ai-edge-torch-nightly 0.3.0.dev20240909__py3-none-any.whl → 0.3.0.dev20240913__py3-none-any.whl

ai_edge_torch/generative/examples/smallm/smallm.py

@@ -0,0 +1,122 @@
+# 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 SmalLM model."""
+
+import copy
+import os
+import pathlib
+
+from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
+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.utilities.loader as loading_utils
+import numpy as np
+import torch
+from torch import nn
+
+TENSOR_NAMES = copy.copy(tiny_llama.TENSOR_NAMES)
+# SmalLM re-uses the embedding as the head projection layer.
+TENSOR_NAMES.lm_head = None
+
+
+class SmalLM(tiny_llama.TinyLlama):
+  """A SmalLM model built from the Edge Generative API layers.
+
+  SmalLM shares the same architecture as TinyLlama, but with different model
+  sizes.
+  """
+
+  def __init__(self, config: cfg.ModelConfig):
+    super().__init__(config)
+    # SmalLM re-uses the embedding as the head projection layer.
+    self.lm_head.weight.data = self.tok_embedding.weight.data
+
+
+def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a SmalLM 135M model.
+
+  Args:
+    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
+      is 1024.
+
+  Returns:
+    The model config for a SmalLM model.
+  """
+  attn_config = cfg.AttentionConfig(
+      num_heads=9,
+      head_dim=64,
+      num_query_groups=3,
+      rotary_percentage=1.0,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.GATED,
+      activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
+      intermediate_size=1536,
+  )
+  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)
+  block_config = cfg.TransformerBlockConfig(
+      attn_config=attn_config,
+      ff_config=ff_config,
+      pre_attention_norm_config=norm_config,
+      post_attention_norm_config=norm_config,
+  )
+  config = cfg.ModelConfig(
+      vocab_size=49152,
+      num_layers=30,
+      max_seq_len=2048,
+      embedding_dim=576,
+      kv_cache_max_len=kv_cache_max_len,
+      block_configs=block_config,
+      final_norm_config=norm_config,
+      enable_hlfb=True,
+  )
+  return config
+
+
+def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
+  config = get_model_config(**kwargs)
+  model = SmalLM(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 define_and_run(checkpoint_path: str) -> None:
+  """Instantiates and runs a SmalLM model."""
+
+  current_dir = pathlib.Path(__file__).parent.resolve()
+  smallm_goldens = torch.load(current_dir / "smallm_lm_logits.pt")
+  kv_cache_max_len = 1024
+  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)
+  kv = kv_utils.KVCache.from_model_config(model.config)
+  output = model.forward(tokens, input_pos, kv)
+  assert torch.allclose(
+      smallm_goldens, output["logits"][0, idx.shape[1] - 1, :], atol=1e-05
+  )
+
+
+if __name__ == "__main__":
+  input_checkpoint_path = os.path.join(
+      pathlib.Path.home(), "Downloads/llm_data/smallm"
+  )
+  define_and_run(input_checkpoint_path)

ai_edge_torch/generative/examples/stable_diffusion/clip.py

@@ -61,8 +61,10 @@ class CLIP(nn.Module):
     )
 
     self.config = config
+    # CLIP has only one block config.
+    block_config = config.block_config(0)
     self.transformer_blocks = nn.ModuleList(
-        TransformerBlock(config) for _ in range(config.num_layers)
+        TransformerBlock(block_config, config) for _ in range(config.num_layers)
     )
     self.final_norm = builder.build_norm(
         config.embedding_dim, config.final_norm_config
@@ -112,15 +114,19 @@ def get_model_config() -> cfg.ModelConfig:
 
   norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.LAYER_NORM)
 
