ai-edge-torch-nightly 0.3.0.dev20241218__py3-none-any.whl → 0.3.0.dev20241224__py3-none-any.whl

ai_edge_torch/generative/examples/gemma/gemma1.py

@@ -72,12 +72,13 @@ def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       pre_attention_norm_config=norm_config,
       post_attention_norm_config=norm_config,
   )
+  embedding_dim = 2048
   config = cfg.ModelConfig(
       vocab_size=256000,
       num_layers=18,
       max_seq_len=8192,
-      embedding_dim=2048,
-      embedding_scale=2048**0.5,
+      embedding_dim=embedding_dim,
+      embedding_scale=embedding_dim**0.5,
       kv_cache_max_len=kv_cache_max_len,
       block_configs=block_config,
       final_norm_config=norm_config,

ai_edge_torch/generative/examples/gemma/gemma2.py

@@ -15,13 +15,14 @@
 
 """Example of building a Gemma2 model."""
 
-from typing import Optional, Tuple
+from typing import List, Optional, 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
+import ai_edge_torch.generative.layers.rotary_position_embedding as rotary_pos_emb
 from ai_edge_torch.generative.utilities import model_builder
 import ai_edge_torch.generative.utilities.loader as loading_utils
 import torch
@@ -103,17 +104,12 @@ class Gemma2(nn.Module):
         config.embedding_dim,
         config.final_norm_config,
     )
-    # Gemma2 has same hyper parameters for each layer except for attention
-    # types. Use the first layer.
-    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,
     )
+    # Gemma2 has same hyper parameters for each layer except for attention
+    # types. Use the first layer.
+    attn_config = config.block_config(0).attn_config
     self.sliding_window_mask_cache = attn_utils.build_sliding_window_mask_cache(
         size=config.kv_cache_max,
         window_size=attn_config.sliding_window_size,
@@ -140,29 +136,48 @@ class Gemma2(nn.Module):
         f"Cannot forward sequence of length {seq_len}, max seq length is only"
         f" {self.config.max_seq_len}"
     )
+
+    # token embeddings of shape (b, t, n_embd)
+    input_embeds = self.tok_embedding(tokens)
+    # RoPE parameters are the same for all blocks. Use the first layer.
+    attn_config = self.config.block_config(0).attn_config
+    n_elem = int(attn_config.rotary_percentage * attn_config.head_dim)
+    rope = rotary_pos_emb.build_rope(
+        input_pos, n_elem, attn_config.head_dim, attn_config.rotary_base
+    )
+    mask = [self.get_attention_mask(
+        self.config.block_config(i).attn_config.attn_type, input_pos
+    ) for i in range(self.config.num_layers)]
+
+    return self._forward_with_embeds(
+        input_embeds, rope, mask, input_pos, kv_cache, export_config
+    )
+
+  def _forward_with_embeds(
+      self,
+      input_embeds: torch.Tensor,
+      rope: Tuple[torch.Tensor, torch.Tensor],
+      mask: List[torch.Tensor],
+      input_pos: torch.Tensor,
+      kv_cache: kv_utils.KVCache,
+      export_config: Optional[model_builder.ExportConfig] = None,
+  ) -> dict[torch.Tensor, kv_utils.KVCache]:
+    """Forwards the model with input embeddings."""
     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)
-
-    # token embeddings of shape (b, t, n_embd)
-    x = self.tok_embedding(tokens)
-    x = x * (self.config.embedding_dim**0.5)
-
-    updated_kv_entires = []
+    if self.config.embedding_scale is not None:
+      input_embeds = input_embeds * self.config.embedding_scale
+    x = input_embeds
+    updated_kv_entries = []
     for i, block in enumerate(self.transformer_blocks):
-      mask = self.get_attention_mask(
-          block.config.attn_config.attn_type, input_pos
-      )
       kv_entry = kv_cache.caches[i] if kv_cache else None
-      x, kv_entry = block(x, (cos, sin), mask, input_pos, kv_entry)
+      x, kv_entry = block(x, rope, mask[i], input_pos, kv_entry)
       if kv_entry:
-        updated_kv_entires.append(kv_entry)
-    updated_kv_cache = kv_utils.KVCache(tuple(updated_kv_entires))
+        updated_kv_entries.append(kv_entry)
+    updated_kv_cache = kv_utils.KVCache(tuple(updated_kv_entries))
 
