ai-edge-torch-nightly 0.3.0.dev20240910__py3-none-any.whl → 0.3.0.dev20240914__py3-none-any.whl

ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py

@@ -12,13 +12,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-# Example of building a TinyLlama model from the Edge Generative API layers.
+
+"""Example of building a TinyLlama model."""
 
 import os
-from pathlib import Path
+import pathlib
 
 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.utilities.loader as loading_utils
@@ -42,13 +44,12 @@ TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
 )
 
 
-class TinyLLamma(nn.Module):
+class TinyLlama(nn.Module):
   """A TinyLlama model built from the Edge Generative API layers."""
 
   def __init__(self, config: cfg.ModelConfig):
     super().__init__()
 
-    self.config = config
     # Construct model layers.
     self.lm_head = nn.Linear(
         config.embedding_dim, config.vocab_size, bias=config.lm_head_use_bias
@@ -56,18 +57,20 @@ class TinyLLamma(nn.Module):
     self.tok_embedding = nn.Embedding(
         config.vocab_size, config.embedding_dim, padding_idx=0
     )
+    # TinyLlama has only one block config.
+    block_config = config.block_config(0)
     self.transformer_blocks = nn.ModuleList(
-        attention.TransformerBlock(config) for _ in range(config.num_layers)
+        attention.TransformerBlock(block_config, config)
+        for _ in range(config.num_layers)
     )
     self.final_norm = builder.build_norm(
         config.embedding_dim,
         config.final_norm_config,
     )
+    attn_config = block_config.attn_config
     self.rope_cache = attn_utils.build_rope_cache(
         size=config.kv_cache_max,
-        dim=int(
-            config.attn_config.rotary_percentage * config.attn_config.head_dim
-        ),
+        dim=int(attn_config.rotary_percentage * attn_config.head_dim),
         base=10_000,
         condense_ratio=1,
         dtype=torch.float32,
@@ -80,16 +83,22 @@ class TinyLLamma(nn.Module):
     )
     self.config = config
 
-  # The model's forward function takes in additional k/v cache tensors
-  # and returns the updated k/v cache tensors to the caller.
-  # This can be eliminated if we handle k/v cache updates inside the model itself.
   @torch.inference_mode
-  def forward(self, idx: torch.Tensor, input_pos: torch.Tensor) -> torch.Tensor:
-    _, seq_len = idx.size()
+  def forward(
+      self,
+      tokens: torch.Tensor,
+      input_pos: torch.Tensor,
+      kv_cache: kv_utils.KVCache,
+  ) -> dict[torch.Tensor, kv_utils.KVCache]:
+    _, seq_len = tokens.size()
     assert self.config.max_seq_len >= seq_len, (
         f"Cannot forward sequence of length {seq_len}, max seq length is only"
         f" {self.config.max_seq_len}"
     )
+    assert len(self.transformer_blocks) == len(kv_cache.caches), (
+        "The number of transformer blocks and the number of KV cache entries"
+        " must be the same."
+    )
 
     cos, sin = self.rope_cache
     cos = cos.index_select(0, input_pos)
@@ -97,16 +106,20 @@ class TinyLLamma(nn.Module):
     mask = self.mask_cache.index_select(2, input_pos)
     mask = mask[:, :, :, : self.config.kv_cache_max]
 
-    # forward the model itself
-    x = self.tok_embedding(idx)  # token embeddings of shape (b, t, n_embd)
+    # token embeddings of shape (b, t, n_embd)
+    x = self.tok_embedding(tokens)
 
