tensor: %s",
                    binding.tensor_name.c_str());
        return Unexpected(
            kLiteRtStatusErrorInvalidArgument,
            absl::StrFormat("Failed to apply external tensor binding for "
                            "signature %s, tensor %s.",
                            binding.signature_name, binding.tensor_name));
      }
    }
  }

  if (profiler_ != nullptr) {
    interp_->SetProfiler(profiler_);
  }

  signature_keys_ = interp_->signature_keys();
  if (signature_keys_.empty()) {
    static auto* default_signature_key =
        new std::string(litert::kDefaultSignatureKey);
    signature_keys_.push_back(default_signature_key);
  }

  signature_needs_allocation_.clear();

  auto get_tensor_id =
      [tflite_interpreter = std::ref(*interp_)](
          const TfLiteOpaqueTensor* target_tensor) -> TfLiteTensorIdentifier {
    auto tensor_id = GetTensorIdentifier(
        tflite_interpreter,
        reinterpret_cast<const TfLiteTensor*>(target_tensor));
    if (!tensor_id) {
      LITERT_LOG(LITERT_ERROR, "Failed to get tensor identifier: %s",
                 tensor_id.Error().Message().c_str());
      constexpr TfLiteTensorIdentifier kInvalidTensorId{-1, -1};
      return kInvalidTensorId;
    }
    return *tensor_id;
  };

  // Register the ExternalLiteRtBufferContext for TensorBuffer handshaking.
  buffer_context_ =
      std::make_unique<LiteRtExternalLiteRtBufferContextT>(env, get_tensor_id);
  interp_->SetExternalContext(kTfLiteLiteRtBufferContext,
                              buffer_context_.get());

  std::unique_ptr<litert::ScopedWeightSource> scoped_weight_source;
  auto* options_impl =
    reinterpret_cast<LiteRtOptionsT*>(jit_compilation_options);
  if (options_impl != nullptr) {
    options_impl->weight_loader = nullptr;
    scoped_weight_source = std::move(options_impl->scoped_weight_source);
  }
  weight_loader_ = weight_loader::CreateLiteRtWeightLoader(
      fb_model_->GetModel(), model_directory_, std::move(scoped_weight_source));
  has_external_weights_ = false;
  if (weight_loader_) {
    auto weight_infos = weight_loader_->GetWeightInfo();
    has_external_weights_ = !weight_infos.empty();
    if (!has_external_weights_) {
      LITERT_LOG(LITERT_DEBUG,
                 "External weight loader: no external weight tensors found");
      return {};
    }
    weight_loader::WeightAccessRequest request;
    request.cpu = true;
    // TODO(b/456318365): Handle weight access request to support multiple
    // backends.
    request.opencl = false;
    absl::Status prepare_status = weight_loader_->PrepareAccess(request, env);
    if (!prepare_status.ok()) {
      return litert::Unexpected(kLiteRtStatusErrorRuntimeFailure,
                                std::string(prepare_status.message()));
    }

    if (options_impl != nullptr) {
      options_impl->weight_loader = weight_loader_.get();
    }

    // Inspect the weight infos to log the available weights for GPU delegates.
    LITERT_LOG(LITERT_DEBUG,
               "External weight loader: %zu weight tensors available for GPU "
               "delegates",
               weight_infos.size());
    for (const auto& info : weight_infos) {
      if (info.packing.empty()) {
        LITERT_LOG(LITERT_DEBUG,
                   "  Weight tensor: external_buffer_id=%u, packing=<none>",
                   info.external_buffer_id);
      } else {
        LITERT_LOG(LITERT_DEBUG,
                   "  Weight tensor: external_buffer_id=%u, packing=%.*s",
                   info.external_buffer_id,
                   static_cast<int>(info.packing.size()), info.packing.data());
      }
    }
  }
  return {};
}

Expected<void> LiteRtCompiledModelT::RestoreExternalWeightsForCpu() {
  if (!weight_loader_ || !has_external_weights_) {
    return {};
  }

  // Iterate through all subgraphs and restore external weights to tensors.
  for (int subgraph_idx = 0; subgraph_idx < interp_->subgraphs_size();
       ++subgraph_idx) {
    auto* subgraph = interp_->subgraph(subgraph_idx);
    const auto& external_buffer_ids =
        subgraph->GetExternalTensorBufferIdentifiers();

    for (const auto& [tensor_index, external_buffer_id] : external_buffer_ids) {
      const weight_loader::WeightAccess* weight_access =
          weight_loader_->GetExternalWeightByBuffer(external_buffer_id);
      if (!weight_access) {
        return litert::Unexpected(
            kLiteRtStatusErrorRuntimeFailure,
            absl::StrFormat(
                "Failed to get external weight for buffer id %u (tensor %zu)",
                external_buffer_id, tensor_index));
      }

      LiteRtTensorBuffer host_buffer = weight_access->GetHostBuffer();
      if (!host_buffer) {
        return litert::Unexpected(
            kLiteRtStatusErrorRuntimeFailure,
            absl::StrFormat(
                "Host tensor buffer is null for buffer id %u (tensor %zu)",
                external_buffer_id, tensor_index));
      }

      void* host_memory_addr = nullptr;
      if (LiteRtGetTensorBufferHostMemory(host_buffer, &host_memory_addr) !=
          kLiteRtStatusOk) {
        return litert::Unexpected(
            kLiteRtStatusErrorRuntimeFailure,
            absl::StrFormat(
                "Failed to get host memory for buffer id %u (tensor %zu)",
                external_buffer_id, tensor_index));
      }

      TfLiteTensor* tensor = subgraph->tensor(tensor_index);
      if (!tensor) {
        return litert::Unexpected(
            kLiteRtStatusErrorRuntimeFailure,
            absl::StrFormat("Tensor %zu not found in subgraph %d", tensor_index,
                            subgraph_idx));
      }

      // Set the tensor data to point to the external weight buffer.
      // We use kTfLiteCustom allocation type so TFLite doesn't try to manage
      // this memory. The memory is owned by weight_loader_ and stays valid
      // until weight_loader_ is destroyed.
      tensor->data.raw = static_cast<char*>(host_memory_addr);
      tensor->allocation_type = kTfLiteCustom;

      LITERT_LOG(LITERT_DEBUG,
                 "Restored external weight: subgraph=%d, tensor=%zu, "
                 "external_buffer_id=%u, addr=%p",
                 subgraph_idx, tensor_index, external_buffer_id,
                 host_memory_addr);
    }
  }
  return {};
}

namespace {

int GetAllocationFd(const tflite::Allocation* allocation) {
  if (allocation != nullptr &&
      allocation->type() == tflite::Allocation::Type::kMMap) {
    auto& mmap_allocation =
        static_cast<const tflite::MMAPAllocation&>(*allocation);
    return mmap_allocation.fd();
  }
  return -1;
}

#if !defined(LITERT_DISABLE_NPU)

Expected<std::vector<litert::internal::CompilerPlugin>> TryGetCompilerPlugins(
    LiteRtOptionsT& options, LiteRtEnvironmentT& env,
    LiteRtHwAcceleratorSet hw_accelerators) {
  auto option = env.GetOption(kLiteRtEnvOptionTagCompilerPluginLibraryDir);
  if (!option.has_value() || option->type != kLiteRtAnyTypeString) {
    return Error(kLiteRtStatusErrorRuntimeFailure,
                 "Compiler plugin is not configured");
  }
  std::string compiler_plugin_lib_path = option->str_value;
  const std::array<const absl::string_view, 1>
      compiler_plugin_lib_search_paths = {compiler_plugin_lib_path};

