ai-edge-torch-nightly 0.3.0.dev20241002__py3-none-any.whl → 0.3.0.dev20241003__py3-none-any.whl

ai_edge_torch/generative/examples/gemma/gemma1.py

@@ -15,14 +15,9 @@
 
 """Example of building a Gemma1 model."""
 
-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
+from ai_edge_torch.generative.utilities import model_builder
 import ai_edge_torch.generative.utilities.loader as loading_utils
-import torch
-from torch import nn
 
 TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
     ff_up_proj="model.layers.{}.mlp.up_proj",
@@ -38,84 +33,6 @@ TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
 )
 
 
-class Gemma(nn.Module):
-  """A Gemma model built from the Edge Generative API layers."""
-
-  def __init__(self, config: cfg.ModelConfig):
-    super().__init__()
-
-    # Construct model layers.
-    self.tok_embedding = nn.Embedding(
-        config.vocab_size, config.embedding_dim, padding_idx=0
-    )
-    self.lm_head = nn.Linear(
-        config.embedding_dim,
-        config.vocab_size,
-        bias=config.lm_head_use_bias,
-    )
-    # Gemma re-uses the embedding as the head projection layer.
-    self.lm_head.weight.data = self.tok_embedding.weight.data
-    # Gemma has only one block config.
-    block_config = config.block_config(0)
-    self.transformer_blocks = nn.ModuleList(
-        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(attn_config.rotary_percentage * attn_config.head_dim),
-        base=attn_config.rotary_base,
-    )
-    self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.kv_cache_max,
-    )
-    self.config = config
-
-  @torch.inference_mode
-  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)
-    sin = sin.index_select(0, input_pos)
-    mask = self.mask_cache.index_select(2, input_pos)
-    mask = mask[:, :, :, : self.config.kv_cache_max]
-
-    # token embeddings of shape (b, t, n_embd)
-    x = self.tok_embedding(tokens)
-    x = x * (self.config.embedding_dim**0.5)
-
-    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)
-    logits = self.lm_head(x)  # (b, t, vocab_size)
-    return {"logits": logits, "kv_cache": updated_kv_cache}
-
-
 def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
   """Returns the model config for a Gemma 2B model.
 
@@ -154,6 +71,7 @@ def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       num_layers=18,
       max_seq_len=8192,
       embedding_dim=2048,
+      embedding_scale=2048**0.5,
       kv_cache_max_len=kv_cache_max_len,
       block_configs=block_config,
       final_norm_config=norm_config,
@@ -173,12 +91,11 @@ def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
   return config
 
 
-def build_2b_model(checkpoint_path: str, **kwargs) -> nn.Module:
-  config = get_model_config_2b(**kwargs)
-  model = Gemma(config)
-  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
-  # Since embedding and lm-head use the same weight, we need to set strict
-  # to False.
-  loader.load(model, strict=False)
-  model.eval()
-  return model
+def build_2b_model(
+    checkpoint_path: str, **kwargs
+) -> model_builder.DecoderOnlyModel:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_model_config_2b(**kwargs),
+      tensor_names=TENSOR_NAMES,
+  )

ai_edge_torch/generative/examples/gemma/gemma2.py

@@ -15,7 +15,6 @@
 
 """Example of building a Gemma2 model."""
 
-import os
 from typing import Optional, Tuple
 
 from ai_edge_torch.generative.layers import attention

ai_edge_torch/generative/examples/llama/convert_to_tflite.py

@@ -23,6 +23,12 @@ from absl import flags
 from ai_edge_torch.generative.examples.llama import llama
 from ai_edge_torch.generative.utilities import converter
 
+_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'),
@@ -49,13 +55,18 @@ _QUANTIZE = flags.DEFINE_bool(
     'Whether the model should be quantized.',
 )
 
