ai-edge-torch-nightly 0.3.0.dev20250114__py3-none-any.whl

ai_edge_torch/generative/examples/llama/convert_to_tflite.py

@@ -0,0 +1,91 @@
+# 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 Llama 3.2 1B model to multi-signature tflite model."""
+
+import os
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.llama import llama
+from ai_edge_torch.generative.utilities import converter
+from ai_edge_torch.generative.utilities.model_builder import ExportConfig
+
+_MODEL_SIZE = flags.DEFINE_enum(
+    'model_size',
+    '1b',
+    ['1b', '3b'],
+    'The size of the model to verify.',
+)
+_CHECKPOINT_PATH = flags.DEFINE_string(
+    'checkpoint_path',
+    os.path.join(pathlib.Path.home(), 'Downloads/llm_data/llama'),
+    'The path to the model checkpoint, or directory holding the checkpoint.',
+)
+_OUTPUT_PATH = flags.DEFINE_string(
+    'output_path',
+    '/tmp/',
+    'The path to export the tflite model.',
+)
+_OUTPUT_NAME_PREFIX = flags.DEFINE_string(
+    'output_name_prefix',
+    'llama',
+    'The prefix of the output tflite model name.',
+)
+_PREFILL_SEQ_LENS = flags.DEFINE_multi_integer(
+    'prefill_seq_lens',
+    (8, 64, 128, 256, 512, 1024),
+    'List of the maximum sizes of prefill input tensors.',
+)
+_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.',
+)
+_LORA_RANKS = flags.DEFINE_multi_integer(
+    'lora_ranks',
+    None,
+    'If set, the model will be converted with the provided list of LoRA ranks.',
+)
+
+_BUILDER = {
+    '1b': llama.build_1b_model,
+    '3b': llama.build_3b_model,
+}
+
+
+def main(_):
+  pytorch_model = _BUILDER[_MODEL_SIZE.value](
+      _CHECKPOINT_PATH.value, kv_cache_max_len=_KV_CACHE_MAX_LEN.value
+  )
+  converter.convert_to_tflite(
+      pytorch_model,
+      output_path=_OUTPUT_PATH.value,
+      output_name_prefix=_OUTPUT_NAME_PREFIX.value,
+      prefill_seq_len=_PREFILL_SEQ_LENS.value,
+      quantize=_QUANTIZE.value,
+      lora_ranks=_LORA_RANKS.value,
+      export_config=ExportConfig(),
+  )
+
+
+if __name__ == '__main__':
+  app.run(main)