  Expected<std::vector<litert::internal::CompilerPlugin>>
      compiler_plugins_expected = litert::internal::CompilerPlugin::LoadPlugins(
          compiler_plugin_lib_search_paths, &env.GetOptions(), &options);

  if (!compiler_plugins_expected) {
    LITERT_LOG(GetLogSeverityForJitCompilationFailure(hw_accelerators),
               "Failed to load compiler plugins: %s",
               compiler_plugins_expected.Error().Message().c_str());
  }
  return compiler_plugins_expected;
}

std::optional<litert::internal::CompilationCache> MaybeCreateCompilationCache(
    LiteRtEnvironmentT& env) {
  std::optional<LiteRtAny> compiler_cache_dir_option =
      env.GetOption(kLiteRtEnvOptionTagCompilerCacheDir);
  if (compiler_cache_dir_option.has_value() &&
      compiler_cache_dir_option->type == kLiteRtAnyTypeString) {
    LITERT_LOG(LITERT_INFO,
               "NPU JIT compilation caching enabled with cache dir: %s",
               compiler_cache_dir_option->str_value);
    auto compilation_cache_expected =
        litert::internal::CompilationCache::Create(
            compiler_cache_dir_option->str_value);
    if (compilation_cache_expected.HasValue()) {
      return compilation_cache_expected.Value();
    }
  }
  return std::nullopt;
}

void TryApplyPluginsImpl(
    LiteRtModel model, LiteRtHwAcceleratorSet selected_hw_accelerators,
    std::vector<litert::internal::CompilerPlugin>& compiler_plugins,
    bool* mutated) {
  LITERT_LOG(LITERT_INFO, "Applying compiler plugins...");
  // TODO: b/409819691 - Pass user provided `LiteRtOptions` down to the
  // vendor code (nullptr are safe for now).
  auto jit_result = litert::internal::ApplyPlugins(
      model, selected_hw_accelerators, compiler_plugins, mutated);
  if (!jit_result) {
    LITERT_LOG(GetLogSeverityForJitCompilationFailure(selected_hw_accelerators),
               "Failed to apply compiler plugins: %s",
               jit_result.Error().Message().c_str());
  } else {
    LITERT_LOG(LITERT_INFO, "%d compiler plugins were applied successfully: %s",
               jit_result->num_applied_plugins,
               jit_result->success_message.c_str());
    LITERT_LOG(LITERT_WARNING, "Plugin errs: %s",
               jit_result->error_message.c_str());
  }
}

#endif  // !defined(LITERT_DISABLE_NPU)

}  // namespace

Expected<void> LiteRtCompiledModelT::InitializeModel(
    LiteRtModelT& model, LiteRtHwAcceleratorSet hw_accelerators,
    LiteRtOptions options, LiteRtEnvironmentT& env) {
  LITERT_RETURN_IF_ERROR(
      litert::internal::ReplaceMagicNumbersIfAny(env, model));

  if (auto source_path = model.SourcePath()) {
    model_directory_ = ExtractDirectory(*source_path);
  } else {
    model_directory_.reset();
  }

  // If hardware acceleration is requested then apply the plugins to the model
  // and initialize the compiled model from the translated LiteRt model.
  if (hw_accelerators != kLiteRtHwAcceleratorNone) {
#if !defined(LITERT_DISABLE_NPU)
    // Check if the model is pre-compiled by looking for LiteRt build stamp.
    // Fall through to the next step if it's pre-compiled, else try to apply
    // plugins to the model.
    if (!IsCompiled(model)) {
      Expected<bool> maybe_initialized_model =
          ApplyPluginsWithCaching(model, hw_accelerators, *options, env);
      if (maybe_initialized_model.HasValue() &&
          maybe_initialized_model.Value() == true) {
        // The compiled model's flatbuffer has initialized by applying the
        // plugins.
        return {};
      }
      // Deliberate fall through, failing to apply plugins is a recoverable
      // error, we will try to initialize the compiled model from the incoming
      // litert model.
    } else {
      if (env.GetOption(kLiteRtEnvOptionTagCompilerPluginLibraryDir)
              .has_value()) {
        LITERT_LOG(LITERT_WARNING,
                   "Compiler plugin path is provided in the environment, but "
                   "the model is "
                   "pre-compiled. Plugins won't be applied.");
      }
    }
#endif
  }

  const auto& tfl_wrapper = litert::internal::GetTflFlatbuffer(model);
  // Currently, in all situations where litert model was import from a
  // flatbuffer, the litert model will own said flatbuffer and stored it in the
  // OwningBufferRef.

  if (auto tfl_buf = tfl_wrapper.Buf(); tfl_buf.Data() != nullptr) {
    LITERT_LOG(
        LITERT_INFO,
        "Flatbuffer model initialized directly from incoming litert model.");
    fb_model_ = tflite::FlatBufferModel::BuildFromBuffer(
        tfl_buf.StrData(), tfl_buf.Size(), error_reporter_.get());
    fb_model_fd_ = GetAllocationFd(tfl_wrapper.FlatbufferModel().allocation());
    return {};
  }

  // If we reach here, it means we weren't able to initialize the compiled
  // model, neither from the incoming litert model nor from a transformed model
  // after applying the plugins.

  return Unexpected(kLiteRtStatusErrorInvalidArgument,
                    "Failed to build flatbuffer from incoming litert model");
}

namespace {

// A utility class that allows appending additional compilation options, but
// only for the duration of a scope.
class ScopedCompilationOptionsModifier {
 public:
  explicit ScopedCompilationOptionsModifier(LiteRtOptions compilation_options)
      : accelerator_options_(&compilation_options->options) {}

  ~ScopedCompilationOptionsModifier() {
    // Remove any option that was appended during the lifetime of this object.
    while (--num_appended_options_ >= 0) {
      LiteRtPopOpaqueOptions(accelerator_options_);
    }
  }

  Expected<void> Append(litert::OpaqueOptions&& accelerator_options) {
    LITERT_RETURN_IF_ERROR(LiteRtAppendOpaqueOptions(
        accelerator_options_, accelerator_options.Release()));
    ++num_appended_options_;
    return {};
  }

 private:
  LiteRtOpaqueOptions* const accelerator_options_;
  int num_appended_options_ = 0;
};

}  // namespace

Expected<LiteRtCompiledModelT::Ptr> LiteRtCompiledModelT::Create(
    LiteRtEnvironmentT* env, LiteRtModel model,
    LiteRtOptions jit_compilation_options) {
  if (!jit_compilation_options) {
    return litert::ErrorStatusBuilder::InvalidArgument()
           << "No compilation options passed.";
  }

  auto compiled_model = std::make_unique<LiteRtCompiledModelT>(env);

  LiteRtHwAcceleratorSet hardware_accelerators = kLiteRtHwAcceleratorNone;
  LITERT_RETURN_IF_ERROR(LiteRtGetOptionsHardwareAccelerators(
      jit_compilation_options, &hardware_accelerators));

  if (hardware_accelerators == kLiteRtHwAcceleratorNone) {
    return litert::ErrorStatusBuilder::InvalidArgument()
           << "No acceleration provided.";
  }

  LITERT_RETURN_IF_ERROR(compiled_model->InitializeModel(
      *model, hardware_accelerators, jit_compilation_options, *env));

  LITERT_RETURN_IF_ERROR(compiled_model->InitializeRuntime(
      env, hardware_accelerators, jit_compilation_options));
  if (compiled_model->GetModelBase() == nullptr) {
    return Error(kLiteRtStatusErrorRuntimeFailure,
                 "Failed to initialize model memory.");
  }

  ScopedCompilationOptionsModifier scoped_modifier(jit_compilation_options);

  {
    // Add information about the model allocation to the opaque chain.
    LITERT_ASSIGN_OR_RETURN(auto dispatch_options,
                            DispatchDelegateOptions::Create());
    LITERT_RETURN_IF_ERROR(
        dispatch_options.SetAllocBase(compiled_model->GetModelBase()));
    LITERT_RETURN_IF_ERROR(
        dispatch_options.SetAllocBaseFd(compiled_model->fb_model_fd_));
    LITERT_RETURN_IF_ERROR(scoped_modifier.Append(std::move(dispatch_options)));
  }

  // Load and restore external weights for CPU execution before delegates are
  // applied. This ensures that XNNPack and other CPU delegates can see the
  // weight data.
  if (hardware_accelerators & kLiteRtHwAcceleratorCpu) {
    LITERT_RETURN_IF_ERROR(compiled_model->RestoreExternalWeightsForCpu());
  }

  // Apply accelerators matching the requested hardware support to the
  // model in the order they were registered.
  for (auto& accelerator : env->GetAcceleratorRegistry()) {
    bool delegate_responsible_for_jit = false;
    LITERT_RETURN_IF_ERROR(
        LiteRtIsAcceleratorDelegateResponsibleForJitCompilation(
            accelerator.get(), &delegate_responsible_for_jit));

