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

ai_edge_torch/generative/examples/stable_diffusion/decoder.py

@@ -0,0 +1,398 @@
+# 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 ai_edge_torch.generative.layers.builder as layers_builder
+import ai_edge_torch.generative.layers.model_config as layers_cfg
+from ai_edge_torch.generative.layers.unet import blocks_2d
+import ai_edge_torch.generative.layers.unet.model_config as unet_cfg
+from ai_edge_torch.generative.utilities import stable_diffusion_loader
+import torch
+from torch import nn
+
+TENSOR_NAMES = stable_diffusion_loader.AutoEncoderModelLoader.TensorNames(
+    post_quant_conv="first_stage_model.post_quant_conv",
+    conv_in="first_stage_model.decoder.conv_in",
+    mid_block_tensor_names=stable_diffusion_loader.MidBlockTensorNames(
+        residual_block_tensor_names=[
+            stable_diffusion_loader.ResidualBlockTensorNames(
+                norm_1="first_stage_model.decoder.mid.block_1.norm1",
+                norm_2="first_stage_model.decoder.mid.block_1.norm2",
+                conv_1="first_stage_model.decoder.mid.block_1.conv1",
+                conv_2="first_stage_model.decoder.mid.block_1.conv2",
+            ),
+            stable_diffusion_loader.ResidualBlockTensorNames(
+                norm_1="first_stage_model.decoder.mid.block_2.norm1",
+                norm_2="first_stage_model.decoder.mid.block_2.norm2",
+                conv_1="first_stage_model.decoder.mid.block_2.conv1",
+                conv_2="first_stage_model.decoder.mid.block_2.conv2",
+            ),
+        ],
+        attention_block_tensor_names=[
+            stable_diffusion_loader.AttentionBlockTensorNames(
+                norm="first_stage_model.decoder.mid.attn_1.norm",
+                q_proj="first_stage_model.decoder.mid.attn_1.q",
+                k_proj="first_stage_model.decoder.mid.attn_1.k",
+                v_proj="first_stage_model.decoder.mid.attn_1.v",
+                output_proj="first_stage_model.decoder.mid.attn_1.proj_out",
+            )
+        ],
+    ),
+    up_decoder_blocks_tensor_names=[
+        stable_diffusion_loader.UpDecoderBlockTensorNames(
+            residual_block_tensor_names=[
+                stable_diffusion_loader.ResidualBlockTensorNames(
+                    norm_1="first_stage_model.decoder.up.3.block.0.norm1",
+                    norm_2="first_stage_model.decoder.up.3.block.0.norm2",
+                    conv_1="first_stage_model.decoder.up.3.block.0.conv1",
+                    conv_2="first_stage_model.decoder.up.3.block.0.conv2",
+                ),
+                stable_diffusion_loader.ResidualBlockTensorNames(
+                    norm_1="first_stage_model.decoder.up.3.block.1.norm1",
+                    norm_2="first_stage_model.decoder.up.3.block.1.norm2",
+                    conv_1="first_stage_model.decoder.up.3.block.1.conv1",
+                    conv_2="first_stage_model.decoder.up.3.block.1.conv2",
+                ),
+                stable_diffusion_loader.ResidualBlockTensorNames(
+                    norm_1="first_stage_model.decoder.up.3.block.2.norm1",
+                    norm_2="first_stage_model.decoder.up.3.block.2.norm2",
+                    conv_1="first_stage_model.decoder.up.3.block.2.conv1",
+                    conv_2="first_stage_model.decoder.up.3.block.2.conv2",
+                ),
+            ],
+            upsample_conv="first_stage_model.decoder.up.3.upsample.conv",
+        ),
+        stable_diffusion_loader.UpDecoderBlockTensorNames(
+            residual_block_tensor_names=[
+                stable_diffusion_loader.ResidualBlockTensorNames(
+                    norm_1="first_stage_model.decoder.up.2.block.0.norm1",
+                    norm_2="first_stage_model.decoder.up.2.block.0.norm2",
+                    conv_1="first_stage_model.decoder.up.2.block.0.conv1",
+                    conv_2="first_stage_model.decoder.up.2.block.0.conv2",
+                ),
