ai-edge-torch-nightly 0.3.0.dev20240910__py3-none-any.whl → 0.3.0.dev20240911__py3-none-any.whl

ai_edge_torch/generative/examples/experimental/tiny_llama/tiny_llama.py

@@ -1,205 +0,0 @@
-# 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 TinyLlama model from the Edge Generative API layers.
-#
-# Note: This is an experimental version of TinyLlama with external KV cache.
-# Please use with caution.
-
-import os
-from pathlib import Path
-from typing import Tuple
-
-from ai_edge_torch.generative.layers import builder
-import ai_edge_torch.generative.layers.attention_utils as attn_utils
-from ai_edge_torch.generative.layers.experimental import attention
-from ai_edge_torch.generative.layers.experimental import ekv_cache as kv_utils
-import ai_edge_torch.generative.layers.model_config as cfg
-import ai_edge_torch.generative.utilities.loader as loading_utils
-import numpy as np
-import torch
-from torch import nn
-
-
-TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
-    ff_up_proj="model.layers.{}.mlp.up_proj",
-    ff_down_proj="model.layers.{}.mlp.down_proj",
-    ff_gate_proj="model.layers.{}.mlp.gate_proj",
-    attn_query_proj="model.layers.{}.self_attn.q_proj",
-    attn_key_proj="model.layers.{}.self_attn.k_proj",
-    attn_value_proj="model.layers.{}.self_attn.v_proj",
-    attn_output_proj="model.layers.{}.self_attn.o_proj",
-    pre_attn_norm="model.layers.{}.input_layernorm",
-    post_attn_norm="model.layers.{}.post_attention_layernorm",
-    embedding="model.embed_tokens",
-    final_norm="model.norm",
-    lm_head="lm_head",
-)
-
-
-class TinyLLamma(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
-    )
-    self.tok_embedding = nn.Embedding(
-        config.vocab_size, config.embedding_dim, padding_idx=0
-    )
-    self.transformer_blocks = nn.ModuleList(
-        attention.TransformerBlock(config) for _ in range(config.num_layers)
-    )
-    self.final_norm = builder.build_norm(
-        config.embedding_dim,
-        config.final_norm_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
-        ),
-        base=10_000,
-        condense_ratio=1,
-        dtype=torch.float32,
-        device=torch.device("cpu"),
-    )
-    self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.kv_cache_max,
-        dtype=torch.float32,
-        device=torch.device("cpu"),
-    )
-    self.config = config
-
-  @torch.inference_mode
-  def forward(
-      self,
-      tokens: torch.Tensor,
-      input_pos: torch.Tensor,
-      kv_cache: kv_utils.EKVCache,
-  ) -> Tuple[torch.Tensor, kv_utils.EKVCache]:
-    _, 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}"
-    )
-
-    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.EKVCache(tuple(updated_kv_entires))
-
-    x = self.final_norm(x)
-    res = self.lm_head(x)  # (b, t, vocab_size)
-    return res, updated_kv_cache
-
-
-def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
-  """Returns the model config for a TinyLlama model.
-
-  Args:
-    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
-      is 1024.
-
-  Returns:
-    The model config for a TinyLlama model.
-  """
-  attn_config = cfg.AttentionConfig(
-      num_heads=32,
-      head_dim=64,
-      num_query_groups=4,
-      rotary_percentage=1.0,
-  )
-  ff_config = cfg.FeedForwardConfig(
-      type=cfg.FeedForwardType.GATED,
-      activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
-      intermediate_size=5632,
-  )
-  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)
-  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,
-      final_norm_config=norm_config,
-      enable_hlfb=True,
-  )
-  return 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 = 256
-  return config
-
-
-def build_model(
-    checkpoint_path: str, test_model: bool = False, **kwargs
-) -> nn.Module:
-  """Instantiates the model instance and load checkpoint if provided."""
-  config = (
-      get_fake_model_config(**kwargs)
-      if test_model
-      else get_model_config(**kwargs)
-  )
-  model = TinyLLamma(config)
-  if checkpoint_path is not None:
-    loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
-    loader.load(model)
-  model.eval()
-  return model
-
-
-def define_and_run(checkpoint_path: str, test_model: bool = False) -> None:
-  """Instantiates and runs a TinyLlama model."""
-
-  kv_cache_max_len = 1024
-  model = build_model(
-      checkpoint_path, test_model=test_model, 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[0, :4] = idx
-  input_pos = torch.arange(0, kv_cache_max_len)
-  kv = kv_utils.EKVCache.from_model_config(model.config)
-  print("running an inference")
-  print(model.forward(tokens, input_pos, kv))
-
-
-if __name__ == "__main__":
-  input_checkpoint_path = os.path.join(Path.home(), "Downloads/tiny_llama")
-  define_and_run(input_checkpoint_path)