     if export_config is not None:
       if (
@@ -228,11 +243,13 @@ def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
     )
 
   num_layers = 26
+  embedding_dim = 2304
   config = cfg.ModelConfig(
       vocab_size=256000,
       num_layers=num_layers,
       max_seq_len=8192,
-      embedding_dim=2304,
+      embedding_dim=embedding_dim,
+      embedding_scale=embedding_dim**0.5,
       kv_cache_max_len=kv_cache_max_len,
       block_configs=[get_block_config(i) for i in range(num_layers)],
       final_norm_config=norm_config,
@@ -249,6 +266,7 @@ def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
   config.num_layers = 2
   config.max_seq_len = 2 * kv_cache_max_len
   config.embedding_dim = 128
+  config.embedding_scale = config.embedding_dim**0.5
   config.block_configs = config.block_configs[: config.num_layers]
   for block_config in config.block_configs:
     block_config.attn_config.num_heads = 4

ai_edge_torch/generative/examples/paligemma/convert_to_tflite.py

@@ -29,9 +29,15 @@ from ai_edge_torch.generative.utilities import converter
 from ai_edge_torch.generative.utilities.model_builder import ExportConfig
 import torch
 
+_VERSION = flags.DEFINE_enum(
+    'version',
+    '2',
+    ['1', '2'],
+    'The version of PaliGemma model to verify.',
+)
 _CHECKPOINT_PATH = flags.DEFINE_string(
     'checkpoint_path',
-    os.path.join(pathlib.Path.home(), 'Downloads/llm_data/paligemma-3b-224'),
+    os.path.join(pathlib.Path.home(), 'Downloads/llm_data/paligemma2-3b-224'),
     'The path to the model checkpoint, or directory holding the checkpoint.',
 )
 _TFLITE_PATH = flags.DEFINE_string(
@@ -63,10 +69,12 @@ _QUANTIZE = flags.DEFINE_bool(
 
 def main(_):
   pytorch_model = paligemma.build_model(
-      _CHECKPOINT_PATH.value, kv_cache_max_len=_KV_CACHE_MAX_LEN.value
+      _CHECKPOINT_PATH.value,
+      version=int(_VERSION.value),
+      kv_cache_max_len=_KV_CACHE_MAX_LEN.value,
   )
   quant_suffix = 'q8' if _QUANTIZE.value else 'f32'
-  output_filename = f'paligemma_{quant_suffix}_seq{_PREFILL_SEQ_LEN.value}_ekv{_KV_CACHE_MAX_LEN.value}.tflite'
+  output_filename = f'paligemma{_VERSION.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/paligemma/decoder.py

@@ -19,6 +19,7 @@ from typing import Optional
 
 from ai_edge_torch.generative.layers import kv_cache as kv_utils
 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.utilities import model_builder
 import ai_edge_torch.generative.utilities.loader as loading_utils
 import torch
@@ -54,6 +55,7 @@ class Decoder(model_builder.DecoderOnlyModel):
       kv_cache: kv_utils.KVCache,
       input_embeds: torch.Tensor = None,
       export_config: Optional[model_builder.ExportConfig] = None,
+      called_by_generate: bool = True,
   ) -> dict[torch.Tensor, kv_utils.KVCache]:
     if input_embeds is None:
       return super().forward(tokens, input_pos, kv_cache)
@@ -61,8 +63,12 @@ class Decoder(model_builder.DecoderOnlyModel):
     assert input_embeds is not None
 
     repo_pos = input_pos + 1  # PaliGemma position is 1-based.
-    cos, sin = self.rope_cache
-    rope = (cos.index_select(0, repo_pos), sin.index_select(0, repo_pos))
+    # ROPE parameters for all attn_configs are the same. Take the first one.
+    attn_config = self.config.block_config(0).attn_config
+    n_elem = int(attn_config.rotary_percentage * attn_config.head_dim)
+    rope = rotary_pos_emb.build_rope(
+        repo_pos, n_elem, attn_config.head_dim, attn_config.rotary_base
+    )
 
