Merge branch 'google-ai-edge:main' into fix_sample_inference_error

ai_edge_torch/generative/examples/README.md

@@ -7,9 +7,13 @@ Gemma is Google's open-source LLM. The model has both a 2B and 7B versions. See
 ## TinyLlama
 [TinyLlama](https://github.com/jzhang38/TinyLlama) is a popular OSS smaller version of Meta's Llama2 model, with only 1.1B parameters. [HuggingFace checkpoint](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0).
 
-## Microsoft Phi-2
-Microsoft Phi-2 is also a decoder-only LLM with 2.7B parameters, see details on
-[Kaggle](https://www.kaggle.com/models/Microsoft/phi/transformers/2).
+## Microsoft Phi-2 and 3.5-mini
+Microsoft Phi-2 and Phi-3.5-mini are also decoder-only LLMs with 2.7B and 3.82B
+parameters each. See details on
+[Kaggle](https://www.kaggle.com/models/Microsoft/phi/transformers/2) for Phi-2
+and [HuggingFace](https://huggingface.co/microsoft/Phi-3.5-mini-instruct) for
+Phi-3.5-mini. Note that the example of Phi-3.5-mini supports up to 4K tokens,
+not to 128K tokens which the original Phi-3.5 supports.
 
 ## Apple OpenELM
 [Apple OpenELM](https://huggingface.co/apple/OpenELM) is also a decoder-only LLM

ai_edge_torch/generative/examples/phi/convert_phi3_to_tflite.py

@@ -0,0 +1,68 @@
+# 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 a Phi-3.5 model to multi-signature tflite model."""
+
+import os
+import pathlib
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.phi import phi3
+from ai_edge_torch.generative.utilities import converter
+
+_CHECKPOINT_PATH = flags.DEFINE_string(
+    'checkpoint_path',
+    os.path.join(pathlib.Path.home(), 'Downloads/llm_data/phi3'),
+    'The path to the model checkpoint, or directory holding the checkpoint.',
+)
+_TFLITE_PATH = flags.DEFINE_string(
+    'tflite_path',
+    '/tmp/',
+    'The tflite file path to export.',
+)
+_PREFILL_SEQ_LEN = flags.DEFINE_integer(
+    'prefill_seq_len',
+    1024,
+    'The maximum size of prefill input tensor.',
+)
+_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.',
+)
+
+
+def main(_):
+  pytorch_model = phi3.build_model(
+      _CHECKPOINT_PATH.value, kv_cache_max_len=_KV_CACHE_MAX_LEN.value
+  )
+  quant_suffix = 'q8' if _QUANTIZE.value else 'f32'
+  output_filename = f'phi3_{quant_suffix}_seq{_PREFILL_SEQ_LEN.value}_ekv{_KV_CACHE_MAX_LEN.value}.tflite'
+  converter.convert_to_tflite(
+      pytorch_model,
+      tflite_path=os.path.join(_TFLITE_PATH.value, output_filename),
+      prefill_seq_len=_PREFILL_SEQ_LEN.value,
+      quantize=_QUANTIZE.value,
+  )
+
+
+if __name__ == '__main__':
+  app.run(main)