+  block_config = cfg.TransformerBlockConfig(
+      attn_config=attn_config,
+      ff_config=ff_config,
+      pre_attention_norm_config=norm_config,
+      post_attention_norm_config=norm_config,
+  )
+
   config = cfg.ModelConfig(
       vocab_size=vocab_size,
       num_layers=num_layers,
       max_seq_len=max_seq_len,
       embedding_dim=embedding_dim,
-      attn_config=attn_config,
-      ff_config=ff_config,
-      pre_attention_norm_config=norm_config,
-      post_attention_norm_config=norm_config,
+      block_configs=block_config,
       final_norm_config=norm_config,
       enable_hlfb=True,
   )

ai_edge_torch/generative/examples/t5/t5.py

@@ -52,9 +52,15 @@ class T5Stack(nn.Module):
     self.config = config
     self.embed_tokens = embed_tokens
     self.is_decoder = config.is_decoder
+    # T5 has only one block config.
+    block_config = config.block_config(0)
     self.transformer_blocks = nn.ModuleList([
-        EncoderDecoderBlock(config, has_relative_attention_bias=bool(i == 0))
-        for i in range(config.num_layers)
+        EncoderDecoderBlock(
+            block_config,
+            config,
+            has_relative_attention_bias=bool(idx == 0),
+        )
+        for idx in range(config.num_layers)
     ])
     self.final_norm = builder.build_norm(
         config.embedding_dim, config.final_norm_config
@@ -73,13 +79,11 @@ class T5Stack(nn.Module):
           torch.Tensor
       ] = None,  # should be for decoder case
   ):
-    input_shape = input_ids.size()
     inputs_embeds = self.embed_tokens(input_ids)
-    batch_size, seq_length = input_shape
     hidden_states = inputs_embeds
     position_bias = None
     encoder_decoder_position_bias = None
-    for i, layer_module in enumerate(self.transformer_blocks):
+    for _, layer_module in enumerate(self.transformer_blocks):
       # EncoderDecoderBlock.forward
       hidden_states, position_bias, encoder_decoder_position_bias = (
           layer_module(
@@ -111,7 +115,8 @@ class T5(nn.Module):
 
     encoder_config = copy.deepcopy(config)
     encoder_config.is_decoder = False
-    encoder_config.attn_config.enable_kv_cache = False
+    # T5 has only one block config.
+    encoder_config.block_config(0).attn_config.enable_kv_cache = False
     self.encoder = T5Stack(encoder_config, self.tok_embedding)
 
     decoder_config = copy.deepcopy(config)
@@ -137,20 +142,22 @@ class T5(nn.Module):
         device=torch.device("cpu"),
     )
 
+    # T5 has only one block config.
+    attn_config = config.block_config(0).attn_config
     self.enc_rel_pos_mask = attn_utils.build_relative_position_buckets(
         bidirectional=True,
         query_length=config.kv_cache_max,
         key_length=config.kv_cache_max,
-        num_buckets=config.attn_config.relative_attention_num_buckets,
-        max_distance=config.attn_config.relative_attention_max_distance,
+        num_buckets=attn_config.relative_attention_num_buckets,
+        max_distance=attn_config.relative_attention_max_distance,
     )
 
     self.dec_rel_pos_mask = attn_utils.build_relative_position_buckets(
         bidirectional=False,
         query_length=config.kv_cache_max,
         key_length=config.kv_cache_max,
-        num_buckets=config.attn_config.relative_attention_num_buckets,
-        max_distance=config.attn_config.relative_attention_max_distance,
+        num_buckets=attn_config.relative_attention_num_buckets,
+        max_distance=attn_config.relative_attention_max_distance,
     )
 
   @torch.inference_mode
@@ -230,7 +237,8 @@ class T5Encoder(nn.Module):
 
     encoder_config = copy.deepcopy(config)
     encoder_config.is_decoder = False
-    encoder_config.attn_config.enable_kv_cache = False
+    # T5 has only one block config.
+    encoder_config.block_config(0).attn_config.enable_kv_cache = False
     self.encoder = T5Stack(encoder_config, self.tok_embedding)
 
     self.enc_attn_mask_cache = (
@@ -243,12 +251,14 @@ class T5Encoder(nn.Module):
         .unsqueeze(0)
     )
 
+    # T5 has only one block config.
+    attn_config = config.block_config(0).attn_config
     self.enc_rel_pos_mask = attn_utils.build_relative_position_buckets(
         bidirectional=True,
         query_length=config.kv_cache_max,
         key_length=config.kv_cache_max,
-        num_buckets=config.attn_config.relative_attention_num_buckets,
-        max_distance=config.attn_config.relative_attention_max_distance,
+        num_buckets=attn_config.relative_attention_num_buckets,
+        max_distance=attn_config.relative_attention_max_distance,
     )
 
