ai-edge-torch-nightly 0.3.0.dev20250114__py3-none-any.whl

ai_edge_torch/generative/examples/tiny_llama/convert_to_tflite.py

@@ -0,0 +1,80 @@
+# 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 converting TinyLlama model to multi-signature tflite model."""
+
+import os
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
+from ai_edge_torch.generative.utilities import converter
+from ai_edge_torch.generative.utilities.model_builder import ExportConfig
+
+_CHECKPOINT_PATH = flags.DEFINE_string(
+    'checkpoint_path',
+    os.path.join(pathlib.Path.home(), 'Downloads/llm_data/tiny_llama'),
+    'The path to the model checkpoint, or directory holding the checkpoint.',
+)
+_OUTPUT_PATH = flags.DEFINE_string(
+    'output_path',
+    '/tmp/',
+    'The path to export the tflite model.',
+)
+_OUTPUT_NAME_PREFIX = flags.DEFINE_string(
+    'output_name_prefix',
+    'tinyllama',
+    'The prefix of the output tflite model name.',
+)
+_PREFILL_SEQ_LENS = flags.DEFINE_multi_integer(
+    'prefill_seq_lens',
+    (8, 64, 128, 256, 512, 1024),
+    'List of the maximum sizes of prefill input tensors.',
+)
+_KV_CACHE_MAX_LEN = flags.DEFINE_integer(
+    'kv_cache_max_len',
+    1280,
+    'The maximum size of KV cache buffer, including both prefill and decode.',
+)
+_QUANTIZE = flags.DEFINE_bool(
+    'quantize',
+    True,
+    'Whether the model should be quantized.',
+)
+_LORA_RANKS = flags.DEFINE_multi_integer(
+    'lora_ranks',
+    None,
+    'If set, the model will be converted with the provided list of LoRA ranks.',
+)
+
+
+def main(_):
+  pytorch_model = tiny_llama.build_model(
+      _CHECKPOINT_PATH.value, kv_cache_max_len=_KV_CACHE_MAX_LEN.value
+  )
+  converter.convert_to_tflite(
+      pytorch_model,
+      output_path=_OUTPUT_PATH.value,
+      output_name_prefix=_OUTPUT_NAME_PREFIX.value,
+      prefill_seq_len=_PREFILL_SEQ_LENS.value,
+      quantize=_QUANTIZE.value,
+      lora_ranks=_LORA_RANKS.value,
+      export_config=ExportConfig(),
+  )
+
+
+if __name__ == '__main__':
+  app.run(main)

ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py

@@ -0,0 +1,88 @@
+# 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."""
+
+import ai_edge_torch.generative.layers.model_config as cfg
+from ai_edge_torch.generative.utilities import model_builder
+from torch import nn
+
+TENSOR_NAMES = model_builder.TENSOR_NAMES_WITH_SEPARATE_LM_HEAD
+
+
+class TinyLlama(model_builder.DecoderOnlyModel):
+  """A TinyLlama model built from the Edge Generative API layers."""
+  pass
+
+
+def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a TinyLlama model.
+
+  Args:
+    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
+      is 1024.
+
+  Returns:
+    The model config for a TinyLlama model.
+  """
+  attn_config = cfg.AttentionConfig(
+      num_heads=32,
+      head_dim=64,
+      num_query_groups=4,
+      rotary_base=10000,
+      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)
+  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=32000,
+      num_layers=22,
+      max_seq_len=2048,
+      embedding_dim=2048,
+      kv_cache_max_len=kv_cache_max_len,
+      block_configs=block_config,
+      final_norm_config=norm_config,
+      lm_head_share_weight_with_embedding=False,
+      enable_hlfb=True,
+  )
+  return config
+
+
+def get_fake_model_config(**kwargs) -> cfg.ModelConfig:
+  config = get_model_config(**kwargs)
+  config.vocab_size = 128
+  config.num_layers = 2
+  # TinyLlama has only one block config.
+  config.block_config(0).ff_config.intermediate_size = 64
+  return config
+
+
+def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+      model_class=TinyLlama,
+  )

