ai-edge-torch-nightly 0.2.0.dev20240719__py3-none-any.whl → 0.2.0.dev20240721__py3-none-any.whl

ai_edge_torch/generative/examples/experimental/__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/gemma/__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/gemma/convert_to_tflite.py

@@ -0,0 +1,87 @@
+# 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 Gemma 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.gemma import gemma
+from ai_edge_torch.generative.layers.experimental import ekv_cache as kv_utils
+from ai_edge_torch.generative.quantize import quant_recipes
+
+
+def convert_gemma_to_tflite(
+    checkpoint_path: str,
+    prefill_seq_len: int = 512,
+    kv_cache_max_len: int = 1024,
+    quantize: bool = True,
+):
+  """An example method for converting a Gemma 2B 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 = gemma.build_2b_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/gemma_seq{prefill_seq_len}_ekv{kv_cache_max_len}.tflite')
+
+
+if __name__ == '__main__':
+  checkpoint_path = os.path.join(Path.home(), 'Downloads/llm_data/gemma-2b')
+  convert_gemma_to_tflite(checkpoint_path)

ai_edge_torch/generative/examples/experimental/gemma/gemma.py

@@ -0,0 +1,195 @@
+# 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 Gemma model.
+#
+# Note: This is an experimental version of Gemma 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=None,
+)
+
+
+class Gemma(nn.Module):
+
+  def __init__(self, config: cfg.ModelConfig):
+    super().__init__()
+
+    self.config = config
+    # Construct model layers.
+    self.tok_embedding = nn.Embedding(
+        config.vocab_size, config.embedding_dim, padding_idx=0
+    )
+    self.lm_head = nn.Linear(
+        config.embedding_dim,
+        config.vocab_size,
+        bias=config.lm_head_use_bias,
+    )
+    # Gemma re-uses the embedding as the head projection layer.
+    self.lm_head.weight.data = self.tok_embedding.weight.data
+    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)
+    x = x * (self.config.embedding_dim**0.5)
+
+    updated_kv_entires = []
+    for i, block in enumerate(self.transformer_blocks):
+      kv_entry = kv_cache.caches[i] if kv_cache else None
+      x, kv_entry = block(x, (cos, sin), mask, input_pos, kv_entry)
+      if kv_entry:
+        updated_kv_entires.append(kv_entry)
+    updated_kv_cache = kv_utils.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_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  attn_config = cfg.AttentionConfig(
+      num_heads=8,
+      num_query_groups=1,
+      rotary_percentage=1.0,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.GATED,
+      activation=cfg.ActivationConfig(cfg.ActivationType.GELU_TANH),
+      intermediate_size=16384,
+  )
+  norm_config = cfg.NormalizationConfig(
+      type=cfg.NormalizationType.RMS_NORM,
+      epsilon=1e-6,
+      zero_centered=True,
+  )
+  config = cfg.ModelConfig(
+      vocab_size=256000,
+      num_layers=18,
+      max_seq_len=8192,
+      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,
+      parallel_residual=False,
+      lm_head_use_bias=False,
+      enable_hlfb=True,
+  )
+  return config
+
+
+def get_fake_model_config_2b_for_test(**kwargs) -> cfg.ModelConfig:
+  config = get_model_config_2b(**kwargs)
+  config.num_layers = 2
+  return config
+
+
+def build_2b_model(checkpoint_path, test_model=False, **kwargs) -> nn.Module:
+  config = (
+      get_fake_model_config_2b_for_test(**kwargs)
+      if test_model
+      else get_model_config_2b(**kwargs)
+  )
+  model = Gemma(config)
+  if checkpoint_path is not None:
+    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_2b(checkpoint_path, test_model=False) -> None:
+  kv_cache_max_len = 1024
+  model = build_2b_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/gemma-2b")
+  define_and_run_2b(checkpoint_path)

ai_edge_torch/generative/examples/experimental/phi/__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/phi/convert_to_tflite.py

@@ -0,0 +1,84 @@
+# 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 phi2 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.phi import phi2
+from ai_edge_torch.generative.layers.experimental import ekv_cache
+from ai_edge_torch.generative.quantize import quant_recipes
+
+
+def convert_phi2_to_tflite(
+    checkpoint_path: str,
+    prefill_seq_len: int = 512,
+    kv_cache_max_len: int = 1024,
+    quantize: bool = True,
+):
+  """An example method for converting 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)
+  kv = ekv_cache.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/phi2_seq{prefill_seq_len}_ekv{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/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()

ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py

@@ -149,6 +149,8 @@ def build_model(checkpoint_path, **kwargs) -> nn.Module:
 
 
 def define_and_run() -> None:
+  current_dir = Path(__file__).parent.resolve()
+  tiny_llama_goldens = torch.load(current_dir / "tiny_llama_lm_logits.pt")
   kv_cache_max_len = 1024
   checkpoint_path = os.path.join(Path.home(), "Downloads/llm_data/tiny_llama")
   model = build_model(checkpoint_path, kv_cache_max_len=kv_cache_max_len)
@@ -156,8 +158,10 @@ 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)
+  assert torch.allclose(
+      tiny_llama_goldens, lm_logits[0, idx.shape[1] - 1, :], atol=1e-05
+  )
 
 
 if __name__ == "__main__":

ai_edge_torch/generative/test/test_experimental_ekv.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.
+# ==============================================================================
+# A suite of tests to validate experimental external KV Cache layers and models.
+
+import unittest
+
+import numpy as np
+import torch
+
+from ai_edge_torch.generative.examples.experimental.gemma import gemma
+from ai_edge_torch.generative.examples.experimental.phi import phi2
+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
+import ai_edge_torch.generative.layers.model_config as cfg
+
+
+class TestExternalKVLayers(unittest.TestCase):
+
+  def _get_test_config(self, num_layers, head_dim, num_query_groups, kv_cache_max_len):
+    attn_config = cfg.AttentionConfig(num_heads=1, num_query_groups=num_query_groups)
+    config = cfg.ModelConfig(
+        kv_cache_max_len=kv_cache_max_len,
+        embedding_dim=head_dim,
+        attn_config=attn_config,
+        num_layers=num_layers,
+        max_seq_len=None,
+        vocab_size=None,
+        ff_config=None,
+    )
+    return config
+
+  def test_cache_udpate(self):
+    N = 1
+    HEAD_DIM = 2
+    NUM_QG = 1
+    KV_LEN = 4
+    config = self._get_test_config(
+        num_layers=N,
+        head_dim=HEAD_DIM,
+        num_query_groups=NUM_QG,
+        kv_cache_max_len=KV_LEN,
+    )
+    kv = kv_utils.EKVCache.from_model_config(config)
+    entry = kv.caches[0]
+    # single-slice update
+    input_pos = torch.tensor([1])
+    k_slice = v_slice = torch.full((1, 1, NUM_QG, HEAD_DIM), 5, dtype=torch.float)
+    updated_entry = kv_utils.update(entry, input_pos, k_slice, v_slice)
+    self.assertEqual(
+        updated_entry.k_cache.numpy().flatten().tolist(), [0, 0, 5, 5, 0, 0, 0, 0]
+    )
+    self.assertEqual(
+        updated_entry.v_cache.numpy().flatten().tolist(), [0, 0, 5, 5, 0, 0, 0, 0]
+    )
+    # multi-slice update
+    input_pos = torch.tensor([0, 3])
+    k_slice = v_slice = torch.full((1, 2, NUM_QG, HEAD_DIM), 7, dtype=torch.float)
+    updated_entry = kv_utils.update(entry, input_pos, k_slice, v_slice)
+    self.assertEqual(
+        updated_entry.k_cache.numpy().flatten().tolist(), [7, 7, 0, 0, 0, 0, 7, 7]
+    )
+    self.assertEqual(
+        updated_entry.v_cache.numpy().flatten().tolist(), [7, 7, 0, 0, 0, 0, 7, 7]
+    )
+
+  def test_serialization(self):
+    class TestModel(torch.nn.Module):
+
+      def forward(self, kv: kv_utils.EKVCache) -> kv_utils.EKVCache:
+        updated_kv_entries = [
+            kv_utils.KVCacheEntry(
+                torch.zeros_like(entry.k_cache), torch.zeros_like(entry.v_cache)
+            )
+            for entry in kv.caches
+        ]
+        return kv_utils.EKVCache(updated_kv_entries)
+
+    N = 1
+    HEAD_DIM = 2
+    NUM_QG = 1
+    KV_LEN = 4
+    config = self._get_test_config(
+        num_layers=N,
+        head_dim=HEAD_DIM,
+        num_query_groups=NUM_QG,
+        kv_cache_max_len=KV_LEN,
+    )
+    kv = kv_utils.EKVCache.from_model_config(config)
+    model = TestModel()
+    exported_program = torch.export.export(model, (kv,))
+    input_specs = exported_program.graph_signature.input_specs
+    self.assertEqual(len(input_specs), 2)
+    self.assertEqual(input_specs[0].arg.name, "kv_k_0")
+    self.assertEqual(input_specs[1].arg.name, "kv_v_0")
+
+
+class TestExternalKVModels(unittest.TestCase):
+
+  def test_can_build_gemma(self):
+    gemma.define_and_run_2b(checkpoint_path=None, test_model=True)
+
+  def test_can_build_phi2(self):
+    phi2.define_and_run(checkpoint_path=None, test_model=True)
+
+  def test_can_build_tinyllama(self):
+    tiny_llama.define_and_run(checkpoint_path=None, test_model=True)
+
+
+if __name__ == "__main__":
+  unittest.main()

