ai-edge-torch-nightly 0.3.0.dev20240928__py3-none-any.whl → 0.3.0.dev20240930__py3-none-any.whl

ai_edge_torch/generative/examples/qwen/__init__.py

@@ -0,0 +1,14 @@
+# 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/qwen/convert_to_tflite.py

@@ -0,0 +1,81 @@
+# 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 converting Qwen 2.5 models to multi-signature tflite model."""
+
+import os
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.qwen import qwen
+from ai_edge_torch.generative.utilities import converter
+
+_MODEL_SIZE = flags.DEFINE_enum(
+    'model_size',
+    '3b',
+    ['0.5b', '1.5b', '3b'],
+    'The size of the model to convert.',
+)
+_CHECKPOINT_PATH = flags.DEFINE_string(
+    'checkpoint_path',
+    os.path.join(pathlib.Path.home(), 'Downloads/llm_data/qwen'),
+    'The path to the model checkpoint, or directory holding the checkpoint.',
+)
+_TFLITE_PATH = flags.DEFINE_string(
+    'tflite_path',
+    '/tmp/',
+    'The tflite file path to export.',
+)
+_PREFILL_SEQ_LEN = flags.DEFINE_integer(
+    'prefill_seq_len',
+    1024,
+    'The maximum size of prefill input tensor.',
+)
+_KV_CACHE_MAX_LEN = flags.DEFINE_integer(
+    'kv_cache_max_len',
+    1280,
+    'The maximum size of KV cache buffer, including both prefill and decode.',
+)
+_QUANTIZE = flags.DEFINE_bool(
+    'quantize',
+    True,
+    'Whether the model should be quantized.',
+)
+
+_BUILDER = {
+    '0.5b': qwen.build_0_5b_model,
+    '1.5b': qwen.build_1_5b_model,
+    '3b': qwen.build_3b_model,
+}
+
+
+def main(_):
+  pytorch_model = _BUILDER[_MODEL_SIZE.value](
+      _CHECKPOINT_PATH.value, kv_cache_max_len=_KV_CACHE_MAX_LEN.value
+  )
+  quant_suffix = 'q8' if _QUANTIZE.value else 'f32'
+  model_size = _MODEL_SIZE.value.replace('.', '_')
+  output_filename = f'qwen_{model_size}_{quant_suffix}_seq{_PREFILL_SEQ_LEN.value}_ekv{_KV_CACHE_MAX_LEN.value}.tflite'
+  converter.convert_to_tflite(
+      pytorch_model,
+      tflite_path=os.path.join(_TFLITE_PATH.value, output_filename),
+      prefill_seq_len=_PREFILL_SEQ_LEN.value,
+      quantize=_QUANTIZE.value,
+  )
+
+
+if __name__ == '__main__':
+  app.run(main)