     # The first part of input_embeds are image embeddings. Diagonal causal mask
     # doesn't work here.
@@ -70,7 +76,7 @@ class Decoder(model_builder.DecoderOnlyModel):
     mask = torch.zeros(embeds_len, self.config.kv_cache_max)
     mask[:, embeds_len:] = float("-inf")
 
-    return self.forward_with_embeds(
+    return self._forward_with_embeds(
         input_embeds, rope, mask, input_pos, kv_cache
     )
 
@@ -108,12 +114,13 @@ def get_decoder_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       pre_attention_norm_config=norm_config,
       post_attention_norm_config=norm_config,
   )
+  embedding_dim = 2048
   config = cfg.ModelConfig(
       vocab_size=257216,
       num_layers=18,
       max_seq_len=8192,
-      embedding_dim=2048,
-      embedding_scale=2048**0.5,
+      embedding_dim=embedding_dim,
+      embedding_scale=embedding_dim**0.5,
       kv_cache_max_len=kv_cache_max_len,
       block_configs=block_config,
       final_norm_config=norm_config,
@@ -130,6 +137,8 @@ def get_fake_decoder_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
   config.vocab_size = 128
   config.num_layers = 2
   config.max_seq_len = 2 * kv_cache_max_len
+  config.embedding_dim = 128
+  config.embedding_scale = 128**0.5
   return config
 
 