   @torch.inference_mode
@@ -313,12 +323,14 @@ class T5Decoder(nn.Module):
         .unsqueeze(0)
     )
 
+    # T5 has only one block config.
+    attn_config = config.block_config(0).attn_config
     self.enc_rel_pos_mask = attn_utils.build_relative_position_buckets(
         bidirectional=True,
         query_length=config.kv_cache_max,
         key_length=config.kv_cache_max,
-        num_buckets=config.attn_config.relative_attention_num_buckets,
-        max_distance=config.attn_config.relative_attention_max_distance,
+        num_buckets=attn_config.relative_attention_num_buckets,
+        max_distance=attn_config.relative_attention_max_distance,
     )
 
     self.dec_attn_mask_cache = attn_utils.build_causal_mask_cache(
@@ -386,19 +398,20 @@ def get_model_config_t5() -> cfg.ModelConfig:
       type=cfg.NormalizationType.RMS_NORM,
       epsilon=1e-6,
   )
-
-  config = cfg.ModelConfig(
-      vocab_size=32128,
-      num_layers=12,
-      max_seq_len=512,
-      embedding_dim=768,
+  block_config = cfg.TransformerBlockConfig(
       attn_config=attn_config,
       relative_attention=True,
       ff_config=ff_config,
       pre_attention_norm_config=norm_config,
       post_attention_norm_config=norm_config,
+  )
+  config = cfg.ModelConfig(
+      vocab_size=32128,
+      num_layers=12,
+      max_seq_len=512,
+      embedding_dim=768,
+      block_configs=block_config,
       final_norm_config=norm_config,
-      parallel_residual=False,
       lm_head_use_bias=False,
       enable_hlfb=True,
   )

ai_edge_torch/generative/examples/t5/t5_attention.py

@@ -24,7 +24,6 @@ from ai_edge_torch.generative.layers.scaled_dot_product_attention import scaled_
 from ai_edge_torch.generative.layers.scaled_dot_product_attention import scaled_dot_product_attention_with_hlfb  # NOQA
 import torch
 from torch import nn
-import torch.nn.functional as F
 
 BATCH_SIZE = 1
 
@@ -32,13 +31,18 @@ BATCH_SIZE = 1
 class EncoderDecoderBlock(nn.Module):
 
   def __init__(
-      self, config: cfg.ModelConfig, has_relative_attention_bias: bool = False
+      self,
+      config: cfg.TransformerBlockConfig,
+      model_config: cfg.ModelConfig,
+      has_relative_attention_bias: bool = False,
   ) -> None:
     """Initialize an instance of the EncoderDecoderBlock.
 
     Args:
-      config (cfg.ModelConfig): the configuration object for this transformer
-        block.
+      config (cfg.TransformerBlockConfig): the configuration object for this
+        transformer block.
+      model_config (cfg.ModelConfig): the configuration object for the model
+        this transformer block belongs to.
       has_relative_attention_bias (bool): whether the self attention block has
         relative bias.
     """
@@ -46,22 +50,22 @@ class EncoderDecoderBlock(nn.Module):
     super().__init__()
     self.atten_func = T5Attention(
         BATCH_SIZE,
-        config.embedding_dim,
+        model_config.embedding_dim,
         config.attn_config,
         config.pre_attention_norm_config,
-        config.kv_cache_max,
-        config.enable_hlfb,
+        model_config.kv_cache_max,
+        model_config.enable_hlfb,
         has_relative_attention_bias=has_relative_attention_bias,
     )
     # For a decoder, we add a cross attention.
-    if config.is_decoder:
+    if model_config.is_decoder:
       self.cross_atten_func = T5Attention(
           BATCH_SIZE,
-          config.embedding_dim,
+          model_config.embedding_dim,
           config.attn_config,
           config.pre_attention_norm_config,
-          config.kv_cache_max,
-          config.enable_hlfb,
+          model_config.kv_cache_max,
+          model_config.enable_hlfb,
           # Cross Attention does not have relative attention bias.
           has_relative_attention_bias=False,
       )
@@ -69,9 +73,10 @@ class EncoderDecoderBlock(nn.Module):
       self.cross_atten_func = None
 