-    for _, block in enumerate(self.transformer_blocks):
-      x = block(x, (cos, sin), mask, input_pos)
+    updated_kv_entires = []
+    for i, block in enumerate(self.transformer_blocks):
+      kv_entry = kv_cache.caches[i] if kv_cache else None
+      x, kv_entry = block(x, (cos, sin), mask, input_pos, kv_entry)
+      if kv_entry:
+        updated_kv_entires.append(kv_entry)
+    updated_kv_cache = kv_utils.KVCache(tuple(updated_kv_entires))
 
     x = self.final_norm(x)
-
-    res = self.lm_head(x)  # (b, t, vocab_size)
-    return res
+    logits = self.lm_head(x)  # (b, t, vocab_size)
+    return {"logits": logits, "kv_cache": updated_kv_cache}
 
 
 def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
@@ -131,55 +144,63 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       intermediate_size=5632,
   )
   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,
+  )
   config = cfg.ModelConfig(
       vocab_size=32000,
       num_layers=22,
       max_seq_len=2048,
       embedding_dim=2048,
       kv_cache_max_len=kv_cache_max_len,
-      attn_config=attn_config,
-      ff_config=ff_config,
-      pre_attention_norm_config=norm_config,
-      post_attention_norm_config=norm_config,
+      block_configs=block_config,
       final_norm_config=norm_config,
       enable_hlfb=True,
   )
   return config
 
 
-def get_fake_model_config() -> cfg.ModelConfig:
-  config = get_model_config()
+def get_fake_model_config(**kwargs) -> cfg.ModelConfig:
+  config = get_model_config(**kwargs)
   config.vocab_size = 128
   config.num_layers = 2
-  config.ff_config.intermediate_size = 64
+  # TinyLlama has only one block config.
+  config.block_config(0).ff_config.intermediate_size = 64
   return config
 
 
 def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
   config = get_model_config(**kwargs)
-  model = TinyLLamma(config)
+  model = TinyLlama(config)
   loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
   loader.load(model)
+  model.eval()
   return model
 
 
-def define_and_run() -> None:
+def define_and_run(checkpoint_path: str) -> None:
   """Instantiates and runs a TinyLlama model."""
 
-  current_dir = Path(__file__).parent.resolve()
+  current_dir = pathlib.Path(__file__).parent.resolve()
   tiny_llama_goldens = torch.load(current_dir / "tiny_llama_lm_logits.pt")
   kv_cache_max_len = 1024
-  checkpoint_path = os.path.join(Path.home(), "Downloads/llm_data/tiny_llama")
   model = build_model(checkpoint_path, kv_cache_max_len=kv_cache_max_len)
   idx = torch.from_numpy(np.array([[1, 2, 3, 4]]))
-  tokens = torch.full((1, kv_cache_max_len), 0, dtype=torch.long, device="cpu")
+  tokens = torch.full((1, kv_cache_max_len), 0, dtype=torch.int, device="cpu")
   tokens[0, :4] = idx
-  input_pos = torch.arange(0, kv_cache_max_len)
-  lm_logits = model.forward(tokens, input_pos)
+  input_pos = torch.arange(0, kv_cache_max_len, dtype=torch.int)
+  kv = kv_utils.KVCache.from_model_config(model.config)
+  output = model.forward(tokens, input_pos, kv)
   assert torch.allclose(
-      tiny_llama_goldens, lm_logits[0, idx.shape[1] - 1, :], atol=1e-05
+      tiny_llama_goldens, output["logits"][0, idx.shape[1] - 1, :], atol=1e-02
   )
 
 
 if __name__ == "__main__":
-  define_and_run()
+  input_checkpoint_path = os.path.join(
+      pathlib.Path.home(), "Downloads/llm_data/tiny_llama"
+  )
+  define_and_run(input_checkpoint_path)

ai_edge_torch/generative/fx_passes/__init__.py

@@ -12,16 +12,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-from ai_edge_torch._convert.fx_passes import CanonicalizePass
-from ai_edge_torch._convert.fx_passes import run_passes
-from ai_edge_torch.generative.fx_passes.remove_sdpa_zero_mask_pass import RemoveSDPACompositeZeroMaskPass  # NOQA
+from ai_edge_torch import fx_pass_base
+from ai_edge_torch.fx_pass_base import CanonicalizePass
+from ai_edge_torch.generative.fx_passes.remove_sdpa_zero_mask_pass import RemoveSDPACompositeZeroMaskPass
 import torch
 