+                stable_diffusion_loader.ResidualBlockTensorNames(
+                    norm_1="first_stage_model.decoder.up.2.block.1.norm1",
+                    norm_2="first_stage_model.decoder.up.2.block.1.norm2",
+                    conv_1="first_stage_model.decoder.up.2.block.1.conv1",
+                    conv_2="first_stage_model.decoder.up.2.block.1.conv2",
+                ),
+                stable_diffusion_loader.ResidualBlockTensorNames(
+                    norm_1="first_stage_model.decoder.up.2.block.2.norm1",
+                    norm_2="first_stage_model.decoder.up.2.block.2.norm2",
+                    conv_1="first_stage_model.decoder.up.2.block.2.conv1",
+                    conv_2="first_stage_model.decoder.up.2.block.2.conv2",
+                ),
+            ],
+            upsample_conv="first_stage_model.decoder.up.2.upsample.conv",
+        ),
+        stable_diffusion_loader.UpDecoderBlockTensorNames(
+            residual_block_tensor_names=[
+                stable_diffusion_loader.ResidualBlockTensorNames(
+                    norm_1="first_stage_model.decoder.up.1.block.0.norm1",
+                    norm_2="first_stage_model.decoder.up.1.block.0.norm2",
+                    conv_1="first_stage_model.decoder.up.1.block.0.conv1",
+                    conv_2="first_stage_model.decoder.up.1.block.0.conv2",
+                    residual_layer=(
+                        "first_stage_model.decoder.up.1.block.0.nin_shortcut"
+                    ),
+                ),
+                stable_diffusion_loader.ResidualBlockTensorNames(
+                    norm_1="first_stage_model.decoder.up.1.block.1.norm1",
+                    norm_2="first_stage_model.decoder.up.1.block.1.norm2",
+                    conv_1="first_stage_model.decoder.up.1.block.1.conv1",
+                    conv_2="first_stage_model.decoder.up.1.block.1.conv2",
+                ),
+                stable_diffusion_loader.ResidualBlockTensorNames(
+                    norm_1="first_stage_model.decoder.up.1.block.2.norm1",
+                    norm_2="first_stage_model.decoder.up.1.block.2.norm2",
+                    conv_1="first_stage_model.decoder.up.1.block.2.conv1",
+                    conv_2="first_stage_model.decoder.up.1.block.2.conv2",
+                ),
+            ],
+            upsample_conv="first_stage_model.decoder.up.1.upsample.conv",
+        ),
+        stable_diffusion_loader.UpDecoderBlockTensorNames(
+            residual_block_tensor_names=[
+                stable_diffusion_loader.ResidualBlockTensorNames(
+                    norm_1="first_stage_model.decoder.up.0.block.0.norm1",
+                    norm_2="first_stage_model.decoder.up.0.block.0.norm2",
+                    conv_1="first_stage_model.decoder.up.0.block.0.conv1",
+                    conv_2="first_stage_model.decoder.up.0.block.0.conv2",
+                    residual_layer=(
+                        "first_stage_model.decoder.up.0.block.0.nin_shortcut"
+                    ),
+                ),
+                stable_diffusion_loader.ResidualBlockTensorNames(
+                    norm_1="first_stage_model.decoder.up.0.block.1.norm1",
+                    norm_2="first_stage_model.decoder.up.0.block.1.norm2",
+                    conv_1="first_stage_model.decoder.up.0.block.1.conv1",
+                    conv_2="first_stage_model.decoder.up.0.block.1.conv2",
+                ),
+                stable_diffusion_loader.ResidualBlockTensorNames(
+                    norm_1="first_stage_model.decoder.up.0.block.2.norm1",
+                    norm_2="first_stage_model.decoder.up.0.block.2.norm2",
+                    conv_1="first_stage_model.decoder.up.0.block.2.conv1",
+                    conv_2="first_stage_model.decoder.up.0.block.2.conv2",
+                ),
+            ],
+        ),
+    ],
+    final_norm="first_stage_model.decoder.norm_out",
+    conv_out="first_stage_model.decoder.conv_out",
+)
+
+
+class Decoder(nn.Module):
+  """The Decoder model used in Stable Diffusion.
