ai-edge-torch-nightly 0.2.0.dev20240714__py3-none-any.whl

ai_edge_torch/generative/layers/unet/builder.py

@@ -0,0 +1,47 @@
+# 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.
+# ==============================================================================
+# Builder utils for individual components.
+
+from torch import nn
+
+import ai_edge_torch.generative.layers.unet.model_config as unet_config
+
+
+def build_upsampling(config: unet_config.UpSamplingConfig):
+  if config.mode == unet_config.SamplingType.NEAREST:
+    return nn.UpsamplingNearest2d(scale_factor=config.scale_factor)
+  elif config.mode == unet_config.SamplingType.BILINEAR:
+    return nn.UpsamplingBilinear2d(scale_factor=config.scale_factor)
+  else:
+    raise ValueError("Unsupported upsampling type.")
+
+
+def build_downsampling(config: unet_config.DownSamplingConfig):
+  if config.mode == unet_config.SamplingType.AVERAGE:
+    return nn.AvgPool2d(config.kernel_size, config.stride, padding=config.padding)
+  elif config.mode == unet_config.SamplingType.CONVOLUTION:
+    out_channels = (
+        config.in_channels if config.out_channels is None else config.out_channels
+    )
+    padding = (0, 1, 0, 1) if config.padding == 0 else config.padding
+    return nn.Conv2d(
+        config.in_channels,
+        out_channels=out_channels,
+        kernel_size=config.kernel_size,
+        stride=config.stride,
+        padding=padding,
+    )
+  else:
+    raise ValueError("Unsupported downsampling type.")

