keras-hub-nightly 0.15.0.dev20240823171555__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (297) hide show
  1. keras_hub/__init__.py +52 -0
  2. keras_hub/api/__init__.py +27 -0
  3. keras_hub/api/layers/__init__.py +47 -0
  4. keras_hub/api/metrics/__init__.py +24 -0
  5. keras_hub/api/models/__init__.py +249 -0
  6. keras_hub/api/samplers/__init__.py +29 -0
  7. keras_hub/api/tokenizers/__init__.py +35 -0
  8. keras_hub/src/__init__.py +13 -0
  9. keras_hub/src/api_export.py +53 -0
  10. keras_hub/src/layers/__init__.py +13 -0
  11. keras_hub/src/layers/modeling/__init__.py +13 -0
  12. keras_hub/src/layers/modeling/alibi_bias.py +143 -0
  13. keras_hub/src/layers/modeling/cached_multi_head_attention.py +137 -0
  14. keras_hub/src/layers/modeling/f_net_encoder.py +200 -0
  15. keras_hub/src/layers/modeling/masked_lm_head.py +239 -0
  16. keras_hub/src/layers/modeling/position_embedding.py +123 -0
  17. keras_hub/src/layers/modeling/reversible_embedding.py +311 -0
  18. keras_hub/src/layers/modeling/rotary_embedding.py +169 -0
  19. keras_hub/src/layers/modeling/sine_position_encoding.py +108 -0
  20. keras_hub/src/layers/modeling/token_and_position_embedding.py +150 -0
  21. keras_hub/src/layers/modeling/transformer_decoder.py +496 -0
  22. keras_hub/src/layers/modeling/transformer_encoder.py +262 -0
  23. keras_hub/src/layers/modeling/transformer_layer_utils.py +106 -0
  24. keras_hub/src/layers/preprocessing/__init__.py +13 -0
  25. keras_hub/src/layers/preprocessing/masked_lm_mask_generator.py +220 -0
  26. keras_hub/src/layers/preprocessing/multi_segment_packer.py +319 -0
  27. keras_hub/src/layers/preprocessing/preprocessing_layer.py +62 -0
  28. keras_hub/src/layers/preprocessing/random_deletion.py +271 -0
  29. keras_hub/src/layers/preprocessing/random_swap.py +267 -0
  30. keras_hub/src/layers/preprocessing/start_end_packer.py +219 -0
  31. keras_hub/src/metrics/__init__.py +13 -0
  32. keras_hub/src/metrics/bleu.py +394 -0
  33. keras_hub/src/metrics/edit_distance.py +197 -0
  34. keras_hub/src/metrics/perplexity.py +181 -0
  35. keras_hub/src/metrics/rouge_base.py +204 -0
  36. keras_hub/src/metrics/rouge_l.py +97 -0
  37. keras_hub/src/metrics/rouge_n.py +125 -0
  38. keras_hub/src/models/__init__.py +13 -0
  39. keras_hub/src/models/albert/__init__.py +20 -0
  40. keras_hub/src/models/albert/albert_backbone.py +267 -0
  41. keras_hub/src/models/albert/albert_classifier.py +202 -0
  42. keras_hub/src/models/albert/albert_masked_lm.py +129 -0
  43. keras_hub/src/models/albert/albert_masked_lm_preprocessor.py +194 -0
  44. keras_hub/src/models/albert/albert_preprocessor.py +206 -0
  45. keras_hub/src/models/albert/albert_presets.py +70 -0
  46. keras_hub/src/models/albert/albert_tokenizer.py +119 -0
  47. keras_hub/src/models/backbone.py +311 -0
  48. keras_hub/src/models/bart/__init__.py +20 -0
  49. keras_hub/src/models/bart/bart_backbone.py +261 -0
  50. keras_hub/src/models/bart/bart_preprocessor.py +276 -0
  51. keras_hub/src/models/bart/bart_presets.py +74 -0
  52. keras_hub/src/models/bart/bart_seq_2_seq_lm.py +490 -0
  53. keras_hub/src/models/bart/bart_seq_2_seq_lm_preprocessor.py +262 -0
  54. keras_hub/src/models/bart/bart_tokenizer.py +124 -0
  55. keras_hub/src/models/bert/__init__.py +23 -0
  56. keras_hub/src/models/bert/bert_backbone.py +227 -0
  57. keras_hub/src/models/bert/bert_classifier.py +183 -0
  58. keras_hub/src/models/bert/bert_masked_lm.py +131 -0
  59. keras_hub/src/models/bert/bert_masked_lm_preprocessor.py +198 -0
  60. keras_hub/src/models/bert/bert_preprocessor.py +184 -0
  61. keras_hub/src/models/bert/bert_presets.py +147 -0
  62. keras_hub/src/models/bert/bert_tokenizer.py +112 -0
  63. keras_hub/src/models/bloom/__init__.py +20 -0
  64. keras_hub/src/models/bloom/bloom_attention.py +186 -0
  65. keras_hub/src/models/bloom/bloom_backbone.py +173 -0
  66. keras_hub/src/models/bloom/bloom_causal_lm.py +298 -0
  67. keras_hub/src/models/bloom/bloom_causal_lm_preprocessor.py +176 -0
  68. keras_hub/src/models/bloom/bloom_decoder.py +206 -0
  69. keras_hub/src/models/bloom/bloom_preprocessor.py +185 -0
  70. keras_hub/src/models/bloom/bloom_presets.py +121 -0
  71. keras_hub/src/models/bloom/bloom_tokenizer.py +116 -0
  72. keras_hub/src/models/causal_lm.py +383 -0
  73. keras_hub/src/models/classifier.py +109 -0
  74. keras_hub/src/models/csp_darknet/__init__.py +13 -0
  75. keras_hub/src/models/csp_darknet/csp_darknet_backbone.py +410 -0
  76. keras_hub/src/models/csp_darknet/csp_darknet_image_classifier.py +133 -0
  77. keras_hub/src/models/deberta_v3/__init__.py +24 -0
  78. keras_hub/src/models/deberta_v3/deberta_v3_backbone.py +210 -0
  79. keras_hub/src/models/deberta_v3/deberta_v3_classifier.py +228 -0
  80. keras_hub/src/models/deberta_v3/deberta_v3_masked_lm.py +135 -0
  81. keras_hub/src/models/deberta_v3/deberta_v3_masked_lm_preprocessor.py +191 -0
  82. keras_hub/src/models/deberta_v3/deberta_v3_preprocessor.py +206 -0