 
 def run_generative_passes(
     exported_program: torch.export.ExportedProgram,
 ) -> torch.export.ExportedProgram:
-  return run_passes(
+  return fx_pass_base.run_passes(
       exported_program,
       [
           RemoveSDPACompositeZeroMaskPass(),

ai_edge_torch/generative/fx_passes/remove_sdpa_zero_mask_pass.py

@@ -12,13 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+from ai_edge_torch import fx_pass_base
 from ai_edge_torch import lowertools
-from ai_edge_torch._convert.fx_passes._pass_base import ExportedProgramPassBase
-from ai_edge_torch._convert.fx_passes._pass_base import ExportedProgramPassResult
 import torch
 
 
-class RemoveSDPACompositeZeroMaskPass(ExportedProgramPassBase):
+class RemoveSDPACompositeZeroMaskPass(fx_pass_base.ExportedProgramPassBase):
 
   def is_zero_tensor_node(self, node: torch.fx.Node):
     return node.target == torch.ops.aten.zeros.default
@@ -48,4 +47,4 @@ class RemoveSDPACompositeZeroMaskPass(ExportedProgramPassBase):
 
     exported_program.graph_module.graph.lint()
     exported_program.graph_module.recompile()
-    return ExportedProgramPassResult(exported_program, True)
+    return fx_pass_base.ExportedProgramPassResult(exported_program, True)

ai_edge_torch/generative/layers/attention.py

@@ -12,16 +12,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-# Common building blocks for Attention layer.
 
-from typing import Optional, Tuple
+"""Common building blocks for Attention layer."""
 
-import ai_edge_torch.generative.layers.builder as builder
-from ai_edge_torch.generative.layers.kv_cache import KVCache
+from typing import Optional, Tuple, Union
+
+from ai_edge_torch.generative.layers import builder
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
+from ai_edge_torch.generative.layers import scaled_dot_product_attention as sdpa
 import ai_edge_torch.generative.layers.model_config as cfg
 import ai_edge_torch.generative.layers.rotary_position_embedding as rotary_pos_emb
-from ai_edge_torch.generative.layers.scaled_dot_product_attention import scaled_dot_product_attention  # NOQA
-from ai_edge_torch.generative.layers.scaled_dot_product_attention import scaled_dot_product_attention_with_hlfb  # NOQA
 import torch
 from torch import nn
 
@@ -55,29 +55,35 @@ def _embed_rope(
 
 class TransformerBlock(nn.Module):
 
-  def __init__(self, config: cfg.ModelConfig) -> None:
+  def __init__(
+      self,
+      config: cfg.TransformerBlockConfig,
+      model_config: cfg.ModelConfig,
+  ) -> None:
     """Initialize an instance of the TransformerBlock.
 
     Args:
-      config (cfg.ModelConfig): the configuration object for this transformer
-        block.
+      config (cfg.TransformerBlockConfig): the configuration object for this
+        transformer block.
+      model_config (cfg.ModelConfig): the configuration object for the model
+        this transformer block belongs to.
     """
-
     super().__init__()
     self.pre_atten_norm = builder.build_norm(
-        config.embedding_dim, config.pre_attention_norm_config
+        model_config.embedding_dim,
+        config.pre_attention_norm_config,
     )
     self.atten_func = CausalSelfAttention(
-        config.batch_size,
-        config.embedding_dim,
+        model_config.batch_size,
+        model_config.embedding_dim,
         config.attn_config,
-        config.kv_cache_max,
-        config.enable_hlfb,
+        model_config.enable_hlfb,
     )
     self.post_atten_norm = builder.build_norm(
-        config.embedding_dim, config.post_attention_norm_config
+        model_config.embedding_dim,
+        config.post_attention_norm_config,
     )
-    self.ff = builder.build_ff(config.embedding_dim, config.ff_config)
+    self.ff = builder.build_ff(model_config.embedding_dim, config.ff_config)
     self.config = config
 
