ai-edge-torch-nightly 0.3.0.dev20240910__py3-none-any.whl → 0.3.0.dev20240911__py3-none-any.whl

{ai_edge_torch_nightly-0.3.0.dev20240910.dist-info → ai_edge_torch_nightly-0.3.0.dev20240911.dist-info}/RECORD

@@ -2,7 +2,7 @@ ai_edge_torch/__init__.py,sha256=48qP37uHT90YPs4eIUQxCiWVwqGEX3idCUs6mQKvX1U,116
 ai_edge_torch/config.py,sha256=PCd9PVrbUNeVIUDFUCnW4goDWU4bjouK28yMYU6VOi0,877
 ai_edge_torch/conftest.py,sha256=r0GTrhMRhlmOGrrkvumHN8hkmyug6WvF60vWq8wRIBI,758
 ai_edge_torch/model.py,sha256=NYV6Mkaje_ditIEI_s_7nLP_-8i4kbGM8nRzieVkbUI,5397
-ai_edge_torch/version.py,sha256=e4sh_RFYgNHGoVuOeICnFZtLu1MQCNv7qpq94nKFarU,706
+ai_edge_torch/version.py,sha256=vCTKdj1Lc6r2UbJhIZpLdXauJSS0KfBLzgy9e3D16AA,706
 ai_edge_torch/_convert/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/_convert/conversion.py,sha256=kcv_QgNgeyDmrqwdzHicGNP68w6zF7GJg7YkMEIXp4Q,3759
 ai_edge_torch/_convert/conversion_utils.py,sha256=Sr8qXVcTwc-ZnZmK7yxVrIOOp1S_vNrwzC0zUvLTI2o,2160
@@ -39,24 +39,14 @@ ai_edge_torch/debug/test/test_search_model.py,sha256=-RuU0QsjqkfzZF2IbeA55MoeVOa
 ai_edge_torch/experimental/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/examples/experimental/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/examples/experimental/gemma/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/examples/experimental/gemma/convert_to_tflite.py,sha256=lpiPFSh3SJd6WwuZ0QegSva3__iSz2tUD7L7QfkAe4I,3085
-ai_edge_torch/generative/examples/experimental/gemma/gemma.py,sha256=aCoD86pf4nuquUMk7MOR-jsN5FqvySSEuMx9Psxjblk,7261
-ai_edge_torch/generative/examples/experimental/phi/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/examples/experimental/phi/convert_to_tflite.py,sha256=DavrdGmqUgoThsGNRv3LXMW5tvJdYEvj66Hf1XRqkXU,3055
-ai_edge_torch/generative/examples/experimental/phi/phi2.py,sha256=Jxf3ZyYDpS78l6uh4_LGGIcHawrOhZ1vHoHFVxRaK40,6789
-ai_edge_torch/generative/examples/experimental/tiny_llama/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/examples/experimental/tiny_llama/convert_to_tflite.py,sha256=xPVvHQjLJHFiRv_-Fy2sDm0Aft7SG8SXiV6o3rF03cQ,3108
-ai_edge_torch/generative/examples/experimental/tiny_llama/tiny_llama.py,sha256=nUm0SQbCTmNAc5u-C9gbQRFPt7GDvUt6UjH6doTvH-I,6817
 ai_edge_torch/generative/examples/gemma/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/examples/gemma/convert_gemma2_to_tflite.py,sha256=pseJExH35lSAK0ZtzSHB1sFtRtF_EuT2xcSpGU0gKVI,2524
-ai_edge_torch/generative/examples/gemma/convert_to_tflite.py,sha256=w589IJETATd6Z9_1XCIWbrlCV3E92X_5ac3VVCVFXG0,2522
-ai_edge_torch/generative/examples/gemma/gemma.py,sha256=lc1-CfIObHj9D5VJy78BOtGTrQM4TYMI6NfVi8KM5qA,6747
-ai_edge_torch/generative/examples/gemma/gemma2.py,sha256=OcUQLFR136e3QRVXRnmtYnRHXyHJS9EYEFlJ1ymXyRY,8859
-ai_edge_torch/generative/examples/phi2/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/examples/phi2/convert_to_tflite.py,sha256=ON6zLO-nFS8eJ2yhyWzT5x2Somr-Ca-VjpjT7OGFU10,2506
-ai_edge_torch/generative/examples/phi2/phi2.py,sha256=FFnhv1kx4fHRhSeOreLGj8kAqPnmkz9pD1RRSDVlM_w,6332
+ai_edge_torch/generative/examples/gemma/convert_gemma2_to_tflite.py,sha256=ZJvw8uFVu7FEJ7eXfpzn-pPKgPELoxkGz4Zg7LKKMSI,3048
+ai_edge_torch/generative/examples/gemma/convert_to_tflite.py,sha256=hM-fwjZG53p1UE_lkovLMmHRDHleJsb6_0ib0_k0v54,3040