  83. keras_hub/src/models/deberta_v3/deberta_v3_presets.py +82 -0
  84. keras_hub/src/models/deberta_v3/deberta_v3_tokenizer.py +155 -0
  85. keras_hub/src/models/deberta_v3/disentangled_attention_encoder.py +227 -0
  86. keras_hub/src/models/deberta_v3/disentangled_self_attention.py +412 -0
  87. keras_hub/src/models/deberta_v3/relative_embedding.py +94 -0
  88. keras_hub/src/models/densenet/__init__.py +13 -0
  89. keras_hub/src/models/densenet/densenet_backbone.py +210 -0
  90. keras_hub/src/models/densenet/densenet_image_classifier.py +131 -0
  91. keras_hub/src/models/distil_bert/__init__.py +26 -0
  92. keras_hub/src/models/distil_bert/distil_bert_backbone.py +187 -0
  93. keras_hub/src/models/distil_bert/distil_bert_classifier.py +208 -0
  94. keras_hub/src/models/distil_bert/distil_bert_masked_lm.py +137 -0
  95. keras_hub/src/models/distil_bert/distil_bert_masked_lm_preprocessor.py +194 -0
  96. keras_hub/src/models/distil_bert/distil_bert_preprocessor.py +175 -0
  97. keras_hub/src/models/distil_bert/distil_bert_presets.py +57 -0
  98. keras_hub/src/models/distil_bert/distil_bert_tokenizer.py +114 -0
  99. keras_hub/src/models/electra/__init__.py +20 -0
  100. keras_hub/src/models/electra/electra_backbone.py +247 -0
  101. keras_hub/src/models/electra/electra_preprocessor.py +154 -0
  102. keras_hub/src/models/electra/electra_presets.py +95 -0
  103. keras_hub/src/models/electra/electra_tokenizer.py +104 -0
  104. keras_hub/src/models/f_net/__init__.py +20 -0
  105. keras_hub/src/models/f_net/f_net_backbone.py +236 -0
  106. keras_hub/src/models/f_net/f_net_classifier.py +154 -0
  107. keras_hub/src/models/f_net/f_net_masked_lm.py +132 -0
  108. keras_hub/src/models/f_net/f_net_masked_lm_preprocessor.py +196 -0
  109. keras_hub/src/models/f_net/f_net_preprocessor.py +177 -0
  110. keras_hub/src/models/f_net/f_net_presets.py +43 -0
  111. keras_hub/src/models/f_net/f_net_tokenizer.py +95 -0
  112. keras_hub/src/models/falcon/__init__.py +20 -0
  113. keras_hub/src/models/falcon/falcon_attention.py +156 -0
  114. keras_hub/src/models/falcon/falcon_backbone.py +164 -0
  115. keras_hub/src/models/falcon/falcon_causal_lm.py +291 -0
  116. keras_hub/src/models/falcon/falcon_causal_lm_preprocessor.py +173 -0
  117. keras_hub/src/models/falcon/falcon_preprocessor.py +187 -0
  118. keras_hub/src/models/falcon/falcon_presets.py +30 -0
  119. keras_hub/src/models/falcon/falcon_tokenizer.py +110 -0
  120. keras_hub/src/models/falcon/falcon_transformer_decoder.py +255 -0
  121. keras_hub/src/models/feature_pyramid_backbone.py +73 -0
  122. keras_hub/src/models/gemma/__init__.py +20 -0
  123. keras_hub/src/models/gemma/gemma_attention.py +250 -0
  124. keras_hub/src/models/gemma/gemma_backbone.py +316 -0
  125. keras_hub/src/models/gemma/gemma_causal_lm.py +448 -0
  126. keras_hub/src/models/gemma/gemma_causal_lm_preprocessor.py +167 -0
  127. keras_hub/src/models/gemma/gemma_decoder_block.py +241 -0
  128. keras_hub/src/models/gemma/gemma_preprocessor.py +191 -0
  129. keras_hub/src/models/gemma/gemma_presets.py +248 -0
  130. keras_hub/src/models/gemma/gemma_tokenizer.py +103 -0
  131. keras_hub/src/models/gemma/rms_normalization.py +40 -0
  132. keras_hub/src/models/gpt2/__init__.py +20 -0
  133. keras_hub/src/models/gpt2/gpt2_backbone.py +199 -0
  134. keras_hub/src/models/gpt2/gpt2_causal_lm.py +437 -0
  135. keras_hub/src/models/gpt2/gpt2_causal_lm_preprocessor.py +173 -0
  136. keras_hub/src/models/gpt2/gpt2_preprocessor.py +187 -0
  137. keras_hub/src/models/gpt2/gpt2_presets.py +82 -0
  138. keras_hub/src/models/gpt2/gpt2_tokenizer.py +110 -0
  139. keras_hub/src/models/gpt_neo_x/__init__.py +13 -0
  140. keras_hub/src/models/gpt_neo_x/gpt_neo_x_attention.py +251 -0
  141. keras_hub/src/models/gpt_neo_x/gpt_neo_x_backbone.py +175 -0
  142. keras_hub/src/models/gpt_neo_x/gpt_neo_x_causal_lm.py +201 -0
  143. keras_hub/src/models/gpt_neo_x/gpt_neo_x_causal_lm_preprocessor.py +141 -0
  144. keras_hub/src/models/gpt_neo_x/gpt_neo_x_decoder.py +258 -0
  145. keras_hub/src/models/gpt_neo_x/gpt_neo_x_preprocessor.py +145 -0
  146. keras_hub/src/models/gpt_neo_x/gpt_neo_x_tokenizer.py +88 -0
  147. keras_hub/src/models/image_classifier.py +90 -0
  148. keras_hub/src/models/llama/__init__.py +20 -0
  149. keras_hub/src/models/llama/llama_attention.py +225 -0
  150. keras_hub/src/models/llama/llama_backbone.py +188 -0
  151. keras_hub/src/models/llama/llama_causal_lm.py +327 -0
  152. keras_hub/src/models/llama/llama_causal_lm_preprocessor.py +170 -0
  153. keras_hub/src/models/llama/llama_decoder.py +246 -0
  154. keras_hub/src/models/llama/llama_layernorm.py +48 -0
  155. keras_hub/src/models/llama/llama_preprocessor.py +189 -0
  156. keras_hub/src/models/llama/llama_presets.py +80 -0
  157. keras_hub/src/models/llama/llama_tokenizer.py +84 -0
  158. keras_hub/src/models/llama3/__init__.py +20 -0
  159. keras_hub/src/models/llama3/llama3_backbone.py +84 -0
  160. keras_hub/src/models/llama3/llama3_causal_lm.py +46 -0