   def forward(
@@ -86,7 +92,8 @@ class TransformerBlock(nn.Module):
       rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
       mask: Optional[torch.Tensor] = None,
       input_pos: Optional[torch.Tensor] = None,
-  ) -> torch.Tensor:
+      kv_cache: kv_utils.KVCacheEntry = None,
+  ) -> Union[torch.Tensor, Tuple[torch.Tensor, kv_utils.KVCacheEntry]]:
     """Forward function of the TransformerBlock.
 
     Args:
@@ -94,24 +101,34 @@ class TransformerBlock(nn.Module):
       rope (Tuple[torch.Tensor, torch.Tensor]): the input rope tensor.
       mask (torch.Tensor): the optional mask tensor.
       input_pos (torch.Tensor): the optional input position tensor.
+      kv_cache (KVCacheEntry): the optional kv cache entry.
 
     Returns:
-      output activation from this transformer block.
+      output activation from this transformer block, and updated kv cache (if
+      passed in).
     """
-
+    kv = None
     if self.config.parallel_residual:
       x_norm = self.pre_atten_norm(x)
-      attn_out = self.atten_func(x_norm, rope, mask, input_pos)
+      atten_func_out = self.atten_func(x_norm, rope, mask, input_pos, kv_cache)
+      if kv_cache is None:
+        attn_out = atten_func_out
+      else:
+        attn_out, kv = atten_func_out
       ff_out = self.ff(x_norm)
       output = x + attn_out + ff_out
     else:
       x_norm = self.pre_atten_norm(x)
-      attn_out = self.atten_func(x_norm, rope, mask, input_pos)
+      atten_func_out = self.atten_func(x_norm, rope, mask, input_pos, kv_cache)
+      if kv_cache is None:
+        attn_out = atten_func_out
+      else:
+        attn_out, kv = atten_func_out
       x = x + attn_out
       x_norm = self.post_atten_norm(x)
       output = x + self.ff(x_norm)
 
-    return output
+    return output if kv is None else (output, kv)
 
 
 class CausalSelfAttention(nn.Module):
@@ -121,7 +138,6 @@ class CausalSelfAttention(nn.Module):
       batch_size: int,
       dim: int,
       config: cfg.AttentionConfig,
-      kv_cache_max: int,
       enable_hlfb: bool,
   ) -> None:
     """Initialize an instance of CausalSelfAttention.
@@ -130,12 +146,9 @@ class CausalSelfAttention(nn.Module):
       batch_size (int): batch size of the input tensor.
       dim (int): causal attention's input/output dimmension.
       config (cfg.AttentionConfig): attention specific configurations.
-      kv_cache_max (int): determines the size of the KV Cache buffer, if
-        enabled.
       enable_hlfb (bool): whether hlfb is enabled or not.
     """
     super().__init__()
-    self.config = config
     self.kv_cache = None
     self.batch_size = batch_size
     qkv_shape = (
@@ -147,21 +160,17 @@ class CausalSelfAttention(nn.Module):
     self.output_projection = nn.Linear(
         output_shape, dim, bias=config.output_proj_use_bias
     )
-
-    # Build a k/v cache with size (batch_size, kv_cache_max, n_heads, head_dim).
-    if config.enable_kv_cache:
-      self.kv_cache = KVCache(
-          batch_size,
-          kv_cache_max,
-          config.num_query_groups,
-          config.head_dim,
-          enable_hlfb,
-      )
-
-    if enable_hlfb:
-      self.sdpa_func = scaled_dot_product_attention_with_hlfb
-    else:
-      self.sdpa_func = scaled_dot_product_attention
+    self.query_norm = builder.build_norm(
+        config.head_dim, config.query_norm_config
+    )
+    self.key_norm = builder.build_norm(config.head_dim, config.key_norm_config)
+    self.config = config
+    self.enable_hlfb = enable_hlfb
+    self.sdpa_func = (
+        sdpa.scaled_dot_product_attention_with_hlfb
+        if enable_hlfb
+        else sdpa.scaled_dot_product_attention
+    )
 