ai_edge_torch/generative/examples/phi2/__init__.py

@@ -1,14 +0,0 @@
-# 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.
-# ==============================================================================

ai_edge_torch/generative/examples/phi2/convert_to_tflite.py

@@ -1,67 +0,0 @@
-# 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 os
-from pathlib import Path
-
-import ai_edge_torch
-from ai_edge_torch.generative.examples.phi2 import phi2
-from ai_edge_torch.generative.quantize import quant_recipes
-import torch
-
-
-def convert_phi2_to_tflite(
-    checkpoint_path: str,
-    prefill_seq_len: int = 512,
-    kv_cache_max_len: int = 1024,
-    quantize: bool = True,
-):
-  """Converts a Phi-2 model to multi-signature tflite model.
-
-  Args:
-      checkpoint_path (str): The filepath to the model checkpoint, or directory
-        holding the checkpoint.
-      prefill_seq_len (int, optional): The maximum size of prefill input tensor.
-        Defaults to 512.
-      kv_cache_max_len (int, optional): The maximum size of KV cache buffer,
-        including both prefill and decode. Defaults to 1024.
-      quantize (bool, optional): Whether the model should be quanized. Defaults
-        to True.
-  """
-  pytorch_model = phi2.build_model(
-      checkpoint_path, kv_cache_max_len=kv_cache_max_len
-  )
-  # Tensors used to trace the model graph during conversion.
-  prefill_tokens = torch.full((1, prefill_seq_len), 0, dtype=torch.long)
-  prefill_input_pos = torch.arange(0, prefill_seq_len)
-  decode_token = torch.tensor([[0]], dtype=torch.long)
-  decode_input_pos = torch.tensor([0], dtype=torch.int64)
-
-  quant_config = quant_recipes.full_int8_dynamic_recipe() if quantize else None
-  edge_model = (
-      ai_edge_torch.signature(
-          'prefill', pytorch_model, (prefill_tokens, prefill_input_pos)
-      )
-      .signature('decode', pytorch_model, (decode_token, decode_input_pos))
-      .convert(quant_config=quant_config)
-  )
-  edge_model.export(
-      f'/tmp/phi2_seq{prefill_seq_len}_kv{kv_cache_max_len}.tflite'
-  )
-
-
-if __name__ == '__main__':
-  checkpoint_path = os.path.join(Path.home(), 'Downloads/llm_data/phi2')
-  convert_phi2_to_tflite(checkpoint_path)