  161. keras_hub/src/models/llama3/llama3_causal_lm_preprocessor.py +173 -0
  162. keras_hub/src/models/llama3/llama3_preprocessor.py +21 -0
  163. keras_hub/src/models/llama3/llama3_presets.py +69 -0
  164. keras_hub/src/models/llama3/llama3_tokenizer.py +63 -0
  165. keras_hub/src/models/masked_lm.py +101 -0
  166. keras_hub/src/models/mistral/__init__.py +20 -0
  167. keras_hub/src/models/mistral/mistral_attention.py +238 -0
  168. keras_hub/src/models/mistral/mistral_backbone.py +203 -0
  169. keras_hub/src/models/mistral/mistral_causal_lm.py +328 -0
  170. keras_hub/src/models/mistral/mistral_causal_lm_preprocessor.py +175 -0
  171. keras_hub/src/models/mistral/mistral_layer_norm.py +48 -0
  172. keras_hub/src/models/mistral/mistral_preprocessor.py +190 -0
  173. keras_hub/src/models/mistral/mistral_presets.py +48 -0
  174. keras_hub/src/models/mistral/mistral_tokenizer.py +82 -0
  175. keras_hub/src/models/mistral/mistral_transformer_decoder.py +265 -0
  176. keras_hub/src/models/mix_transformer/__init__.py +13 -0
  177. keras_hub/src/models/mix_transformer/mix_transformer_backbone.py +181 -0
  178. keras_hub/src/models/mix_transformer/mix_transformer_classifier.py +133 -0
  179. keras_hub/src/models/mix_transformer/mix_transformer_layers.py +300 -0
  180. keras_hub/src/models/opt/__init__.py +20 -0
  181. keras_hub/src/models/opt/opt_backbone.py +173 -0
  182. keras_hub/src/models/opt/opt_causal_lm.py +301 -0
  183. keras_hub/src/models/opt/opt_causal_lm_preprocessor.py +177 -0
  184. keras_hub/src/models/opt/opt_preprocessor.py +188 -0
  185. keras_hub/src/models/opt/opt_presets.py +72 -0
  186. keras_hub/src/models/opt/opt_tokenizer.py +116 -0
  187. keras_hub/src/models/pali_gemma/__init__.py +23 -0
  188. keras_hub/src/models/pali_gemma/pali_gemma_backbone.py +277 -0
  189. keras_hub/src/models/pali_gemma/pali_gemma_causal_lm.py +313 -0
  190. keras_hub/src/models/pali_gemma/pali_gemma_causal_lm_preprocessor.py +147 -0
  191. keras_hub/src/models/pali_gemma/pali_gemma_decoder_block.py +160 -0
  192. keras_hub/src/models/pali_gemma/pali_gemma_presets.py +78 -0
  193. keras_hub/src/models/pali_gemma/pali_gemma_tokenizer.py +79 -0
  194. keras_hub/src/models/pali_gemma/pali_gemma_vit.py +566 -0
  195. keras_hub/src/models/phi3/__init__.py +20 -0
  196. keras_hub/src/models/phi3/phi3_attention.py +260 -0
  197. keras_hub/src/models/phi3/phi3_backbone.py +224 -0
  198. keras_hub/src/models/phi3/phi3_causal_lm.py +218 -0
  199. keras_hub/src/models/phi3/phi3_causal_lm_preprocessor.py +173 -0
  200. keras_hub/src/models/phi3/phi3_decoder.py +260 -0
  201. keras_hub/src/models/phi3/phi3_layernorm.py +48 -0
  202. keras_hub/src/models/phi3/phi3_preprocessor.py +190 -0
  203. keras_hub/src/models/phi3/phi3_presets.py +50 -0
  204. keras_hub/src/models/phi3/phi3_rotary_embedding.py +137 -0
  205. keras_hub/src/models/phi3/phi3_tokenizer.py +94 -0
  206. keras_hub/src/models/preprocessor.py +207 -0
  207. keras_hub/src/models/resnet/__init__.py +13 -0
  208. keras_hub/src/models/resnet/resnet_backbone.py +612 -0
  209. keras_hub/src/models/resnet/resnet_image_classifier.py +136 -0
  210. keras_hub/src/models/roberta/__init__.py +20 -0
  211. keras_hub/src/models/roberta/roberta_backbone.py +184 -0
  212. keras_hub/src/models/roberta/roberta_classifier.py +209 -0
  213. keras_hub/src/models/roberta/roberta_masked_lm.py +136 -0
  214. keras_hub/src/models/roberta/roberta_masked_lm_preprocessor.py +198 -0
  215. keras_hub/src/models/roberta/roberta_preprocessor.py +192 -0
  216. keras_hub/src/models/roberta/roberta_presets.py +43 -0
  217. keras_hub/src/models/roberta/roberta_tokenizer.py +132 -0
  218. keras_hub/src/models/seq_2_seq_lm.py +54 -0
  219. keras_hub/src/models/t5/__init__.py +20 -0
  220. keras_hub/src/models/t5/t5_backbone.py +261 -0
  221. keras_hub/src/models/t5/t5_layer_norm.py +35 -0
  222. keras_hub/src/models/t5/t5_multi_head_attention.py +324 -0
  223. keras_hub/src/models/t5/t5_presets.py +95 -0
  224. keras_hub/src/models/t5/t5_tokenizer.py +100 -0
  225. keras_hub/src/models/t5/t5_transformer_layer.py +178 -0
  226. keras_hub/src/models/task.py +419 -0
  227. keras_hub/src/models/vgg/__init__.py +13 -0
  228. keras_hub/src/models/vgg/vgg_backbone.py +158 -0
  229. keras_hub/src/models/vgg/vgg_image_classifier.py +124 -0
  230. keras_hub/src/models/vit_det/__init__.py +13 -0
  231. keras_hub/src/models/vit_det/vit_det_backbone.py +204 -0
  232. keras_hub/src/models/vit_det/vit_layers.py +565 -0
  233. keras_hub/src/models/whisper/__init__.py +20 -0
  234. keras_hub/src/models/whisper/whisper_audio_feature_extractor.py +260 -0
  235. keras_hub/src/models/whisper/whisper_backbone.py +305 -0
  236. keras_hub/src/models/whisper/whisper_cached_multi_head_attention.py +153 -0
  237. keras_hub/src/models/whisper/whisper_decoder.py +141 -0
  238. keras_hub/src/models/whisper/whisper_encoder.py +106 -0
  239. keras_hub/src/models/whisper/whisper_preprocessor.py +326 -0
  240. keras_hub/src/models/whisper/whisper_presets.py +148 -0