+_BUILDER = {
+    '1b': llama.build_1b_model,
+    '3b': llama.build_3b_model,
+}
+
 
 def main(_):
-  pytorch_model = llama.build_model(
+  pytorch_model = _BUILDER[_MODEL_SIZE.value](
       _CHECKPOINT_PATH.value, kv_cache_max_len=_KV_CACHE_MAX_LEN.value
   )
   quant_suffix = 'q8' if _QUANTIZE.value else 'f32'
-  output_filename = f'llama_{quant_suffix}_seq{_PREFILL_SEQ_LEN.value}_ekv{_KV_CACHE_MAX_LEN.value}.tflite'
+  output_filename = f'llama_{_MODEL_SIZE.value}_{quant_suffix}_seq{_PREFILL_SEQ_LEN.value}_ekv{_KV_CACHE_MAX_LEN.value}.tflite'
   converter.convert_to_tflite(
       pytorch_model,
       tflite_path=os.path.join(_TFLITE_PATH.value, output_filename),

ai_edge_torch/generative/examples/llama/llama.py

@@ -15,19 +15,15 @@
 
 """Example of building Llama 3.2 models."""
 
-import copy
 import math
 from typing import Tuple
 
-from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
 import ai_edge_torch.generative.layers.model_config as cfg
+from ai_edge_torch.generative.utilities import model_builder
 import ai_edge_torch.generative.utilities.loader as loading_utils
 import torch
-from torch import nn
 
-TENSOR_NAMES = copy.copy(tiny_llama.TENSOR_NAMES)
-# SmolLM re-uses the embedding as the head projection layer.
-TENSOR_NAMES.lm_head = None
+TENSOR_NAMES = model_builder.TENSOR_NAMES
 
 
 def _build_llama3_rope_cache(
@@ -93,7 +89,7 @@ def _build_llama3_rope_cache(
   return cos, sin
 
 
-class Llama(tiny_llama.TinyLlama):
+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.
@@ -101,9 +97,6 @@ class Llama(tiny_llama.TinyLlama):
 
   def __init__(self, config: cfg.ModelConfig):
     super().__init__(config)
-    # Llama 3.2 re-uses the embedding as the head projection layer.
-    self.lm_head.weight.data = self.tok_embedding.weight.data
-    # Llama has only one block config.
     attn_config = self.config.block_config(0).attn_config
     self.rope_cache = _build_llama3_rope_cache(
         size=self.config.kv_cache_max,
@@ -119,7 +112,7 @@ class Llama(tiny_llama.TinyLlama):
     )
 
 
-def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+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:
@@ -163,7 +156,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
 
 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_model_config(kv_cache_max_len)
+  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
@@ -174,7 +167,7 @@ def get_3b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
 
 
 def get_fake_model_config(**kwargs) -> cfg.ModelConfig:
-  config = get_model_config(**kwargs)
+  config = get_1b_model_config(**kwargs)
   config.vocab_size = 128
   config.num_layers = 2
   # SmolLM has only one block config.
@@ -182,8 +175,9 @@ def get_fake_model_config(**kwargs) -> cfg.ModelConfig:
   return config
 
 
-def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
-  config = get_model_config(**kwargs)
+def _build_model(
+    checkpoint_path: str, config: cfg.ModelConfig
+) -> model_builder.DecoderOnlyModel:
   model = Llama(config)
   loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
   # Since embedding and lm-head use the same weight, we need to set strict
@@ -193,12 +187,13 @@ def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
   return model
 
 
-def build_3b_model(checkpoint_path: str, **kwargs) -> nn.Module:
-  config = get_3b_model_config(**kwargs)
-  model = Llama(config)
-  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
-  # Since embedding and lm-head use the same weight, we need to set strict
-  # to False.
-  loader.load(model, strict=False)
-  model.eval()
-  return model
+def build_1b_model(
+    checkpoint_path: str, **kwargs
+) -> model_builder.DecoderOnlyModel:
+  return _build_model(checkpoint_path, get_1b_model_config(**kwargs))
+
+
+def build_3b_model(
+    checkpoint_path: str, **kwargs
+) -> model_builder.DecoderOnlyModel:
+  return _build_model(checkpoint_path, get_3b_model_config(**kwargs))

ai_edge_torch/generative/examples/llama/verify.py

@@ -25,7 +25,12 @@ 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?",
@@ -37,9 +42,19 @@ _MAX_NEW_TOKENS = flags.DEFINE_integer(
     "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 = "meta-llama/Llama-3.2-1B-Instruct"
+  checkpoint = _CHECKPOINT[_MODEL_SIZE.value]
   logging.info("Loading the original model from: %s", checkpoint)
   original_model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint)
 