    LiteRtHwAcceleratorSet accelerator_supported_hardware;
    LITERT_RETURN_IF_ERROR(accelerator->GetHardwareSupport(
        accelerator.get(), &accelerator_supported_hardware));

    // We don't apply the delegate if:
    //   - the delegate is responsible for JIT compilation
    //   - and JIT has not been requested for the hardware it supports.
    if (delegate_responsible_for_jit &&
        !(hardware_accelerators & accelerator_supported_hardware)) {
      continue;
    }

    LITERT_DEBUG_CODE({
      const char* accelerator_name = nullptr;
      if (accelerator->GetName(accelerator.get(), &accelerator_name) !=
              kLiteRtStatusOk ||
          !accelerator_name) {
        LITERT_LOG(LITERT_WARNING, "Failed to get name for accelerator");
      } else {
        LITERT_LOG(LITERT_DEBUG, "Apply accelerator %s", accelerator_name);
      }
    });

    LiteRtDelegateWrapper delegate_wrapper = nullptr;
    LITERT_RETURN_IF_ERROR(accelerator->CreateDelegate(
        LrtGetRuntimeContext(), env, accelerator.get(), jit_compilation_options,
        &delegate_wrapper));

    TfLiteOpaqueDelegate* delegate_ptr = nullptr;
    LrtGetRuntimeContext()->unwrap_delegate(delegate_wrapper, &delegate_ptr);

    auto delegate = std::unique_ptr<LiteRtDelegateWrapperT,
                                    std::function<void(LiteRtDelegateWrapper)>>{
        delegate_wrapper, [destroy_fn = accelerator->DestroyDelegate](
                              LiteRtDelegateWrapper wrapper) {
          if (destroy_fn) destroy_fn(LrtGetRuntimeContext(), wrapper);
        }};

    if (compiled_model->interp_->ModifyGraphWithDelegate(delegate_ptr) !=
        kTfLiteOk) {
      return Unexpected(kLiteRtStatusErrorRuntimeFailure,
                        "Failed to modify graph with delegate");
    }

    compiled_model->RegisterDelegate({std::move(delegate),
                                      accelerator->StartMetricsCollection,
                                      accelerator->StopMetricsCollection});
  }

  LITERT_ASSIGN_OR_RETURN(bool has_non_delegated_ops,
                          compiled_model->HasNonDelegatedOps());
  if (!(hardware_accelerators & kLiteRtHwAcceleratorCpu) &&
      has_non_delegated_ops) {
    return Error(
        kLiteRtStatusErrorCompilation,
        "Some ops are not accelerated. Add kLiteRtHwAcceleratorCpu to the "
        "compilation accelerator set to allow using the CPU to run those.");
  }
  compiled_model->CheckCpuTensors();
  return compiled_model;
}

Expected<bool> LiteRtCompiledModelT::HasNonDelegatedOps() {
  for (int subgraph_no = 0; subgraph_no < interp_->subgraphs_size();
       ++subgraph_no) {
    const auto* const subgraph = interp_->subgraph(subgraph_no);
    if (subgraph->IsDelegationSkippable()) {
      continue;
    }
    const auto& execution_plan = subgraph->execution_plan();
    const auto& nodes_and_registration = subgraph->nodes_and_registration();
    for (int node_index : execution_plan) {
      const TfLiteRegistration& registration =
          nodes_and_registration[node_index].second;
      if (registration.builtin_code != kTfLiteBuiltinDelegate &&
          (registration.builtin_code != kTfLiteBuiltinCustom ||
           litert::internal::kLiteRtDispatchOpCustomName !=
               registration.custom_name)) {
        return true;
      }
    }
  }
  return false;
}

void LiteRtCompiledModelT::CheckCpuTensors() {
  cpu_tensors_.clear();
  for (int subgraph_no = 0; subgraph_no < interp_->subgraphs_size();
       ++subgraph_no) {
    auto* subgraph = interp_->subgraph(subgraph_no);
    auto& execution_plan = subgraph->execution_plan();
    auto& nodes_and_registration = subgraph->nodes_and_registration();
    for (int node_index : execution_plan) {
      const TfLiteNode& node = nodes_and_registration[node_index].first;
      const TfLiteRegistration& registration =
          nodes_and_registration[node_index].second;
      // Don't mark delegate nodes as CPU nodes except for XNNPack ones.
      if (registration.builtin_code == kTfLiteBuiltinDelegate &&
          !(registration.custom_name &&
            registration.custom_name ==
                absl::string_view("TfLiteXNNPackDelegate"))) {
        continue;
      }
      // Don't mark AOT compiled NPU custom ops as CPU nodes.
      if (registration.builtin_code == kTfLiteBuiltinCustom &&
          registration.custom_name &&
          absl::StrContains(registration.custom_name,
                            litert::internal::kLiteRtDispatchOpCustomName)) {
        continue;
      }
      // Mark input of node as CPU tensors.
      for (int i = 0; i < node.inputs->size; ++i) {
        int input_tensor_index = node.inputs->data[i];
        if (input_tensor_index == kTfLiteOptionalTensor) continue;
        cpu_tensors_.insert({subgraph_no, input_tensor_index});
      }
    }
  }
}

#if !defined(LITERT_DISABLE_NPU)
Expected<bool> LiteRtCompiledModelT::ApplyPluginsWithCaching(
    LiteRtModelT& model, LiteRtHwAcceleratorSet hw_accelerators,
    LiteRtOptionsT& options, LiteRtEnvironmentT& env) {
  bool need_reserialization = false;
  compilation_cache_ = MaybeCreateCompilationCache(env);
  std::optional<uint64_t> model_hash = std::nullopt;
  // Load the plugins before JIT compilation attempt, so that we can check the
  // cache first.
  auto maybe_compiled_plugins =
      TryGetCompilerPlugins(options, env, hw_accelerators);
  // If we have a cache (user provided
  // 'kLiteRtEnvOptionTagCompilerPluginLibraryDir'), and we loaded the plugins
  // successfully, we can try to load the model from the cache.
  if (compilation_cache_.has_value()) {
    Expected<uint64_t> maybe_model_hash =
        litert::internal::CompilationCache::TryGetModelHash(
            model, &options, maybe_compiled_plugins);
    if (maybe_model_hash.HasValue()) {
      model_hash = maybe_model_hash.Value();
      if (TryLoadingFromCache(model_hash.value())) {
        LITERT_LOG(LITERT_INFO,
                   "Flatbuffer model initialized from cached model.");
        return true;
      }
    }
  }
  // Cache miss, we need to continue with JIT compilation.
  if (maybe_compiled_plugins.HasValue()) {
    TryApplyPluginsImpl(&model, hw_accelerators, maybe_compiled_plugins.Value(),
                        &need_reserialization);
    // Store the compiler plugins to a member variable to postpone its
    // destruction and reduce initialization time.
    maybe_compiled_plugins_ = std::move(maybe_compiled_plugins.Value());
  }
  if (!need_reserialization) {
    return false;
  }
  LITERT_LOG(LITERT_INFO, "JIT compilation changed model, reserializing...");

  LITERT_ASSIGN_OR_RETURN(auto serialized, SerializeModel(std::move(model)));
  if (model_hash.has_value()) {
    LITERT_LOG(LITERT_DEBUG, "Saving JIT compiled model to cache.");
    LITERT_RETURN_IF_ERROR(
        compilation_cache_.value().SaveModel(serialized, model_hash.value()));
  }

  model_buf_ = std::move(serialized);
  fb_model_ = tflite::FlatBufferModel::BuildFromBuffer(
      reinterpret_cast<const char*>(model_buf_.Data()), model_buf_.Size(),
      error_reporter_.get());
  if (fb_model_ == nullptr) {
    return Unexpected(kLiteRtStatusErrorFileIO,
                      "Failed to build flatbuffer from buffer");
  }
  LITERT_LOG(LITERT_DEBUG,
             "Plugins applied, flatbuffer model initialized from  JIT compiled "
             "model.");
  fb_model_fd_ = GetAllocationFd(fb_model_->allocation());
  return true;
}

bool LiteRtCompiledModelT::TryLoadingFromCache(uint64_t model_hash) {
  if (!compilation_cache_.has_value()) {
    return false;
  }
  // Check if we compiled this model before.
  Expected<std::optional<LiteRtModelT::Ptr>> maybe_cached_model =
      compilation_cache_.value().TryLoadModel(model_hash);
  if (!maybe_cached_model) {
    // The model was found in the cache, but failed to load.
    LITERT_LOG(LITERT_WARNING, "Failed to load model from cache: %s",
               maybe_cached_model.Error().Message().c_str());
    return false;
  }
  std::optional<LiteRtModelT::Ptr> cached_model =
      std::move(maybe_cached_model.Value());
  if (!cached_model.has_value()) {
    // The model was not found in the cache, don't log anything because this is
    // expected when the cache is cleared, or when the model is new.
    return false;
  }