ai_edge_torch/generative/examples/paligemma/decoder2.py

@@ -0,0 +1,174 @@
+# 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.
+# ==============================================================================
+
+"""Example of building a decoder of PaliGemma2 3B model which is Gemma2."""
+
+from typing import Optional
+
+from ai_edge_torch.generative.examples.gemma import gemma2
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
+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.utilities import model_builder
+import ai_edge_torch.generative.utilities.loader as loading_utils
+import torch
+
+TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
+    ff_up_proj="language_model.model.layers.{}.mlp.up_proj",
+    ff_down_proj="language_model.model.layers.{}.mlp.down_proj",
+    ff_gate_proj="language_model.model.layers.{}.mlp.gate_proj",
+    attn_query_proj="language_model.model.layers.{}.self_attn.q_proj",
+    attn_key_proj="language_model.model.layers.{}.self_attn.k_proj",
+    attn_value_proj="language_model.model.layers.{}.self_attn.v_proj",
+    attn_output_proj="language_model.model.layers.{}.self_attn.o_proj",
+    pre_attn_norm="language_model.model.layers.{}.input_layernorm",
+    post_attn_norm="language_model.model.layers.{}.post_attention_layernorm",
+    pre_ff_norm="language_model.model.layers.{}.pre_feedforward_layernorm",
+    post_ff_norm="language_model.model.layers.{}.post_feedforward_layernorm",
+    embedding="language_model.model.embed_tokens",
+    final_norm="language_model.model.norm",
+    lm_head=None,
+)
+
+
+class Decoder2(gemma2.Gemma2):
+  """A decoder of PaliGemma2 3B model which is Gemma2.
+
+  Besides a tensor of text token IDs, forward() can also take a tensor of
+  embeddings which may include text or image or both.
+  """
+
+  @torch.inference_mode
+  def forward(
+      self,
+      tokens: torch.Tensor,
+      input_pos: torch.Tensor,
+      kv_cache: kv_utils.KVCache,
+      input_embeds: torch.Tensor = None,
+      export_config: Optional[model_builder.ExportConfig] = None,
+      called_by_generate: bool = True,
+  ) -> dict[torch.Tensor, kv_utils.KVCache]:
+    if input_embeds is None:
+      return super().forward(tokens, input_pos, kv_cache)
+
+    assert input_embeds is not None
+
+    repo_pos = input_pos + 1  # PaliGemma2 position is 1-based.
+    # ROPE parameters for all attn_configs are the same. Take the first one.
+    attn_config = self.config.block_config(0).attn_config
+    n_elem = int(attn_config.rotary_percentage * attn_config.head_dim)
+    rope = rotary_pos_emb.build_rope(
+        repo_pos, n_elem, attn_config.head_dim, attn_config.rotary_base
+    )
+
+    if called_by_generate:
+      # PaliGemma2 generate() use a diagonal causal mask even with image embeds.
+      mask = [self.get_attention_mask(
+          self.config.block_config(i).attn_config.attn_type, input_pos
+      ) for i in range(self.config.num_layers)]
+    else:
+      # By default, don't mask image embeds with a diagonal causal mask.
+      embeds_len = input_embeds.shape[1]
+      mask = torch.zeros(embeds_len, self.config.kv_cache_max)
+      mask[:, embeds_len:] = float("-inf")
+      mask = [mask] * self.config.num_layers
+
+    return self._forward_with_embeds(
+        input_embeds, rope, mask, input_pos, kv_cache, export_config
+    )
+
+
+def get_decoder2_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for the decoder of a PaliGemma 3B model.
+
+  Args:
+    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
+      is 1024.
+
+  Returns:
+    The model config for the decoder of a PaliGemma 3B model.
+  """
+  norm_config = cfg.NormalizationConfig(
+      type=cfg.NormalizationType.RMS_NORM,
+      epsilon=1e-6,
+      zero_centered=True,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.GATED,
+      activation=cfg.ActivationConfig(cfg.ActivationType.GELU_TANH),
+      intermediate_size=9216,
+      pre_ff_norm_config=norm_config,
+      post_ff_norm_config=norm_config,
+  )
+
+  def get_block_config(idx: int) -> cfg.TransformerBlockConfig:
+    attn_config = cfg.AttentionConfig(
+        num_heads=8,
+        head_dim=256,
+        num_query_groups=4,
+        rotary_base=10000,
+        rotary_percentage=1.0,
+        logit_softcap=50.0,
+        sliding_window_size=4096,
+        attn_type=(
+            cfg.AttentionType.GLOBAL
+            if idx % 2 == 0
+            else cfg.AttentionType.LOCAL_SLIDING
+        ),
+    )
+    return cfg.TransformerBlockConfig(
+        attn_config=attn_config,
+        ff_config=ff_config,
+        pre_attention_norm_config=norm_config,
+        post_attention_norm_config=norm_config,
+    )
+
+  num_layers = 26
+  embedding_dim = 2304
+  config = cfg.ModelConfig(
+      vocab_size=257216,
+      num_layers=num_layers,
+      max_seq_len=8192,
+      embedding_dim=embedding_dim,
+      embedding_scale=embedding_dim**0.5,
+      kv_cache_max_len=kv_cache_max_len,
+      block_configs=[get_block_config(i) for i in range(num_layers)],
+      final_norm_config=norm_config,
+      lm_head_use_bias=False,
+      enable_hlfb=True,
+      final_logit_softcap=30.0,
+  )
+  return config
+
+
+def get_fake_decoder2_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
+  config = get_decoder2_config(kv_cache_max_len)
+  # PaliGemma2 decoder has only one block config.
+  config.block_config(0).ff_config.intermediate_size = 128
+  config.vocab_size = 128
+  config.num_layers = 2
+  config.max_seq_len = 2 * kv_cache_max_len
+  config.embedding_dim = 128
+  config.embedding_scale = 128**0.5
+  return config
+
+
+def build_decoder2(checkpoint_path: str, **kwargs) -> torch.nn.Module:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_decoder2_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+      model_class=Decoder2,
+  )

ai_edge_torch/generative/examples/paligemma/paligemma.py

@@ -19,6 +19,7 @@ from dataclasses import dataclass
 from typing import Optional
 
 from ai_edge_torch.generative.examples.paligemma import decoder
+from ai_edge_torch.generative.examples.paligemma import decoder2
 from ai_edge_torch.generative.examples.paligemma import image_encoder
 import ai_edge_torch.generative.layers.kv_cache as kv_utils
 import ai_edge_torch.generative.layers.model_config as cfg
@@ -38,13 +39,14 @@ class PaliGemmaConfig:
   decoder_config: cfg.ModelConfig
 
   image_token_id: int
+  image_projection_scale: float
   image_projection_use_bias: bool = False
 
 
 class PaliGemma(nn.Module):
   """PaliGemma model from the Edge Generative API."""
 