@@ -49,7 +64,7 @@ def main(_):
   )
   reauthored_checkpoint = pathlib.Path(cached_config_file).parent
   logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
-  reauthored_model = llama.build_model(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,

ai_edge_torch/generative/examples/openelm/openelm.py

@@ -15,14 +15,9 @@
 
 """Example of building an OpenELM model."""
 
-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
+from ai_edge_torch.generative.utilities import model_builder
 import ai_edge_torch.generative.utilities.loader as loading_utils
-import torch
-from torch import nn
 
 TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
     ff_up_proj="transformer.layers.{}.ffn.proj_1",
@@ -39,81 +34,6 @@ TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
 )
 
 
-class OpenELM(nn.Module):
-  """An OpenELM model built from the Edge Generative API layers."""
-
-  def __init__(self, config: cfg.ModelConfig):
-    super().__init__()
-
-    # Construct model layers.
-    self.tok_embedding = nn.Embedding(
-        config.vocab_size, config.embedding_dim, padding_idx=0
-    )
-    self.lm_head = nn.Linear(
-        config.embedding_dim, config.vocab_size, bias=config.lm_head_use_bias
-    )
-    # OpenELM re-uses the embedding as the head projection layer.
-    self.lm_head.weight.data = self.tok_embedding.weight.data
-    self.transformer_blocks = nn.ModuleList(
-        attention.TransformerBlock(config.block_config(idx), config)
-        for idx in range(config.num_layers)
-    )
-    self.final_norm = builder.build_norm(
-        config.embedding_dim,
-        config.final_norm_config,
-    )
-    # OpenELM has same hyper parameters for rotary_percentage and head_dim for
-    # each layer block. Use the first block.
-    attn_config = config.block_config(0).attn_config
-    self.rope_cache = attn_utils.build_rope_cache(
-        size=config.kv_cache_max,
-        dim=int(attn_config.rotary_percentage * attn_config.head_dim),
-        base=attn_config.rotary_base,
-    )
-    self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.kv_cache_max,
-    )
-    self.config = config
-
-  @torch.inference_mode
-  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)
-    sin = sin.index_select(0, input_pos)
-    mask = self.mask_cache.index_select(2, input_pos)
-    mask = mask[:, :, :, : self.config.kv_cache_max]
-
-    # token embeddings of shape (b, t, n_embd)
-    x = self.tok_embedding(tokens)
-
-    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)
-    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:
   """Returns the model config for an OpenELM model.
 
@@ -191,12 +111,11 @@ def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
   return config
 
 
-def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
-  config = get_model_config(**kwargs)
-  model = OpenELM(config)
-  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
-  # Since embedding and lm-head use the same weight, we need to set strict
-  # to False.
-  loader.load(model, strict=False)
-  model.eval()
-  return model
+def build_model(
+    checkpoint_path: str, **kwargs
+) -> model_builder.DecoderOnlyModel:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+  )

ai_edge_torch/generative/examples/phi/phi2.py

@@ -15,14 +15,9 @@
 
 """Example of building a Phi-2 model."""
 
-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
+from ai_edge_torch.generative.utilities import model_builder
 import ai_edge_torch.generative.utilities.loader as loading_utils
-import torch
-from torch import nn
 
 TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
     ff_up_proj="model.layers.{}.mlp.fc1",
@@ -38,78 +33,6 @@ TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
 )
 