  // Cache hit and model loaded successfully, initialize the compiled model
  // with the cached model.
  const auto& tfl_wrapper_from_cached_model =
      litert::internal::GetTflFlatbuffer(*cached_model.value());

  auto tfl_buf_from_cached_model = tfl_wrapper_from_cached_model.Buf();
  fb_model_ = tflite::FlatBufferModel::BuildFromBuffer(
      tfl_buf_from_cached_model.StrData(), tfl_buf_from_cached_model.Size(),
      error_reporter_.get());
  fb_model_fd_ = GetAllocationFd(
      tfl_wrapper_from_cached_model.FlatbufferModel().allocation());
  cached_model_ = std::move(cached_model.value());
  return true;
}
#endif  // !defined(LITERT_DISABLE_NPU)

Expected<const LiteRtTensorBufferRequirementsT*>
LiteRtCompiledModelT::GetTensorBufferRequirements(const TfLiteTensor* tensor) {
  LITERT_ASSIGN_OR_RETURN(const auto tensor_id,
                          GetTensorIdentifier(*interp_, tensor));
  // Use the buffer context to get the buffer requirements only if the tensor
  // is not a CPU tensor.
  if (cpu_tensors_.find(tensor_id) == cpu_tensors_.end()) {
    if (auto requirements = buffer_context_->GetBufferRequirements(tensor)) {
      return *requirements;
    }
  } else {
    LITERT_LOG(LITERT_DEBUG, "Tensor %s is shared with CPU.\n", tensor->name);
  }
  // Check if we have a cached CPU buffer requirement.
  auto cached_req = cpu_buffer_requirements_.find(tensor_id);
  if (cached_req != cpu_buffer_requirements_.end()) {
    return cached_req->second.get();
  }
  LiteRtTensorBufferRequirements litert_cpu_buffer_requirements;
  LiteRtTensorBufferType cpu_buffer_type[] = {
      kLiteRtTensorBufferTypeHostMemory};
  uint32_t cpu_buffer_strides[] = {0};
  LITERT_RETURN_IF_ERROR(LiteRtCreateTensorBufferRequirements(
      /*num_supported_tensor_buffer_types=*/1, cpu_buffer_type, tensor->bytes,
      /*num_strides=*/1, cpu_buffer_strides, &litert_cpu_buffer_requirements));
  cpu_buffer_requirements_[tensor_id] =
      LiteRtTensorBufferRequirementsPtr(litert_cpu_buffer_requirements);
  return const_cast<LiteRtTensorBufferRequirementsT*>(
      litert_cpu_buffer_requirements);
}

Expected<const LiteRtTensorBufferRequirementsT*>
LiteRtCompiledModelT::GetInputBufferRequirements(
    absl::string_view signature_key, size_t input_index) {
  auto runner = GetSignatureRunner(signature_key);
  if (runner == nullptr) {
    return Unexpected(kLiteRtStatusErrorNotFound,
                      "Failed to get signature runner");
  }
  auto input_names = runner->subgraph_input_names();
  if (input_index >= input_names.size()) {
    return Unexpected(kLiteRtStatusErrorIndexOOB, "Input index out of range");
  }
  auto input_name = input_names[input_index];
  auto* input_tensor = runner->input_tensor(input_name);
  if (input_tensor == nullptr) {
    return Unexpected(kLiteRtStatusErrorNotFound, "Failed to get input tensor");
  }

  return GetTensorBufferRequirements(input_tensor);
}

Expected<const LiteRtTensorBufferRequirementsT*>
LiteRtCompiledModelT::GetOutputBufferRequirements(
    absl::string_view signature_key, size_t output_index) {
  auto runner = GetSignatureRunner(signature_key);
  if (runner == nullptr) {
    return Unexpected(kLiteRtStatusErrorNotFound,
                      "Failed to get signature runner");
  }
  auto output_names = runner->subgraph_output_names();
  if (output_index >= output_names.size()) {
    return Unexpected(kLiteRtStatusErrorIndexOOB, "Output index out of range");
  }
  auto output_name = output_names[output_index];
  auto* output_tensor = runner->output_tensor(output_name);
  if (output_tensor == nullptr) {
    return Unexpected(kLiteRtStatusErrorNotFound,
                      "Failed to get output tensor");
  }

  return GetTensorBufferRequirements(output_tensor);
}

Expected<LiteRtLayout> LiteRtCompiledModelT::GetInputTensorLayout(
    size_t signature_index, size_t input_index) {
  if (signature_index >= signature_keys_.size()) {
    return Unexpected(kLiteRtStatusErrorIndexOOB,
                      "Signature index is out of range of signature keys");
  }
  auto* runner = GetSignatureRunner(*signature_keys_[signature_index]);
  if (runner == nullptr) {
    return Unexpected(kLiteRtStatusErrorInvalidArgument,
                      "Failed to get signature runner");
  }
  const auto& input_names = runner->subgraph_input_names();
  if (input_index >= input_names.size()) {
    return Unexpected(kLiteRtStatusErrorIndexOOB, "Input index out of range");
  }
  auto* input_tensor = runner->input_tensor(input_names[input_index]);
  if (input_tensor == nullptr) {
    return Unexpected(kLiteRtStatusErrorNotFound, "Failed to get input tensor");
  }

  const TfLiteIntArray* dims = input_tensor->dims;
  if ((!dims || dims->size == 0) && input_tensor->dims_signature &&
      input_tensor->dims_signature->size > 0) {
    dims = input_tensor->dims_signature;
  }

  const size_t rank = dims ? dims->size : 0;
  if (rank > LITERT_TENSOR_MAX_RANK) {
    return Unexpected(kLiteRtStatusErrorInvalidArgument,
                      "Input tensor rank exceeds maximum supported "
                      "rank for layouts");
  }

  LiteRtLayout layout{};
  layout.rank = static_cast<unsigned int>(rank);
  for (size_t i = 0; i < rank; ++i) {
    layout.dimensions[i] = dims->data[i];
  }
  return layout;
}

Expected<void> LiteRtCompiledModelT::GetOutputTensorShapes(
    absl::string_view signature_key, absl::Span<LiteRtLayout>& output_layouts,
    bool update_allocation) {
  auto runner = GetSignatureRunner(signature_key);
  if (runner == nullptr) {
    return Unexpected(kLiteRtStatusErrorNotFound,
                      "Failed to get signature runner");
  }
  if (update_allocation) {
    LITERT_ASSIGN_OR_RETURN(bool needs_allocation,
                            SignatureNeedsAllocation(runner));
    if (needs_allocation) {
      if (auto res = runner->AllocateTensors(); res != kTfLiteOk) {
        if (error_reporter_) {
          error_reporter_->Report("Failed to allocate tensors for execution");
        }
        return Unexpected(kLiteRtStatusErrorRuntimeFailure,
                          "Failed to allocate tensors");
      }
      LITERT_RETURN_IF_ERROR(MarkSignatureAllocationUpToDate(runner));
    }
  }
  auto output_names = runner->subgraph_output_names();
  // Check whether output_layouts has enough space to store all output tensors
  if (output_layouts.size() != output_names.size()) {
    return Unexpected(
        kLiteRtStatusErrorInvalidArgument,
        absl::StrFormat("Output layout size is incorrect, expected "
                        "%d but got %d",
                        output_names.size(), output_layouts.size()));
  }
  for (int i = 0; i < output_names.size(); ++i) {
    const TfLiteIntArray* dims = runner->output_tensor(output_names[i])->dims;
    output_layouts[i].rank = dims->size;
    for (int j = 0; j < dims->size; ++j) {
      output_layouts[i].dimensions[j] = dims->data[j];
    }
  }
  return {};
}

tflite::SignatureRunner* LiteRtCompiledModelT::GetSignatureRunner(
    absl::string_view signature_key) {
  if (signature_runners_.contains(signature_key)) {
    return signature_runners_[signature_key];
  }
  auto runner =
      interp_->GetSignatureRunner(signature_key == litert::kDefaultSignatureKey
                                      ? nullptr
                                      : std::string(signature_key).c_str());
  signature_runners_[signature_key] = runner;
  return runner;
}

