Add GeGLU activation and change activation type to activation config …

…for model_config. (#42)

ai_edge_torch/generative/examples/gemma/gemma.py

@@ -116,7 +116,7 @@ def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
   )
   ff_config = cfg.FeedForwardConfig(
       type=cfg.FeedForwardType.GATED,
-      activation=cfg.ActivationType.GELU_TANH,
+      activation=cfg.ActivationConfig(cfg.ActivationType.GELU_TANH),
       intermediate_size=16384,
   )
   norm_config = cfg.NormalizationConfig(

ai_edge_torch/generative/examples/phi2/phi2.py

@@ -112,7 +112,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
   )
   ff_config = cfg.FeedForwardConfig(
       type=cfg.FeedForwardType.SEQUENTIAL,
-      activation=cfg.ActivationType.GELU_TANH,
+      activation=cfg.ActivationConfig(cfg.ActivationType.GELU_TANH),
       intermediate_size=10240,
       use_bias=True,
   )

ai_edge_torch/generative/examples/stable_diffusion/clip.py

@@ -90,7 +90,7 @@ def get_model_config() -> cfg.ModelConfig:
 
   ff_config = cfg.FeedForwardConfig(
       type=cfg.FeedForwardType.SEQUENTIAL,
-      activation=cfg.ActivationType.GELU_QUICK,
+      activation=cfg.ActivationConfig(cfg.ActivationType.GELU_QUICK),
       intermediate_size=embedding_dim * 4,
       use_bias=True,
   )

ai_edge_torch/generative/examples/stable_diffusion/decoder.py

@@ -221,7 +221,7 @@ def __init__(self, config: unet_cfg.AutoEncoderConfig):
                   in_channels=prev_output_channel,
                   out_channels=block_out_channels,
                   normalization_config=config.normalization_config,
-                  activation_type=config.activation_type,
+                  activation_config=config.activation_config,
                   num_layers=config.layers_per_block,
                   add_upsample=not_final_block,
                   upsample_conv=True,
@@ -235,7 +235,7 @@ def __init__(self, config: unet_cfg.AutoEncoderConfig):
     self.final_norm = layers_builder.build_norm(
         block_out_channels, config.normalization_config
     )
-    self.act_fn = layers_builder.get_activation(config.activation_type)
+    self.act_fn = layers_builder.get_activation(config.activation_config)
     self.conv_out = nn.Conv2d(
         block_out_channels,
         config.out_channels,
@@ -287,7 +287,7 @@ def get_model_config() -> unet_cfg.AutoEncoderConfig:
   mid_block_config = unet_cfg.MidBlock2DConfig(
       in_channels=block_out_channels[-1],
       normalization_config=norm_config,
-      activation_type=layers_cfg.ActivationType.SILU,
+      activation_config=layers_cfg.ActivationConfig(layers_cfg.ActivationType.SILU),
       num_layers=1,
       attention_block_config=att_config,
   )
@@ -296,7 +296,7 @@ def get_model_config() -> unet_cfg.AutoEncoderConfig:
       in_channels=in_channels,
       latent_channels=latent_channels,
       out_channels=out_channels,
-      activation_type=layers_cfg.ActivationType.SILU,
+      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,

ai_edge_torch/generative/examples/stable_diffusion/diffusion.py

@@ -130,7 +130,7 @@ def __init__(self, channels):
     self.conv = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
 
   def forward(self, x):
-    x = F.interpolate(x, scale_factor=2, mode='nearest')
+    x = F.interpolate(x, scale_factor=2, mode="nearest")
     return self.conv(x)
 
 
@@ -237,3 +237,8 @@ def forward(self, latent, context, time):
     output = self.unet(latent, context, time)
     output = self.final(output)
     return output
+
+
+if __name__ == "__main__":
+  diffusion = Diffusion()
+  print(diffusion.state_dict().keys())

ai_edge_torch/generative/examples/t5/t5.py

@@ -349,7 +349,7 @@ def get_model_config_t5() -> cfg.ModelConfig:
   )
   ff_config = cfg.FeedForwardConfig(
       type=cfg.FeedForwardType.SEQUENTIAL,
-      activation=cfg.ActivationType.RELU,
+      activation=cfg.ActivationConfig(cfg.ActivationType.RELU),
       intermediate_size=3072,
   )
   # T5 Confirmed as RMS Norm and eps = 1e-6 TJA.