     self.post_atten_norm = builder.build_norm(
-        config.embedding_dim, config.post_attention_norm_config
+        model_config.embedding_dim,
+        config.post_attention_norm_config,
     )
-    self.ff = builder.build_ff(config.embedding_dim, config.ff_config)
+    self.ff = builder.build_ff(model_config.embedding_dim, config.ff_config)
     self.config = config
 
   def forward(

ai_edge_torch/generative/examples/test_models/toy_model.py

@@ -20,7 +20,6 @@ from ai_edge_torch.generative.layers.attention import TransformerBlock
 import ai_edge_torch.generative.layers.attention_utils as attn_utils
 import ai_edge_torch.generative.layers.builder as builder
 import ai_edge_torch.generative.layers.model_config as cfg
-import numpy as np
 import torch
 import torch.nn as nn
 
@@ -36,16 +35,16 @@ class ToySingleLayerModel(torch.nn.Module):
         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_block = TransformerBlock(config)
+    self.transformer_block = TransformerBlock(config.block_config(0), config)
     self.final_norm = builder.build_norm(
         config.embedding_dim,
         config.final_norm_config,
     )
+    # Toy model has only one block config.
+    attn_config = config.block_config(0).attn_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
-        ),
+        dim=int(attn_config.rotary_percentage * attn_config.head_dim),
         base=10_000,
         condense_ratio=1,
         dtype=torch.float32,
@@ -85,16 +84,16 @@ class ToySingleLayerModelWeightSharing(torch.nn.Module):
         bias=config.lm_head_use_bias,
     )
     self.lm_head.weight.data = self.tok_embedding.weight.data
-    self.transformer_block = TransformerBlock(config)
+    self.transformer_block = TransformerBlock(config.block_config(0), config)
     self.final_norm = builder.build_norm(
         config.embedding_dim,
         config.final_norm_config,
     )
+    # Toy model has only one block config.
+    attn_config = config.block_config(0).attn_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
-        ),
+        dim=int(attn_config.rotary_percentage * attn_config.head_dim),
         base=10_000,
         condense_ratio=1,
         dtype=torch.float32,
@@ -135,15 +134,18 @@ def get_model_config() -> cfg.ModelConfig:
       intermediate_size=256,
   )
   norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)
+  block_config = cfg.TransformerBlockConfig(
+      attn_config=attn_config,
+      ff_config=ff_config,
+      pre_attention_norm_config=norm_config,
+      post_attention_norm_config=norm_config,
+  )
   config = cfg.ModelConfig(
       vocab_size=400,
       num_layers=1,
       max_seq_len=KV_CACHE_MAX_LEN,
       embedding_dim=128,
-      attn_config=attn_config,
-      ff_config=ff_config,
-      pre_attention_norm_config=norm_config,
-      post_attention_norm_config=norm_config,
+      block_configs=block_config,
       final_norm_config=norm_config,
   )
   return config

ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py

@@ -12,14 +12,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-# A toy example which has basic transformer block (w/ KV-Cache).
+
+"""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
-from ai_edge_torch.generative.layers.attention import TransformerBlock
+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.builder as builder
 import ai_edge_torch.generative.layers.model_config as cfg
 import torch
 import torch.nn as nn
@@ -27,7 +30,7 @@ import torch.nn as nn
 RoPECache = Tuple[torch.Tensor, torch.Tensor]
 