ai_edge_torch/generative/layers/unet/model_config.py

@@ -0,0 +1,269 @@
+# 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.
+# ==============================================================================
+
+# UNet configuration class.
+from dataclasses import dataclass
+from dataclasses import field
+import enum
+from typing import List, Optional
+
+import ai_edge_torch.generative.layers.model_config as layers_cfg
+
+
+@enum.unique
+class SamplingType(enum.Enum):
+  NEAREST = enum.auto()
+  BILINEAR = enum.auto()
+  AVERAGE = enum.auto()
+  CONVOLUTION = enum.auto()
+
+
+@dataclass
+class UpSamplingConfig:
+  mode: SamplingType
+  scale_factor: float
+
+
+@dataclass
+class DownSamplingConfig:
+  mode: SamplingType
+  in_channels: int
+  kernel_size: int
+  stride: int
+  padding: int
+  out_channels: Optional[int] = None
+
+
+@dataclass
+class ResidualBlock2DConfig:
+  in_channels: int
+  out_channels: int
+  normalization_config: layers_cfg.NormalizationConfig
+  activation_config: layers_cfg.ActivationConfig
+  # Optional time embedding channels if the residual block takes a time embedding context as input
+  time_embedding_channels: Optional[int] = None
+
+
+@dataclass
+class AttentionBlock2DConfig:
+  dim: int
+  normalization_config: layers_cfg.NormalizationConfig
+  attention_config: layers_cfg.AttentionConfig
+  enable_hlfb: bool = True
+  attention_batch_size: int = 1
+
+
+@dataclass
+class CrossAttentionBlock2DConfig:
+  query_dim: int
+  cross_dim: int
+  normalization_config: layers_cfg.NormalizationConfig
+  attention_config: layers_cfg.AttentionConfig
+  enable_hlfb: bool = True
+  attention_batch_size: int = 1
+
+
+@dataclass
+class FeedForwardBlock2DConfig:
+  dim: int
+  hidden_dim: int
+  normalization_config: layers_cfg.NormalizationConfig
+  activation_config: layers_cfg.ActivationConfig
+  use_bias: bool
+
+
+@dataclass
+class TransformerBlock2DConfig:
+  pre_conv_normalization_config: layers_cfg.NormalizationConfig
+  attention_block_config: AttentionBlock2DConfig
+  cross_attention_block_config: CrossAttentionBlock2DConfig
+  feed_forward_block_config: FeedForwardBlock2DConfig
+
+
+@dataclass
+class UpDecoderBlock2DConfig:
+  in_channels: int
+  out_channels: int
+  normalization_config: layers_cfg.NormalizationConfig
+  activation_config: layers_cfg.ActivationConfig
+  num_layers: int
+  # Optional time embedding channels if the residual blocks take a time embedding as input
+  time_embedding_channels: Optional[int] = None
+  # Whether to add upsample operation after residual blocks
+  add_upsample: bool = True
+  # Whether to add a conv2d layer after upsample
+  upsample_conv: bool = True
+  # Optional sampling config if add_upsample is True.
+  sampling_config: Optional[UpSamplingConfig] = None
+  # Optional config of transformer blocks interleaved with residual blocks
+  transformer_block_config: Optional[TransformerBlock2DConfig] = None
+  # Optional dimension of context tensor if context tensor is given as input.
+  context_dim: Optional[int] = None
+
+
+@dataclass
+class SkipUpDecoderBlock2DConfig:
+  in_channels: int
+  out_channels: int
+  # The dimension of output channels of previous connected block
+  prev_out_channels: int
+  normalization_config: layers_cfg.NormalizationConfig
+  activation_config: layers_cfg.ActivationConfig
+  num_layers: int
+  # Optional time embedding channels if the residual blocks take a time embedding as input
+  time_embedding_channels: Optional[int] = None
+  # Whether to add upsample operation after residual blocks
+  add_upsample: bool = True
+  # Whether to add a conv2d layer after upsample
+  upsample_conv: bool = True
+  # Optional sampling config if add_upsample is True.
+  sampling_config: Optional[UpSamplingConfig] = None
+  # Optional config of transformer blocks interleaved with residual blocks
+  transformer_block_config: Optional[TransformerBlock2DConfig] = None
+  # Optional dimension of context tensor if context tensor is given as input.
+  context_dim: Optional[int] = None
+
+
+@dataclass
+class DownEncoderBlock2DConfig:
+  in_channels: int
+  out_channels: int
+  normalization_config: layers_cfg.NormalizationConfig
+  activation_config: layers_cfg.ActivationConfig
+  num_layers: int
+  # Padding for the downsampling convolution.
+  padding: int = 1
+  # Optional time embedding channels if the residual blocks take a time embedding as input