Expected<void> LiteRtCompiledModelT::RegisterBuffer(
    tflite::SignatureRunner* runner, TfLiteTensor* tensor,
    const char* tensor_name, LiteRtTensorBufferT* buffer, bool is_input,
    std::vector<LiteRtTensorBuffer>& locked_buffers,
    std::vector<ConstantOutputInfo>& constant_outputs) {
  LITERT_DEBUG_CODE({
    absl::string_view io = is_input ? "input" : "output";
    auto buffer_type = litert::BufferTypeToString(buffer->buffer_type());
    LITERT_LOG(LITERT_DEBUG,
               "Registering %s tensor from TfliteTensor %p to "
               "LiteRtTensorBuffer %p of type %s",
               io.data(), tensor, buffer, buffer_type.data());
  });

  bool is_constant_output = !is_input &&
                            tensor->allocation_type == kTfLiteMmapRo &&
                            tensor->data.raw != nullptr;

  // Automatic shape detection for input tensors.
  if (is_input) {
    auto [_, layout] = buffer->tensor_type();
    absl::Span<const int> buffer_shape =
        absl::MakeConstSpan(layout.dimensions, layout.rank);

    LITERT_ASSIGN_OR_RETURN(bool needs_auto_resize,
                            InputTensorNeedsResize(tensor, buffer_shape));
    if (needs_auto_resize) {
      // When an input tensor is resized, output and intermediate tensors may
      // also be resized. This can invalidate previously cached buffer
      // requirements, so we clear the cache.
      cpu_buffer_requirements_.clear();
      // Shape change detected - perform automatic resize.
      if (runner->ResizeInputTensor(
              tensor_name, std::vector(buffer_shape.begin(),
                                       buffer_shape.end())) == kTfLiteOk) {
        LITERT_RETURN_IF_ERROR(MarkSignatureNeedsAllocation(runner));
        LITERT_LOG(LITERT_INFO, "Automatically resized input tensor %s",
                   tensor_name ? tensor_name : "<unnamed>");
      } else {
        LITERT_LOG(LITERT_WARNING, "Automatic resize failed for tensor %s",
                   tensor_name ? tensor_name : "<unnamed>");
      }
    } else {
      // Get current tensor shape.
      absl::Span<const int> current_shape =
          absl::MakeConstSpan(tensor->dims->data, tensor->dims->size);
      LITERT_RETURN_IF_ERROR(current_shape == buffer_shape,
                             Unexpected(kLiteRtStatusErrorInvalidArgument,
                                        "Input tensor shape mismatch"));
    }
  }

  bool buffer_requires_cpu_sync = false;

  Expected<LiteRtTensorBufferRequirementsConst> requirements =
      buffer_context_->GetBufferRequirements(tensor);
  if (requirements) {
    const auto& supported_types = (*requirements)->SupportedBufferTypes();

    for (auto& type : supported_types) {
      if (type == buffer->buffer_type()) {
        // Register tensor buffer if it can be used by the backend.
        buffer->Duplicate();
        LiteRtTensorBufferPtr duplicated_buffer(buffer);
        if (auto status = buffer_context_->RegisterTensorBuffer(
                tensor, std::move(duplicated_buffer));
            status != kLiteRtStatusOk) {
          return Unexpected(kLiteRtStatusErrorRuntimeFailure,
                            "Failed to register tensor buffer");
        }

        // Mark the tensor as non-CPU to avoid TFLite from allocating it.
        tensor->allocation_type = kTfLiteNonCpu;
        tensor->data.data = nullptr;
        return {};
      }
    }
    // At this point, none of the supported buffer types of the backend matches
    // the buffer type of the tensor. Also, the backend does not support host
    // memory buffer.
    // TODO: b/452064364 - refactor register buffer logic.
    buffer_requires_cpu_sync =
        buffer->buffer_type() == kLiteRtTensorBufferTypeHostMemory;
  } else {
    // If the BufferRequirement is not registered, assumes the backend requires
    // CPU buffer.
    buffer_requires_cpu_sync = true;
  }

  if (buffer_requires_cpu_sync) {
    // When backend requires CPU buffer.
    bool buffer_is_cpu_compatible =
        buffer->buffer_type() == kLiteRtTensorBufferTypeHostMemory ||
        IsUserCustomBuffer(buffer->buffer_type()) || buffer->is_opencl_memory();
#if defined(__ANDROID__)
    if (buffer->buffer_type() == kLiteRtTensorBufferTypeAhwb) {
      if (__builtin_available(android 26, *)) {
        if (auto ahwb = buffer->GetAhwbBuffer()) {
          // TODO: b/382330322 - Update logic to check if the AHWB (stride) is
          // CPU compatible.
          // AHardwareBuffer_Desc desc;
          // AHardwareBuffer_describe(*ahwb, &desc);
          buffer_is_cpu_compatible = true;
        }
      }
    } else if (buffer->buffer_type() == kLiteRtTensorBufferTypeFastRpc ||
               buffer->buffer_type() == kLiteRtTensorBufferTypeDmaBuf) {
      buffer_is_cpu_compatible = true;
    }
#endif
    if (buffer_is_cpu_compatible) {
      void* host_mem_addr;
      LiteRtTensorBufferLockMode lock_mode =
          is_input ? kLiteRtTensorBufferLockModeRead
                   : kLiteRtTensorBufferLockModeWrite;
      if (auto status =
              LiteRtLockTensorBuffer(buffer, &host_mem_addr, lock_mode);
          status != kLiteRtStatusOk) {
        return Unexpected(
            status, absl::StrFormat("Failed to lock the tensor buffer: %s",
                                    tensor->name ? tensor->name : "<unnamed>"));
      }
      // For the following tflite custom allocation call, we explicitly use
      // buffer_size instead of tensor->bytes. This is because tensor->bytes
      // might be updated to the latest value if resize didn't trigger tensor
      // allocation.
      TfLiteCustomAllocation custom_allocation{host_mem_addr,
                                               buffer->buffer_size()};
      // If this is a constant output, save the locked address for later data
      // copying
      if (is_constant_output) {
        constant_outputs.push_back(
            {buffer, host_mem_addr, tensor_name, buffer->buffer_size()});
        LITERT_LOG(LITERT_INFO,
                   "Tracked constant output tensor %s with locked address",
                   tensor_name);
      }
      if (is_input) {
        runner->SetCustomAllocationForInputTensor(tensor_name,
                                                  custom_allocation,
                                                  /*flags=*/0);
        // TODO: b/419350199 - Ad-hoc solution to unlock input buffers.
        LITERT_RETURN_IF_ERROR(LiteRtUnlockTensorBuffer(buffer));
      } else {
        locked_buffers.push_back(buffer);

        // Skip SetCustomAllocationForOutputTensor for constant tensors
        // TFLite doesn't allow custom allocation for read-only memory-mapped
        // tensors
        if (!is_constant_output) {
          runner->SetCustomAllocationForOutputTensor(tensor_name,
                                                     custom_allocation,
                                                     /*flags=*/0);
        }
      }
      return {};
    }
  }

  // If the tensor is shared with CPU, register tensor buffer as is and let
  // accelerator handle the conversion.
  LITERT_ASSIGN_OR_RETURN(const auto tensor_id,
                          GetTensorIdentifier(*interp_, tensor));
  if (cpu_tensors_.find(tensor_id) != cpu_tensors_.end()) {
    void* host_mem_addr;
    if (auto status = LiteRtLockTensorBuffer(
            buffer, &host_mem_addr, kLiteRtTensorBufferLockModeReadWrite);
        status != kLiteRtStatusOk) {
      return Unexpected(
          status, absl::StrFormat("Failed to lock the tensor buffer: %s",
                                  tensor->name ? tensor->name : "<unnamed>"));
    }
    // If this is a constant output, save the locked address for later data
    // copying
    if (is_constant_output) {
      constant_outputs.push_back(
          {buffer, host_mem_addr, tensor_name, buffer->buffer_size()});
      LITERT_LOG(LITERT_INFO,
                 "Tracked CPU constant output tensor %s with locked address",
                 tensor_name);
    }
    TfLiteCustomAllocation custom_allocation{host_mem_addr,
                                             buffer->buffer_size()};
    if (is_input) {
      runner->SetCustomAllocationForInputTensor(tensor_name, custom_allocation,
                                                /*flags=*/0);
    } else {
      // Skip SetCustomAllocationForOutputTensor for constant tensors
      if (!is_constant_output) {
        runner->SetCustomAllocationForOutputTensor(tensor_name,
                                                   custom_allocation,
                                                   /*flags=*/0);
      }
    }
    return {};
  }
  // TODO: b/382330322 - Add buffer conversion logic instead of returning error.
  return Unexpected(kLiteRtStatusErrorRuntimeFailure,
                    "The given buffer type is not supported.");
}

