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

ai_edge_torch/convert/conversion_utils.py

@@ -20,7 +20,7 @@ import gc
 import itertools
 import logging
 import tempfile
-from typing import Any, Dict, Optional, Tuple, Union
+from typing import Any, Dict, List, Optional, Tuple, Union
 
 import torch
 import torch.utils._pytree as pytree
@@ -79,28 +79,49 @@ class Signature:
     for i in range(args_spec.num_leaves):
       names.append(f"args_{i}")
 
-    dict_context = (
-        kwargs_spec.context
-        if kwargs_spec.type is not collections.defaultdict
-        # ignore mismatch of `default_factory` for defaultdict
-        else kwargs_spec.context[1]
+    kwargs_names = self._flat_kwarg_names(
+        kwargs_spec.children_specs, kwargs_spec.context
     )
+    names.extend(kwargs_names)
+    return names
 
-    for name, value_spec in zip(dict_context, kwargs_spec.children_specs):
-      if value_spec.num_leaves == 1:
-        names.append(name)
+  def _flat_kwarg_names(self, specs, context) -> List[str]:
+    flat_names = []
+    if context is None:
+      for i, spec in enumerate(specs):
+        if spec.children_specs:
+          flat_names.extend(
+              [
+                  f"{i}_{name}"
+                  for name in self._flat_kwarg_names(spec.children_specs, spec.context)
+              ]
+          )
+        else:
+          flat_names.append(f"{i}")
+    else:
+      flat_ctx = self._flatten_list(context)
+      for prefix, spec in zip(flat_ctx, specs):
+        leaf_flat_names = self._flat_kwarg_names(spec.children_specs, spec.context)
+        if leaf_flat_names:
+          flat_names.extend([f"{prefix}_{name}" for name in leaf_flat_names])
+        else:
+          flat_names.append(prefix)
+
+    return flat_names
+
+  def _flatten_list(self, l: List) -> List:
+    flattened = []
+    for item in l:
+      if isinstance(item, list):
+        flattened.extend(self._flatten_list(item))
       else:
-        # value_spec.num_leaves may be greater than 1 when the value is a (nested)
-        # tuple of tensors. We haven't decided how we should support flattenable
-        # tensor containers as inputs.
-        # TODO(b/352584188): Decide the behavior of tensor container as input (flatten or reject)
-        for i in range(value_spec.num_leaves):
-          names.append(f"{name}_{i}")
-    return names
+        flattened.append(item)
+    return flattened
 
   @property
-  def flat_args(self) -> tuple[torch.Tensor]:
-    return tuple(pytree.tree_flatten(self._normalized_sample_args_kwargs)[0])
+  def flat_args(self) -> tuple[Any]:
+    args, kwargs = self._normalized_sample_args_kwargs
+    return tuple([*args, *kwargs.values()])
 
 
 def exported_program_to_stablehlo_bundle(

ai_edge_torch/convert/test/test_convert.py

@@ -14,10 +14,14 @@
 # ==============================================================================
 
 
+from dataclasses import dataclass
 import os
 import tempfile
+from typing import Tuple
 import unittest
 
+import numpy as np
+import tensorflow as tf
 import torch
 import torchvision
 
@@ -26,6 +30,15 @@ from ai_edge_torch.convert import conversion_utils as cutils
 from ai_edge_torch.testing import model_coverage
 
 
+@dataclass
+class TestContainer1:
+  data_1: torch.Tensor
+  data_2: Tuple[torch.Tensor, torch.Tensor]
+
+
+torch.export.register_dataclass(TestContainer1, serialized_type_name="TestContainer1")
+
+
 class TestConvert(unittest.TestCase):
   """Tests conversion of various modules."""
 
