ai-edge-torch-nightly 0.2.0.dev20240718__py3-none-any.whl → 0.2.0.dev20240720__py3-none-any.whl

ai_edge_torch/generative/examples/experimental/phi/phi2.py

@@ -0,0 +1,184 @@
+# 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.
+#
+# Note: This is an experimental version of phi2 with external KV cache.
+# Please use with caution.
+
+
+import os
+from pathlib import Path
+from typing import Tuple
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+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 ai_edge_torch.generative.utilities.loader as loading_utils
+
+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):
+
+  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(
+        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.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]:
+    B, T = tokens.size()
+    assert (
+        self.config.max_seq_len >= T
+    ), f"Cannot forward sequence of length {T}, max seq length is only {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]
+
+    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:
+  attn_config = cfg.AttentionConfig(
+      num_heads=32,
+      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_for_test(**kwargs) -> cfg.ModelConfig:
+  config = get_model_config(**kwargs)
+  config.num_layers = 2
+  return config
+
+
+def build_model(checkpoint_path, test_model=False, **kwargs) -> nn.Module:
+  config = (
+      get_fake_model_config_for_test(**kwargs)
+      if test_model
+      else get_model_config(**kwargs)
+  )
+  model = Phi2(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, test_model=False) -> None:
+  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__":
+  checkpoint_path = os.path.join(Path.home(), "Downloads/phi2")
+  define_and_run(checkpoint_path)

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

@@ -0,0 +1,14 @@
+# 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/experimental/tiny_llama/convert_to_tflite.py

@@ -0,0 +1,89 @@
+# 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.
+# ==============================================================================
+#
+# Note: This is an experimental version of TinyLlama with external KV cache.
+# Please use with caution.
+
+
+import os
+from pathlib import Path
+
+import torch
+
+import ai_edge_torch
+from ai_edge_torch.generative.examples.experimental.tiny_llama import tiny_llama  # NOQA
+from ai_edge_torch.generative.layers.experimental import ekv_cache as kv_utils
+from ai_edge_torch.generative.quantize import quant_recipes
+
+
+def convert_tiny_llama_to_tflite(
+    checkpoint_path: str,
+    prefill_seq_len: int = 512,
+    kv_cache_max_len: int = 1024,
+    quantize: bool = True,
+):
+  """An example method for converting TinyLlama 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 = tiny_llama.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)
+  kv = kv_utils.EKVCache.from_model_config(pytorch_model.config)
+
+  quant_config = quant_recipes.full_int8_dynamic_recipe() if quantize else None
+  edge_model = (
+      ai_edge_torch.signature(
+          'prefill',
+          pytorch_model,
+          sample_kwargs={
+              'tokens': prefill_tokens,
+              'input_pos': prefill_input_pos,
+              'kv_cache': kv,
+          },
+      )
+      .signature(
+          'decode',
+          pytorch_model,
+          sample_kwargs={
+              'tokens': decode_token,
+              'input_pos': decode_input_pos,
+              'kv_cache': kv,
+          },
+      )
+      .convert(quant_config=quant_config)
+  )
+  edge_model.export(
+      f'/tmp/tiny_llama_seq{prefill_seq_len}_ekv{kv_cache_max_len}.tflite'
+  )
+
+
+if __name__ == '__main__':
+  checkpoint_path = os.path.join(Path.home(), 'Downloads/llm_data/tiny_llama')
+  convert_tiny_llama_to_tflite(checkpoint_path)

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

@@ -0,0 +1,185 @@
+# 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
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+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 ai_edge_torch.generative.utilities.loader as loading_utils
+
+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",
+    pre_ff_norm="model.layers.{}.post_attention_layernorm",
+    embedding="model.embed_tokens",
+    final_norm="model.norm",
+    lm_head="lm_head",
+)
+
+
+class TinyLLamma(nn.Module):
+
+  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(
+        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.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]:
+    B, T = tokens.size()
+    assert (
+        self.config.max_seq_len >= T
+    ), f"Cannot forward sequence of length {T}, max seq length is only {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:
+  attn_config = cfg.AttentionConfig(
+      num_heads=32,
+      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,
+      pre_ff_norm_config=norm_config,
+      final_norm_config=norm_config,
+      enable_hlfb=True,
+  )
+  return config
+
+
+def get_fake_model_config_for_test(**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, test_model=False, **kwargs) -> nn.Module:
+  config = (
+      get_fake_model_config_for_test(**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, test_model=False) -> None:
+  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__":
+  checkpoint_path = os.path.join(Path.home(), "Downloads/tiny_llama")
+  define_and_run(checkpoint_path)

ai_edge_torch/generative/examples/gemma/gemma.py

@@ -159,6 +159,9 @@ def build_2b_model(checkpoint_path, **kwargs) -> nn.Module:
 
 
 def define_and_run_2b() -> None:
+  current_dir = Path(__file__).parent.resolve()
+  gemma_goldens = torch.load(current_dir / "gemma_lm_logits.pt")
+
   kv_cache_max_len = 1024
   checkpoint_path = os.path.join(Path.home(), "Downloads/llm_data/gemma-2b")
   model = build_2b_model(checkpoint_path, kv_cache_max_len=kv_cache_max_len)
@@ -166,8 +169,9 @@ def define_and_run_2b() -> None:
   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)
-  print("running an inference")
-  print(model.forward(tokens, input_pos))
+  lm_logits = model.forward(tokens, input_pos)
+  print("comparing with goldens..")
+  assert torch.allclose(gemma_goldens, lm_logits[0, idx.shape[1] - 1, :], atol=1e-05)
 