Expected<void> LiteRtCompiledModelT::SetSchedulingInfo(
    const LiteRtSchedulingInfo* scheduling_info) {
  if (scheduling_info == nullptr) {
    model_debug_feature_id_.clear();
    model_scheduling_info_ = LiteRtSchedulingInfo{};
    return {};
  }

  model_scheduling_info_ = *scheduling_info;
  model_scheduling_info_.fields_mask &= kAllSchedulingInfoFields;
  LITERT_RETURN_IF_ERROR(ValidateSchedulingInfo(model_scheduling_info_));
  if (model_scheduling_info_.fields_mask &
      kLiteRtSchedulingInfoFieldDebugFeatureId) {
    model_debug_feature_id_ = model_scheduling_info_.debug_feature_id;
    model_scheduling_info_.debug_feature_id = model_debug_feature_id_.c_str();
  } else {
    model_debug_feature_id_.clear();
    model_scheduling_info_.debug_feature_id = nullptr;
  }
  return {};
}

Expected<void> LiteRtCompiledModelT::Run(
    absl::string_view signature_key,
    const std::vector<LiteRtTensorBuffer>& input_buffers,
    const std::vector<LiteRtTensorBuffer>& output_buffers, bool& async,
    LiteRtOptions run_options, const LiteRtSchedulingInfo* scheduling_info) {
  uint64_t event_handle = std::numeric_limits<uint64_t>::max();
  if (profiler_ && profiler_->IsProfiling()) {
    profiler_->SetCurrentEventSource(LITERT);
    event_handle =
        profiler_->BeginEvent("LiteRT::Run[buffer registration]",
                              tflite::Profiler::EventType::DEFAULT, 0, 0);
  }
  auto runner = GetSignatureRunner(signature_key);
  if (runner == nullptr) {
    return Unexpected(kLiteRtStatusErrorNotFound,
                      "Failed to get signature runner");
  }
  size_t num_inputs = input_buffers.size();
  if (num_inputs != runner->subgraph_input_names().size()) {
    std::string error_message = absl::StrCat(
        "Input buffer size mismatch: number of inputs:",
        runner->subgraph_input_names().size(), " vs buffers:", num_inputs);

    return Unexpected(kLiteRtStatusErrorRuntimeFailure, error_message);
  }
  size_t num_outputs = output_buffers.size();
  if (num_outputs != runner->subgraph_output_names().size()) {
    return Unexpected(kLiteRtStatusErrorRuntimeFailure,
                      "Output buffer size mismatch");
  }

  // In general output buffer events are assigned by the runtime and not the
  // caller; here we check for any violation of that condition.
  for (auto litert_output_buffer : output_buffers) {
    if (litert_output_buffer->HasEvent()) {
      return Error(kLiteRtStatusErrorInvalidArgument,
                   "Output buffers cannot have events attached");
    }
  }

  // The collection of locked buffers. It is used to unlock the buffers after
  // the inference is done.
  std::vector<LiteRtTensorBuffer> locked_buffers;
  locked_buffers.reserve(num_inputs + num_outputs);
  // Vector to track only constant output tensors.
  std::vector<ConstantOutputInfo> constant_outputs;
  auto unlock_buffers = absl::MakeCleanup([&locked_buffers]() {
    for (auto locked_buffer : locked_buffers) {
      if (LiteRtUnlockTensorBuffer(locked_buffer) != kLiteRtStatusOk) {
        LITERT_LOG(LITERT_ERROR, "Failed to unlock buffer %p", locked_buffer);
        ABSL_DCHECK(false);
      }
    }
  });
  for (int i = 0; i < num_inputs; ++i) {
    const auto& input_name = runner->subgraph_input_names()[i];
    auto* input_tensor = runner->input_tensor(input_name);
    if (input_buffers[i] == nullptr) {
      // skip if the input buffer is set to nullptr, indicating the input has
      // been bound to an external buffer.
      continue;
    }
    auto res =
        RegisterBuffer(runner, input_tensor, input_name, input_buffers[i],
                       /*is_input=*/true, locked_buffers, constant_outputs);

    if (!res) {
      return Unexpected(kLiteRtStatusErrorRuntimeFailure,
                        absl::StrCat("Failed to register input tensor buffer: ",
                                     res.Error().Message()));
    }
  }

  for (int i = 0; i < runner->subgraph_output_names().size(); ++i) {
    const auto& output_name = runner->subgraph_output_names()[i];
    auto* output_tensor = runner->output_tensor(output_name);
    auto res =
        RegisterBuffer(runner, const_cast<TfLiteTensor*>(output_tensor),
                       output_name, output_buffers[i],
                       /*is_input=*/false, locked_buffers, constant_outputs);

    if (!res) {
      return Unexpected(
          kLiteRtStatusErrorRuntimeFailure,
          absl::StrCat("Failed to register output tensor buffer: ",
                       res.Error().Message()));
    }
  }
  if (profiler_ && profiler_->IsProfiling() &&
      event_handle != std::numeric_limits<uint64_t>::max()) {
    profiler_->SetCurrentEventSource(LITERT);
    profiler_->EndEvent(event_handle);
  }

  if (auto res = runner->AllocateTensors(); res != kTfLiteOk) {
    if (error_reporter_) {
      error_reporter_->Report("Failed to allocate tensors for execution");
    }
    return Unexpected(kLiteRtStatusErrorRuntimeFailure,
                      "Failed to allocate tensors");
  }
  LITERT_RETURN_IF_ERROR(MarkSignatureAllocationUpToDate(runner));

  // Relay scheduling information to DelegateKernel of Accelerator. This is
  // used by vendor delegates (e.g. NPU dispatch implementations) to pass
  // scheduling hints/metadata down to the hardware runtime.
  LiteRtSchedulingInfo effective_scheduling_info = GetDefaultSchedulingInfo();
  ApplySchedulingInfoOverrides(model_scheduling_info_,
                               effective_scheduling_info);
  if (scheduling_info != nullptr) {
    LiteRtSchedulingInfo request_scheduling_info = *scheduling_info;
    request_scheduling_info.fields_mask &= kAllSchedulingInfoFields;
    LITERT_RETURN_IF_ERROR(ValidateSchedulingInfo(request_scheduling_info));
    ApplySchedulingInfoOverrides(request_scheduling_info,
                                 effective_scheduling_info);
  }
  buffer_context_->SetCurrentSchedulingInfo(effective_scheduling_info);
  auto clear_scheduling_info = absl::MakeCleanup(
      [this]() { buffer_context_->ClearCurrentSchedulingInfo(); });