  241. keras_hub/src/models/whisper/whisper_tokenizer.py +163 -0
  242. keras_hub/src/models/xlm_roberta/__init__.py +26 -0
  243. keras_hub/src/models/xlm_roberta/xlm_roberta_backbone.py +81 -0
  244. keras_hub/src/models/xlm_roberta/xlm_roberta_classifier.py +225 -0
  245. keras_hub/src/models/xlm_roberta/xlm_roberta_masked_lm.py +141 -0
  246. keras_hub/src/models/xlm_roberta/xlm_roberta_masked_lm_preprocessor.py +195 -0
  247. keras_hub/src/models/xlm_roberta/xlm_roberta_preprocessor.py +205 -0
  248. keras_hub/src/models/xlm_roberta/xlm_roberta_presets.py +43 -0
  249. keras_hub/src/models/xlm_roberta/xlm_roberta_tokenizer.py +191 -0
  250. keras_hub/src/models/xlnet/__init__.py +13 -0
  251. keras_hub/src/models/xlnet/relative_attention.py +459 -0
  252. keras_hub/src/models/xlnet/xlnet_backbone.py +222 -0
  253. keras_hub/src/models/xlnet/xlnet_content_and_query_embedding.py +133 -0
  254. keras_hub/src/models/xlnet/xlnet_encoder.py +378 -0
  255. keras_hub/src/samplers/__init__.py +13 -0
  256. keras_hub/src/samplers/beam_sampler.py +207 -0
  257. keras_hub/src/samplers/contrastive_sampler.py +231 -0
  258. keras_hub/src/samplers/greedy_sampler.py +50 -0
  259. keras_hub/src/samplers/random_sampler.py +77 -0
  260. keras_hub/src/samplers/sampler.py +237 -0
  261. keras_hub/src/samplers/serialization.py +97 -0
  262. keras_hub/src/samplers/top_k_sampler.py +92 -0
  263. keras_hub/src/samplers/top_p_sampler.py +113 -0
  264. keras_hub/src/tests/__init__.py +13 -0
  265. keras_hub/src/tests/test_case.py +608 -0
  266. keras_hub/src/tokenizers/__init__.py +13 -0
  267. keras_hub/src/tokenizers/byte_pair_tokenizer.py +638 -0
  268. keras_hub/src/tokenizers/byte_tokenizer.py +299 -0
  269. keras_hub/src/tokenizers/sentence_piece_tokenizer.py +267 -0
  270. keras_hub/src/tokenizers/sentence_piece_tokenizer_trainer.py +150 -0
  271. keras_hub/src/tokenizers/tokenizer.py +235 -0
  272. keras_hub/src/tokenizers/unicode_codepoint_tokenizer.py +355 -0
  273. keras_hub/src/tokenizers/word_piece_tokenizer.py +544 -0
  274. keras_hub/src/tokenizers/word_piece_tokenizer_trainer.py +176 -0
  275. keras_hub/src/utils/__init__.py +13 -0
  276. keras_hub/src/utils/keras_utils.py +130 -0
  277. keras_hub/src/utils/pipeline_model.py +293 -0
  278. keras_hub/src/utils/preset_utils.py +621 -0
  279. keras_hub/src/utils/python_utils.py +21 -0
  280. keras_hub/src/utils/tensor_utils.py +206 -0
  281. keras_hub/src/utils/timm/__init__.py +13 -0
  282. keras_hub/src/utils/timm/convert.py +37 -0
  283. keras_hub/src/utils/timm/convert_resnet.py +171 -0
  284. keras_hub/src/utils/transformers/__init__.py +13 -0
  285. keras_hub/src/utils/transformers/convert.py +101 -0
  286. keras_hub/src/utils/transformers/convert_bert.py +173 -0
  287. keras_hub/src/utils/transformers/convert_distilbert.py +184 -0
  288. keras_hub/src/utils/transformers/convert_gemma.py +187 -0
  289. keras_hub/src/utils/transformers/convert_gpt2.py +186 -0
  290. keras_hub/src/utils/transformers/convert_llama3.py +136 -0
  291. keras_hub/src/utils/transformers/convert_pali_gemma.py +303 -0
  292. keras_hub/src/utils/transformers/safetensor_utils.py +97 -0
  293. keras_hub/src/version_utils.py +23 -0
  294. keras_hub_nightly-0.15.0.dev20240823171555.dist-info/METADATA +34 -0
  295. keras_hub_nightly-0.15.0.dev20240823171555.dist-info/RECORD +297 -0
  296. keras_hub_nightly-0.15.0.dev20240823171555.dist-info/WHEEL +5 -0
  297. keras_hub_nightly-0.15.0.dev20240823171555.dist-info/top_level.txt +1 -0
@@ -0,0 +1,298 @@
1
+ # Copyright 2024 The KerasHub Authors
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # https://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+
16
+ from keras import ops
17
+
18
+ from keras_hub.src.api_export import keras_hub_export
19
+ from keras_hub.src.models.bloom.bloom_backbone import BloomBackbone
20
+ from keras_hub.src.models.bloom.bloom_causal_lm_preprocessor import (
21
+ BloomCausalLMPreprocessor,
22
+ )
23
+ from keras_hub.src.models.causal_lm import CausalLM
24
+ from keras_hub.src.utils.tensor_utils import any_equal
25
+
26
+
27
+ @keras_hub_export("keras_hub.models.BloomCausalLM")
28
+ class BloomCausalLM(CausalLM):
29
+ """An end-to-end BLOOM model for causal language modeling.
30
+
31
+ A causal language model (LM) predicts the next token based on previous
32
+ tokens. This task setup can be used to train the model unsupervised on
33
+ plain text input, or to autoregressively generate plain text similar to
34
+ the data used for training. This task can be used for pre-training or
35
+ fine-tuning a BLOOM model, simply by calling `fit()`.
36
+
37
+ This model has a `generate()` method, which generates text based on a
38
+ prompt. The generation strategy used is controlled by an additional
39
+ `sampler` argument on `compile()`. You can recompile the model with
40
+ different `keras_hub.samplers` objects to control the generation. By
41
+ default, `"greedy"` sampling will be used.