-  def __init__(self, config: PaliGemmaConfig):
+  def __init__(self, config: PaliGemmaConfig, decoder_class: nn.Module):
     super().__init__()
 
     self.image_encoder = image_encoder.SiglipVisionEncoder(
@@ -55,7 +57,7 @@ class PaliGemma(nn.Module):
         config.decoder_config.embedding_dim,
         bias=config.image_projection_use_bias,
     )
-    self.decoder = decoder.Decoder(config.decoder_config)
+    self.decoder = decoder_class(config.decoder_config)
     image_embedding_config = config.image_encoder_config.image_embedding
     self.num_patches = (
         image_embedding_config.image_size // image_embedding_config.patch_size
@@ -70,6 +72,7 @@ class PaliGemma(nn.Module):
       kv_cache: kv_utils.KVCache,
       pixel_values: torch.Tensor = None,
       export_config: Optional[model_builder.ExportConfig] = None,
+      called_by_generate: bool = True,
   ) -> dict[torch.Tensor, kv_utils.KVCache]:
     if pixel_values is None:
       return self.decoder(
@@ -77,15 +80,15 @@ class PaliGemma(nn.Module):
           input_pos=input_pos,
           kv_cache=kv_cache,
           input_embeds=None,
-          export_config=export_config
+          export_config=export_config,
+          called_by_generate=called_by_generate,
       )
 
     input_embeds = self.decoder.tok_embedding(tokens)
 
     image_encoded = self.image_encoder(pixel_values=pixel_values)
     image_embeds = self.image_projection(image_encoded)
-    if self.config.decoder_config.embedding_scale is not None:
-      image_embeds = image_embeds / self.config.decoder_config.embedding_scale
+    image_embeds = image_embeds / self.config.image_projection_scale
 
     # Merging image_embeds into text_embeds as PaliGemmaForConditionalGeneration
     # can be done like:
@@ -110,10 +113,11 @@ class PaliGemma(nn.Module):
         kv_cache=kv_cache,
         input_embeds=input_embeds,
         export_config=export_config,
+        called_by_generate=called_by_generate,
     )
 
 
-def get_model_config(**kwargs) -> PaliGemmaConfig:
+def get_model_config(get_decoder_config, **kwargs) -> PaliGemmaConfig:
   """Returns the model config for a PaliGemma 3B-224 model.
 
   Returns:
@@ -121,31 +125,42 @@ def get_model_config(**kwargs) -> PaliGemmaConfig:
   """
   return PaliGemmaConfig(
       image_encoder_config=image_encoder.get_image_encoder_config(),
-      decoder_config=decoder.get_decoder_config(**kwargs),
-      image_projection_use_bias=True,
+      decoder_config=get_decoder_config(**kwargs),
       image_token_id=257152,
+      image_projection_scale=2048**0.5,
+      image_projection_use_bias=True,
   )
 
 
-def get_fake_model_config() -> PaliGemmaConfig:
+def get_fake_model_config(get_decoder_config, **kwargs) -> PaliGemmaConfig:
   return PaliGemmaConfig(
       image_encoder_config=image_encoder.get_fake_image_encoder_config(),
-      decoder_config=decoder.get_fake_decoder_config(),
+      decoder_config=get_decoder_config(**kwargs),
+      image_token_id=127,
+      image_projection_scale=128**0.5,
       image_projection_use_bias=True,
-      image_token_id=257152,
   )
 
 
-def build_model(checkpoint_path: str, **kwargs) -> PaliGemma:
-  config = get_model_config(**kwargs)
-  model = PaliGemma(config)
+def build_model(checkpoint_path: str, version: int = 2, **kwargs) -> PaliGemma:
+  if version == 1:
+    decoder_class = decoder.Decoder
+    decoder_tensor_names = decoder.TENSOR_NAMES
+    get_decoder_config = decoder.get_decoder_config
+  else:
+    decoder_class = decoder2.Decoder2
+    decoder_tensor_names = decoder2.TENSOR_NAMES
+    get_decoder_config = decoder2.get_decoder2_config
+
+  config = get_model_config(get_decoder_config, **kwargs)
+  model = PaliGemma(config, decoder_class)
   # Load the parameters of image encoder.
   loader = loading_utils.ModelLoader(
       checkpoint_path, image_encoder.TENSOR_NAMES
   )
   loader.load(model.image_encoder, strict=False)
   # Load the parameters of decoder.
-  loader = loading_utils.ModelLoader(checkpoint_path, decoder.TENSOR_NAMES)
+  loader = loading_utils.ModelLoader(checkpoint_path, decoder_tensor_names)
   loader.load(model.decoder, strict=False)
 