+ai_edge_torch/generative/examples/gemma/gemma.py,sha256=oVV1lXgi9cMPES6JmiV8fJOgBQruRdHpyJL7MmXU09M,7283
+ai_edge_torch/generative/examples/gemma/gemma2.py,sha256=X6WfUCDJDEqyyEAYGq1lmKtlDXcYLzy-n2moQPLJA_U,9769
+ai_edge_torch/generative/examples/phi/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/phi/convert_to_tflite.py,sha256=vqEpZVmB0_wMKcAl6RXm7W57DqPTzEdVVN6W2Z-QYzI,3011
+ai_edge_torch/generative/examples/phi/phi2.py,sha256=BzvUrClFx5HKf6PYzJc7ba2O3AwYUJE485u5GSOiPy4,6851
 ai_edge_torch/generative/examples/stable_diffusion/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/stable_diffusion/attention.py,sha256=kDWG6MlIGa89zC5KSRcJlw2c4ITuw8KcchtfmF55f4g,3545
 ai_edge_torch/generative/examples/stable_diffusion/clip.py,sha256=0WniBWQ6_NcQc5WycX3YRRX7Os9AGQSxfc1m2HKBqg8,4479
@@ -78,19 +68,18 @@ ai_edge_torch/generative/examples/t5/t5.py,sha256=Zobw5BV-PC0nlU9Z6fzb2O07rMeU8v
 ai_edge_torch/generative/examples/t5/t5_attention.py,sha256=1lvbSlzyBwmd5Bs7-Up_v4iJQkCPIJx2RmMkLgy7l2Q,8508
 ai_edge_torch/generative/examples/test_models/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/test_models/toy_model.py,sha256=5wj2RmQRIwD6O_R_pp-A_7gKGSdHWDSXyis97r1ELVI,5622
-ai_edge_torch/generative/examples/test_models/toy_model_with_external_kv_cache.py,sha256=l9swUKTcDtnTibNSNExaMgLvDeJ4Er2tVh5ZW1EtRgk,5809
-ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py,sha256=mQkcpSe6HlRLMkIRCEHc9ZXL7jxEp9RWSGUQjjd-r2w,4841
+ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py,sha256=PbWpfg3AOEZjI1FlnZCxRD-kIKtdkR9AOZ6l-9-TpRA,5664
 ai_edge_torch/generative/examples/tiny_llama/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/examples/tiny_llama/convert_to_tflite.py,sha256=CLRqO7ycMbpy7J3_Czp1sLx6hcdwGD9zVq04yRba0e8,2550
-ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py,sha256=4ku0ni3MOWamhPrzLap0BmtdNFk7CH0hwjPNoRAKpvQ,6278
+ai_edge_torch/generative/examples/tiny_llama/convert_to_tflite.py,sha256=y4LiWhwgflqrg4WWh3wq5ei3VOT_cV0A62x62qptQiM,3070
+ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py,sha256=RK7oisSwIPqUWwwE1P-hDJlEnRJJ_V29UjUCxt4xETE,6780
 ai_edge_torch/generative/fx_passes/__init__.py,sha256=fmNNXawJ722M4cTUuTx289rT0NHxBEsOy_k8baqCOms,1173
 ai_edge_torch/generative/fx_passes/remove_sdpa_zero_mask_pass.py,sha256=sXis0U4u-RoIp_NyrmWJNnqFqpqRuZOrhfsJIO6rMps,2028
 ai_edge_torch/generative/layers/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/layers/attention.py,sha256=2UujQePRJ1LK02PN-hGcuMu0ooCJC6ETfPvzEYVFyho,12284
+ai_edge_torch/generative/layers/attention.py,sha256=ee0KHRakhjLjawP32FY2EntxOkyPvjiEZChLnBn_HPc,12601
 ai_edge_torch/generative/layers/attention_utils.py,sha256=68GXGR2HSWBFViTxX7cHifzVG-kcLS2IL2tQJPIpupg,7344
 ai_edge_torch/generative/layers/builder.py,sha256=xb7rjADv3Jm4qfmlYtg6oLLe7ReDE9UjsEqiejPpDD8,4346
 ai_edge_torch/generative/layers/feed_forward.py,sha256=uto7xtwx6jPkk1GZ2x7pSTentQzRrPSKw4_PSE12ahA,3525
-ai_edge_torch/generative/layers/kv_cache.py,sha256=Ob8QeXWW5xt-6hcGA0uoC48eRQ8lfvKca8JbWtFx2CE,3082
+ai_edge_torch/generative/layers/kv_cache.py,sha256=WDu03NQwkDCrrrT9Du_3ZOxlURZz3XDbS1PLzFozhMI,6013
 ai_edge_torch/generative/layers/model_config.py,sha256=WpZ9djUBAZddyeSODHDaVMG37EQqfzGGrlMPi8AA-Hc,5752