 
-class Phi2(nn.Module):
-  """A Phi-2 model built from the Edge Generative API layers."""
-
-  def __init__(self, config: cfg.ModelConfig):
-    super().__init__()
-
-    # Construct model layers.
-    self.lm_head = nn.Linear(
-        config.embedding_dim, config.vocab_size, bias=config.lm_head_use_bias
-    )
-    self.tok_embedding = nn.Embedding(
-        config.vocab_size, config.embedding_dim, padding_idx=0
-    )
-    # Phi-2 has only one block config.
-    block_config = config.block_config(0)
-    self.transformer_blocks = nn.ModuleList(
-        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(attn_config.rotary_percentage * attn_config.head_dim),
-        base=attn_config.rotary_base,
-    )
-    self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.kv_cache_max,
-    )
-    self.config = config
-
-  @torch.inference_mode
-  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)
-    sin = sin.index_select(0, input_pos)
-    mask = self.mask_cache.index_select(2, input_pos)
-    mask = mask[:, :, :, : self.config.kv_cache_max]
-
-    x = self.tok_embedding(tokens)
-
-    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)
-    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:
   """Returns the model config for a Phi-2 model.
 
@@ -154,6 +77,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       block_configs=block_config,
       final_norm_config=norm_config,
       lm_head_use_bias=True,
+      lm_head_share_weight_with_embedding=False,
       enable_hlfb=True,
   )
   return config
@@ -169,11 +93,11 @@ def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
   return config
 
 
-def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
-  """Instantiates the model instance and load checkpoint if provided."""
-  config = get_model_config(**kwargs)
-  model = Phi2(config)
-  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
-  loader.load(model)
-  model.eval()
-  return model
+def build_model(
+    checkpoint_path: str, **kwargs
+) -> model_builder.DecoderOnlyModel:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+  )

ai_edge_torch/generative/examples/phi/phi3.py

@@ -18,14 +18,10 @@
 import math
 from typing import Tuple
 
-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
+from ai_edge_torch.generative.utilities import model_builder
 import ai_edge_torch.generative.utilities.loader as loading_utils
 import torch
-from torch import nn
 
 TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
     ff_up_proj="model.layers.{}.mlp.gate_up_proj",
@@ -137,32 +133,14 @@ def _build_rope_cache(
   return cos, sin
 
 
-class Phi3_5Mini(nn.Module):
+class Phi3_5Mini(model_builder.DecoderOnlyModel):
   """A Phi-3.5 model built from the Edge Generative API layers."""
 
   def __init__(self, config: cfg.ModelConfig):
-    super().__init__()
-
-    # Construct model layers.
-    self.lm_head = nn.Linear(
-        config.embedding_dim, config.vocab_size, bias=config.lm_head_use_bias
-    )
-    self.tok_embedding = nn.Embedding(
-        config.vocab_size, config.embedding_dim, padding_idx=0
-    )
-    # Phi-3.5 has only one block config.
-    block_config = config.block_config(0)
-    self.transformer_blocks = nn.ModuleList(
-        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
+    super().__init__(config)
+    attn_config = self.config.block_config(0).attn_config
     self.rope_cache = _build_rope_cache(
-        size=config.kv_cache_max,
+        size=self.config.kv_cache_max,
         dim=int(attn_config.rotary_percentage * attn_config.head_dim),
         base=attn_config.rotary_base,
         condense_ratio=1,
@@ -173,47 +151,6 @@ class Phi3_5Mini(nn.Module):
             1 + math.log(ROPE_SCALE_FACTOR) / math.log(config.max_seq_len)
         ),
     )
-    self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.kv_cache_max,
-    )
-    self.config = config
-
-  @torch.inference_mode
-  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)
-    sin = sin.index_select(0, input_pos)
-    mask = self.mask_cache.index_select(2, input_pos)
-    mask = mask[:, :, :, : self.config.kv_cache_max]
-
-    x = self.tok_embedding(tokens)
-
-    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)
-    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:
@@ -254,6 +191,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       embedding_dim=3072,
       block_configs=block_config,
       final_norm_config=norm_config,
+      lm_head_share_weight_with_embedding=False,
       enable_hlfb=True,
   )
   return config
@@ -269,7 +207,9 @@ def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
   return config
 
 
-def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
+def build_model(
+    checkpoint_path: str, **kwargs
+) -> model_builder.DecoderOnlyModel:
   """Instantiates the model instance and load checkpoint if provided."""
   config = get_model_config(**kwargs)
   model = Phi3_5Mini(config)