ai_edge_torch/generative/examples/llama/llama.py

@@ -0,0 +1,196 @@
+# 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 Llama 3.2 models."""
+
+from functools import partial
+import math
+from typing import Tuple
+
+import ai_edge_torch.generative.layers.model_config as cfg
+from ai_edge_torch.generative.utilities import model_builder
+import torch
+
+TENSOR_NAMES = model_builder.TENSOR_NAMES
+
+
+def _build_llama3_rope_cache(
+    input_pos: torch.Tensor,
+    n_elem: int,
+    base: int,
+    condense_ratio: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    factor: float,
+    low_freq_factor: float,
+    high_freq_factor: float,
+    max_seq_len: int,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+  """Computes Rotary Positional Embeddings for Llama 3.2 model.
+
+  It's a modified version of attn_utils.build_rope_cache with additional
+  arguments for Llama 3.2 model. It precomputes Rotary Positional Embedding Sin
+  and Cos values with scaling factors for quick lookup during the inference.
+
+  Reference:
+  https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_rope_utils.py#L307
+
+  Args:
+      input_pos (torch.Tensor): the given input sequence positions
+      n_elem (int): Each sequence's dimmension.
+      base (int): Rope base value.
+      condense_ratio (int): The ratio by which sequence indicies are condensed.
+      dtype (torch.dtype): Output tensor's data type.
+      device (torch.device): Output tensor's data type.
+      factor (float): Factor to scale theta down for tokens in long range in the
+        sequence.
+      low_freq_factor (float): Factor to determine if tokens are in long range
+        in the sequence.
+      high_freq_factor (float): Factor to determine if tokens are in short range
+        in the sequence.
+      max_seq_len (int): The original token sequence length before extending
+        ROPE to support longer sequence.
+
+  Returns:
+      Tuple[torch.Tensor, torch.Tensor]: Rope's Cosine and Sine waves.
+  """
+  theta = 1.0 / (base ** (torch.arange(0, n_elem, 2).float() / n_elem))
+  low_freq_wavelen = max_seq_len / low_freq_factor
+  high_freq_wavelen = max_seq_len / high_freq_factor
+  wavelen = 2 * math.pi / theta
+  # wavelen < high_freq_wavelen: do nothing
+  # wavelen > low_freq_wavelen: divide by factor
+  theta = torch.where(wavelen > low_freq_wavelen, theta / factor, theta)
+  # otherwise: interpolate between the two, using a smooth factor
+  smooth_factor = (max_seq_len / wavelen - low_freq_factor) / (
+      high_freq_factor - low_freq_factor
+  )
+  smoothed_theta = (1 - smooth_factor) * theta / factor + smooth_factor * theta
+  is_medium = ~(wavelen < high_freq_wavelen) * ~(wavelen > low_freq_wavelen)
+  theta = torch.where(is_medium, smoothed_theta, theta)
+
+  seq_idx = input_pos / condense_ratio
+  idx_theta = torch.outer(seq_idx, theta)
+  cos = torch.cos(idx_theta).to(dtype=dtype, device=device)
+  sin = torch.sin(idx_theta).to(dtype=dtype, device=device)
+  return cos, sin
+
+
+class Llama(model_builder.DecoderOnlyModel):
+  """A Llama model built from the Edge Generative API layers.
+
+  Llama 3.2 shares the same architecture as TinyLlama except ROPE calculation.
+  """
+
+  def __init__(self, config: cfg.ModelConfig):
+    super().__init__(config)
+    attn_config = self.config.block_config(0).attn_config
+
+
+def get_1b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Llama 3.2-1B 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=32,
+      head_dim=64,
+      num_query_groups=8,
+      rotary_base=500000,
+      rotary_percentage=1.0,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.GATED,
+      activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
+      intermediate_size=8192,
+  )
+  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)
+  block_config = cfg.TransformerBlockConfig(
+      attn_config=attn_config,
+      ff_config=ff_config,
+      pre_attention_norm_config=norm_config,
+      post_attention_norm_config=norm_config,
+  )
+
+  max_seq_len = 8192
+  # Create the RoPE callable
+  build_rope = partial(
+      _build_llama3_rope_cache,
+      condense_ratio=1,
+      dtype=torch.float32,
+      device=torch.device("cpu"),
+      factor=32.0,
+      low_freq_factor=1.0,
+      high_freq_factor=4.0,
+      max_seq_len=max_seq_len,
+  )
+
+  config = cfg.ModelConfig(
+      vocab_size=128256,
+      num_layers=16,
+      max_seq_len=max_seq_len,
+      embedding_dim=2048,
+      kv_cache_max_len=kv_cache_max_len,
+      block_configs=block_config,
+      final_norm_config=norm_config,
+      enable_hlfb=True,
+      build_rope=build_rope,
+  )
+  return config
+
+
+def get_3b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Llama 3.2-3B model."""
+  config = get_1b_model_config(kv_cache_max_len)
+  # Llama 3.2 has only one block config.
+  attn_config = config.block_config(0).attn_config
+  attn_config.num_heads = 24
+  attn_config.head_dim = 128
+  config.num_layers = 28
+  config.embedding_dim = 3072
+  return config
+
+
+def get_fake_model_config(**kwargs) -> cfg.ModelConfig:
+  config = get_1b_model_config(**kwargs)
+  config.vocab_size = 128
+  config.num_layers = 2
+  # SmolLM 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
+) -> torch.nn.Module:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=config,
+      tensor_names=TENSOR_NAMES,
+      model_class=Llama,
+  )
+
+
+def build_1b_model(checkpoint_path: str, **kwargs) -> torch.nn.Module:
+  return _build_model(checkpoint_path, get_1b_model_config(**kwargs))
+
+
+def build_3b_model(checkpoint_path: str, **kwargs) -> torch.nn.Module:
+  return _build_model(checkpoint_path, get_3b_model_config(**kwargs))