 ai_edge_torch/generative/layers/normalization.py,sha256=u8lv0p-ktKcRqCDlOqZQa9WQcfDK9JM2IaUQFQdn7xs,1860
 ai_edge_torch/generative/layers/rotary_position_embedding.py,sha256=CZqOoibLcHvUgrgaIIWAlmk3XgE2inzx340MN-npLoU,1347
@@ -107,11 +96,12 @@ ai_edge_torch/generative/quantize/quant_recipe_utils.py,sha256=4fgmP_GgeiFUOkIaC
 ai_edge_torch/generative/quantize/quant_recipes.py,sha256=0Kvr_o7pbMnE8VMe6Ml0FBxkHM6RJ3C14B2I1mjItjc,2030
 ai_edge_torch/generative/quantize/supported_schemes.py,sha256=FjdycEOvxRgBmQdZVufetPvkDoD7rUowIOSKV9oV5Kk,1418
 ai_edge_torch/generative/test/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/generative/test/test_experimental_ekv.py,sha256=8qv_eVtJW9GPvBEf2hPQe3tpdJ33XShya6MCX1FqrZM,4355
+ai_edge_torch/generative/test/test_kv_cache.py,sha256=FU2rmU03Lp-vZ5wWXXCao1WEw7xbpqebFMANL_O2chA,3713
 ai_edge_torch/generative/test/test_loader.py,sha256=_y5EHGgoNOmCuYonsB81UJScHVsTAQXUVd44czMAw6k,3379
-ai_edge_torch/generative/test/test_model_conversion.py,sha256=b3InJ8Rx03YtHpE9h-j0pSXAY1cCf-dLlx4Y5LSJnRQ,5174
-ai_edge_torch/generative/test/test_model_conversion_large.py,sha256=9JXcd-rX8MpsYeEWUFEXf783GOwYOLY64KzDfFdmRJ8,4484
+ai_edge_torch/generative/test/test_model_conversion.py,sha256=OmAHSGkxTNzDX5_kYjK7pxlPk0YZLqL9YiVIJQfuvPc,5889
+ai_edge_torch/generative/test/test_model_conversion_large.py,sha256=F3q3K9ZgWBzlLy4WpE8-w6UWSuJ-UoJwMm3N6Zb3Y14,5016
 ai_edge_torch/generative/test/test_quantize.py,sha256=kY_NRpF-v1i4clqI1CFFWEagJv-5PzBDkeJ2fInl9_w,5913
+ai_edge_torch/generative/test/utils.py,sha256=YvEhO2HIj1LkBs5du1UxY-cGRW9HMyAYsOUhgsTrTpA,1796
 ai_edge_torch/generative/utilities/__init__.py,sha256=-_jxnnFnCgnTU4oTm4MnRsvL5lqhomBNdFBbqfmfHPo,720
 ai_edge_torch/generative/utilities/loader.py,sha256=6J0aAP6-6LySeqeYIHKcchr5T9cVtSO34aoDr3V9gxY,12726
 ai_edge_torch/generative/utilities/stable_diffusion_loader.py,sha256=pKp3AMSbS3otCvgwJRF5M1l4JRNKk-aCKimXzIMSrds,35679
@@ -161,8 +151,8 @@ ai_edge_torch/quantize/quant_config.py,sha256=U0KisSW-uZkoMJcy-ZP9W57p3tsa594fr9
 ai_edge_torch/testing/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/testing/model_coverage/__init__.py,sha256=5P8J6Zk5YYtDvTBucFvB9NGSRI7Gw_24WnrbhXgycEE,765
 ai_edge_torch/testing/model_coverage/model_coverage.py,sha256=UPB448aMDUyC0HNYVqio2rcJPnDN0tBQMP08J6vPYew,4718
-ai_edge_torch_nightly-0.3.0.dev20240910.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
-ai_edge_torch_nightly-0.3.0.dev20240910.dist-info/METADATA,sha256=WFNExTO6eF-tAWPmDdQDlr9dvplcoNB0uPdVxSNXYHk,1859
-ai_edge_torch_nightly-0.3.0.dev20240910.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-ai_edge_torch_nightly-0.3.0.dev20240910.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
-ai_edge_torch_nightly-0.3.0.dev20240910.dist-info/RECORD,,
+ai_edge_torch_nightly-0.3.0.dev20240911.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.3.0.dev20240911.dist-info/METADATA,sha256=caHeAQX6pxEskue_BvgwkTfZEsG55rXHFwPDcV9oCN8,1859
+ai_edge_torch_nightly-0.3.0.dev20240911.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+ai_edge_torch_nightly-0.3.0.dev20240911.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.3.0.dev20240911.dist-info/RECORD,,