@@ -306,6 +319,99 @@ class TestConvert(unittest.TestCase):
         model_coverage.compare_tflite_torch(edge_model, model, args_gen, kwargs_gen)
     )
 
+  def test_convert_model_with_args_nested_kwargs_1(self):
+    """
+    Test converting a simple model with both sample_args and nested sample_kwargs.
+    """
+
+    class SampleModel(torch.nn.Module):
+
+      def forward(self, x: torch.Tensor, y: torch.Tensor, z: TestContainer1):
+        return x + y + z.data_1 + z.data_2[0] + z.data_2[1]
+
+    args = (torch.randn(10, 10),)
+    kwargs = dict(
+        y=torch.randn(10, 10),
+        z=TestContainer1(
+            data_1=torch.randn(10, 10),
+            data_2=(torch.randn(10, 10), torch.randn(10, 10)),
+        ),
+    )
+    flat_inputs = {
+        "args_0": args[0].numpy(),
+        "y": kwargs["y"].numpy(),
+        "z_data_1": kwargs["z"].data_1.numpy(),
+        "z_data_2_0": kwargs["z"].data_2[0].numpy(),
+        "z_data_2_1": kwargs["z"].data_2[1].numpy(),
+    }
+    self._compare_tflite_torch_args_kwargs(SampleModel(), args, kwargs, flat_inputs)
+
+  def test_convert_model_with_args_nested_kwargs_2(self):
+    """
+    Test converting a simple model with both sample_args and nested sample_kwargs.
+    """
+
+    class SampleModel(torch.nn.Module):
+
+      def forward(self, x, y, z):
+        return x + y + z.data_1 + z.data_2[0][0] + z.data_2[1]
+
+    args = (torch.randn(10, 10),)
+    kwargs = dict(
+        y=torch.randn(10, 10),
+        z=TestContainer1(
+            data_1=torch.randn(10, 10),
+            data_2=[(torch.randn(10, 10),), torch.randn(10, 10)],
+        ),
+    )
+    flat_inputs = {
+        "args_0": args[0].numpy(),
+        "y": kwargs["y"].numpy(),
+        "z_data_1": kwargs["z"].data_1.numpy(),
+        "z_data_2_0_0": kwargs["z"].data_2[0][0].numpy(),
+        "z_data_2_1": kwargs["z"].data_2[1].numpy(),
+    }
+    self._compare_tflite_torch_args_kwargs(SampleModel(), args, kwargs, flat_inputs)
+
+  def test_convert_model_with_args_nested_kwargs_3(self):
+    """
+    Test converting a simple model with both sample_args and nested sample_kwargs.
+    """
+
+    class SampleModel(torch.nn.Module):
+
+      def forward(self, x, y, z):
+        return x + y + z.data_1 + z.data_2[0]["foo"] + z.data_2[1]
+
+    args = (torch.randn(10, 10),)
+    kwargs = dict(
+        y=torch.randn(10, 10),
+        z=TestContainer1(
+            data_1=torch.randn(10, 10),
+            data_2=(dict(foo=torch.randn(10, 10)), torch.randn(10, 10)),
+        ),
+    )
+    flat_inputs = {
+        "args_0": args[0].numpy(),
+        "y": kwargs["y"].numpy(),
+        "z_data_1": kwargs["z"].data_1.numpy(),
+        "z_data_2_0_foo": kwargs["z"].data_2[0]["foo"].numpy(),
+        "z_data_2_1": kwargs["z"].data_2[1].numpy(),
+    }
+    self._compare_tflite_torch_args_kwargs(SampleModel(), args, kwargs, flat_inputs)
+
+  def _compare_tflite_torch_args_kwargs(self, model, args, kwargs, flat_inputs):
+    model.eval()
+    edge_model = ai_edge_torch.convert(model, args, kwargs)
+    interpreter = tf.lite.Interpreter(model_content=edge_model._tflite_model)
+    runner = interpreter.get_signature_runner("serving_default")
+    input_details = runner.get_input_details()
+    self.assertEqual(input_details.keys(), flat_inputs.keys())
+
+    reference_output = model(*args, **kwargs)
+    tflite_output = edge_model(**flat_inputs)
+    np.testing.assert_almost_equal(reference_output, tflite_output)
+
 
 if __name__ == "__main__":
   unittest.main()

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)