   # Load the parameters of image projection.

ai_edge_torch/generative/examples/paligemma/verify.py

@@ -22,11 +22,18 @@ from absl import flags
 from ai_edge_torch.generative.examples.paligemma import paligemma
 from ai_edge_torch.generative.layers import kv_cache
 from ai_edge_torch.generative.utilities import verifier
+import kagglehub
 from PIL import Image
 import requests
 import torch
 import transformers
 
+_VERSION = flags.DEFINE_enum(
+    "version",
+    "2",
+    ["1", "2"],
+    "The version of PaliGemma model to verify.",
+)
 _IMAGE_URL = flags.DEFINE_string(
     "image_url",
     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg?download=true",
@@ -34,7 +41,7 @@ _IMAGE_URL = flags.DEFINE_string(
 )
 _PROMPTS = flags.DEFINE_string(
     "prompts",
-    "Caption en",
+    "describe en",
     "The input prompts to generate answers.",
 )
 _MAX_NEW_TOKENS = flags.DEFINE_integer(
@@ -43,28 +50,47 @@ _MAX_NEW_TOKENS = flags.DEFINE_integer(
     "The maximum size of the generated tokens.",
 )
 
+_CHECKPOINT = {
+    "1": "google/paligemma-3b-mix-224",
+    "2": "google/paligemma-2/transformers/paligemma2-3b-pt-224",
+}
+
 
 class ReauthoredPaliGemmaWrapper(verifier.ReauthoredModelWrapper):
   """Reauthored PaliGemma model wrapper."""
 
+  def __init__(self, model: torch.nn.Module):
+    super().__init__(model)
+    self.forward_called_by_generate = False
+
   def _init_kv_cache(self):
     return kv_cache.KVCache.from_model_config(self.model.config.decoder_config)
 
+  def _get_extra_args_for_forward(self):
+    return {"called_by_generate": self.forward_called_by_generate}
+
 
 def main(_):
-  checkpoint = "google/paligemma-3b-mix-224"
+  if _VERSION.value == "1":
+    checkpoint = _CHECKPOINT[_VERSION.value]
+    # Locate the cached dir.
+    cached_config_file = transformers.utils.cached_file(
+        checkpoint, transformers.utils.CONFIG_NAME
+    )
+    reauthored_checkpoint = pathlib.Path(cached_config_file).parent
+  else:
+    checkpoint = kagglehub.model_download(_CHECKPOINT[_VERSION.value])
+    reauthored_checkpoint = checkpoint
+
   logging.info("Loading the original model from: %s", checkpoint)
   original_model = (
       transformers.PaliGemmaForConditionalGeneration.from_pretrained(checkpoint)
   )
 
-  # Locate the cached dir.
-  cached_config_file = transformers.utils.cached_file(
-      checkpoint, transformers.utils.CONFIG_NAME
-  )
-  reauthored_checkpoint = pathlib.Path(cached_config_file).parent
   logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
-  reauthored_model = paligemma.build_model(reauthored_checkpoint)
+  reauthored_model = paligemma.build_model(
+      reauthored_checkpoint, version=int(_VERSION.value)
+  )
 
   logging.info("Loading the processor from: %s", checkpoint)
   # It works only when GemmaTokenizerFast is available. In some environments,
@@ -93,7 +119,7 @@ def main(_):
   logging.info("outputs_reauthored: %s", outputs_reauthored)
 
   try:
-    assert torch.allclose(outputs_original, outputs_reauthored, atol=1e-03)
+    assert torch.allclose(outputs_original, outputs_reauthored, atol=1e-02)
   except AssertionError as e:
     logging.error("*** FAILED *** verify with forward()")
     raise e
@@ -111,6 +137,7 @@ def main(_):
   logging.info("outputs_from_original_model: [[%s]]", response_original)
 
   logging.info("Generating answer with the reauthored model...")
+  wrapped_reauthored_model.forward_called_by_generate = True
   outputs_reauthored = wrapped_reauthored_model.generate(
       prompts=inputs["input_ids"],
       pixel_values=inputs["pixel_values"],