ai_edge_torch/generative/examples/llama/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 Llama 3.2-1B model."""
+
+import logging
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.llama import llama
+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",
+    "1b",
+    ["1b", "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 = {
+    "1b": "meta-llama/Llama-3.2-1B-Instruct",
+    "3b": "meta-llama/Llama-3.2-3B-Instruct",
+}
+
+_BUILDER = {
+    "1b": llama.build_1b_model,
+    "3b": llama.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)
+  # Llama tokenizer_config.json sets a fast tokenizer class explicitly,
+  # "PreTrainedTokenizerFast". It works only when the fast tokenizer is
+  # available.
+  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/examples/moonshine/__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/moonshine/convert_moonshine_to_tflite.py

@@ -0,0 +1,50 @@
+# 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 a Moonshine model to multi-signature tflite model."""
+
+import os
+import pathlib
+
+from absl import app
+from absl import flags
+import ai_edge_torch
+from ai_edge_torch.generative.examples.moonshine import moonshine
+from ai_edge_torch.generative.utilities import converter
+import torch
+
+_CHECKPOINT_PATH = flags.DEFINE_string(
+    'checkpoint_path',
+    os.path.join(pathlib.Path.home(), 'Downloads/llm_data/moonshine'),
+    '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.',
+)
+
+
+def main(_):
+  p_model = moonshine.build_preprocessor(_CHECKPOINT_PATH.value)
+  output_filename = f'moonshine_preprocessor.tflite'
+  _input = torch.randn((1, 1, 159414), dtype=torch.float)
+  edge_model = ai_edge_torch.convert(p_model, (_input,), quant_config=None)
+  tflite_path = os.path.join(_TFLITE_PATH.value, output_filename)
+  edge_model.export(tflite_path)
+
+
+if __name__ == '__main__':
+  app.run(main)