  // Relay the intended async execution mode to DelegateKernel of Accelerator.
  buffer_context_->SetAsyncExecutionMode(async);
  const LiteRtOptions previous_run_options = buffer_context_->GetRunOptions();
  buffer_context_->SetRunOptions(run_options);
  absl::Cleanup restore_run_options = [this, previous_run_options]() {
    if (buffer_context_ != nullptr) {
      buffer_context_->SetRunOptions(previous_run_options);
    }
  };

  if (auto res = runner->Invoke(); res != kTfLiteOk) {
    if (res == kTfLiteCancelled) {
      return Unexpected(kLiteRtStatusCancelled, "Execution was cancelled");
    }
    return Unexpected(kLiteRtStatusErrorRuntimeFailure, "Failed to invoke");
  }
  // Copy constant data to constant output tensors after invoke
  // This only iterates through constant outputs that were identified during
  // RegisterBuffer
  for (const auto& constant_output : constant_outputs) {
    // Get the constant tensor to access its data
    auto* output_tensor = runner->output_tensor(constant_output.tensor_name);
    if (output_tensor && output_tensor->data.raw != nullptr) {
      const void* const_data_ptr = output_tensor->data.raw;
      if (constant_output.locked_address != nullptr) {
        memcpy(constant_output.locked_address, const_data_ptr,
               constant_output.data_size);
      } else {
        LITERT_LOG(LITERT_WARNING,
                   "Failed to obtain CPU view for constant output tensor %s",
                   constant_output.tensor_name);
      }
    }
  }

  if (profiler_ && profiler_->IsProfiling()) {
    profiler_->SetCurrentEventSource(LITERT);
    event_handle = profiler_->BeginEvent(
        "LiteRT::Run[Buffer sync]", tflite::Profiler::EventType::DEFAULT, 0, 0);
  }

  if (async) {
    // If the caller requested async execution, then set async to true if any
    // of the output buffers have been assigned a synchronization event.
    async = false;
    for (auto& tb : output_buffers) {
      async |= tb->HasEvent();
    }
  } else {
    // If the caller has not requested async execution, then wait on
    // synchronization events that have been attached to the outputs.
    for (auto& tb : output_buffers) {
      if (tb->HasEvent()) {
        LITERT_ASSIGN_OR_RETURN(LiteRtEventT * event, tb->GetEvent());
        LITERT_RETURN_IF_ERROR(event->Wait(/*timeout_in_ms=*/-1));
      }
    }
  }
  if (profiler_ && profiler_->IsProfiling() &&
      event_handle != std::numeric_limits<uint64_t>::max()) {
    profiler_->SetCurrentEventSource(LITERT);
    profiler_->EndEvent(event_handle);
  }

  return {};
}

Expected<void> LiteRtCompiledModelT::RunCApi(
    size_t signature_index, size_t num_input_buffers,
    const LiteRtTensorBuffer* input_buffers, size_t num_output_buffers,
    const LiteRtTensorBuffer* output_buffers, bool* async,
    LiteRtOptions run_options) {
  return RunCApi(signature_index, num_input_buffers, input_buffers,
                 num_output_buffers, output_buffers, async, run_options,
                 /*scheduling_info=*/nullptr);
}

Expected<void> LiteRtCompiledModelT::RunCApi(
    size_t signature_index, size_t num_input_buffers,
    const LiteRtTensorBuffer* input_buffers, size_t num_output_buffers,
    const LiteRtTensorBuffer* output_buffers, bool* async,
    const LiteRtSchedulingInfo* scheduling_info) {
  return RunCApi(signature_index, num_input_buffers, input_buffers,
                 num_output_buffers, output_buffers, async,
                 /*run_options=*/nullptr, scheduling_info);
}

Expected<void> LiteRtCompiledModelT::RunCApi(
    size_t signature_index, size_t num_input_buffers,
    const LiteRtTensorBuffer* input_buffers, size_t num_output_buffers,
    const LiteRtTensorBuffer* output_buffers, bool* async,
    LiteRtOptions run_options, const LiteRtSchedulingInfo* scheduling_info) {
  if (signature_index >= signature_keys_.size()) {
    return Unexpected(kLiteRtStatusErrorIndexOOB,
                      "Signature index is out of range of signature keys");
  }
  std::vector<LiteRtTensorBuffer> input_buffers_vec;
  input_buffers_vec.reserve(num_input_buffers);
  for (int i = 0; i < num_input_buffers; ++i) {
    input_buffers_vec.push_back(input_buffers[i]);
  }
  std::vector<LiteRtTensorBuffer> output_buffers_vec;
  output_buffers_vec.reserve(num_output_buffers);
  for (int i = 0; i < num_output_buffers; ++i) {
    output_buffers_vec.push_back(output_buffers[i]);
  }
  bool async_ = async ? *async : false;
  auto result = Run(*signature_keys_[signature_index], input_buffers_vec,
                    output_buffers_vec, async_, run_options, scheduling_info);
  if (async) {
    *async = async_;
  }
  return result;
}

Expected<void> LiteRtCompiledModelT::StartMetricsCollection(int detail_level) const {
  if (detail_level < 0) {
    return Unexpected(kLiteRtStatusErrorInvalidArgument,
                      "Detail level must be >= 0");
  }
  for (auto& delegate : delegates_) {
    if (delegate.StartMetricsCollection) {
      LITERT_RETURN_IF_ERROR(delegate.StartMetricsCollection(
          LrtGetRuntimeContext(), delegate.delegate.get(), detail_level));
    }
  }
  return {};
}

Expected<LiteRtMetricsT> LiteRtCompiledModelT::StopMetricsCollection() const {
  std::vector<LiteRtMetricsT::Metric> metrics;
  for (auto& delegate : delegates_) {
    if (delegate.StopMetricsCollection) {
      LiteRtMetricsT accelerator_metrics;
      LITERT_RETURN_IF_ERROR(delegate.StopMetricsCollection(
          LrtGetRuntimeContext(), delegate.delegate.get(),
          &accelerator_metrics));
      metrics.insert(
          metrics.end(),
          std::make_move_iterator(accelerator_metrics.metrics.begin()),
          std::make_move_iterator(accelerator_metrics.metrics.end()));
    }
  }
  return LiteRtMetricsT{.metrics = std::move(metrics)};
}

// Three cases are handled in this function:
// 1. The input tensor has the same shape as the new shape, then no resize is
// needed.
// 2. The input tensor has dynamic dimensions and the new shape is compatible,
// then resize is needed.
// 3. The input tensor has static dimensions or the new shape is not compatible,
// then return error.
Expected<bool> LiteRtCompiledModelT::InputTensorNeedsResize(
    const TfLiteTensor* tensor, absl::Span<const int> new_shape) {
  const TfLiteIntArray* shape_array = tensor->dims;
  if (!shape_array || shape_array->size == 0 || new_shape.empty()) {
    return false;
  }

  // Get current tensor shape.
  absl::Span<const int> current_shape =
      absl::MakeConstSpan(shape_array->data, shape_array->size);

  // Check if shapes are already the same.
  if (current_shape == new_shape) {
    return false;
  }

  if (!tensor->dims_signature || tensor->dims_signature->size == 0) {
    return Unexpected(kLiteRtStatusErrorInvalidArgument,
                      absl::StrCat("Cannot auto-resize tensor ",
                                   tensor->name ? tensor->name : "<unnamed>",
                                   ": no dims_signature exists"));
  }
  // Validate that the tensor has dynamic dimensions (contains -1).
  absl::Span<const int> signature_shape = absl::MakeConstSpan(
      tensor->dims_signature->data, tensor->dims_signature->size);

  LITERT_RETURN_IF_ERROR(
      std::find(signature_shape.begin(), signature_shape.end(), -1) !=
          signature_shape.end(),
      Unexpected(kLiteRtStatusErrorInvalidArgument,
                 absl::StrCat("Cannot auto-resize tensor ",
                              tensor->name ? tensor->name : "<unnamed>",
                              ": no dynamic dimensions found")));

  // Validate that new shape is compatible with tensor structure.
  LITERT_RETURN_IF_ERROR(
      signature_shape.size() == new_shape.size(),
      Unexpected(
          kLiteRtStatusErrorInvalidArgument,
          absl::StrCat("Cannot auto-resize tensor ",
                       tensor->name ? tensor->name : "<unnamed>",
                       ": rank mismatch (current: ", signature_shape.size(),
                       ", new: ", new_shape.size(), ")")));

  // Check that static dimensions match and dynamic dimensions are reasonable.
  for (size_t i = 0; i < signature_shape.size(); ++i) {
    if (signature_shape[i] != -1) {
      // Static dim ⇒ must be identical.
      LITERT_RETURN_IF_ERROR(
          signature_shape[i] == new_shape[i],
          Unexpected(kLiteRtStatusErrorInvalidArgument,
                     absl::StrCat("Cannot auto-resize tensor ",
                                  tensor->name ? tensor->name : "<unnamed>",
                                  ": static dimension mismatch at index ", i,
                                  " (current: ", signature_shape[i],
                                  ", new: ", new_shape[i], ")")));
    } else {
      // Dynamic dim ⇒ new value must be positive.
      LITERT_RETURN_IF_ERROR(
          new_shape[i] > 0,
          Unexpected(kLiteRtStatusErrorInvalidArgument,
                     absl::StrCat("Cannot auto-resize tensor ",
                                  tensor->name ? tensor->name : "<unnamed>",
                                  ": invalid dimension size ", new_shape[i],
                                  " at index ", i)));
    }
  }