ai_edge_torch/generative/examples/tiny_llama/verify.py

@@ -0,0 +1,72 @@
+# 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.
+# ==============================================================================
+
+"""Verifies the reauthored TinyLlama-1.1B model."""
+
+import logging
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
+from ai_edge_torch.generative.utilities import transformers_verifier
+from ai_edge_torch.generative.utilities import verifier
+import transformers
+
+
+_PROMPTS = flags.DEFINE_multi_string(
+    "prompts",
+    "Show me the program to add 2 and 3.",
+    "The input prompts to generate answers.",
+)
+_MAX_NEW_TOKENS = flags.DEFINE_integer(
+    "max_new_tokens",
+    30,
+    "The maximum size of the generated tokens.",
+)
+
+
+def main(_):
+  checkpoint = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
+  logging.info("Loading the original model from: %s", checkpoint)
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(
+      checkpoint, trust_remote_code=True
+  )
+
+  # Locate the cached dir.
+  cached_config_file = transformers.utils.cached_file(
+      checkpoint, transformers.utils.CONFIG_NAME
+  )
+  reauthored_checkpoint = pathlib.Path(cached_config_file).parent
+  logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
+  reauthored_model = tiny_llama.build_model(reauthored_checkpoint)
+
+  logging.info("Loading the tokenizer from: %s", checkpoint)
+  tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint)
+
+  verifier.verify_reauthored_model(
+      original_model=transformers_verifier.TransformersModelWrapper(
+          original_model
+      ),
+      reauthored_model=verifier.ReauthoredModelWrapper(reauthored_model),
+      tokenizer=verifier.TokenizerWrapper(tokenizer),
+      generate_prompts=_PROMPTS.value,
+      max_new_tokens=_MAX_NEW_TOKENS.value,
+      atol=1e-04,
+  )
+
+
+if __name__ == "__main__":
+  app.run(main)

ai_edge_torch/generative/fx_passes/__init__.py

@@ -0,0 +1,30 @@
+# 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.
+# ==============================================================================
+from ai_edge_torch import fx_pass_base
+from ai_edge_torch.fx_pass_base import CanonicalizePass
+from ai_edge_torch.generative.fx_passes.remove_sdpa_zero_mask_pass import RemoveSDPACompositeZeroMaskPass
+import torch
+
+
+def run_generative_passes(
+    exported_program: torch.export.ExportedProgram,
+) -> torch.export.ExportedProgram:
+  return fx_pass_base.run_passes(
+      exported_program,
+      [
+          RemoveSDPACompositeZeroMaskPass(),
+          CanonicalizePass(),
+      ],
+  )

ai_edge_torch/generative/fx_passes/remove_sdpa_zero_mask_pass.py

@@ -0,0 +1,50 @@
+# 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.
+# ==============================================================================
+from ai_edge_torch import fx_pass_base
+from ai_edge_torch import lowertools
+import torch
+
+
+class RemoveSDPACompositeZeroMaskPass(fx_pass_base.ExportedProgramPassBase):
+
+  def is_zero_tensor_node(self, node: torch.fx.Node):
+    return node.target == torch.ops.aten.zeros.default
+
+  def call(self, exported_program: torch.export.ExportedProgram):
+    graph = exported_program.graph_module.graph
+    for node in graph.nodes:
+      if not (
+          node.op == "call_function"
+          and node.target == lowertools.mark_tensor_op
+      ):
+        continue
+
+      source, name, io_position, id, is_input = node.args[:5]
+      # Composite info:
+      # - name: odml.scaled_dot_product_attention
+      # - inputs: q, k, v, mask
+      if (
+          name == "odml.scaled_dot_product_attention"
+          and is_input
+          and io_position == 3
+      ):
+        if self.is_zero_tensor_node(source):
+          # Remove the mark_tensor call on the mask input by
+          # replacing the target with an identity function.
+          node.target = lambda *args, **kwargs: torch.zeros_like(args[0])
+
+    exported_program.graph_module.graph.lint()
+    exported_program.graph_module.recompile()
+    return fx_pass_base.ExportedProgramPassResult(exported_program, True)

ai_edge_torch/generative/layers/__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.
+# ==============================================================================