ai_edge_torch/generative/examples/moonshine/moonshine.py

@@ -0,0 +1,103 @@
+# 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 the Moonshine model."""
+
+import os
+import pathlib
+from typing import Optional, Tuple
+from absl import app
+from ai_edge_torch.generative.layers import attention
+from ai_edge_torch.generative.layers import builder
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
+import ai_edge_torch.generative.layers.attention_utils as attn_utils
+import ai_edge_torch.generative.layers.model_config as cfg
+import ai_edge_torch.generative.layers.normalization as normalization
+import ai_edge_torch.generative.utilities.moonshine_loader as loading_utils
+import h5py
+import torch
+from torch import nn
+import torch.nn as nn
+
+TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
+    conv1D_0="layers/sequential/layers/conv1d/vars",
+    conv1D_1="layers/sequential/layers/conv1d_1/vars",
+    conv1D_2="layers/sequential/layers/conv1d_2/vars",
+    group_norm="layers/sequential/layers/group_normalization/vars",
+)
+
+
+class AudioPreprocessor(nn.Module):
+
+  def __init__(self, dim):
+    super(AudioPreprocessor, self).__init__()
+    self.conv1 = nn.Conv1d(
+        in_channels=1, out_channels=dim, kernel_size=127, stride=64, bias=False
+    )
+    self.tanh = nn.Tanh()
+    self.group_norm = normalization.GroupNorm(group_num=1, dim=dim, eps=1e-5)
+    self.conv2 = nn.Conv1d(
+        in_channels=dim,
+        out_channels=2 * dim,
+        kernel_size=7,
+        stride=3,
+        padding=0,  # Equivalent to padding="valid"
+    )
+    self.gelu1 = nn.GELU()
+    self.conv3 = nn.Conv1d(
+        in_channels=2 * dim,
+        out_channels=dim,
+        kernel_size=3,
+        stride=2,
+        padding=0,  # Equivalent to padding="valid"
+    )
+    self.gelu2 = nn.GELU()
+
+  def forward(self, inputs):
+    x = self.conv1(inputs)
+    x = self.tanh(x)
+    x = self.group_norm(x)
+    x = self.conv2(x)
+    x = self.gelu1(x)
+    x = self.conv3(x)
+    x = self.gelu2(x)
+    return x
+
+
+def build_preprocessor(checkpoint_path: str, **kwargs) -> nn.Module:
+  ap = AudioPreprocessor(dim=416)
+  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
+  loader.load(ap, strict=True)
+  return ap
+
+
+def main(_):
+  # TODO(b/375421767) Remove golden checks once full model is implemented.
+  HF_PATH = os.path.join(pathlib.Path.home(), "Downloads/llm_data/moonshine")
+
+  test_data_path = pathlib.Path(__file__).parent.resolve()
+  INPUT_PATH = test_data_path / "data" / "pp_input.pt")
+  GOLDEN_PATH = test_data_path / "data" / "pp_output.pt")
+
+  ap = build_preprocessor(HF_PATH)
+  ap.eval()
+  inputs = torch.load(INPUT_PATH).reshape((1, 1, 159414))
+  out = ap(inputs)
+  golden = torch.load(GOLDEN_PATH).transpose(1, 2)
+  assert torch.allclose(out, golden, atol=1e-4, rtol=1e-4)
+
+
+if __name__ == "__main__":
+  app.run(main)

ai_edge_torch/generative/examples/openelm/__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/openelm/convert_to_tflite.py

@@ -0,0 +1,80 @@
+# 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 OpenELM model to multi-signature tflite model."""
+
+import os
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.openelm import openelm
+from ai_edge_torch.generative.utilities import converter
+from ai_edge_torch.generative.utilities.model_builder import ExportConfig
+
+_CHECKPOINT_PATH = flags.DEFINE_string(
+    'checkpoint_path',
+    os.path.join(pathlib.Path.home(), 'Downloads/llm_data/openelm'),
+    'The path to the model checkpoint, or directory holding the checkpoint.',
+)
+_OUTPUT_PATH = flags.DEFINE_string(
+    'output_path',
+    '/tmp/',
+    'The path to export the tflite model.',
+)
+_OUTPUT_NAME_PREFIX = flags.DEFINE_string(
+    'output_name_prefix',
+    'openelm',
+    'The prefix of the output tflite model name.',
+)
+_PREFILL_SEQ_LENS = flags.DEFINE_multi_integer(
+    'prefill_seq_lens',
+    (8, 64, 128, 256, 512, 1024),
+    'List of the maximum sizes of prefill input tensors.',
+)
+_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.',
+)
+_LORA_RANKS = flags.DEFINE_multi_integer(
+    'lora_ranks',
+    None,
+    'If set, the model will be converted with the provided list of LoRA ranks.',
+)
+
+
+def main(_):
+  pytorch_model = openelm.build_model(
+      _CHECKPOINT_PATH.value, kv_cache_max_len=_KV_CACHE_MAX_LEN.value
+  )
+  converter.convert_to_tflite(
+      pytorch_model,
+      output_path=_OUTPUT_PATH.value,
+      output_name_prefix=_OUTPUT_NAME_PREFIX.value,
+      prefill_seq_len=_PREFILL_SEQ_LENS.value,
+      quantize=_QUANTIZE.value,
+      lora_ranks=_LORA_RANKS.value,
+      export_config=ExportConfig(),
+  )
+
+
+if __name__ == '__main__':
+  app.run(main)