ai_edge_torch/generative/examples/experimental/gemma/__init__.py

@@ -1,14 +0,0 @@
-# Copyright 2024 The AI Edge Torch 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.
-# ==============================================================================

ai_edge_torch/generative/examples/experimental/gemma/convert_to_tflite.py

@@ -1,88 +0,0 @@
-# Copyright 2024 The AI Edge Torch 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.
-# ==============================================================================
-#
-# Note: This is an experimental version of Gemma with external KV cache.
-# Please use with caution.
-
-
-import os
-from pathlib import Path
-
-import ai_edge_torch
-from ai_edge_torch.generative.examples.experimental.gemma import gemma
-from ai_edge_torch.generative.layers.experimental import ekv_cache as kv_utils
-from ai_edge_torch.generative.quantize import quant_recipes
-import torch
-
-
-def convert_gemma_to_tflite(
-    checkpoint_path: str,
-    prefill_seq_len: int = 512,
-    kv_cache_max_len: int = 1024,
-    quantize: bool = True,
-):
-  """An example method for converting a Gemma 2B model to multi-signature
-
-  tflite model.
-  Args:
-      checkpoint_path (str): The filepath to the model checkpoint, or directory
-        holding the checkpoint.
-      prefill_seq_len (int, optional): The maximum size of prefill input tensor.
-        Defaults to 512.
-      kv_cache_max_len (int, optional): The maximum size of KV cache buffer,
-        including both prefill and decode. Defaults to 1024.
-      quantize (bool, optional): Whether the model should be quanized. Defaults
-        to True.
-  """
-  pytorch_model = gemma.build_2b_model(
-      checkpoint_path, kv_cache_max_len=kv_cache_max_len
-  )
-  # Tensors used to trace the model graph during conversion.
-  prefill_tokens = torch.full((1, prefill_seq_len), 0, dtype=torch.long)
-  prefill_input_pos = torch.arange(0, prefill_seq_len)
-  decode_token = torch.tensor([[0]], dtype=torch.long)
-  decode_input_pos = torch.tensor([0], dtype=torch.int64)
-  kv = kv_utils.EKVCache.from_model_config(pytorch_model.config)
-
-  quant_config = quant_recipes.full_int8_dynamic_recipe() if quantize else None
-  edge_model = (
-      ai_edge_torch.signature(
-          'prefill',
-          pytorch_model,
-          sample_kwargs={
-              'tokens': prefill_tokens,
-              'input_pos': prefill_input_pos,
-              'kv_cache': kv,
-          },
-      )
-      .signature(
-          'decode',
-          pytorch_model,
-          sample_kwargs={
-              'tokens': decode_token,
-              'input_pos': decode_input_pos,
-              'kv_cache': kv,
-          },
-      )
-      .convert(quant_config=quant_config)
-  )
-  edge_model.export(
-      f'/tmp/gemma_seq{prefill_seq_len}_ekv{kv_cache_max_len}.tflite'
-  )
-
-
-if __name__ == '__main__':
-  checkpoint_path = os.path.join(Path.home(), 'Downloads/llm_data/gemma-2b')
-  convert_gemma_to_tflite(checkpoint_path)