ai_edge_torch/generative/examples/phi2/phi2.py

@@ -1,189 +0,0 @@
-# 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 phi-2 model from the Edge Generative API layers.
-
-
-import os
-from pathlib import Path
-
-from ai_edge_torch.generative.layers import attention
-from ai_edge_torch.generative.layers import builder
-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
-import numpy as np
-import torch
-from torch import nn
-
-TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
-    ff_up_proj="model.layers.{}.mlp.fc1",
-    ff_down_proj="model.layers.{}.mlp.fc2",
-    attn_query_proj="model.layers.{}.self_attn.q_proj",
-    attn_key_proj="model.layers.{}.self_attn.k_proj",
-    attn_value_proj="model.layers.{}.self_attn.v_proj",
-    attn_output_proj="model.layers.{}.self_attn.dense",
-    pre_attn_norm="model.layers.{}.input_layernorm",
-    embedding="model.embed_tokens",
-    final_norm="model.final_layernorm",
-    lm_head="lm_head",
-)
-
-
-class Phi2(nn.Module):
-  """A Phi-2 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
-    )
-    self.tok_embedding = nn.Embedding(
-        config.vocab_size, config.embedding_dim, padding_idx=0
-    )
-    self.transformer_blocks = nn.ModuleList(
-        attention.TransformerBlock(config) for _ in range(config.num_layers)
-    )
-    self.final_norm = builder.build_norm(
-        config.embedding_dim,
-        config.final_norm_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
-        ),
-        base=10_000,
-        condense_ratio=1,
-        dtype=torch.float32,
-        device=torch.device("cpu"),
-    )
-    self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.kv_cache_max,
-        dtype=torch.float32,
-        device=torch.device("cpu"),
-    )
-    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()
-    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}"
-    )
-
-    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]
-
-    # forward the model itself
-    x = self.tok_embedding(idx)  # token embeddings of shape (b, t, n_embd)
-
-    for _, block in enumerate(self.transformer_blocks):
-      x = block(x, (cos, sin), mask, input_pos)
-
-    x = self.final_norm(x)
-    res = self.lm_head(x)  # (b, t, vocab_size)
-    return res
-
-
-def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
-  """Returns the model config for a Phi-2 model.
-
-  Args:
-    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
-      is 1024.
-
-  Returns:
-    The model config for a Phi-2 model.
-  """
-  attn_config = cfg.AttentionConfig(
-      num_heads=32,
-      head_dim=80,
-      num_query_groups=32,
-      rotary_percentage=0.4,
-      qkv_use_bias=True,
-      output_proj_use_bias=True,
-  )
-  ff_config = cfg.FeedForwardConfig(
-      type=cfg.FeedForwardType.SEQUENTIAL,
-      activation=cfg.ActivationConfig(cfg.ActivationType.GELU_TANH),
-      intermediate_size=10240,
-      use_bias=True,
-  )
-  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.LAYER_NORM)
-  config = cfg.ModelConfig(
-      vocab_size=51200,
-      num_layers=32,
-      max_seq_len=2048,
-      kv_cache_max_len=kv_cache_max_len,
-      embedding_dim=2560,
-      attn_config=attn_config,
-      ff_config=ff_config,
-      pre_attention_norm_config=norm_config,
-      final_norm_config=norm_config,
-      parallel_residual=True,
-      lm_head_use_bias=True,
-      enable_hlfb=True,
-  )
-  return config
-
-
-def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
-  config = get_model_config(kv_cache_max_len)
-  config.vocab_size = 128
-  config.num_layers = 2
-  config.max_seq_len = 2 * kv_cache_max_len
-  config.ff_config.intermediate_size = 128
-  return config
-
-
-def build_model(checkpoint_path, **kwargs) -> nn.Module:
-  config = get_model_config(**kwargs)
-  model = Phi2(config)
-  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
-  loader.load(model)
-  return model
-
-
-def define_and_run() -> None:
-  """Instantiates and runs a Phi-2 model."""
-
-  current_dir = Path(__file__).parent.resolve()
-  phi2_goldens = torch.load(current_dir / "phi2_lm_logits.pt")
-  kv_cache_max_len = 1024
-  checkpoint_path = os.path.join(Path.home(), "Downloads/llm_data/phi2")
-  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[0, :4] = idx
-  input_pos = torch.arange(0, kv_cache_max_len)
-  lm_logits = model.forward(tokens, input_pos)
-  print("comparing with goldens..")
-  assert torch.allclose(
-      phi2_goldens, lm_logits[0, idx.shape[1] - 1, :], atol=1e-05
-  )
-
-
-if __name__ == "__main__":
-  define_and_run()

ai_edge_torch/generative/examples/test_models/toy_model_with_external_kv_cache.py