+  time_embedding_channels: Optional[int] = None
+  # Whether to add downsample operation after residual blocks
+  add_downsample: bool = True
+  # Optional sampling config if add_upsample is True.
+  sampling_config: Optional[DownSamplingConfig] = None
+  # Optional config of transformer blocks interleaved with residual blocks
+  transformer_block_config: Optional[TransformerBlock2DConfig] = None
+  # Optional dimension of context tensor if context tensor is given as input.
+  context_dim: Optional[int] = None
+
+
+@dataclass
+class MidBlock2DConfig:
+  in_channels: int
+  normalization_config: layers_cfg.NormalizationConfig
+  activation_config: layers_cfg.ActivationConfig
+  num_layers: int
+  # Optional time embedding channels if the residual blocks take a time embedding context as input
+  time_embedding_channels: Optional[int] = None
+  # Optional config of attention blocks interleaved with residual blocks
+  attention_block_config: Optional[AttentionBlock2DConfig] = None
+  # Optional config of transformer blocks interleaved with residual blocks
+  transformer_block_config: Optional[TransformerBlock2DConfig] = None
+  # Optional dimension of context tensor if context tensor is given as input.
+  context_dim: Optional[int] = None
+
+
+@dataclass
+class AutoEncoderConfig:
+  """Configurations of encoder/decoder in the autoencoder model."""
+
+  # The activation type of encoder/decoder blocks.
+  activation_config: layers_cfg.ActivationConfig
+
+  # The output channels of each block.
+  block_out_channels: List[int]
+
+  # Number of channels in the input image.
+  in_channels: int
+
+  # Number of channels in the output.
+  out_channels: int
+
+  # Number of channels in the latent space.
+  latent_channels: int
+
+  # The component-wise standard deviation of the trained latent space computed using the first batch of the
+  # training set. This is used to scale the latent space to have unit variance when training the diffusion
+  # model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+  # diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+  # / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+  # Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+  scaling_factor: float
+
+  # The layesr number of each encoder/decoder block.
+  layers_per_block: int
+
+  # The normalization config.
+  normalization_config: layers_cfg.NormalizationConfig
+
+  # The configuration of middle blocks, that is, after the last block of encoder and before the first block of decoder.
+  mid_block_config: MidBlock2DConfig
+
+
+@dataclass
+class DiffusionModelConfig:
+  """Configurations of Diffusion model."""
+
+  # Number of channels in the input tensor.
+  in_channels: int
+
+  # Number of channels in the output tensor.
+  out_channels: int
+
+  # The output channels of each block.
+  block_out_channels: List[int]
+
+  # The layesr number of each block.
+  layers_per_block: int
+
+  # The padding to use for the downsampling.
+  downsample_padding: int
+
+  # Normalization config used in residual blocks.
+  residual_norm_config: layers_cfg.NormalizationConfig
+
+  # Activation config used in residual blocks
+  residual_activation_type: layers_cfg.ActivationType
+
+  # The batch size used in transformer blocks, for attention layers.
+  transformer_batch_size: int
+
+  # The number of attention heads used in transformer blocks.
+  transformer_num_attention_heads: int
+
+  # The dimension of cross attention used in transformer blocks.
+  transformer_cross_attention_dim: int
+
+  # Normalization config used in prev conv layer of transformer blocks.
+  transformer_pre_conv_norm_config: layers_cfg.NormalizationConfig
+
+  # Normalization config used in transformer blocks.
+  transformer_norm_config: layers_cfg.NormalizationConfig
+
+  # Activation type of feed forward used in transformer blocks.
+  transformer_ff_activation_type: layers_cfg.ActivationType
+
+  # Number of layers in mid block.
+  mid_block_layers: int
+
+  # Dimension of time embedding.
+  time_embedding_dim: int
+
+  # Time embedding dimensions for blocks.
+  time_embedding_blocks_dim: int
+
+  # Normalization config used for final layer
+  final_norm_config: layers_cfg.NormalizationConfig
+
+  # Activation type used in final layer
+  final_activation_type: layers_cfg.ActivationType