ai_edge_torch/generative/examples/experimental/gemma/gemma.py

@@ -1,219 +0,0 @@
-# Copyright 2024 The AI Edge Torch 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.
-# ==============================================================================
-# Example of building a Gemma model.
-#
-# Note: This is an experimental version of Gemma with external KV cache.
-# Please use with caution.
-
-import os
-from pathlib import Path
-from typing import Tuple
-
-from ai_edge_torch.generative.layers import builder
-import ai_edge_torch.generative.layers.attention_utils as attn_utils
-from ai_edge_torch.generative.layers.experimental import attention
-from ai_edge_torch.generative.layers.experimental import ekv_cache as kv_utils
-import ai_edge_torch.generative.layers.model_config as cfg
-import ai_edge_torch.generative.utilities.loader as loading_utils
-import numpy as np
-import torch
-from torch import nn
-
-
-TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
-    ff_up_proj="model.layers.{}.mlp.up_proj",
-    ff_down_proj="model.layers.{}.mlp.down_proj",
-    ff_gate_proj="model.layers.{}.mlp.gate_proj",
-    attn_query_proj="model.layers.{}.self_attn.q_proj",
-    attn_key_proj="model.layers.{}.self_attn.k_proj",
-    attn_value_proj="model.layers.{}.self_attn.v_proj",
-    attn_output_proj="model.layers.{}.self_attn.o_proj",
-    pre_attn_norm="model.layers.{}.input_layernorm",
-    post_attn_norm="model.layers.{}.post_attention_layernorm",
-    embedding="model.embed_tokens",
-    final_norm="model.norm",
-    lm_head=None,
-)
-
-
-class Gemma(nn.Module):
-  """A Gemma model built from the Edge Generative API layers."""
-
-  def __init__(self, config: cfg.ModelConfig):
-    super().__init__()
-
-    self.config = config
-    # Construct model layers.
-    self.tok_embedding = nn.Embedding(
-        config.vocab_size, config.embedding_dim, padding_idx=0
-    )
-    self.lm_head = nn.Linear(
-        config.embedding_dim,
-        config.vocab_size,
-        bias=config.lm_head_use_bias,
-    )
-    # Gemma re-uses the embedding as the head projection layer.
-    self.lm_head.weight.data = self.tok_embedding.weight.data
-    self.transformer_blocks = nn.ModuleList(
-        attention.TransformerBlock(config) for _ in range(config.num_layers)
-    )
-    self.final_norm = builder.build_norm(
-        config.embedding_dim,
-        config.final_norm_config,
-    )
-    self.rope_cache = attn_utils.build_rope_cache(
-        size=config.kv_cache_max,
-        dim=int(
-            config.attn_config.rotary_percentage * config.attn_config.head_dim
-        ),
-        base=10_000,
-        condense_ratio=1,
-        dtype=torch.float32,
-        device=torch.device("cpu"),
-    )
-    self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.kv_cache_max,
-        dtype=torch.float32,
-        device=torch.device("cpu"),
-    )
-    self.config = config
-
-  @torch.inference_mode
-  def forward(
-      self,
-      tokens: torch.Tensor,
-      input_pos: torch.Tensor,
-      kv_cache: kv_utils.EKVCache,
-  ) -> Tuple[torch.Tensor, kv_utils.EKVCache]:
-    _, seq_len = tokens.size()
-    assert self.config.max_seq_len >= seq_len, (
-        f"Cannot forward sequence of length {seq_len}, max seq length is only"
-        f" {self.config.max_seq_len}"
-    )
-
-    cos, sin = self.rope_cache
-    cos = cos.index_select(0, input_pos)
-    sin = sin.index_select(0, input_pos)
-    mask = self.mask_cache.index_select(2, input_pos)