ai_edge_torch/generative/examples/qwen/qwen.py

@@ -0,0 +1,141 @@
+# 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 Qwen 2.5 models."""
+
+import copy
+
+from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
+import ai_edge_torch.generative.layers.model_config as cfg
+import ai_edge_torch.generative.utilities.loader as loading_utils
+from torch import nn
+
+TENSOR_NAMES = copy.copy(tiny_llama.TENSOR_NAMES)
+# Qwen re-uses the embedding as the head projection layer.
+TENSOR_NAMES.lm_head = None
+
+
+class Qwen(tiny_llama.TinyLlama):
+  """A Qwen model built from the Edge Generative API layers.
+
+  Qwen 2.5 shares the same architecture as TinyLlama.
+  """
+
+  def __init__(self, config: cfg.ModelConfig):
+    super().__init__(config)
+    # Qwen re-uses the embedding as the head projection layer.
+    self.lm_head.weight.data = self.tok_embedding.weight.data
+
+
+def get_3b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Qwen 2.5 3B model.
+
+  Args:
+    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
+      is 1024.
+
+  Returns:
+    The model config for a SmolLM model.
+  """
+  attn_config = cfg.AttentionConfig(
+      num_heads=16,
+      head_dim=128,
+      num_query_groups=2,
+      rotary_base=1000000,
+      rotary_percentage=1.0,
+      qkv_use_bias=True,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.GATED,
+      activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
+      intermediate_size=11008,
+  )
+  norm_config = cfg.NormalizationConfig(
+      type=cfg.NormalizationType.RMS_NORM,
+      epsilon=1e-06,
+  )
+  block_config = cfg.TransformerBlockConfig(
+      attn_config=attn_config,
+      ff_config=ff_config,
+      pre_attention_norm_config=norm_config,
+      post_attention_norm_config=norm_config,
+  )
+  config = cfg.ModelConfig(
+      vocab_size=151936,
+      num_layers=36,
+      max_seq_len=32768,
+      embedding_dim=2048,
+      kv_cache_max_len=kv_cache_max_len,
+      block_configs=block_config,
+      final_norm_config=norm_config,
+      enable_hlfb=True,
+  )
+  return config
+
+
+def get_1_5b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Qwen 2.5 1B model."""
+  config = get_3b_model_config(kv_cache_max_len)
+  # Qwen has only one block config.
+  block_config = config.block_config(0)
+  block_config.attn_config.num_heads = 12
+  block_config.ff_config.intermediate_size = 8960
+  config.num_layers = 28
+  config.embedding_dim = 1536
+  return config
+
+
+def get_0_5b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Qwen 2.5 0.5B model."""
+  config = get_3b_model_config(kv_cache_max_len)
+  # Qwen has only one block config.
+  block_config = config.block_config(0)
+  block_config.attn_config.num_heads = 14
+  block_config.attn_config.head_dim = 64
+  block_config.ff_config.intermediate_size = 4864
+  config.num_layers = 24
+  config.embedding_dim = 896
+  return config
+
+
+def get_fake_model_config(**kwargs) -> cfg.ModelConfig:
+  config = get_3b_model_config(**kwargs)
+  config.vocab_size = 128
+  config.num_layers = 2
+  # Qwen has only one block config.
+  config.block_config(0).ff_config.intermediate_size = 64
+  return config
+
+
+def _build_model(checkpoint_path: str, config: cfg.ModelConfig) -> nn.Module:
+  model = Qwen(config)
+  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 build_3b_model(checkpoint_path: str, **kwargs) -> nn.Module:
+  return _build_model(checkpoint_path, get_3b_model_config(**kwargs))
+
+
+def build_1_5b_model(checkpoint_path: str, **kwargs) -> nn.Module:
+  return _build_model(checkpoint_path, get_1_5b_model_config(**kwargs))
+
+
+def build_0_5b_model(checkpoint_path: str, **kwargs) -> nn.Module:
+  return _build_model(checkpoint_path, get_0_5b_model_config(**kwargs))

ai_edge_torch/generative/examples/qwen/verify.py

@@ -0,0 +1,88 @@
+# 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.
+# ==============================================================================
+
+"""Verifies the reauthored Qwen 2.5 0.5B, 1.5B, and 3B models."""
+
+import logging
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.qwen import qwen
+from ai_edge_torch.generative.utilities import transformers_verifier
+from ai_edge_torch.generative.utilities import verifier
+import transformers
+
+
+_MODEL_SIZE = flags.DEFINE_enum(
+    "model_size",
+    "3b",
+    ["0.5b", "1.5b", "3b"],
+    "The size of the model to verify.",
+)
+_PROMPTS = flags.DEFINE_multi_string(
+    "prompts",
+    "What is the meaning of life?",
+    "The input prompts to generate answers.",
+)
+_MAX_NEW_TOKENS = flags.DEFINE_integer(
+    "max_new_tokens",
+    30,
+    "The maximum size of the generated tokens.",
+)
+
+_CHECKPOINT = {
+    "0.5b": "Qwen/Qwen2.5-0.5B-Instruct",
+    "1.5b": "Qwen/Qwen2.5-1.5B-Instruct",
+    "3b": "Qwen/Qwen2.5-3B-Instruct",
+}
+
+_BUILDER = {
+    "0.5b": qwen.build_0_5b_model,
+    "1.5b": qwen.build_1_5b_model,
+    "3b": qwen.build_3b_model,
+}
+
+
+def main(_):
+  checkpoint = _CHECKPOINT[_MODEL_SIZE.value]
+  logging.info("Loading the original model from: %s", checkpoint)
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint)
+
+  # Locate the cached dir.
+  cached_config_file = transformers.utils.cached_file(
+      checkpoint, transformers.utils.CONFIG_NAME
+  )
+  reauthored_checkpoint = pathlib.Path(cached_config_file).parent
+  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
+  reauthored_model = _BUILDER[_MODEL_SIZE.value](reauthored_checkpoint)
+
+  logging.info("Loading the tokenizer from: %s", checkpoint)
+  tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint)
+
+  verifier.verify_reauthored_model(
+      original_model=transformers_verifier.TransformersModelWrapper(
+          original_model
+      ),
+      reauthored_model=verifier.ReauthoredModelWrapper(reauthored_model),
+      tokenizer=verifier.TokenizerWrapper(tokenizer),
+      generate_prompts=_PROMPTS.value,
+      max_new_tokens=_MAX_NEW_TOKENS.value,
+      atol=1e-04,
+  )
+
+
+if __name__ == "__main__":
+  app.run(main)