42
+
43
+ This model can optionally be configured with a `preprocessor` layer, in
44
+ which case it will automatically apply preprocessing to string inputs during
45
+ `fit()`, `predict()`, `evaluate()` and `generate()`. This is done by default
46
+ when creating the model with `from_preset()`.
47
+
48
+ Args:
49
+ backbone: A `keras_hub.models.BloomBackbone` instance.
50
+ preprocessor: A `keras_hub.models.BloomCausalLMPreprocessor` or `None`.
51
+ If `None`, this model will not apply preprocessing, and inputs
52
+ should be preprocessed before calling the model.
53
+
54
+ Examples:
55
+
56
+ Use `generate()` to do text generation.
57
+ ```python
58
+ bloom_lm = keras_hub.models.BloomCausalLM.from_preset("bloom_560m_multi")
59
+ bloom_lm.generate("I want to say", max_length=30)
60
+
61
+ # Generate with batched prompts.
62
+ bloom_lm.generate(["This is a", "Where are you"], max_length=30)
63
+ ```
64
+
65
+ Compile the `generate()` function with a custom sampler.
66
+ ```python
67
+ bloom_lm = keras_hub.models.BloomCausalLM.from_preset("bloom_560m_multi")
68
+ bloom_lm.compile(sampler="top_k")
69
+ bloom_lm.generate("I want to say", max_length=30)
70
+
71
+ bloom_lm.compile(sampler=keras_hub.samplers.BeamSampler(num_beams=2))
72
+ bloom_lm.generate("I want to say", max_length=30)
73
+ ```
74
+
75
+ Use `generate()` without preprocessing.
76
+ ```python
77
+ prompt = {
78
+ # Token ids for "<s> Keras is".
79
+ "token_ids": np.array([[1, 46, 15762, 632, 3, 3, 3, 3, 3]] * 2),
80
+ # Use `"padding_mask"` to indicate values that should not be overridden.
81
+ "padding_mask": np.array([[1, 1, 1, 1, 0, 0, 0, 0, 0]] * 2),
82
+ }
83
+
84
+ bloom_lm = keras_hub.models.BloomCausalLM.from_preset(
85
+ "bloom_560m_multi",
86
+ preprocessor=None,
87
+ )
88
+ bloom_lm.generate(prompt)
89
+ ```
90
+
91
+ Call `fit()` on a single batch.
92
+ ```python
93
+ features = ["The quick brown fox jumped.", "I forgot my homework."]
94
+ bloom_lm = keras_hub.models.BloomCausalLM.from_preset("bloom_560m_multi")
95
+ bloom_lm.fit(x=features, batch_size=2)
96
+ ```
97
+
98
+ Call `fit()` without preprocessing.
99
+ ```python
100
+ x = {
101
+ # Token ids for "<bos> Keras is deep learning library<eos>"
102
+ "token_ids": np.array([[2, 214064, 603, 5271, 6044, 9581, 1, 0]] * 2),
103
+ "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 0]] * 2),
104
+ }
105
+ y = np.array([[214064, 603, 5271, 6044, 9581, 3, 0, 0]] * 2)
106
+ sw = np.array([[1, 1, 1, 1, 1, 1, 0, 0]] * 2)
107
+
108
+ bloom_lm = keras_hub.models.BloomCausalLM.from_preset(
109
+ "bloom_560m_multi",
110
+ preprocessor=None,
111
+ )
112
+ bloom_lm.fit(x=x, y=y, sample_weight=sw, batch_size=2)
113
+ ```
114
+
115
+ Custom backbone and vocabulary.
116
+ ```python
117
+ features = [
118
+ " airplane at airport",
119
+ " airplane airport",
120
+ ]
121
+ vocab = ["<unk>", "<s>", "</s>", "<pad>"]
122
+ vocab += ["!", "air", "Ġair", "plane", "Ġat", "port"]
123
+ vocab = dict([(token, i) for i, token in enumerate(vocab)])
124
+ merges = ["Ġ a", "Ġ t", "Ġ i", "Ġ b", "a i", "p l", "n e"]
125
+ merges += ["Ġa t", "p o", "r t", "Ġt h", "ai r", "pl a", "po rt"]
126
+ merges += ["Ġai r", "Ġa i", "pla ne"]
127
+ tokenizer = keras_hub.models.BloomTokenizer(vocabulary=vocab, merges=merges)
128
+ preprocessor = keras_hub.models.BloomCausalLMPreprocessor(
129
+ tokenizer=tokenizer,
130
+ sequence_length=128,
131
+ )
132
+ backbone = keras_hub.models.BloomBackbone(
133
+ vocabulary_size=tokenizer.vocabulary_size(),
134
+ num_layers=4,
135
+ num_heads=4,
136
+ hidden_dim=32,
137
+ intermediate_dim=128,
138
+ )
139
+ bloom_lm = keras_hub.models.BloomCausalLM(
140
+ backbone=backbone,
141
+ preprocessor=preprocessor,
142
+ )
143
+ bloom_lm.fit(x=features, batch_size=2)
144
+ ```
145
+ """
146
+
147
+ backbone_cls = BloomBackbone
148
+ preprocessor_cls = BloomCausalLMPreprocessor
149
+
150
+ def __init__(
151
+ self,
152
+ backbone,
153
+ preprocessor=None,
154
+ **kwargs,
155
+ ):
156
+ # === Layers ===
157
+ self.backbone = backbone
158
+ self.preprocessor = preprocessor
159
+
160
+ # === Functional Model ===
161
+ inputs = backbone.input
162
+ hidden_states = backbone(inputs)
163
+ outputs = backbone.token_embedding(hidden_states, reverse=True)
164
+ super().__init__(
165
+ inputs=inputs,
166
+ outputs=outputs,
167
+ **kwargs,
168
+ )
169
+
170
+ def call_with_cache(
171
+ self,
172
+ token_ids,
173
+ cache,
174
+ cache_update_index,
175
+ ):
176
+ """Forward pass of `BloomCausalLM` with cache.
177
+
178
+ `call_with_cache` adds an additional forward pass for the model for
179
+ autoregressive inference. Unlike calling the model directly, this method
180
+ allows caching previous key/value Tensors in multi-head attention layer,
181
+ and avoids recomputing the outputs of seen tokens.
182
+
183
+ Args:
184
+ token_ids: a dense int Tensor with shape `(batch_size, max_length)`.
185
+ cache: a dense float Tensor, the cache of key and value.