ai_edge_torch/generative/examples/test_models/toy_model.py

@@ -76,7 +76,7 @@ def define_and_run() -> None:
   )
   ff_config = cfg.FeedForwardConfig(
       type=cfg.FeedForwardType.GATED,
-      activation=cfg.ActivationType.SILU,
+      activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
       intermediate_size=256,
   )
   norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)

ai_edge_torch/generative/examples/test_models/toy_model_with_external_kv_cache.py

@@ -95,7 +95,7 @@ def get_model_config() -> cfg.ModelConfig:
   )
   ff_config = cfg.FeedForwardConfig(
       type=cfg.FeedForwardType.GATED,
-      activation=cfg.ActivationType.SILU,
+      activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
       intermediate_size=256,
   )
   norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)

ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py

@@ -83,7 +83,7 @@ def get_model_config() -> cfg.ModelConfig:
   )
   ff_config = cfg.FeedForwardConfig(
       type=cfg.FeedForwardType.GATED,
-      activation=cfg.ActivationType.SILU,
+      activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
       intermediate_size=256,
   )
   norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)

ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py

@@ -112,7 +112,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
   )
   ff_config = cfg.FeedForwardConfig(
       type=cfg.FeedForwardType.GATED,
-      activation=cfg.ActivationType.SILU,
+      activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
       intermediate_size=5632,
   )
   norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)

ai_edge_torch/generative/layers/builder.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 # ==============================================================================
 # Builder class for individual components.
+import torch
 from torch import nn
 import torch.nn.functional as F
 
@@ -21,6 +22,23 @@
 import ai_edge_torch.generative.layers.normalization as normalization
 
 
+class GeGLU(nn.Module):
+  """GeGLU is an activation function which is a variant of GELU.
+
+  GeGLU(x) = (xW+b) * GELU(xV+c)
+  See: https://arxiv.org/abs/2002.05202v1
+
+  """
+
+  def __init__(self, d_in: int, d_out: int):
+    super().__init__()
+    self.proj = nn.Linear(d_in, d_out * 2)
+
+  def forward(self, x: torch.Tensor):
+    x, gate = self.proj(x).chunk(2, dim=-1)
+    return x * F.gelu(gate)
+
+
 def build_norm(dim: int, config: cfg.NormalizationConfig):
   """Builder function for normalizers.
 
@@ -81,29 +99,33 @@ def build_ff(dim: int, config: cfg.FeedForwardConfig):
   )
 
 
-def get_activation(type_: cfg.ActivationType):
-  """Get pytorch callable activation from the name.
+def get_activation(config: cfg.ActivationConfig):
+  """Get pytorch callable activation from the activation config.
 
   Args:
-    name (string): activation's name.
+    config (cfg.ActivationConfig): activation config.
 
   Returns:
     Activation function.
 
   Raises:
-    ValueError: If activation name is not supported.
+    ValueError: If activation config is not supported.
   """
-  if type_ == cfg.ActivationType.SILU:
+  if config.type == cfg.ActivationType.LINEAR:
+    return lambda x: x
+  elif config.type == cfg.ActivationType.SILU:
     return F.silu
-  elif type_ == cfg.ActivationType.GELU:
+  elif config.type == cfg.ActivationType.GELU:
     return F.gelu
-  elif type_ == cfg.ActivationType.GELU_TANH:
+  elif config.type == cfg.ActivationType.GELU_TANH:
     return lambda x: F.gelu(x, approximate="tanh")
-  elif type_ == cfg.ActivationType.GELU_QUICK:
+  elif config.type == cfg.ActivationType.GELU_QUICK:
     # GELU approximation that is fast but somewhat inaccurate.
     # See: https://github.com/hendrycks/GELUs
     return lambda x: x * F.sigmoid(1.702 * x)
-  elif type_ == cfg.ActivationType.RELU:
+  elif config.type == cfg.ActivationType.GE_GLU:
+    return GeGLU(config.dim_in, config.dim_out)
+  elif config.type == cfg.ActivationType.RELU:
     return F.relu
   else:
     raise ValueError("Unsupported activation type.")