ai_edge_torch/generative/test/test_model_conversion_large.py

@@ -23,6 +23,7 @@ from ai_edge_torch.generative.examples.llama import llama
 from ai_edge_torch.generative.examples.openelm import openelm
 from ai_edge_torch.generative.examples.phi import phi2
 from ai_edge_torch.generative.examples.phi import phi3
+from ai_edge_torch.generative.examples.qwen import qwen
 from ai_edge_torch.generative.examples.smollm import smollm
 from ai_edge_torch.generative.examples.stable_diffusion import clip as sd_clip
 from ai_edge_torch.generative.examples.stable_diffusion import decoder as sd_decoder
@@ -152,6 +153,15 @@ class TestModelConversion(googletest.TestCase):
     pytorch_model = openelm.OpenELM(config).eval()
     self._test_model(config, pytorch_model, "prefill", atol=1e-4, rtol=1e-5)
 
+  @googletest.skipIf(
+      ai_edge_config.Config.use_torch_xla,
+      reason="tests with custom ops are not supported on oss",
+  )
+  def test_qwen(self):
+    config = qwen.get_fake_model_config()
+    pytorch_model = qwen.Qwen(config).eval()
+    self._test_model(config, pytorch_model, "prefill", atol=1e-3, rtol=1e-5)
+
   @googletest.skipIf(
       ai_edge_config.Config.use_torch_xla,
       reason="tests with custom ops are not supported on oss",

ai_edge_torch/version.py

@@ -13,4 +13,4 @@
 # limitations under the License.
 # ==============================================================================
 
-__version__ = "0.3.0.dev20240928"
+__version__ = "0.3.0.dev20240930"

{ai_edge_torch_nightly-0.3.0.dev20240928.dist-info → ai_edge_torch_nightly-0.3.0.dev20240930.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.3.0.dev20240928
+Version: 0.3.0.dev20240930
 Summary: Supporting PyTorch models with the Google AI Edge TFLite runtime.
 Home-page: https://github.com/google-ai-edge/ai-edge-torch
 Keywords: On-Device ML,AI,Google,TFLite,PyTorch,LLMs,GenAI

{ai_edge_torch_nightly-0.3.0.dev20240928.dist-info → ai_edge_torch_nightly-0.3.0.dev20240930.dist-info}/RECORD