   def forward(
       self,
@@ -169,7 +178,8 @@ class CausalSelfAttention(nn.Module):
       rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
       mask: Optional[torch.Tensor] = None,
       input_pos: Optional[torch.Tensor] = None,
-  ) -> torch.Tensor:
+      kv_cache: Optional[kv_utils.KVCacheEntry] = None,
+  ) -> Union[torch.Tensor, Tuple[torch.Tensor, kv_utils.KVCacheEntry]]:
     """Forward function of the CausalSelfAttention layer, which can support
 
        MQA, GQA and MHA.
@@ -179,9 +189,11 @@ class CausalSelfAttention(nn.Module):
       rope (Tuple[torch.Tensor, torch.Tensor]): the input rope tensor.
       mask (torch.Tensor): the optional mask tensor.
       input_pos (torch.Tensor): the optional input position tensor.
+      kv_cache (KVCacheEntry): The KV cache entry corresponding to this module.
 
     Returns:
-      output activation from this self attention layer.
+      output activation from this self attention layer, and the updated
+        KV Cach Entry (if passed in).
     """
     # Batch size, sequence length, embedding dimensionality.
     B, T, E = x.size()
@@ -216,6 +228,9 @@ class CausalSelfAttention(nn.Module):
           dim=-1,
       )
 
+    q = self.query_norm(q)
+    k = self.key_norm(k)
+
     q = q.reshape(B, T, -1, self.config.head_dim)
     k = k.reshape(B, T, -1, self.config.head_dim)
     v = v.reshape(B, T, -1, self.config.head_dim)
@@ -224,9 +239,11 @@ class CausalSelfAttention(nn.Module):
     n_elem = int(self.config.rotary_percentage * self.config.head_dim)
     q, k = _embed_rope(q, k, n_elem, rope)
 
-    if self.kv_cache is not None:
-      # TODO(haoliang): Handle when execeeding max sequence length.
-      k, v = self.kv_cache.update_cache(input_pos, k, v)
+    if kv_cache is not None:
+      kv_cache = kv_utils.update(
+          kv_cache, input_pos, k, v, enable_hlfb=self.enable_hlfb
+      )
+      k, v = kv_cache.k_cache, kv_cache.v_cache
 
     y = self.sdpa_func(
         q,
@@ -240,7 +257,7 @@ class CausalSelfAttention(nn.Module):
 
     # Compute the output projection.
     y = self.output_projection(y)
-    return y
+    return y if kv_cache is None else (y, kv_cache)
 
 
 class SelfAttention(CausalSelfAttention):
@@ -251,16 +268,19 @@ class SelfAttention(CausalSelfAttention):
       x: torch.Tensor,
       rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
       input_pos: Optional[torch.Tensor] = None,
-  ) -> torch.Tensor:
+      kv_cache: Optional[kv_utils.KVCacheEntry] = None,
+  ) -> Union[torch.Tensor, Tuple[torch.Tensor, kv_utils.KVCacheEntry]]:
     """Forward function of the SelfAttention layer, which can support MQA, GQA and MHA.
 
     Args:
       x (torch.Tensor): the input tensor.
       rope (Tuple[torch.Tensor, torch.Tensor]): the input rope tensor.
       input_pos (torch.Tensor): the optional input position tensor.
+      kv_cache (KVCacheEntry): The KV cache entry corresponding to this module.
 