+
+  For details, see https://arxiv.org/abs/2103.00020
+
+  Sturcture of the Decoder:
+
+      latents tensor
+            |
+            ▼
+  ┌───────────────────┐
+  │  Post Quant Conv  │
+  └─────────┬─────────┘
+            │
+  ┌─────────▼─────────┐
+  │      ConvIn       │
+  └─────────┬─────────┘
+            │
+  ┌─────────▼─────────┐
+  │     MidBlock2D    │
+  └─────────┬─────────┘
+            │
+  ┌─────────▼─────────┐
+  │    UpDecoder2D    │ x 4
+  └─────────┬─────────┘
+            │
+  ┌─────────▼─────────┐
+  │     FinalNorm     │
+  └─────────┬─────────┘
+            |
+  ┌─────────▼─────────┐
+  │    Activation     │
+  └─────────┬─────────┘
+            |
+  ┌─────────▼─────────┐
+  │      ConvOut      │
+  └─────────┬─────────┘
+            |
+            ▼
+      Output Image
+  """
+
+  def __init__(self, config: unet_cfg.AutoEncoderConfig):
+    super().__init__()
+    self.config = config
+    self.post_quant_conv = nn.Conv2d(
+        config.latent_channels,
+        config.latent_channels,
+        kernel_size=1,
+        stride=1,
+        padding=0,
+    )
+    reversed_block_out_channels = list(reversed(config.block_out_channels))
+    self.conv_in = nn.Conv2d(
+        config.latent_channels,
+        reversed_block_out_channels[0],
+        kernel_size=3,
+        stride=1,
+        padding=1,
+    )
+    self.mid_block = blocks_2d.MidBlock2D(config.mid_block_config)
+    up_decoder_blocks = []
+    block_out_channels = reversed_block_out_channels[0]
+    for i, out_channels in enumerate(reversed_block_out_channels):
+      prev_output_channel = block_out_channels
+      block_out_channels = out_channels
+      not_final_block = i < len(reversed_block_out_channels) - 1
+      up_decoder_blocks.append(
+          blocks_2d.UpDecoderBlock2D(
+              unet_cfg.UpDecoderBlock2DConfig(
+                  in_channels=prev_output_channel,
+                  out_channels=block_out_channels,
+                  normalization_config=config.normalization_config,
+                  activation_config=config.activation_config,
+                  num_layers=config.layers_per_block,
+                  add_upsample=not_final_block,
+                  upsample_conv=True,
+                  sampling_config=unet_cfg.UpSamplingConfig(
+                      mode=unet_cfg.SamplingType.NEAREST, scale_factor=2
+                  ),
+              )
+          )
+      )
+    self.up_decoder_blocks = nn.ModuleList(up_decoder_blocks)
+    self.final_norm = layers_builder.build_norm(
+        block_out_channels, config.normalization_config
+    )
+    self.act_fn = layers_builder.get_activation(config.activation_config)
+    self.conv_out = nn.Conv2d(
+        block_out_channels,
+        config.out_channels,
+        kernel_size=3,
+        stride=1,
+        padding=1,
+    )
+
+  def forward(self, latents_tensor: torch.Tensor) -> torch.Tensor:
+    """Forward function of decoder model.
+
+    Args:
+        latents (torch.Tensor): latents space tensor.
+
+    Returns:
+        output decoded image tensor from decoder model.
+    """
+    x = latents_tensor / self.config.scaling_factor
+    x = self.post_quant_conv(x)
+    x = self.conv_in(x)
+    x = self.mid_block(x)
+    for up_decoder_block in self.up_decoder_blocks:
+      x = up_decoder_block(x)
+    x = self.final_norm(x)
+    x = self.act_fn(x)
+    x = self.conv_out(x)
+    return x
+
+
+def get_model_config(device_type: str = "cpu") -> unet_cfg.AutoEncoderConfig:
+  """Get configs for the Decoder of Stable Diffusion v1.5."""
+  in_channels = 3
+  latent_channels = 4
+  out_channels = 3
+  block_out_channels = [128, 256, 512, 512]
+  scaling_factor = 0.18215
+  layers_per_block = 3
+
+  # For now, only turns on StableHLO composite ops on GPU backend for better
+  # performance. CPU should also switch to it once the support is done.