186
+ cache_update_index: int, or int Tensor. The index of current inputs
187
+ in the whole sequence.
188
+
189
+ Returns:
190
+ A (logits, hidden_states, cache) tuple. Where `logits` is the
191
+ language model logits for the input token_ids, `hidden_states` is
192
+ the final hidden representation of the input tokens, and `cache` is
193
+ the decoding cache.
194
+ """
195
+ x = self.backbone.token_embedding(token_ids)
196
+ x = self.backbone.embeddings_layer_norm(x)
197
+ # Each decoder layer has a cache; we update them separately.
198
+ caches = []
199
+ for i, transformer_layer in enumerate(self.backbone.transformer_layers):
200
+ current_cache = cache[:, i, ...]
201
+ x, next_cache = transformer_layer(
202
+ x,
203
+ cache=current_cache,
204
+ cache_update_index=cache_update_index,
205
+ )
206
+ caches.append(next_cache)
207
+ cache = ops.stack(caches, axis=1)
208
+ hidden_states = x = self.backbone.layer_norm(x)
209
+ logits = self.backbone.token_embedding(x, reverse=True)
210
+ return logits, hidden_states, cache
211
+
212
+ def _build_cache(self, token_ids):
213
+ """Build an empty cache for use with `call_with_cache()`."""
214
+ batch_size = ops.shape(token_ids)[0]
215
+ max_length = ops.shape(token_ids)[1]
216
+ num_layers = self.backbone.num_layers
217
+ num_heads = self.backbone.num_heads
218
+ head_dim = self.backbone.hidden_dim // num_heads
219
+ shape = [batch_size, num_layers, 2, max_length, num_heads, head_dim]
220
+ cache = ops.zeros(shape, dtype=self.compute_dtype)
221
+ # Seed the cache.
222
+ _, hidden_states, cache = self.call_with_cache(token_ids, cache, 0)
223
+ return hidden_states, cache
224
+
225
+ def generate_step(
226
+ self,
227
+ inputs,
228
+ stop_token_ids=None,
229
+ ):
230
+ """A compilable generation function for a single batch of inputs.
231
+
232
+ This function represents the inner, XLA-compilable, generation function
233
+ for a single batch of inputs. Inputs should have the same structure as
234
+ model inputs, a dictionary with keys `"token_ids"` and `"padding_mask"`.
235
+
236
+ Args:
237
+ inputs: A dictionary with two keys `"token_ids"` and
238
+ `"padding_mask"` and batched tensor values.
239
+ stop_token_ids: Tuple of id's of end token's to stop on. If all
240
+ sequences have produced a new stop token, generation
241
+ will stop.
242
+ """
243
+ token_ids, padding_mask = inputs["token_ids"], inputs["padding_mask"]
244
+ # Create and seed cache with a single forward pass.
245
+ hidden_states, cache = self._build_cache(token_ids)
246
+ # Compute the lengths of all user inputted tokens ids.
247
+ row_lengths = ops.sum(ops.cast(padding_mask, "int32"), axis=-1)
248
+ # Start at the first index that has no user inputted id.
249
+ index = ops.min(row_lengths)
250
+
251
+ def next(prompt, cache, index):
252
+ # The cache index is the index of our previous token.
253
+ cache_update_index = index - 1
254
+ batch_size = ops.shape(prompt)[0]
255
+ prompt = ops.slice(prompt, [0, cache_update_index], [batch_size, 1])
256
+ logits, hidden_states, cache = self.call_with_cache(
257
+ prompt,
258
+ cache,
259
+ cache_update_index,
260
+ )
261
+ return (
262
+ ops.squeeze(logits, axis=1),
263
+ ops.squeeze(hidden_states, axis=1),
264
+ cache,
265
+ )
266
+
267
+ token_ids = self.sampler(
268
+ next=next,
269
+ prompt=token_ids,
270
+ cache=cache,
271
+ index=index,
272
+ mask=padding_mask,
273
+ stop_token_ids=stop_token_ids,
274
+ hidden_states=hidden_states,
275
+ model=self,
276
+ )
277
+
278
+ # Compute an output padding mask with the token ids we updated.
279
+ if stop_token_ids is not None:
280
+ # Build a mask of stop token locations not in the original
281
+ # prompt (not in locations where `padding_mask` is True).
282
+ end_locations = any_equal(
283
+ token_ids, stop_token_ids, ops.logical_not(padding_mask)
284
+ )
285
+
286
+ end_locations = ops.cast(end_locations, "int32")
287
+ # Use cumsum to get ones in all locations after end_locations.
288
+ cumsum = ops.cast(ops.cumsum(end_locations, axis=-1), "int32")
289
+ overflow = cumsum - end_locations
290
+ # Our padding mask is the inverse of these overflow locations.
291
+ padding_mask = ops.logical_not(ops.cast(overflow, "bool"))
292
+ else:
293
+ # Without early stopping, all locations will have been updated.
294
+ padding_mask = ops.ones_like(token_ids, dtype="bool")
295
+ return {
296
+ "token_ids": token_ids,
297
+ "padding_mask": padding_mask,
298
+ }
@@ -0,0 +1,176 @@
1
+ # Copyright 2024 The KerasHub Authors
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # https://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ import keras
16
+ from absl import logging
17
+
18
+ from keras_hub.src.api_export import keras_hub_export
19
+ from keras_hub.src.models.bloom.bloom_preprocessor import BloomPreprocessor
20
+ from keras_hub.src.utils.keras_utils import (
21
+ convert_inputs_to_list_of_tensor_segments,
22
+ )
23
+ from keras_hub.src.utils.tensor_utils import strip_to_ragged
24
+
25
+
26
+ @keras_hub_export("keras_hub.models.BloomCausalLMPreprocessor")
27
+ class BloomCausalLMPreprocessor(BloomPreprocessor):
28
+ """BLOOM Causal LM preprocessor.
29
+
30
+ This preprocessing layer is meant for use with
31
+ `keras_hub.models.BloomCausalLM`. By default, it will take in batches of
32
+ strings, and return outputs in a `(x, y, sample_weight)` format, where the
33
+ `y` label is the next token id in the `x` sequence.
34
+
35
+ For use with generation, the layer also exposes two methods
36
+ `generate_preprocess()` and `generate_postprocess()`. When this preprocessor
37
+ is attached to a `keras_hub.models.BloomCausalLM` instance, these methods
38
+ will be called implicitly in `generate()`. They can also be called
39
+ standalone (e.g. to precompute preprocessing inputs for generation in a
40
+ separate process).