     Returns:
-      output activation from this self attention layer.
+      output activation from this self attention layer, and the updated
+        KV Cach Entry (if passed in).
     """
     B, T, _ = x.size()
     return super().forward(
@@ -279,9 +299,8 @@ class CrossAttention(nn.Module):
       query_dim: int,
       cross_dim: int,
       config: cfg.AttentionConfig,
-      kv_cache_max: int,
       enable_hlfb: bool,
-  ) -> None:
+  ):
     """Initialize an instance of CrossAttention.
 
     Args:
@@ -289,8 +308,6 @@ class CrossAttention(nn.Module):
       query_dim (int): query tensor's dimension.
       cross_dim (int): cross attention's dimensions, for key and value tensors.
       config (cfg.AttentionConfig): attention specific configurations.
-      kv_cache_max (int): determines the size of the KV Cache buffer, if
-        enabled.
       enable_hlfb (bool): whether hlfb is enabled or not.
     """
     super().__init__()
@@ -309,21 +326,11 @@ class CrossAttention(nn.Module):
         query_dim, query_dim, bias=config.output_proj_use_bias
     )
 
-    self.kv_cache = None
-    # Build a k/v cache with size (batch_size, kv_cache_max, n_heads, head_dim).
-    if config.enable_kv_cache:
-      self.kv_cache = KVCache(
-          batch_size,
-          kv_cache_max,
-          config.num_query_groups,
-          self.config.head_dim,
-          enable_hlfb,
-      )
-
-    if enable_hlfb:
-      self.sdpa_func = scaled_dot_product_attention_with_hlfb
-    else:
-      self.sdpa_func = scaled_dot_product_attention
+    self.sdpa_func = (
+        sdpa.scaled_dot_product_attention_with_hlfb
+        if enable_hlfb
+        else sdpa.scaled_dot_product_attention
+    )
 
   def forward(
       self,
@@ -332,6 +339,7 @@ class CrossAttention(nn.Module):
       rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
       mask: Optional[torch.Tensor] = None,
       input_pos: Optional[torch.Tensor] = None,
+      kv_cache: Optional[kv_utils.KVCacheEntry] = None,
   ):
     """Forward function of the CrossAttention layer.
 
@@ -342,6 +350,7 @@ class CrossAttention(nn.Module):
       mask (torch.Tensor): the optional mask tensor can be broadcaseted to shape
         [B, n_heads, target_seq_len, source_seq_len].
       input_pos (torch.Tensor): the optional input position tensor.
+      kv_cache (KVCacheEntry): The KV cache entry corresponding to this module.
 
     Returns:
       output activation from this cross attention layer.
@@ -363,9 +372,11 @@ class CrossAttention(nn.Module):
     n_elem = int(self.config.rotary_percentage * self.config.head_dim)
     q, k = _embed_rope(q, k, n_elem, rope)
 
-    if self.kv_cache is not None:
-      # TODO(haoliang): Handle when execeeding max sequence length.
-      k, v = self.kv_cache.update_cache(input_pos, k, v)
+    if kv_cache is not None:
+      kv_cache = kv_utils.update(
+          kv_cache, input_pos, k, v, enable_hlfb=self.enable_hlfb
+      )
+      k, v = kv_cache.k_cache, kv_cache.v_cache
     if mask is None:
       mask = torch.zeros(
           (batch_size, 1, target_seq_len, source_seq_len), dtype=torch.float32
@@ -375,4 +386,4 @@ class CrossAttention(nn.Module):
 
     # Compute the output projection.
     y = self.output_projection(y)
-    return y
+    return y if kv_cache is None else (y, kv_cache)

ai_edge_torch/generative/layers/builder.py

@@ -13,6 +13,8 @@
 # limitations under the License.
 # ==============================================================================
 # Builder class for individual components.
+from typing import Callable
+
 import ai_edge_torch.generative.layers.feed_forward as feed_forward
 import ai_edge_torch.generative.layers.model_config as cfg
 import ai_edge_torch.generative.layers.normalization as normalization
@@ -21,20 +23,34 @@ from torch import nn
 import torch.nn.functional as F
 