  LITERT_LOG(LITERT_INFO,
             "Detected shape change for tensor %s - validation passed",
             tensor->name ? tensor->name : "<unnamed>");

  return true;
}

Expected<void> LiteRtCompiledModelT::ResizeInputTensor(
    size_t signature_index, size_t input_index, absl::Span<const int> dims) {
  return ResizeInputTensorImpl(signature_index, input_index, dims,
                               /*strict_mode=*/true);
}

Expected<void> LiteRtCompiledModelT::ResizeInputTensorNonStrict(
    size_t signature_index, size_t input_index, absl::Span<const int> dims) {
  return ResizeInputTensorImpl(signature_index, input_index, dims,
                               /*strict_mode=*/false);
}

Expected<void> LiteRtCompiledModelT::ResizeInputTensorImpl(
    size_t signature_index, size_t input_index, absl::Span<const int> dims,
    bool strict_mode) {
  if (signature_index >= signature_keys_.size()) {
    return Unexpected(kLiteRtStatusErrorIndexOOB,
                      "Signature index is out of range of signature keys");
  }

  auto* runner = GetSignatureRunner(*signature_keys_[signature_index]);
  if (runner == nullptr) {
    return Unexpected(kLiteRtStatusErrorInvalidArgument,
                      "Failed to get signature runner");
  }

  const auto& input_names = runner->subgraph_input_names();
  if (input_index >= input_names.size()) {
    return Unexpected(kLiteRtStatusErrorIndexOOB, "Input index out of range");
  }

  const auto& input_name = input_names[input_index];
  auto* input_tensor = runner->input_tensor(input_name);
  if (input_tensor == nullptr) {
    return Unexpected(kLiteRtStatusErrorNotFound, "Failed to get input tensor");
  }

  // Get current tensor shape.
  if (dims.empty()) {
    return Unexpected(kLiteRtStatusErrorInvalidArgument,
                      "New shape must not be empty.");
  }
  for (int dim : dims) {
    if (dim <= 0) {
      return Unexpected(kLiteRtStatusErrorInvalidArgument,
                        "Dimensions must be positive.");
    }
  }

  // Fast path: nothing to do if the runtime shape already matches the request.
  const TfLiteIntArray* runtime_shape = input_tensor->dims;
  if (runtime_shape && runtime_shape->size == static_cast<int>(dims.size())) {
    bool identical = true;
    for (size_t i = 0; i < dims.size(); ++i) {
      if (runtime_shape->data[i] != dims[i]) {
        identical = false;
        break;
      }
    }
    if (identical) {
      return {};
    }
  }

  const TfLiteIntArray* signature_shape =
      (input_tensor->dims_signature && input_tensor->dims_signature->size > 0)
          ? input_tensor->dims_signature
          : input_tensor->dims;

  if (!signature_shape) {
    return Unexpected(kLiteRtStatusErrorInvalidArgument,
                      "Failed to get current shape.");
  }

  if (strict_mode) {
    if (signature_shape->size != dims.size()) {
      return Unexpected(kLiteRtStatusErrorInvalidArgument,
                        "New shape rank does not match current shape rank.");
    }

    bool has_dynamic_shape = false;
    for (size_t i = 0; i < dims.size(); ++i) {
      const int signature_dim = signature_shape->data[i];
      if (signature_dim == -1) {
        has_dynamic_shape = true;
      } else if (signature_dim != dims[i]) {
        return Unexpected(kLiteRtStatusErrorInvalidArgument,
                          "New shape is not compatible with current shape.");
      }
    }
    if (!has_dynamic_shape) {
      return Unexpected(kLiteRtStatusErrorInvalidArgument,
                        "Tensor does not have a dynamic shape.");
    }
  }

  // Resize the input tensor using TFLite's SignatureRunner API
  const auto status = runner->ResizeInputTensor(
      input_name, std::vector(dims.begin(), dims.end()));
  if (status != kTfLiteOk) {
    return Unexpected(kLiteRtStatusErrorRuntimeFailure,
                      "Failed to resize input tensor");
  }

  cpu_buffer_requirements_.clear();
  return MarkSignatureNeedsAllocation(runner);
}

Expected<void> LiteRtCompiledModelT::MarkSignatureNeedsAllocation(
    const tflite::SignatureRunner* runner) {
  signature_needs_allocation_[runner] = true;
  return {};
}

Expected<void> LiteRtCompiledModelT::MarkSignatureAllocationUpToDate(
    const tflite::SignatureRunner* runner) {
  signature_needs_allocation_[runner] = false;
  return {};
}

Expected<bool> LiteRtCompiledModelT::SignatureNeedsAllocation(
    const tflite::SignatureRunner* runner) const {
  auto iter = signature_needs_allocation_.find(runner);
  if (iter == signature_needs_allocation_.end()) {
    return true;
  }
  return iter->second;
}

// Error reporter APIs implementation

void LiteRtCompiledModelT::ReportError(const char* format, ...) {
  if (!error_reporter_) {
    return;  // No error reporter configured
  }

  va_list args;
  va_start(args, format);
  error_reporter_->Report(format, args);
  va_end(args);
}

Expected<void> LiteRtCompiledModelT::ClearErrors() const {
  if (!error_reporter_) {
    return Unexpected(kLiteRtStatusErrorInvalidArgument,
                      "No error reporter configured");
  }

  auto* buffer_reporter =
      dynamic_cast<litert::BufferErrorReporter*>(error_reporter_.get());
  if (!buffer_reporter) {
    return Unexpected(
        kLiteRtStatusErrorUnsupported,
        "Clear errors is only available with buffer error reporter");
  }

  buffer_reporter->Clear();
  return {};
}

Expected<std::string> LiteRtCompiledModelT::GetErrorMessages() const {
  if (!error_reporter_) {
    return Unexpected(kLiteRtStatusErrorInvalidArgument,
                      "No error reporter configured");
  }

  auto* buffer_reporter =
      dynamic_cast<litert::BufferErrorReporter*>(error_reporter_.get());
  if (!buffer_reporter) {
    return Unexpected(
        kLiteRtStatusErrorUnsupported,
        "Get error messages is only available with buffer error reporter");
  }

  return buffer_reporter->message();
}

bool LiteRtCompiledModelT::CheckCancelledWrapper(void* data) {
  auto* model = static_cast<LiteRtCompiledModelT*>(data);
  if (model && model->check_cancelled_func_cpp_) {
    return model->check_cancelled_func_cpp_();
  }
  return false;
}

void LiteRtCompiledModelT::SetCancellationFunction(
    absl::AnyInvocable<bool()> check_cancelled_func) {
  check_cancelled_func_cpp_ = std::move(check_cancelled_func);
  check_cancelled_func_ = nullptr;
  interp_->SetCancellationFunction(this, &CheckCancelledWrapper);
}

void LiteRtCompiledModelT::SetCancellationFunction(
    void* data, bool (*check_cancelled_func)(void*)) {
  check_cancelled_func_ = check_cancelled_func;
  check_cancelled_func_cpp_ = nullptr;

  // Set the cancellation function on the underlying TFLite interpreter
  interp_->SetCancellationFunction(data, check_cancelled_func);
}

// -----------------------------------------------------------------------------
// Friend APIs
// -----------------------------------------------------------------------------

Expected<::tflite::Interpreter*> GetInterpreter(
    LiteRtCompiledModelT* compiled_model) {
  if (compiled_model == nullptr) {
    return Unexpected(kLiteRtStatusErrorInvalidArgument,
                      "Compiled model is null");
  }
  if (compiled_model->interp_ == nullptr) {
    return Unexpected(kLiteRtStatusErrorInvalidArgument, "Interpreter is null");
  }
  return compiled_model->interp_.get();
}

Expected<bool> InputTensorNeedsResize(LiteRtCompiledModelT* compiled_model,
                                      const TfLiteTensor* tensor,
                                      absl::Span<const int> new_shape) {
  return compiled_model->InputTensorNeedsResize(tensor, new_shape);
}