@@ -1,176 +0,0 @@
-# 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.
-# ==============================================================================
-# A toy example which has basic transformer block (w/ externalized KV-Cache).
-
-from typing import Tuple
-
-import ai_edge_torch
-from ai_edge_torch import lowertools
-import ai_edge_torch.generative.layers.attention_utils as attn_utils
-import ai_edge_torch.generative.layers.builder as builder
-from ai_edge_torch.generative.layers.experimental import ekv_cache as kv_utils
-from ai_edge_torch.generative.layers.experimental.attention import TransformerBlock  # NOQA
-import ai_edge_torch.generative.layers.model_config as cfg
-import torch
-import torch.nn as nn
-
-RoPECache = Tuple[torch.Tensor, torch.Tensor]
-
-
-class ToyModelWithExternalKV(torch.nn.Module):
-
-  def __init__(self, config: cfg.ModelConfig) -> None:
-    super().__init__()
-    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)
-    self.transformer_blocks = nn.ModuleList(
-        TransformerBlock(config) for _ in range(config.num_layers)
-    )
-    self.final_norm = builder.build_norm(
-        config.embedding_dim,
-        config.final_norm_config,
-    )
-    self.rope_cache = attn_utils.build_rope_cache(
-        size=config.max_seq_len,
-        dim=int(
-            config.attn_config.rotary_percentage * config.attn_config.head_dim
-        ),
-        base=10_000,
-        condense_ratio=1,
-        dtype=torch.float32,
-        device=torch.device('cpu'),
-    )
-    self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.max_seq_len, dtype=torch.float32, device=torch.device('cpu')
-    )
-    self.config = config
-
-  def forward(
-      self,
-      tokens: torch.Tensor,
-      input_pos: torch.Tensor,
-      kv_cache: kv_utils.EKVCache,
-  ) -> Tuple[torch.Tensor, kv_utils.EKVCache]:
-    x = self.tok_embedding(tokens)
-    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.max_seq_len]
-
-    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)
-
-    x = self.final_norm(x)
-    updated_kv_cache = kv_utils.EKVCache(tuple(updated_kv_entires))
-    return self.lm_head(x), updated_kv_cache
-
-
-def _export_stablehlo_mlir(model, args):
-  ep = torch.export.export(model, args)
-  return lowertools.exported_program_to_mlir_text(ep)
-
-
-def get_model_config() -> cfg.ModelConfig:
-  attn_config = cfg.AttentionConfig(
-      num_heads=32, head_dim=4, num_query_groups=4, rotary_percentage=1.0
-  )
-  ff_config = cfg.FeedForwardConfig(
-      type=cfg.FeedForwardType.GATED,
-      activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
-      intermediate_size=256,
-  )
-  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)
-  config = cfg.ModelConfig(
-      vocab_size=150,
-      num_layers=2,
-      max_seq_len=100,
-      embedding_dim=128,
-      attn_config=attn_config,
-      ff_config=ff_config,
-      pre_attention_norm_config=norm_config,
-      post_attention_norm_config=norm_config,
-      final_norm_config=norm_config,
-      enable_hlfb=True,
-  )
-  return config
-
-
-def get_sample_prefill_inputs() -> Tuple[torch.Tensor, torch.Tensor]:
-  tokens = torch.unsqueeze(torch.arange(0, 100), 0)
-  input_pos = torch.arange(0, 100)
-  return tokens, input_pos
-
-
-def get_sample_decode_inputs() -> Tuple[torch.Tensor, torch.Tensor]:
-  tokens = torch.tensor([[1]], dtype=torch.long)
-  input_pos = torch.tensor([10])
-  return tokens, input_pos
-
-
-def define_and_run() -> None:
-  dump_mlir = False
-
-  config = get_model_config()
-  model = ToyModelWithExternalKV(config)
-  model.eval()
-  print('running an inference')
-  kv = kv_utils.EKVCache.from_model_config(config)
-
-  tokens, input_pos = get_sample_prefill_inputs()
-  decode_token, decode_input_pos = get_sample_decode_inputs()
-  print(model.forward(tokens, input_pos, kv))
-
-  if dump_mlir:
-    mlir_text = _export_stablehlo_mlir(model, (tokens, input_pos, kv))
-    with open('/tmp/toy_model_with_external_kv.stablehlo.mlir', 'w') as f:
-      f.write(mlir_text)
-
-  # Convert model to tflite with 2 signatures (prefill + decode).
-  # TODO(b/344014416): currently conversion will fail, because we generate int64 index
-  # in dynamic update slice op.
-  print('converting toy model to tflite with 2 signatures (prefill + decode)')
-  edge_model = (
-      ai_edge_torch.signature(
-          'prefill',
-          model,
-          sample_kwargs={
-              'tokens': tokens,
-              'input_pos': input_pos,
-              'kv_cache': kv,
-          },
-      )
-      .signature(
-          'decode',
-          model,
-          sample_kwargs={
-              'tokens': decode_token,
-              'input_pos': decode_input_pos,
-              'kv_cache': kv,
-          },
-      )
-      .convert()
-  )
-  edge_model.export('/tmp/toy_external_kv_cache.tflite')
-
-
-if __name__ == '__main__':
-  define_and_run()