+  enable_hlfb = True if device_type == "gpu" else False
+
+  norm_config = layers_cfg.NormalizationConfig(
+      layers_cfg.NormalizationType.GROUP_NORM,
+      group_num=32,
+      enable_hlfb=enable_hlfb,
+  )
+
+  att_config = unet_cfg.AttentionBlock2DConfig(
+      dim=block_out_channels[-1],
+      normalization_config=norm_config,
+      attention_config=layers_cfg.AttentionConfig(
+          num_heads=1,
+          head_dim=block_out_channels[-1],
+          num_query_groups=1,
+          qkv_use_bias=True,
+          output_proj_use_bias=True,
+          enable_kv_cache=False,
+          qkv_transpose_before_split=True,
+          qkv_fused_interleaved=False,
+          rotary_base=0,
+          rotary_percentage=0.0,
+      ),
+      enable_hlfb=enable_hlfb,
+  )
+
+  mid_block_config = unet_cfg.MidBlock2DConfig(
+      in_channels=block_out_channels[-1],
+      normalization_config=norm_config,
+      activation_config=layers_cfg.ActivationConfig(
+          layers_cfg.ActivationType.SILU
+      ),
+      num_layers=1,
+      attention_block_config=att_config,
+  )
+
+  config = unet_cfg.AutoEncoderConfig(
+      in_channels=in_channels,
+      latent_channels=latent_channels,
+      out_channels=out_channels,
+      activation_config=layers_cfg.ActivationConfig(
+          layers_cfg.ActivationType.SILU
+      ),
+      block_out_channels=block_out_channels,
+      scaling_factor=scaling_factor,
+      layers_per_block=layers_per_block,
+      normalization_config=norm_config,
+      mid_block_config=mid_block_config,
+  )
+  return config
+
+
+def get_fake_model_config(
+    device_type: str = "cpu",
+) -> unet_cfg.AutoEncoderConfig:
+  """Get fake configs for the Decoder of Stable Diffusion v1.5 for testing."""
+  in_channels = 3
+  latent_channels = 4
+  out_channels = 3
+  block_out_channels = [2, 4]
+  scaling_factor = 0.18215
+  layers_per_block = 2
+
+  # For now, only turns on StableHLO composite ops on GPU backend for better
+  # performance. CPU should also switch to it once the support is done.
+  enable_hlfb = True if device_type == "gpu" else False
+
+  norm_config = layers_cfg.NormalizationConfig(
+      layers_cfg.NormalizationType.GROUP_NORM,
+      group_num=2,
+      enable_hlfb=enable_hlfb,
+  )
+
+  att_config = unet_cfg.AttentionBlock2DConfig(
+      dim=block_out_channels[-1],
+      normalization_config=norm_config,
+      attention_config=layers_cfg.AttentionConfig(
+          num_heads=1,
+          head_dim=block_out_channels[-1],
+          num_query_groups=1,
+          qkv_use_bias=True,
+          output_proj_use_bias=True,
+          enable_kv_cache=False,
+          qkv_transpose_before_split=True,
+          qkv_fused_interleaved=False,
+          rotary_base=0,
+          rotary_percentage=0.0,
+      ),
+      enable_hlfb=enable_hlfb,
+  )
+
+  mid_block_config = unet_cfg.MidBlock2DConfig(
+      in_channels=block_out_channels[-1],
+      normalization_config=norm_config,
+      activation_config=layers_cfg.ActivationConfig(
+          layers_cfg.ActivationType.SILU
+      ),
+      num_layers=1,
+      attention_block_config=att_config,
+  )
+
+  config = unet_cfg.AutoEncoderConfig(
+      in_channels=in_channels,
+      latent_channels=latent_channels,
+      out_channels=out_channels,
+      activation_config=layers_cfg.ActivationConfig(
+          layers_cfg.ActivationType.SILU
+      ),
+      block_out_channels=block_out_channels,
+      scaling_factor=scaling_factor,
+      layers_per_block=layers_per_block,
+      normalization_config=norm_config,
+      mid_block_config=mid_block_config,
+  )
+  return config