File size: 9,933 Bytes
5f923cd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
// Copyright 2025 The ODML Authors.
//
// 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.

#ifndef THIRD_PARTY_ODML_LITE_RT_LLM_EXECUTOR_EXECUTOR_SETTINGS_BASE_H_
#define THIRD_PARTY_ODML_LITE_RT_LLM_EXECUTOR_EXECUTOR_SETTINGS_BASE_H_

#include <iostream>
#include <memory>
#include <optional>
#include <string>
#include <utility>

#include "absl/status/status.h"  // from @com_google_absl
#include "absl/status/statusor.h"  // from @com_google_absl
#include "absl/strings/string_view.h"  // from @com_google_absl
#include "runtime/util/memory_mapped_file.h"
#include "runtime/util/scoped_file.h"

namespace litert::lm {

enum class Backend {
  // Unspecified backend.
  UNSPECIFIED,

  // CPU hand-written path backend.
  CPU_ARTISAN,

  // GPU hand-written path backend.
  GPU_ARTISAN,

  // CPU LiteRT backend.
  CPU,

  // GPU LiteRT backend.
  GPU,

  // Google Tensor Emission Graph backend.
  GOOGLE_TENSOR_ARTISAN,

  // NPU backend.
  NPU,
};
std::ostream& operator<<(std::ostream& os, const Backend& backend);
// Returns the backend enum from the string. Case-insensitive.
absl::StatusOr<Backend> GetBackendFromString(absl::string_view backend_str);
// Returns the string representation of the backend enum.
std::string GetBackendString(Backend backend);

enum class ActivationDataType {
  // Use float32 as the activation data type.
  FLOAT32,

  // Use float16 as the activation data type.
  FLOAT16,

  // Use int16 as the activation data type.
  INT16,

  // Use int8 as the activation data type.
  INT8,
};
std::ostream& operator<<(std::ostream& os,
                         const ActivationDataType& activation);

absl::StatusOr<ActivationDataType> GetActivationDataTypeFromString(
    const std::string& activation_data_type);

// Fake weights mode.
enum class FakeWeightsMode {
  // Don't use fake weights, read real weights from disk.
  FAKE_WEIGHTS_NONE,

  // Replace all weights with INT8 fakes.
  FAKE_WEIGHTS_8BITS_ALL_LAYERS,

  // Replace feedforward and embedding weights with INT4 fakes and replace
  // attention weights with INT8 fakes.
  FAKE_WEIGHTS_ATTN_8_FFN_4_EMB_4,
};
std::ostream& operator<<(std::ostream& os,
                         const FakeWeightsMode& fake_weights_mode);

enum class FileFormat {
  // .tflite file format.
  TFLITE,

  // .task file format.
  TASK,

  // .litert_lm file format.
  LITERT_LM,
};
std::ostream& operator<<(std::ostream& os, const FileFormat& file_format);

// Class to host the model assets, including base models and lora models.
class ModelAssets {
 public:
  static absl::StatusOr<ModelAssets> Create(
      std::shared_ptr<ScopedFile> model_file);
  static absl::StatusOr<ModelAssets> Create(absl::string_view model_path);
  static absl::StatusOr<ModelAssets> Create(
      std::shared_ptr<MemoryMappedFile> model_file);
  static absl::StatusOr<ModelAssets> Create(
      std::shared_ptr<MemoryMappedFile> model_file,
      absl::string_view model_path);

  // Convenience factory function to create a ModelAssets with both a model
  // path and file. Will use the scoped file if both are provided.
  static absl::StatusOr<ModelAssets> Create(
      std::shared_ptr<ScopedFile> model_file, absl::string_view model_path);

  bool HasScopedFile() const { return scoped_file_ != nullptr; }
  bool HasMemoryMappedFile() const { return memory_mapped_file_ != nullptr; }

  // Returns the model file if it was created with the respective variant,
  // otherwise returns an error.
  absl::StatusOr<absl::string_view> GetPath() const;
  absl::StatusOr<std::shared_ptr<ScopedFile>> GetScopedFile() const;
  absl::StatusOr<std::shared_ptr<MemoryMappedFile>> GetMemoryMappedFile() const;

  // Convenience method to get a read-only scoped file to the model file
  // regardless of whether this instance was created from a path or scoped file.
  absl::StatusOr<std::shared_ptr<ScopedFile>> GetOrCreateScopedFile() const;

  FakeWeightsMode fake_weights_mode() const { return fake_weights_mode_; }

  void SetFakeWeightsMode(FakeWeightsMode fake_weights_mode) {
    fake_weights_mode_ = fake_weights_mode;
  }

 private:
  explicit ModelAssets(std::shared_ptr<ScopedFile> model_file,
                       absl::string_view model_path);
  explicit ModelAssets(absl::string_view model_path);
  explicit ModelAssets(std::shared_ptr<MemoryMappedFile> model_file);
  explicit ModelAssets(std::shared_ptr<MemoryMappedFile> model_file,
                       absl::string_view model_path);