{ai_edge_torch_nightly-0.2.0.dev20240719.dist-info → ai_edge_torch_nightly-0.2.0.dev20240721.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.2.0.dev20240719
+Version: 0.2.0.dev20240721
 Summary: Supporting PyTorch models with the Google AI Edge TFLite runtime.
 Home-page: https://github.com/google-ai-edge/ai-edge-torch
 Keywords: On-Device ML,AI,Google,TFLite,PyTorch,LLMs,GenAI

{ai_edge_torch_nightly-0.2.0.dev20240719.dist-info → ai_edge_torch_nightly-0.2.0.dev20240721.dist-info}/RECORD

@@ -35,12 +35,22 @@ ai_edge_torch/debug/test/test_search_model.py,sha256=0guAEon5cvwBpPXk6J0wVOKj7TX
 ai_edge_torch/experimental/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/experimental/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/experimental/gemma/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/experimental/gemma/convert_to_tflite.py,sha256=bW_KOj_3fcZAggfST3zHWcMcNJs70b0pld-vvauAOgo,3076
+ai_edge_torch/generative/examples/experimental/gemma/gemma.py,sha256=u4DNsZRnN7whDoK8yQet9Yahb01ToVqTuFQmWV1__1g,6606
+ai_edge_torch/generative/examples/experimental/phi/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/experimental/phi/convert_to_tflite.py,sha256=sLU35tpQ-PEbhZbLfC1vSqM-HamKREVBpIoywWh9O3M,3036
+ai_edge_torch/generative/examples/experimental/phi/phi2.py,sha256=zgxB2JSFAevyS28C6-wIBaQeeKTUejUJY4dnR4BqRBI,6150
+ai_edge_torch/generative/examples/experimental/tiny_llama/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/experimental/tiny_llama/convert_to_tflite.py,sha256=PChEhBotZ8k6GZiq9e_AYnn3RyhNIVm_U96QhVjx3jY,3126
+ai_edge_torch/generative/examples/experimental/tiny_llama/tiny_llama.py,sha256=1vL0u6Pkd8SV8uei9BGzSAIokclT_RaE3K0IczoPfeI,6291
 ai_edge_torch/generative/examples/gemma/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/gemma/convert_to_tflite.py,sha256=UMEZGDGhFvAX4eT5KHAE1Xbxw-qtQWEMxgvB8cSH6wY,2531
-ai_edge_torch/generative/examples/gemma/gemma.py,sha256=1lZfXGHmbII4rFu0U2B9NzlJCRhphxtmQtkCHQ39_uw,5935
+ai_edge_torch/generative/examples/gemma/gemma.py,sha256=YyGGsgEByIg_tIysMBqaBztf_csthZIjah8mmH5o7UA,6144
 ai_edge_torch/generative/examples/phi2/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/phi2/convert_to_tflite.py,sha256=uF1A2EX8xYie30-T2Z7s1WZCtFhp5CEwRV8SCd7Umrc,2505
-ai_edge_torch/generative/examples/phi2/phi2.py,sha256=PMhKC6JCAMYSj2F3UmWHWK4rTcXD-B6PuehaoDccRqk,5562
+ai_edge_torch/generative/examples/phi2/phi2.py,sha256=KjfTrD2OBzOfq83-XvJ6ZhmXLuP_VqugSOwyj-M5YY4,5767
 ai_edge_torch/generative/examples/stable_diffusion/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/stable_diffusion/attention.py,sha256=Lo4Dq7a3Kg-lyH56iqGtqCo5UaClQHRCTDdNagXGTo8,3535
 ai_edge_torch/generative/examples/stable_diffusion/clip.py,sha256=P-cUUQaQKGKV2p-7hvLJ--RpCIA7gk8WCDRgg0pNtd0,4331
@@ -58,15 +68,15 @@ ai_edge_torch/generative/examples/stable_diffusion/samplers/k_lms.py,sha256=iPYX
 ai_edge_torch/generative/examples/stable_diffusion/samplers/sampler.py,sha256=5iRfU5MO6GR6K3WrdddIU_9U7ZZGEEb7zGKVY1WFl-8,1340
 ai_edge_torch/generative/examples/t5/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/t5/convert_to_tflite.py,sha256=7RwaZQaKhFt3zKAUbFjq95CSYhL1nd9BVSbSRNJp4-4,4529
-ai_edge_torch/generative/examples/t5/t5.py,sha256=L6YrVzUEzP-Imb8W28LdukFGrx1aWSzz1kyYK_9RFZM,21087
+ai_edge_torch/generative/examples/t5/t5.py,sha256=fVtJ0S8v2bMtvEuDqD6Orw7CTyXqnRIqZfKcz7DBeJc,21212
 ai_edge_torch/generative/examples/t5/t5_attention.py,sha256=KaGzCAViNOpJIQbRF-ItouuVPqI9nroWRRGN-KFYKZs,8357