-    mask = mask[:, :, :, : self.config.kv_cache_max]
-
-    # token embeddings of shape (b, t, n_embd)
-    x = self.tok_embedding(tokens)
-    x = x * (self.config.embedding_dim**0.5)
-
-    updated_kv_entires = []
-    for i, block in enumerate(self.transformer_blocks):
-      kv_entry = kv_cache.caches[i] if kv_cache else None
-      x, kv_entry = block(x, (cos, sin), mask, input_pos, kv_entry)
-      if kv_entry:
-        updated_kv_entires.append(kv_entry)
-    updated_kv_cache = kv_utils.EKVCache(tuple(updated_kv_entires))
-
-    x = self.final_norm(x)
-    res = self.lm_head(x)  # (b, t, vocab_size)
-    return res, updated_kv_cache
-
-
-def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
-  """Returns the model config for a Gemma 2B model.
-
-  Args:
-    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
-      is 1024.
-
-  Returns:
-    The model config for a Gemma 2B model.
-  """
-  attn_config = cfg.AttentionConfig(
-      num_heads=8,
-      head_dim=256,
-      num_query_groups=1,
-      rotary_percentage=1.0,
-  )
-  ff_config = cfg.FeedForwardConfig(
-      type=cfg.FeedForwardType.GATED,
-      activation=cfg.ActivationConfig(cfg.ActivationType.GELU_TANH),
-      intermediate_size=16384,
-  )
-  norm_config = cfg.NormalizationConfig(
-      type=cfg.NormalizationType.RMS_NORM,
-      epsilon=1e-6,
-      zero_centered=True,
-  )
-  config = cfg.ModelConfig(
-      vocab_size=256000,
-      num_layers=18,
-      max_seq_len=8192,
-      embedding_dim=2048,
-      kv_cache_max_len=kv_cache_max_len,
-      attn_config=attn_config,
-      ff_config=ff_config,
-      pre_attention_norm_config=norm_config,
-      post_attention_norm_config=norm_config,
-      final_norm_config=norm_config,
-      parallel_residual=False,
-      lm_head_use_bias=False,
-      enable_hlfb=True,
-  )
-  return config
-
-
-def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
-  config = get_model_config_2b(kv_cache_max_len)
-  config.ff_config.intermediate_size = 128
-  config.vocab_size = 128
-  config.num_layers = 2
-  config.max_seq_len = 2 * kv_cache_max_len
-  return config
-
-
-def build_2b_model(
-    checkpoint_path: str, test_model: bool = False, **kwargs
-) -> nn.Module:
-  """Instantiates the model instance and load checkpoint if provided."""
-  config = (
-      get_fake_model_config(**kwargs)
-      if test_model
-      else get_model_config_2b(**kwargs)
-  )
-  model = Gemma(config)
-  if checkpoint_path is not None:
-    loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
-    # since embedding and lm-head use the same weight, we need to set strict
-    # to False.
-    loader.load(model, strict=False)
-  model.eval()
-  return model
-
-
-def define_and_run_2b(checkpoint_path: str, test_model: bool = False) -> None:
-  """Instantiates and runs a Gemma 2B model."""
-
-  kv_cache_max_len = 1024
-  model = build_2b_model(
-      checkpoint_path, test_model=test_model, kv_cache_max_len=kv_cache_max_len
-  )
-  idx = torch.from_numpy(np.array([[1, 2, 3, 4]]))
-  tokens = torch.full((1, kv_cache_max_len), 0, dtype=torch.long, device="cpu")
-  tokens[0, :4] = idx
-  input_pos = torch.arange(0, kv_cache_max_len)
-  kv = kv_utils.EKVCache.from_model_config(model.config)
-  print("running an inference")
-  print(model.forward(tokens, input_pos, kv))
-
-
-if __name__ == "__main__":
-  input_checkpoint_path = os.path.join(Path.home(), "Downloads/gemma-2b")
-  define_and_run_2b(input_checkpoint_path)