  // TODO: b/417814685 - Consider supporting multiple model files if the need
  // case arises.
  std::string path_;
  std::shared_ptr<ScopedFile> scoped_file_;
  std::shared_ptr<MemoryMappedFile> memory_mapped_file_;

  FakeWeightsMode fake_weights_mode_ = FakeWeightsMode::FAKE_WEIGHTS_NONE;
};
std::ostream& operator<<(std::ostream& os, const ModelAssets& model_assets);

// Base Settings for the executor modules.
class ExecutorSettingsBase {
 public:
  virtual ~ExecutorSettingsBase() = default;

  // Getter APIs.
  const ModelAssets& GetModelAssets() const { return model_assets_; }

  // Backend APIs.
  const Backend& GetBackend() const { return backend_; }
  virtual absl::Status SetBackend(const Backend& backend) {
    backend_ = backend;
    return absl::OkStatus();
  }

  // Activation data type APIs.
  const std::optional<ActivationDataType>& GetActivationDataType() const {
    return activation_data_type_;
  }
  void SetActivationDataType(const ActivationDataType& activation_data_type) {
    activation_data_type_ = activation_data_type;
  }

  // Should be used by consumers who want to write to a single weight cache
  // file. Returns, in order of preference:
  //   1. an open file descriptor to the weight cache file,
  //   2. the file path of the weight cache file, based on the given cache
  //      directory and/or model path. Will append `suffix`.
  //   3. an error if a weight cache file could not be determined.
  absl::StatusOr<
      std::variant<std::string, std::shared_ptr<litert::lm::ScopedFile>>>
  GetWeightCacheFile(absl::string_view suffix = ".cache") const;
  // Prefer to use `GetWeightCacheFile()` if possible.
  const std::string& GetCacheDir() const { return cache_dir_; }
  // Prefer to use `GetWeightCacheFile()` if possible.
  std::shared_ptr<litert::lm::ScopedFile> GetScopedCacheFile() const {
    return scoped_cache_file_;
  }
  const std::string& GetLitertDispatchLibDir() const {
    return litert_dispatch_lib_dir_;
  }

  // Should be used by consumers who want to write to a single program cache
  // file. Returns, in order of preference:
  //   1. an open file descriptor to the program cache file,
  //   2. the file path of the program cache file, based on the given cache
  //      directory and/or model path. Will append `suffix`.
  //   3. an error if a program cache file could not be determined.
  absl::StatusOr<
      std::variant<std::string, std::shared_ptr<litert::lm::ScopedFile>>>
  GetProgramCacheFile(absl::string_view suffix = ".program_cache") const;
  // Prefer to use `GetProgramCacheFile()` if possible.
  std::shared_ptr<litert::lm::ScopedFile> GetScopedProgramCacheFile() const {
    return scoped_program_cache_file_;
  }

  // Setter APIs.
  void SetCacheDir(const std::string& cache_dir) { cache_dir_ = cache_dir; }
  void SetScopedCacheFile(std::shared_ptr<litert::lm::ScopedFile> cache_file) {
    scoped_cache_file_ = std::move(cache_file);
  }
  void SetLitertDispatchLibDir(const std::string& litert_dispatch_lib_dir) {
    litert_dispatch_lib_dir_ = litert_dispatch_lib_dir;
  }

  void SetScopedProgramCacheFile(
      std::shared_ptr<litert::lm::ScopedFile> cache_file) {
    scoped_program_cache_file_ = std::move(cache_file);
  }

 protected:
  explicit ExecutorSettingsBase(ModelAssets model_assets)
      : model_assets_(std::move(model_assets)) {}
  // Optional setting to use LLM executor backend.
  Backend backend_ = Backend::CPU;

 private:
  // Path to the LiteRT model file.
  ModelAssets model_assets_;

  // Directory for saving the weight cache file. If this is set and the
  // backend supports it, the re-arranged weights will be stored in the
  // directory after the 1st initialization, making the future initialization
  // to be much faster.
  //
  // Consumers should prefer to use the `cache_file_` if set.
  std::string cache_dir_;

  // Open file for writing the weight cache to and later loading cache from.
  // If set, this should be preferred over the `cache_dir_`.
  std::shared_ptr<litert::lm::ScopedFile> scoped_cache_file_;

  // Open file for writing the program cache to and later loading cache from.
  // If set, this should be preferred over the `cache_dir_`.
  std::shared_ptr<litert::lm::ScopedFile> scoped_program_cache_file_;

  // Optional setting for specific activation data type. If not set, the
  // default activation data type for each OS & backend will be used. Setting
  // this field will override the default activation data type, for example,
  // OpenCL backend only support fp32 on Linux.
  std::optional<ActivationDataType> activation_data_type_;

  // Optional LoRA model assets.
  std::optional<ModelAssets> lora_model_assets_;

  // LiteRT dispatch library directory. If not set, the runtime will look for
  // the library in the path defined as the environment variables.
  std::string litert_dispatch_lib_dir_;
};

}  // namespace litert::lm

#endif  // THIRD_PARTY_ODML_LITE_RT_LLM_EXECUTOR_LLM_EXECUTOR_SETTINGS_H_