41
+
42
+ Args:
43
+ tokenizer: A `keras_hub.models.BloomTokenizer` instance.
44
+ sequence_length: The length of the packed inputs.
45
+ add_start_token: If `True`, the preprocessor will prepend the tokenizer
46
+ start token to each input sequence.
47
+ add_end_token: If `True`, the preprocessor will append the tokenizer
48
+ end token to each input sequence.
49
+
50
+ Call arguments:
51
+ x: A string, `tf.Tensor` or list of python strings.
52
+ y: Label data. Should always be `None` as the layer generates labels.
53
+ sample_weight: Label weights. Should always be `None` as the layer
54
+ generates label weights.
55
+ sequence_length: Pass to override the configured `sequence_length` of
56
+ the layer.
57
+
58
+ Examples:
59
+ ```python
60
+ # Load the preprocessor from a preset.
61
+ preprocessor = keras_hub.models.BloomCausalLMPreprocessor.from_preset(
62
+ "bloom_560m_multi"
63
+ )
64
+
65
+ # Tokenize and pack a single sentence.
66
+ sentence = tf.constant("League of legends")
67
+ preprocessor(sentence)
68
+ # Same output.
69
+ preprocessor("League of legends")
70
+
71
+ # Tokenize a batch of sentences.
72
+ sentences = tf.constant(["Taco tuesday", "Fish taco please!"])
73
+ preprocessor(sentences)
74
+ # Same output.
75
+ preprocessor(["Taco tuesday", "Fish taco please!"])
76
+
77
+ # Map a dataset to preprocess a single sentence.
78
+ features = tf.constant(
79
+ [
80
+ "Avatar 2 is amazing!",
81
+ "Well, I am not sure.",
82
+ ]
83
+ )
84
+ labels = tf.constant([1, 0])
85
+ ds = tf.data.Dataset.from_tensor_slices((features, labels))
86
+ ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
87
+
88
+ # Map a dataset to preprocess unlabled sentences.
89
+ ds = tf.data.Dataset.from_tensor_slices(features)
90
+ ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
91
+ ```
92
+ """
93
+
94
+ def call(
95
+ self,
96
+ x,
97
+ y=None,
98
+ sample_weight=None,
99
+ sequence_length=None,
100
+ ):
101
+ if y is not None or sample_weight is not None:
102
+ logging.warning(
103
+ "`BloomCausalLMPreprocessor` generates `y` and `sample_weight` "
104
+ "based on your input data, but your data already contains `y` "
105
+ "or `sample_weight`. Your `y` and `sample_weight` will be "
106
+ "ignored."
107
+ )
108
+ sequence_length = sequence_length or self.sequence_length
109
+
110
+ x = convert_inputs_to_list_of_tensor_segments(x)[0]
111
+ x = self.tokenizer(x)
112
+ # Pad with one extra token to account for the truncation below.
113
+ token_ids, padding_mask = self.packer(
114
+ x,
115
+ sequence_length=sequence_length + 1,
116
+ add_start_value=self.add_start_token,
117
+ add_end_value=self.add_end_token,
118
+ )
119
+ # The last token does not have a next token, so we truncate it out.
120
+ x = {
121
+ "token_ids": token_ids[..., :-1],
122
+ "padding_mask": padding_mask[..., :-1],
123
+ }
124
+ # Target `y` will be the next token.
125
+ y, sample_weight = token_ids[..., 1:], padding_mask[..., 1:]
126
+ return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
127
+
128
+ def generate_preprocess(
129
+ self,
130
+ x,
131
+ sequence_length=None,
132
+ ):
133
+ """Convert strings to integer token input for generation.
134
+
135
+ Similar to calling the layer for training, this method takes in strings
136
+ or tensor strings, tokenizes and packs the input, and computes a padding
137
+ mask masking all inputs not filled in with a padded value.
138
+
139
+ Unlike calling the layer for training, this method does not compute
140
+ labels and will never append a `tokenizer.end_token_id` to the end of
141
+ the sequence (as generation is expected to continue at the end of the
142
+ inputted prompt).
143
+ """
144
+ if not self.built:
145
+ self.build(None)
146
+
147
+ x = convert_inputs_to_list_of_tensor_segments(x)[0]
148
+ x = self.tokenizer(x)
149
+ token_ids, padding_mask = self.packer(
150
+ x, sequence_length=sequence_length, add_end_value=False
151
+ )
152
+ return {
153
+ "token_ids": token_ids,
154
+ "padding_mask": padding_mask,
155
+ }
156
+
157
+ def generate_postprocess(
158
+ self,
159
+ x,
160
+ ):
161
+ """Convert integer token output to strings for generation.
162
+
163
+ This method reverses `generate_preprocess()`, by first removing all
164
+ padding and start/end tokens, and then converting the integer sequence
165
+ back to a string.
166
+ """
167
+ if not self.built:
168
+ self.build(None)
169
+
170
+ token_ids, padding_mask = x["token_ids"], x["padding_mask"]
171
+ ids_to_strip = (
172
+ self.tokenizer.start_token_id,
173
+ self.tokenizer.end_token_id,
174
+ )
175
+ token_ids = strip_to_ragged(token_ids, padding_mask, ids_to_strip)
176
+ return self.tokenizer.detokenize(token_ids)
@@ -0,0 +1,206 @@
1
+ # Copyright 2024 The KerasHub Authors
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # https://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+ # import keras
15
+ # from keras import ops
16
+ import keras
17
+ from keras import ops
18
+
19
+ from keras_hub.src.layers.modeling.transformer_layer_utils import (
20
+ compute_causal_mask,
21
+ )
22
+ from keras_hub.src.layers.modeling.transformer_layer_utils import (
23
+ merge_padding_and_attention_mask,
24
+ )
25
+ from keras_hub.src.models.bloom.bloom_attention import BloomAttention
26
+ from keras_hub.src.utils.keras_utils import clone_initializer
27
+
28
+
29
+ class BloomDecoder(keras.layers.Layer):
30
+ def __init__(
31
+ self,
32
+ num_heads,
33
+ intermediate_dim,
34
+ dropout=0.0,
35
+ layer_norm_epsilon=1e-5,
36
+ kernel_initializer="glorot_uniform",
37
+ bias_initializer="zeros",
38
+ **kwargs,
39
+ ):
40
+ super().__init__(**kwargs)
41
+
42
+ self.num_heads = num_heads
43
+ self.intermediate_dim = intermediate_dim
44
+ self.dropout = dropout
45
+ self.layer_norm_epsilon = layer_norm_epsilon
46
+ self.kernel_initializer = keras.initializers.get(kernel_initializer)
47
+ self.bias_initializer = keras.initializers.get(bias_initializer)
48
+
49
+ def build(self, decoder_sequence_shape):
50
+ hidden_dim = decoder_sequence_shape[-1]
51
+ head_dim = int(hidden_dim // self.num_heads)
52
+
53
+ if head_dim * self.num_heads != hidden_dim:
54
+ raise ValueError(
55
+ f"`hidden_dim` must be divisible by num_heads (got `hidden_dim`"
56
+ f": {hidden_dim} and `num_heads`: {self.num_heads})."