 ai_edge_torch/generative/examples/test_models/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/test_models/toy_model.py,sha256=Sf3ZMYv-iuMRKAKLow47qth8vTF1zl6i8TxJ9uT_StU,3885
-ai_edge_torch/generative/examples/test_models/toy_model_with_external_kv_cache.py,sha256=zwCmCnhr-vhBwHqv9i7xMasdBGVNqAGxZvWsncsJn58,5543
+ai_edge_torch/generative/examples/test_models/toy_model_with_external_kv_cache.py,sha256=jmucKpWY_nHEOAh7G62IxpReNmrKWo4PxfELul_h9xQ,5796
 ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py,sha256=lfYUiem_Pbn3vGgPx84BeI8n7rN3-1fImwCLm8Eo2U8,4853
 ai_edge_torch/generative/examples/tiny_llama/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 ai_edge_torch/generative/examples/tiny_llama/convert_to_tflite.py,sha256=nT7Fh-f5ZdwaK3dPoCvZflpJ4fRHjLdFMjk1_uw3-b8,2559
-ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py,sha256=IFRLPG9wz_aLl_zV_6CETCjSM03ukA6bZqqyDLVACuw,5651
+ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py,sha256=to9IlF-X_uIJvO-roZOW1ZMUhmkYbvFjc-tUVaQr6TE,5848
 ai_edge_torch/generative/fx_passes/__init__.py,sha256=aXvYiaHDvETIrh0Q9DDZA_ZBiazGk80DT6nt7lLtC1o,1172
 ai_edge_torch/generative/fx_passes/remove_sdpa_zero_mask_pass.py,sha256=BCAcc_OcEjvbaXQSbc8vlKeMad7E3gCA4BNsUdWRwBI,1966
 ai_edge_torch/generative/layers/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
@@ -94,6 +104,7 @@ ai_edge_torch/generative/quantize/ai_edge_quantizer_glue/__init__.py,sha256=47DE
 ai_edge_torch/generative/quantize/ai_edge_quantizer_glue/translate_recipe.py,sha256=iTNPrlubmq9ia7C3zHl50J2YEMsc4o33GwL5tr5VkkE,5229
 ai_edge_torch/generative/test/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/test/loader_test.py,sha256=N88CbrLW7Q2x1EyurwdXQ6YjsA-ySQcPxpZH3QOGp-M,3317
+ai_edge_torch/generative/test/test_experimental_ekv.py,sha256=qMR0r7Pr_t2bn-cyeA7Qw_Rl94H1NmFcqM2ua8gpDDw,4230
 ai_edge_torch/generative/test/test_model_conversion.py,sha256=LsPTrLC1I4JW2GowTS3V9Eu257vLHr2Yj5f_qaFUX84,7589
 ai_edge_torch/generative/test/test_quantize.py,sha256=QbF7LC9olJFGXqlAVGciac7xXc4rDtCSr71tTIYuqPk,5230
 ai_edge_torch/generative/utilities/__init__.py,sha256=-_jxnnFnCgnTU4oTm4MnRsvL5lqhomBNdFBbqfmfHPo,720
@@ -114,8 +125,8 @@ ai_edge_torch/quantize/quant_config.py,sha256=eO9Ra160ITjQSyRBEGy6nNIVH3gYacSWDd
 ai_edge_torch/testing/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/testing/model_coverage/__init__.py,sha256=5P8J6Zk5YYtDvTBucFvB9NGSRI7Gw_24WnrbhXgycEE,765
 ai_edge_torch/testing/model_coverage/model_coverage.py,sha256=kzIulTldq8R9E-lAZsvfSTvLu3FYEX7b9DyYM3qisXM,4485
-ai_edge_torch_nightly-0.2.0.dev20240719.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
-ai_edge_torch_nightly-0.2.0.dev20240719.dist-info/METADATA,sha256=X9TaI_3Rxn0rk89P3ZcXJlNtEIUBOhOIIKAncN3Xpos,1745
-ai_edge_torch_nightly-0.2.0.dev20240719.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-ai_edge_torch_nightly-0.2.0.dev20240719.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
-ai_edge_torch_nightly-0.2.0.dev20240719.dist-info/RECORD,,
+ai_edge_torch_nightly-0.2.0.dev20240721.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.2.0.dev20240721.dist-info/METADATA,sha256=TJYFNAxXQkRwt9I_0OqpUOS3opWBU5i-ioMwsicD7cY,1745
+ai_edge_torch_nightly-0.2.0.dev20240721.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+ai_edge_torch_nightly-0.2.0.dev20240721.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.2.0.dev20240721.dist-info/RECORD,,