ai_edge_torch/generative/quantize/__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/quantize/ai_edge_quantizer_glue/translate_recipe.py

@@ -0,0 +1,148 @@
+# 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.
+# ==============================================================================
+
+import json
+
+from ai_edge_quantizer import quantizer
+
+from ai_edge_torch.generative.quantize import quant_attrs
+from ai_edge_torch.generative.quantize import quant_recipe
+
+_OpExecutionMode = quantizer.qtyping.OpExecutionMode
+_OpName = quantizer.qtyping.TFLOperationName
+_TensorQuantConfig = quantizer.qtyping.TensorQuantizationConfig
+_OpQuantConfig = quantizer.qtyping.OpQuantizationConfig
+
+_DEFAULT_REGEX_STR = '.*'
+_SINGULAR_TRANSFORMER_BLOCK_REGEX_STR = 'transformer_block'
+_IDX_TRANSFORMER_BLOCKS_REGEX_STR = 'transformer_blocks\[{}\]'
+_ATTENTION_REGEX_STR = 'ai_edge_torch.generative.layers.attention'
+_FEEDFORWARD_REGEX_STR = 'ai_edge_torch.generative.layers.feed_forward'
+_EMBEDDING_REGEX_STR = 'Embedding_tok_embedding'
+_ANY_TWO_DIGITS_REGEX_STR = '\d{1,2}'
+
+
+def _get_nbits_from_dtype(dtype: quant_attrs.Dtype) -> int:
+  if dtype == quant_attrs.Dtype.FP32:
+    return 32
+  elif dtype == quant_attrs.Dtype.FP16:
+    return 16
+  elif dtype == quant_attrs.Dtype.INT8:
+    return 8
+  raise ValueError('Unimplemented number of bits')
+
+
+def _get_dtype_from_dtype(dtype: quant_attrs.Dtype) -> quantizer.qtyping.TensorDataType:
+  if dtype == quant_attrs.Dtype.FP32 or dtype == quant_attrs.Dtype.FP16:
+    return quantizer.qtyping.TensorDataType.FLOAT
+  else:
+    return quantizer.qtyping.TensorDataType.INT
+
+
+def _get_execution_mode_from_mode(mode: quant_attrs.Mode) -> _OpExecutionMode:
+  if mode == quant_attrs.Mode.DYNAMIC_RANGE:
+    return _OpExecutionMode.DRQ
+  elif mode == quant_attrs.Mode.WEIGHT_ONLY:
+    return _OpExecutionMode.WEIGHT_ONLY
+  raise ValueError('Unimplemented execution mode')
+
+
+def _get_channelwise_from_granularity(granularity: quant_attrs.Granularity) -> bool:
+  if granularity == quant_attrs.Granularity.CHANNELWISE:
+    return True
+  elif granularity == quant_attrs.Granularity.NONE:
+    return False
+  raise ValueError('Unimplemented granularity')
+
+
+def _get_algorithm_key_from_algorithm(algo: quant_attrs.Algorithm) -> str:
+  if algo == quant_attrs.Algorithm.MIN_MAX:
+    return quantizer.algorithm_manager.AlgorithmName.MIN_MAX_UNIFORM_QUANT
+  elif algo == quant_attrs.Algorithm.FLOAT_CAST:
+    return quantizer.algorithm_manager.AlgorithmName.FLOAT_CASTING
+  raise ValueError('Unimplemented algorithm')
+
+
+def _set_quant_config(
+    rm: quantizer.recipe_manager.RecipeManager,
+    layer_recipe: quant_recipe.LayerQuantRecipe,
+    regex: str,
+):
+  rm.add_quantization_config(
+      regex=regex,
+      operation_name=_OpName.ALL_SUPPORTED,
+      op_config=_OpQuantConfig(
+          weight_tensor_config=_TensorQuantConfig(
+              num_bits=_get_nbits_from_dtype(layer_recipe.weight_dtype),
+              symmetric=True,
+              channel_wise=_get_channelwise_from_granularity(layer_recipe.granularity),
+              dtype=_get_dtype_from_dtype(layer_recipe.weight_dtype),
+          ),
+          execution_mode=_get_execution_mode_from_mode(layer_recipe.mode),
+      ),
+      algorithm_key=_get_algorithm_key_from_algorithm(layer_recipe.algorithm),
+  )
+
+
+def translate_to_ai_edge_recipe(
+    recipe: quant_recipe.GenerativeQuantRecipe,
+) -> quantizer.recipe_manager.ModelQuantizationRecipe:
+  rm = quantizer.recipe_manager.RecipeManager()
+
+  if recipe.default is not None:
+    _set_quant_config(rm, recipe.default, _DEFAULT_REGEX_STR)
+
+  if recipe.embedding is not None:
+    _set_quant_config(rm, recipe.embedding, _EMBEDDING_REGEX_STR)
+
+  if recipe.attention is not None:
+    if isinstance(recipe.attention, dict):
+      for idx, layer in recipe.attention.items():
+        _set_quant_config(
+            rm,
+            layer,
+            f'{_IDX_TRANSFORMER_BLOCKS_REGEX_STR.format(idx)}/{_ATTENTION_REGEX_STR}',
+        )
+    else:
+      _set_quant_config(
+          rm,
+          recipe.attention,
+          f'{_SINGULAR_TRANSFORMER_BLOCK_REGEX_STR}/{_ATTENTION_REGEX_STR}',
+      )
+
+  if recipe.feedforward is not None:
+    if isinstance(recipe.feedforward, dict):
+      for idx, layer in recipe.feedforward.items():
+        _set_quant_config(
+            rm,
+            layer,
+            f'{_IDX_TRANSFORMER_BLOCKS_REGEX_STR.format(idx)}/{_FEEDFORWARD_REGEX_STR}',
+        )
+    else:
+      _set_quant_config(
+          rm,
+          recipe.feedforward,
+          f'{_SINGULAR_TRANSFORMER_BLOCK_REGEX_STR}/{_FEEDFORWARD_REGEX_STR}',
+      )
+
+  return rm.get_quantization_recipe()
+
+
+def quantize_model(
+    model: bytearray, recipe: quantizer.recipe_manager.ModelQuantizationRecipe
+) -> bytearray:
+  qt = quantizer.Quantizer(bytearray(model), recipe)
+  result = qt.quantize()
+  return result.quantized_model