@@ -3,7 +3,7 @@ ai_edge_torch/config.py,sha256=FMWeCH2b7HYILBvaI1iZNnYCO4WAhDOwBZBmIE-xrF0,909
 ai_edge_torch/conftest.py,sha256=r0GTrhMRhlmOGrrkvumHN8hkmyug6WvF60vWq8wRIBI,758
 ai_edge_torch/fx_pass_base.py,sha256=D86Gw3pIRcpnTebUPKlnPbPGJae1S6Fw4DZZ3ZkD0zw,3730
 ai_edge_torch/model.py,sha256=N-pNpTxzhaFGhWhnSGd70lBzb9VlEhTOq5mddU7bvvI,5542
-ai_edge_torch/version.py,sha256=YiCjdglLzSPYyRq64U8zJSgWDFqJs-t2JSzuA0bYYzA,706
+ai_edge_torch/version.py,sha256=sc79FL0Fo_EhFaJ-4XlyIRT3Kjmn_oUXzMtvcUeXZDo,706
 ai_edge_torch/_convert/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/_convert/conversion.py,sha256=5uPwHhmc6kwiIz-CqaiHDejf2SOWMHrb-rYEHm69wKc,3801
 ai_edge_torch/_convert/conversion_utils.py,sha256=Sr8qXVcTwc-ZnZmK7yxVrIOOp1S_vNrwzC0zUvLTI2o,2160
@@ -63,6 +63,10 @@ ai_edge_torch/generative/examples/phi/phi2.py,sha256=82SEKRwtKfT9VcNQaykGmemiov_
 ai_edge_torch/generative/examples/phi/phi3.py,sha256=Xh-l7TQdXYZJ9PViRVk2_y91Ec7Yntn0UpkuzRIG3T8,9231
 ai_edge_torch/generative/examples/phi/verify.py,sha256=YPFCdbnfmvq38fbpBNr0kHPfSZo4p3_6WkLJAW3pLPo,2177
 ai_edge_torch/generative/examples/phi/verify_phi3.py,sha256=kVYaBVvddfQng0IyZGxyTJEzhiPO0G4VFJm2WOc2Q94,2360
+ai_edge_torch/generative/examples/qwen/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/qwen/convert_to_tflite.py,sha256=QAAVoSKDVf2rHAChzumGloVCWIU0Oe5UYKgv3T192Iw,2496
+ai_edge_torch/generative/examples/qwen/qwen.py,sha256=b03q1On6JzPhJzTs1dQwT_tJjO7C9NYmyzrzV2kQ_yo,4579
+ai_edge_torch/generative/examples/qwen/verify.py,sha256=9_AyEJTeUfvhhID64Rto2bflFPyXMFokdQLsseLUMiI,2775
 ai_edge_torch/generative/examples/smollm/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/smollm/convert_to_tflite.py,sha256=zPrDTDeRVWFi9DS32uNi-RLpzOStFOk5MhNla4ixeew,2179
 ai_edge_torch/generative/examples/smollm/smollm.py,sha256=dal8vnZjQd6vR7sc76-FYGDKUlVjOlfUALV-pwbXJGc,3264
@@ -121,7 +125,7 @@ ai_edge_torch/generative/test/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudj
 ai_edge_torch/generative/test/test_kv_cache.py,sha256=W6Bh0gYDzmwb0j9HdD5_D7Z7FPToP2HSyFrmwIXuFqo,3793
 ai_edge_torch/generative/test/test_loader.py,sha256=8y74ChO3CZCfEi1eCf3-w47kRgAI4qPYCXpi8rTQXMA,3378
 ai_edge_torch/generative/test/test_model_conversion.py,sha256=s-EVLOQGjIeVtgNI8Ggs37pkRdErAliT6NhrrFigPOE,5459
-ai_edge_torch/generative/test/test_model_conversion_large.py,sha256=kCm-L3rWbPj25E_QEbkSLiaCk3y23SjKJs-MG-EwKug,8545
+ai_edge_torch/generative/test/test_model_conversion_large.py,sha256=SBGHbY8-k7kSEEv-WQQlxGIYtJEVBIbjJPygGdDg9Qg,8921
 ai_edge_torch/generative/test/test_quantize.py,sha256=8geJhKwYBU20m0mdGPD1BUFwQ0lZKNtCB04SOLO18y4,5980
 ai_edge_torch/generative/test/utils.py,sha256=YvEhO2HIj1LkBs5du1UxY-cGRW9HMyAYsOUhgsTrTpA,1796
 ai_edge_torch/generative/utilities/__init__.py,sha256=-_jxnnFnCgnTU4oTm4MnRsvL5lqhomBNdFBbqfmfHPo,720
@@ -177,8 +181,8 @@ 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.dev20240928.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
-ai_edge_torch_nightly-0.3.0.dev20240928.dist-info/METADATA,sha256=3HuAFZTfvmU787dVypwpmUvo4DdZSekGsqGimO-oPfM,1897
-ai_edge_torch_nightly-0.3.0.dev20240928.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-ai_edge_torch_nightly-0.3.0.dev20240928.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
-ai_edge_torch_nightly-0.3.0.dev20240928.dist-info/RECORD,,
+ai_edge_torch_nightly-0.3.0.dev20240930.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.3.0.dev20240930.dist-info/METADATA,sha256=sFwzdW-SmFri3T15doM8okOIAWKw_GX4xWzAlO6GK7o,1897
+ai_edge_torch_nightly-0.3.0.dev20240930.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+ai_edge_torch_nightly-0.3.0.dev20240930.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.3.0.dev20240930.dist-info/RECORD,,