ai_edge_torch/generative/examples/phi/phi3.py

@@ -0,0 +1,286 @@
+# 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 Phi-3.5 model up to 4K tokens, not to 128K tokens."""
+
+import math
+from typing import Tuple
+
+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.model_config as cfg
+import ai_edge_torch.generative.utilities.loader as loading_utils
+import torch
+from torch import nn
+
+TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
+    ff_up_proj="model.layers.{}.mlp.gate_up_proj",
+    ff_down_proj="model.layers.{}.mlp.down_proj",
+    attn_fused_qkv_proj="model.layers.{}.self_attn.qkv_proj",
+    attn_output_proj="model.layers.{}.self_attn.o_proj",
+    pre_attn_norm="model.layers.{}.input_layernorm",
+    post_attn_norm="model.layers.{}.post_attention_layernorm",
+    embedding="model.embed_tokens",
+    final_norm="model.norm",
+    lm_head="lm_head",
+)
+
+# max_position_embeddings / original_max_position_embeddings in Phi-3.5 config.
+ROPE_SCALE_FACTOR = 32
+
+# ROPE short factor in Phi-3.5 config. According to LOPE paper and its code in
+# https://github.com/microsoft/LongRoPE, these values had been searched with
+# min=1.0, step-0.01 to optimize the errors of sample dataset.
+ROPE_SHORT_FACTOR = [
+    1.0,
+    1.0199999809265137,
+    1.0299999713897705,
+    1.0299999713897705,
+    1.0499999523162842,
+    1.0499999523162842,
+    1.0499999523162842,
+    1.0499999523162842,
+    1.0499999523162842,
+    1.0699999332427979,
+    1.0999999046325684,
+    1.1099998950958252,
+    1.1599998474121094,
+    1.1599998474121094,
+    1.1699998378753662,
+    1.2899998426437378,
+    1.339999794960022,
+    1.679999828338623,
+    1.7899998426437378,
+    1.8199998140335083,
+    1.8499997854232788,
+    1.8799997568130493,
+    1.9099997282028198,
+    1.9399996995925903,
+    1.9899996519088745,
+    2.0199997425079346,
+    2.0199997425079346,
+    2.0199997425079346,
+    2.0199997425079346,
+    2.0199997425079346,
+    2.0199997425079346,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0299997329711914,
+    2.0799996852874756,
+    2.0899996757507324,
+    2.189999580383301,
+    2.2199995517730713,
+    2.5899994373321533,
+    2.729999542236328,
+    2.749999523162842,
+    2.8399994373321533,
+]
+
+
+def build_rope_cache(
+    size: int,
+    dim: int,
+    base: int = 10000,
+    condense_ratio: int = 1,
+    dtype: torch.dtype = torch.float32,
+    device: torch.device = None,
+    theta_factors: torch.Tensor = None,
+    scale: float = 1.0,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+  """Precomputes Rotary Positional Embeddings for Phi-3.5 model.
+
+  It's a modified version of attn_utils.build_rope_cache with additional
+  arguments for Phi-3.5 model. It precompute Rotary Positional Embedding Sin and
+  Cos values with scaling factors for quick lookup during the inference.
+
+  Args:
+      size (int): The size of the built cache.
+      dim (int): Each sequence's dimmension.
+      base (int, optional): Rope base value. Defaults to 10000.
+      condense_ratio (int, optional): The ratio by which sequence indicies are
+        condensed. Defaults to 1.
+      dtype (torch.dtype, optional): Output tensor's data type. Defaults to
+        torch.float32.
+      device (torch.device, optional): Output tensor's data type. Defaults to
+        None in which case "cpu" is used.
+      theta_factors (torch.Tensor, optional): A tensor of shape (dim,) used to
+        scale the theta values. Defaults to None.
+      scale (float, optional): A float used to scale the rope values. Defaults
+        to 1.0.
+
+  Returns:
+      Tuple[torch.Tensor, torch.Tensor]: Rope's Cosine and Sine waves.
+  """
+  if device is None:
+    device = torch.device('cpu')
+  theta = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
+  if theta_factors is not None:
+    theta = theta / theta_factors
+  seq_idx = torch.arange(size) / condense_ratio
+  idx_theta = torch.outer(seq_idx, theta)
+  cos = torch.cos(idx_theta).to(dtype=dtype, device=device) * scale
+  sin = torch.sin(idx_theta).to(dtype=dtype, device=device) * scale
+  return cos, sin
+
+
+class Phi3_5Mini(nn.Module):
+  """A Phi-3.5 model built from the Edge Generative API layers."""
+
+  def __init__(self, config: cfg.ModelConfig):
+    super().__init__()
+
+    # 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
+    )
+    # Phi-3.5 has only one block config.
+    block_config = config.block_config(0)
+    self.transformer_blocks = nn.ModuleList(
+        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 = build_rope_cache(
+        size=config.kv_cache_max,
+        dim=int(attn_config.rotary_percentage * attn_config.head_dim),
+        base=10_000,
+        condense_ratio=1,
+        dtype=torch.float32,
+        device=torch.device("cpu"),
+        theta_factors=torch.tensor(ROPE_SHORT_FACTOR),
+        scale=math.sqrt(
+            1 + math.log(ROPE_SCALE_FACTOR) / math.log(config.max_seq_len)
+        ),
+    )
+    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.KVCache,
+  ) -> dict[torch.Tensor, kv_utils.KVCache]:
+    _, seq_len = tokens.size()
+    assert self.config.max_seq_len >= seq_len, (
+        f"Cannot forward sequence of length {seq_len}, max seq length is only"
+        f" {self.config.max_seq_len}"
+    )
+    assert len(self.transformer_blocks) == len(kv_cache.caches), (
+        "The number of transformer blocks and the number of KV cache entries"
+        " must be the same."
+    )
+
+    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.KVCache(tuple(updated_kv_entires))
+
+    x = self.final_norm(x)
+    logits = self.lm_head(x)  # (b, t, vocab_size)
+    return {"logits": logits, "kv_cache": updated_kv_cache}
+
+
+def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Phi-3.5 model.
+
+  Args:
+    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
+      is 1024.
+
+  Returns:
+    The model config for a Phi-2 model.
+  """
+  attn_config = cfg.AttentionConfig(
+      num_heads=32,
+      head_dim=96,
+      num_query_groups=32,
+      rotary_percentage=1.0,
+      qkv_transpose_before_split=True,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.SEQUENTIAL,
+      activation=cfg.ActivationConfig(cfg.ActivationType.SILU_GLU),
+      intermediate_size=8192,
+  )
+  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=32064,
+      num_layers=32,
+      max_seq_len=4096,
+      kv_cache_max_len=kv_cache_max_len,
+      embedding_dim=3072,
+      block_configs=block_config,
+      final_norm_config=norm_config,
+      enable_hlfb=True,
+  )
+  return config
+
+
+def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
+  config = get_model_config(kv_cache_max_len)
+  config.vocab_size = 128
+  config.num_layers = 2
+  config.max_seq_len = 2 * kv_cache_max_len
+  # Phi-3.5 has only one block config.
+  config.block_config(0).ff_config.intermediate_size = 128
+  return config
+
+
+def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
+  """Instantiates the model instance and load checkpoint if provided."""
+  config = get_model_config(**kwargs)
+  model = Phi3_5Mini(config)
+  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
+  loader.load(model)
+  model.eval()
+  return model

ai_edge_torch/generative/examples/phi/verify.py

@@ -27,7 +27,6 @@
     "Instruct: Write an email about the weather Output:",
     "The input prompts to generate answers.",
 )
-
 _MAX_NEW_TOKENS = flags.DEFINE_integer(
     "max_new_tokens",
     30,