 
-class ToyModelWithKV(torch.nn.Module):
+class ToyModelWithKVCache(torch.nn.Module):
 
   def __init__(self, config: cfg.ModelConfig) -> None:
     super().__init__()
@@ -35,18 +38,20 @@ class ToyModelWithKV(torch.nn.Module):
         config.embedding_dim, config.vocab_size, bias=config.lm_head_use_bias
     )
     self.tok_embedding = nn.Embedding(config.vocab_size, config.embedding_dim)
+    # Toy model has only one block config.
+    block_config = config.block_config(0)
     self.transformer_blocks = nn.ModuleList(
-        TransformerBlock(config) for _ in range(config.num_layers)
+        attention.TransformerBlock(block_config, config)
+        for _ in range(config.num_layers)
     )
     self.final_norm = builder.build_norm(
         config.embedding_dim,
         config.final_norm_config,
     )
+    attn_config = block_config.attn_config
     self.rope_cache = attn_utils.build_rope_cache(
         size=config.max_seq_len,
-        dim=int(
-            config.attn_config.rotary_percentage * config.attn_config.head_dim
-        ),
+        dim=int(attn_config.rotary_percentage * attn_config.head_dim),
         base=10_000,
         condense_ratio=1,
         dtype=torch.float32,
@@ -57,18 +62,29 @@ class ToyModelWithKV(torch.nn.Module):
     )
     self.config = config
 
-  @torch.inference_mode
-  def forward(self, idx: torch.Tensor, input_pos: torch.Tensor) -> torch.Tensor:
-    x = self.tok_embedding(idx)
+  def forward(
+      self,
+      tokens: torch.Tensor,
+      input_pos: torch.Tensor,
+      kv_cache: kv_utils.KVCache,
+  ) -> Tuple[torch.Tensor, kv_utils.KVCache]:
+    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):
-      x = block(x, (cos, sin), mask, 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)
+      if kv_entry:
+        updated_kv_entires.append(kv_entry)
+
     x = self.final_norm(x)
-    return self.lm_head(x)
+    updated_kv_cache = kv_utils.KVCache(tuple(updated_kv_entires))
+    return {'logits': self.lm_head(x), 'kv_cache': updated_kv_cache}
 
 
 def _export_stablehlo_mlir(model, args):
@@ -78,7 +94,10 @@ def _export_stablehlo_mlir(model, args):
 
 def get_model_config() -> cfg.ModelConfig:
   attn_config = cfg.AttentionConfig(
-      num_heads=32, head_dim=4, num_query_groups=4, rotary_percentage=1.0
+      num_heads=32,
+      head_dim=4,
+      num_query_groups=4,
+      rotary_percentage=1.0,
   )
   ff_config = cfg.FeedForwardConfig(
       type=cfg.FeedForwardType.GATED,
@@ -86,15 +105,18 @@ def get_model_config() -> cfg.ModelConfig:
       intermediate_size=256,
   )
   norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)
-  config = cfg.ModelConfig(
-      vocab_size=150,
-      num_layers=2,
-      max_seq_len=500,
-      embedding_dim=128,
+  block_config = cfg.TransformerBlockConfig(
       attn_config=attn_config,
       ff_config=ff_config,
       pre_attention_norm_config=norm_config,
       post_attention_norm_config=norm_config,
+  )
+  config = cfg.ModelConfig(
+      vocab_size=150,
+      num_layers=2,
+      max_seq_len=100,
+      embedding_dim=128,
+      block_configs=block_config,
       final_norm_config=norm_config,
       enable_hlfb=True,
   )
@@ -102,40 +124,59 @@ 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)
-  input_pos = torch.tensor([10], dtype=torch.int64)
-  return idx, input_pos
+  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 = ToyModelWithKV(config)
+  model = ToyModelWithExternalKV(config)
+  model.eval()
   print('running an inference')
-  idx, input_pos = get_sample_prefill_inputs()
-  decode_idx, decode_input_pos = get_sample_decode_inputs()
-  print(model.forward(idx, input_pos))
+  kv = kv_utils.KVCache.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, (idx, input_pos))
-    with open('/tmp/toy_model_with_kv.stablehlo.mlir', 'w') as f:
+    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).
   print('converting toy model to tflite with 2 signatures (prefill + decode)')
   edge_model = (
-      ai_edge_torch.signature('prefill', model, (idx, input_pos))
-      .signature('decode', model, (decode_idx, decode_input_pos))
+      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_kv_cache.tflite')
+  edge_model.export('/tmp/toy_external_kv_cache.tflite')
 
 
 if __name__ == '__main__':