 
-class GeGLU(nn.Module):
-  """GeGLU is an activation function which is a variant of GELU.
+def build_glu(
+    act: Callable[[torch.Tensor], torch.Tensor], gate_is_front: bool = False
+) -> Callable[[torch.Tensor], torch.Tensor]:
+  """Builds an activation function with GLU (Gated Linear Unit).
+
+  If gate_is_front is True,
+    f(x) = act(x) * y
+  otherwise,
+    f(x) = x * act(y),
+  where x is the first half of the input and y is the second half of the input.
+
+  Args:
+    act (Callable[[torch.Tensor], torch.Tensor]): activation function to apply
+      to the gate.
+    gate_is_front: whether the gate is in front half of the input. Other part is
+      the output in GLU.
 
-  GeGLU(x) = (xW+b) * GELU(xV+c)
-  See: https://arxiv.org/abs/2002.05202v1
+  Returns:
+    A callable activation function with GLU.
   """
 
-  def __init__(self, d_in: int, d_out: int):
-    super().__init__()
-    self.proj = nn.Linear(d_in, d_out * 2)
+  def _glu(x):
+    x, y = x.chunk(2, dim=-1)
+    if gate_is_front:
+      return act(x) * y
+    return x * act(y)
 
-  def forward(self, x: torch.Tensor):
-    x, gate = self.proj(x).chunk(2, dim=-1)
-    return x * F.gelu(gate)
+  return _glu
 
 
 def build_norm(dim: int, config: cfg.NormalizationConfig):
@@ -59,9 +75,11 @@ def build_norm(dim: int, config: cfg.NormalizationConfig):
         zero_centered_gamma=config.zero_centered,
     )
   elif config.type == cfg.NormalizationType.LAYER_NORM:
-    return nn.LayerNorm(dim, eps=config.epsilon)
+    return normalization.LayerNorm(dim, config.epsilon, config.enable_hlfb)
   elif config.type == cfg.NormalizationType.GROUP_NORM:
-    return nn.GroupNorm(config.group_num, dim, config.epsilon)
+    return normalization.GroupNorm(
+        config.group_num, dim, config.epsilon, config.enable_hlfb
+    )
   else:
     raise ValueError("Unsupported norm type.")
 
@@ -71,7 +89,7 @@ def build_ff(dim: int, config: cfg.FeedForwardConfig):
 
   Args:
     dim (int): dimension of the input tensor.
-    config (`ModelConfig` object): the model configuration.
+    config (`FeedForwardConfig` object): the model configuration.
 
   Returns:
     The constructed `nn.Module` feedforward layer.
@@ -97,6 +115,10 @@ def build_ff(dim: int, config: cfg.FeedForwardConfig):
       hidden_dim=config.intermediate_size,
       activation=activation,
       use_bias=config.use_bias,
+      use_glu=(
+          config.activation.type == cfg.ActivationType.GE_GLU
+          or config.activation.type == cfg.ActivationType.SILU_GLU
+      ),
       pre_ff_norm=pre_ff_norm,
       post_ff_norm=post_ff_norm,
   )
@@ -127,8 +149,10 @@ def get_activation(config: cfg.ActivationConfig):
     # See: https://github.com/hendrycks/GELUs
     return lambda x: x * F.sigmoid(1.702 * x)
   elif config.type == cfg.ActivationType.GE_GLU:
-    return GeGLU(config.dim_in, config.dim_out)
+    return build_glu(F.gelu, config.gate_is_front)
   elif config.type == cfg.ActivationType.RELU:
     return F.relu
+  elif config.type == cfg.ActivationType.SILU_GLU:
+    return build_glu(F.silu, config.gate_is_front)
   else:
     raise ValueError("Unsupported activation type.")