ai_edge_torch/generative/quantize/example.py

@@ -0,0 +1,45 @@
+# Copyright 2024 The AI Edge Torch Authors. All Rights Reserved.
+#
+# 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.
+# ==============================================================================
+
+import numpy as np
+import torch
+
+import ai_edge_torch
+from ai_edge_torch.generative.examples.gemma import gemma
+from ai_edge_torch.generative.quantize import quant_recipes
+
+
+def main():
+  # Build a PyTorch model as usual
+  config = gemma.get_fake_model_config_2b_for_test()
+  model = gemma.Gemma(config)
+  idx = torch.from_numpy(np.array([[1, 2, 3, 4]]))
+  tokens = torch.full((1, 10), 0, dtype=torch.long, device="cpu")
+  tokens[0, :4] = idx
+  input_pos = torch.arange(0, 10)
+
+  # Create a quantization recipe to be applied to the model
+  quant_config = quant_recipes.full_int8_dynamic_recipe()
+  print(quant_config)
+
+  # Convert with quantization
+  edge_model = ai_edge_torch.convert(
+      model, (tokens, input_pos), quant_config=quant_config
+  )
+  edge_model.export("/tmp/gemma_2b_quantized.tflite")
+
+
+if __name__ == "__main__":
+  main()

ai_edge_torch/generative/quantize/quant_attrs.py

@@ -0,0 +1,68 @@
+# Copyright 2024 The AI Edge Torch Authors. All Rights Reserved.
+#
+# 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.
+# ==============================================================================
+
+import enum
+
+
+@enum.unique
+class Dtype(enum.Enum):
+  """Data types and precision of tensors."""
+
+  FP32 = enum.auto()
+  FP16 = enum.auto()
+  INT8 = enum.auto()
+
+
+@enum.unique
+class Algorithm(enum.Enum):
+  """Algorithm used to calculate quantization parameters.
+
+  Attributes:
+    MIN_MAX: Maps the min/max of floating point space to the min/max of
+      quantized space and quantize uniformly.
+    FLOAT_CAST: Casts a float to another float of a different type.
+  """
+
+  MIN_MAX = enum.auto()
+  FLOAT_CAST = enum.auto()
+
+
+@enum.unique
+class Mode(enum.Enum):
+  """Mode of quantization.
+
+  Attributes:
+    DYNAMIC_RANGE: Quantize activations during runtime and weights statically to
+      perform computation in integers.
+    WEIGHT_ONLY: Quantize weights statically and dequantize during runtime to
+      perform computation in floating points.
+  """
+
+  DYNAMIC_RANGE = enum.auto()
+  WEIGHT_ONLY = enum.auto()
+
+
+@enum.unique
+class Granularity(enum.Enum):
+  """Granularity of quantization parameters.
+
+  Attributes:
+    NONE: Granularity not applicable to this quantization scheme.
+    CHANNELWISE: Or per-channel quantization. Each channel of relevant tensors
+      is quantized independently of one another.
+  """
+
+  NONE = enum.auto()
+  CHANNELWISE = enum.auto()