ai_edge_torch/generative/layers/model_config.py

@@ -27,7 +27,7 @@ class ActivationType(enum.Enum):
   SILU = enum.auto()
   GELU = enum.auto()
   GELU_TANH = enum.auto()
-  GELU_QUICK = enum.auto()
+  GE_GLU = enum.auto()
   RELU = enum.auto()
 
 
@@ -74,12 +74,20 @@ class AttentionConfig:
   relative_attention_max_distance: int = 0
 
 
+@dataclass
+class ActivationConfig:
+  type: ActivationType = ActivationType.LINEAR
+  # Dimension of input and output, used in GeGLU.
+  dim_in: Optional[int] = None
+  dim_out: Optional[int] = None
+
+
 @dataclass
 class FeedForwardConfig:
   """FeedForward module's parameters."""
 
   type: FeedForwardType
-  activation: ActivationType
+  activation: ActivationConfig
   intermediate_size: int
   use_bias: bool = False
 

ai_edge_torch/generative/layers/unet/blocks_2d.py

@@ -53,7 +53,7 @@ def __init__(self, config: unet_cfg.ResidualBlock2DConfig):
     self.conv_2 = nn.Conv2d(
         config.out_channels, config.out_channels, kernel_size=3, stride=1, padding=1
     )
-    self.act_fn = layers_builder.get_activation(config.activation_type)
+    self.act_fn = layers_builder.get_activation(config.activation_config)
     if config.in_channels == config.out_channels:
       self.residual_layer = nn.Identity()
     else:
@@ -167,7 +167,7 @@ def __init__(self, config: unet_cfg.UpDecoderBlock2DConfig):
                   out_channels=config.out_channels,
                   time_embedding_channels=config.time_embedding_channels,
                   normalization_config=config.normalization_config,
-                  activation_type=config.activation_type,
+                  activation_config=config.activation_config,
               )
           )
       )
@@ -244,7 +244,7 @@ def __init__(self, config: unet_cfg.MidBlock2DConfig):
                 out_channels=config.in_channels,
                 time_embedding_channels=config.time_embedding_channels,
                 normalization_config=config.normalization_config,
-                activation_type=config.activation_type,
+                activation_config=config.activation_config,
             )
         )
     ]
@@ -259,7 +259,7 @@ def __init__(self, config: unet_cfg.MidBlock2DConfig):
                   out_channels=config.in_channels,
                   time_embedding_channels=config.time_embedding_channels,
                   normalization_config=config.normalization_config,
-                  activation_type=config.activation_type,
+                  activation_config=config.activation_config,
               )
           )
       )

ai_edge_torch/generative/layers/unet/model_config.py

@@ -39,7 +39,7 @@ class ResidualBlock2DConfig:
   in_channels: int
   out_channels: int
   normalization_config: layers_cfg.NormalizationConfig
-  activation_type: layers_cfg.ActivationType
+  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
 
@@ -56,7 +56,7 @@ class UpDecoderBlock2DConfig:
   in_channels: int
   out_channels: int
   normalization_config: layers_cfg.NormalizationConfig
-  activation_type: layers_cfg.ActivationType
+  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
@@ -72,7 +72,7 @@ class UpDecoderBlock2DConfig:
 class MidBlock2DConfig:
   in_channels: int
   normalization_config: layers_cfg.NormalizationConfig
-  activation_type: layers_cfg.ActivationType
+  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
@@ -85,7 +85,7 @@ class AutoEncoderConfig:
   """Configurations of encoder/decoder in the autoencoder model."""
 
   # The activation type of encoder/decoder blocks.
-  activation_type: layers_cfg.ActivationType
+  activation_config: layers_cfg.ActivationConfig
 
   # The output channels of each block.
   block_out_channels: List[int]