 
 if __name__ == "__main__":

ai_edge_torch/generative/examples/phi2/phi2.py

@@ -149,6 +149,8 @@ def build_model(checkpoint_path, **kwargs) -> nn.Module:
 
 
 def define_and_run() -> None:
+  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)
@@ -156,8 +158,9 @@ def define_and_run() -> None:
   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)
-  print("running an inference")
-  print(model.forward(tokens, input_pos))
+  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__":

ai_edge_torch/generative/examples/t5/t5.py

@@ -557,7 +557,8 @@ def get_sample_encoder_input_ids() -> torch.Tensor:
 
 
 def define_and_run_t5(checkpoint_path: str) -> None:
-  t5_goldens = torch.load("t5_lm_logits.pt")
+  current_dir = Path(__file__).parent.resolve()
+  t5_goldens = torch.load(current_dir / "t5_lm_logits.pt")
 
   model = build_t5_model(checkpoint_path)
 
@@ -579,7 +580,9 @@ def define_and_run_t5(checkpoint_path: str) -> None:
 
 # TODO(haoliang): Move those tests.
 def define_and_run_t5_split(checkpoint_path: str) -> None:
-  t5_goldens = torch.load("t5_lm_logits.pt")
+  current_dir = Path(__file__).parent.resolve()
+  t5_goldens = torch.load(current_dir / "t5_lm_logits.pt")
+
   config = get_model_config_t5()
   embedding_layer = nn.Embedding(config.vocab_size, config.embedding_dim, padding_idx=0)
   t5_encoder_model = build_t5_encoder_model(config, embedding_layer, checkpoint_path)

ai_edge_torch/generative/examples/test_models/toy_model_with_external_kv_cache.py

@@ -14,9 +14,8 @@
 # ==============================================================================
 # A toy example which has basic transformer block (w/ externalized KV-Cache).
 
-from typing import List, Tuple
+from typing import Tuple
 
-import numpy as np
 import torch
 import torch.nn as nn
 import torch_xla
@@ -24,6 +23,7 @@ import torch_xla
 import ai_edge_torch
 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
 
@@ -60,27 +60,27 @@ class ToyModelWithExternalKV(torch.nn.Module):
 
   def forward(
       self,
-      idx: torch.Tensor,
+      tokens: torch.Tensor,
       input_pos: torch.Tensor,
-      k_caches: torch.Tensor,
-      v_caches: torch.Tensor,
-  ) -> (torch.Tensor, torch.Tensor, torch.Tensor):
-    x = self.tok_embedding(idx)
+      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):
-      input_k, input_v = k_caches[i], v_caches[i]
-      x, (updated_k, updated_v) = block(
-          x, (cos, sin), mask, input_pos, (input_k, input_v)
-      )
-      k_caches[i], v_caches[i] = updated_k, updated_v
+      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)
-    return self.lm_head(x), k_caches, v_caches
+    updated_kv_cache = kv_utils.EKVCache(tuple(updated_kv_entires))
+    return self.lm_head(x), updated_kv_cache
 
 
 def _export_stablehlo_mlir(model, args):
@@ -115,15 +115,15 @@ def get_model_config() -> cfg.ModelConfig:
 
 
 def get_sample_prefill_inputs() -> Tuple[torch.Tensor, torch.Tensor]:
-  idx = torch.unsqueeze(torch.arange(0, 100), 0)
+  tokens = torch.unsqueeze(torch.arange(0, 100), 0)
   input_pos = torch.arange(0, 100)
-  return idx, input_pos
+  return tokens, input_pos
 
 
 def get_sample_decode_inputs() -> Tuple[torch.Tensor, torch.Tensor]:
-  idx = torch.tensor([[1]], dtype=torch.long)
+  tokens = torch.tensor([[1]], dtype=torch.long)
   input_pos = torch.tensor([10])
-  return idx, input_pos
+  return tokens, input_pos
 
 
 def define_and_run() -> None:
@@ -131,16 +131,16 @@ def define_and_run() -> None:
 
   config = get_model_config()
   model = ToyModelWithExternalKV(config)
+  model.eval()
   print('running an inference')
-  k_caches = torch.zeros((2, 1, 100, 4, 4), dtype=torch.float32)
-  v_caches = torch.zeros((2, 1, 100, 4, 4), dtype=torch.float32)
+  kv = kv_utils.EKVCache.from_model_config(config)
 
-  idx, input_pos = get_sample_prefill_inputs()
-  decode_idx, decode_input_pos = get_sample_decode_inputs()
-  print(model.forward(idx, input_pos, k_caches, v_caches))
+  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, (idx, input_pos, k_caches, v_caches))
+    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)
 
@@ -149,13 +149,28 @@ def define_and_run() -> None:
   # in dynamic update slice op.
   print('converting toy model to tflite with 2 signatures (prefill + decode)')
   edge_model = (
-      ai_edge_torch.signature('prefill', model, (idx, input_pos, k_caches, v_caches))
-      .signature('decode', model, (decode_idx, decode_input_pos, k_caches, v_caches))
+      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__':
-  with torch.inference_mode():
-    define_and_run()
+  define_and_run()