ai_edge_torch/generative/examples/experimental/phi/__init__.py

@@ -1,14 +0,0 @@
-# Copyright 2024 The AI Edge Torch 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.
-# ==============================================================================

ai_edge_torch/generative/examples/experimental/tiny_llama/__init__.py

@@ -1,14 +0,0 @@
-# Copyright 2024 The AI Edge Torch 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.
-# ==============================================================================

ai_edge_torch/generative/examples/experimental/tiny_llama/convert_to_tflite.py

@@ -1,87 +0,0 @@
-# Copyright 2024 The AI Edge Torch 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.
-# ==============================================================================
-#
-# Note: This is an experimental version of TinyLlama with external KV cache.
-# Please use with caution.
-
-
-import os
-from pathlib import Path
-
-import ai_edge_torch
-from ai_edge_torch.generative.examples.experimental.tiny_llama import tiny_llama
-from ai_edge_torch.generative.layers.experimental import ekv_cache as kv_utils
-from ai_edge_torch.generative.quantize import quant_recipes
-import torch
-
-
-def convert_tiny_llama_to_tflite(
-    checkpoint_path: str,
-    prefill_seq_len: int = 512,
-    kv_cache_max_len: int = 1024,
-    quantize: bool = True,
-):
-  """An example for converting TinyLlama model to multi-signature tflite model.
-
-  Args:
-      checkpoint_path (str): The filepath to the model checkpoint, or directory
-        holding the checkpoint.
-      prefill_seq_len (int, optional): The maximum size of prefill input tensor.
-        Defaults to 512.
-      kv_cache_max_len (int, optional): The maximum size of KV cache buffer,
-        including both prefill and decode. Defaults to 1024.
-      quantize (bool, optional): Whether the model should be quanized. Defaults
-        to True.
-  """
-  pytorch_model = tiny_llama.build_model(
-      checkpoint_path, kv_cache_max_len=kv_cache_max_len
-  )
-  # Tensors used to trace the model graph during conversion.
-  prefill_tokens = torch.full((1, prefill_seq_len), 0, dtype=torch.long)
-  prefill_input_pos = torch.arange(0, prefill_seq_len)
-  decode_token = torch.tensor([[0]], dtype=torch.long)
-  decode_input_pos = torch.tensor([0], dtype=torch.int64)
-  kv = kv_utils.EKVCache.from_model_config(pytorch_model.config)
-
-  quant_config = quant_recipes.full_int8_dynamic_recipe() if quantize else None
-  edge_model = (
-      ai_edge_torch.signature(
-          'prefill',
-          pytorch_model,
-          sample_kwargs={
-              'tokens': prefill_tokens,
-              'input_pos': prefill_input_pos,
-              'kv_cache': kv,
-          },
-      )
-      .signature(
-          'decode',
-          pytorch_model,
-          sample_kwargs={
-              'tokens': decode_token,
-              'input_pos': decode_input_pos,
-              'kv_cache': kv,
-          },
-      )
-      .convert(quant_config=quant_config)
-  )
-  edge_model.export(
-      f'/tmp/tiny_llama_seq{prefill_seq_len}_ekv{kv_cache_max_len}.tflite'
-  )
-
-
-if __name__ == '__main__':
-  checkpoint_path = os.path.join(Path.home(), 'Downloads/llm_data/tiny_llama')
-  convert_tiny_llama_to_tflite(checkpoint_path)