57
+ )
58
+
59
+ self._pre_attention_layernorm = keras.layers.LayerNormalization(
60
+ epsilon=self.layer_norm_epsilon,
61
+ dtype=self.dtype_policy,
62
+ name="pre_attention_layernorm",
63
+ )
64
+ self._pre_attention_layernorm.build(decoder_sequence_shape)
65
+
66
+ self._self_attention_layer = BloomAttention(
67
+ num_heads=self.num_heads,
68
+ dropout=self.dropout,
69
+ kernel_initializer=clone_initializer(self.kernel_initializer),
70
+ bias_initializer=clone_initializer(self.bias_initializer),
71
+ dtype=self.dtype_policy,
72
+ name="self_attention",
73
+ )
74
+ self._self_attention_layer.build(decoder_sequence_shape)
75
+
76
+ self._post_attention_layernorm = keras.layers.LayerNormalization(
77
+ epsilon=self.layer_norm_epsilon,
78
+ dtype=self.dtype_policy,
79
+ name="post_attention_layernorm",
80
+ )
81
+ self._post_attention_layernorm.build(decoder_sequence_shape)
82
+
83
+ self._mlp_intermediate_dense = keras.layers.Dense(
84
+ self.intermediate_dim,
85
+ kernel_initializer=clone_initializer(self.kernel_initializer),
86
+ bias_initializer=clone_initializer(self.bias_initializer),
87
+ dtype=self.dtype_policy,
88
+ name="mlp_intermediate_dense",
89
+ )
90
+ self._mlp_intermediate_dense.build(decoder_sequence_shape)
91
+
92
+ self._mlp_output_dense = keras.layers.Dense(
93
+ hidden_dim,
94
+ kernel_initializer=clone_initializer(self.kernel_initializer),
95
+ bias_initializer=clone_initializer(self.bias_initializer),
96
+ dtype=self.dtype_policy,
97
+ name="mlp_output_dense",
98
+ )
99
+ intermediate_shape = list(decoder_sequence_shape)
100
+ intermediate_shape[-1] = self.intermediate_dim
101
+ self._mlp_output_dense.build(tuple(intermediate_shape))
102
+
103
+ self._dropout_layer = keras.layers.Dropout(
104
+ rate=self.dropout, dtype=self.dtype_policy, name="dropout"
105
+ )
106
+
107
+ self.built = True
108
+
109
+ def call(
110
+ self,
111
+ decoder_sequence,
112
+ decoder_padding_mask=None,
113
+ decoder_attention_mask=None,
114
+ cache=None,
115
+ cache_update_index=None,
116
+ use_causal_mask=True,
117
+ ):
118
+ self_attention_mask = self._compute_attention_mask(
119
+ decoder_sequence=decoder_sequence,
120
+ decoder_padding_mask=decoder_padding_mask,
121
+ decoder_attention_mask=decoder_attention_mask,
122
+ use_causal_mask=use_causal_mask,
123
+ cache=cache,
124
+ cache_update_index=cache_update_index,
125
+ )
126
+
127
+ residual = decoder_sequence
128
+ x = self._pre_attention_layernorm(decoder_sequence)
129
+
130
+ attention_output = self._self_attention_layer(
131
+ hidden_states=x,
132
+ attention_mask=self_attention_mask,
133
+ cache=cache,
134
+ cache_update_index=cache_update_index,
135
+ )
136
+
137
+ if cache is None:
138
+ x = attention_output
139
+ else:
140
+ x, cache = attention_output
141
+
142
+ x = x + residual
143
+ residual = x
144
+ x = self._post_attention_layernorm(x)
145
+ x = self._mlp_intermediate_dense(x)
146
+ x = keras.activations.gelu(x, approximate=True)
147
+ x = self._mlp_output_dense(x)
148
+ x = self._dropout_layer(x)
149
+ x = x + residual
150
+
151
+ if cache is not None:
152
+ return x, cache
153
+ else:
154
+ return x
155
+
156
+ def _compute_attention_mask(
157
+ self,
158
+ decoder_sequence,
159
+ decoder_padding_mask,
160
+ decoder_attention_mask,
161
+ use_causal_mask,
162
+ cache,
163
+ cache_update_index,
164
+ ):
165
+ decoder_mask = merge_padding_and_attention_mask(
166
+ decoder_sequence, decoder_padding_mask, decoder_attention_mask
167
+ )
168
+ if use_causal_mask:
169
+ batch_size = ops.shape(decoder_sequence)[0]
170
+ input_length = output_length = ops.shape(decoder_sequence)[1]
171
+ if cache is not None:
172
+ input_length = ops.shape(cache)[2]
173
+
174
+ causal_mask = compute_causal_mask(
175
+ batch_size,
176
+ input_length,
177
+ output_length,
178
+ (0 if cache_update_index is None else cache_update_index),
179
+ )
180
+ return (
181
+ ops.minimum(decoder_mask, causal_mask)
182
+ if decoder_mask is not None
183
+ else causal_mask
184
+ )
185
+ return decoder_mask
186
+
187
+ def get_config(self):
188
+ config = super().get_config()
189
+ config.update(
190
+ {
191
+ "num_heads": self.num_heads,
192
+ "intermediate_dim": self.intermediate_dim,
193
+ "dropout": self.dropout,
194
+ "layer_norm_epsilon": self.layer_norm_epsilon,
195
+ "kernel_initializer": keras.initializers.serialize(
196
+ self.kernel_initializer
197
+ ),
198
+ "bias_initializer": keras.initializers.serialize(
199
+ self.bias_initializer
200
+ ),
201
+ }
202
+ )
203
+ return config
204
+
205
+ def compute_output_shape(self, decoder_sequence_shape):
206
+ return decoder_sequence_shape