ai-edge-torch-nightly 0.3.0.dev20240910__py3-none-any.whl → 0.3.0.dev20240914__py3-none-any.whl

ai_edge_torch/generative/examples/{experimental/phi → phi}/phi2.py

@@ -12,26 +12,22 @@
 # 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.
+
+"""Example of building a Phi-2 model."""
 
 import os
-from pathlib import Path
-from typing import Tuple
+import pathlib
 
+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
-from ai_edge_torch.generative.layers.experimental import attention
-from ai_edge_torch.generative.layers.experimental import ekv_cache as kv_utils
 import ai_edge_torch.generative.layers.model_config as cfg
 import ai_edge_torch.generative.utilities.loader as loading_utils
 import numpy as np
 import torch
 from torch import nn
 
-
 TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
     ff_up_proj="model.layers.{}.mlp.fc1",
     ff_down_proj="model.layers.{}.mlp.fc2",
@@ -52,7 +48,6 @@ 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
@@ -60,18 +55,20 @@ class Phi2(nn.Module):
     self.tok_embedding = nn.Embedding(
         config.vocab_size, config.embedding_dim, padding_idx=0
     )
+    # Phi-2 has only one block config.
+    block_config = config.block_config(0)
     self.transformer_blocks = nn.ModuleList(
-        attention.TransformerBlock(config) for _ in range(config.num_layers)
+        attention.TransformerBlock(block_config, config)
+        for _ in range(config.num_layers)
     )
     self.final_norm = builder.build_norm(
         config.embedding_dim,
         config.final_norm_config,
     )
+    attn_config = block_config.attn_config
     self.rope_cache = attn_utils.build_rope_cache(
         size=config.kv_cache_max,
-        dim=int(
-            config.attn_config.rotary_percentage * config.attn_config.head_dim
-        ),
+        dim=int(attn_config.rotary_percentage * attn_config.head_dim),
         base=10_000,
         condense_ratio=1,
         dtype=torch.float32,
@@ -89,13 +86,17 @@ class Phi2(nn.Module):
       self,
       tokens: torch.Tensor,
       input_pos: torch.Tensor,
-      kv_cache: kv_utils.EKVCache,
-  ) -> Tuple[torch.Tensor, kv_utils.EKVCache]:
+      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)
@@ -111,11 +112,11 @@ class Phi2(nn.Module):
       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))
+    updated_kv_cache = kv_utils.KVCache(tuple(updated_kv_entires))
 
     x = self.final_norm(x)
-    res = self.lm_head(x)  # (b, t, vocab_size)
-    return res, updated_kv_cache
+    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:
@@ -143,17 +144,20 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       use_bias=True,
   )
   norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.LAYER_NORM)
+  block_config = cfg.TransformerBlockConfig(
+      attn_config=attn_config,
+      ff_config=ff_config,
+      pre_attention_norm_config=norm_config,
+      parallel_residual=True,
+  )
   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,
+      block_configs=block_config,
       final_norm_config=norm_config,
-      parallel_residual=True,
       lm_head_use_bias=True,
       enable_hlfb=True,
   )
@@ -165,43 +169,42 @@ def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
   config.vocab_size = 128
   config.num_layers = 2
   config.max_seq_len = 2 * kv_cache_max_len
-  config.ff_config.intermediate_size = 128
+  # Phi-2 has only one block config.
+  config.block_config(0).ff_config.intermediate_size = 128
   return config
 
 
-def build_model(
-    checkpoint_path: str, test_model: bool = False, **kwargs
-) -> nn.Module:
+def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
   """Instantiates the model instance and load checkpoint if provided."""
-  config = (
-      get_fake_model_config(**kwargs)
-      if test_model
-      else get_model_config(**kwargs)
-  )
+  config = get_model_config(**kwargs)
   model = Phi2(config)
-  if checkpoint_path is not None:
-    loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
-    loader.load(model)
+  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
+  loader.load(model)
   model.eval()
   return model
 
 
-def define_and_run(checkpoint_path: str, test_model: bool = False) -> None:
+def define_and_run(checkpoint_path: str) -> None:
   """Instantiates and runs a Phi-2 model."""
 
+  current_dir = pathlib.Path(__file__).parent.resolve()
+  phi2_goldens = torch.load(current_dir / "phi2_lm_logits.pt")
   kv_cache_max_len = 1024
-  model = build_model(
-      checkpoint_path, test_model=test_model, kv_cache_max_len=kv_cache_max_len
-  )
+  model = build_model(checkpoint_path, kv_cache_max_len=kv_cache_max_len)
   idx = torch.from_numpy(np.array([[1, 2, 3, 4]]))
-  tokens = torch.full((1, kv_cache_max_len), 0, dtype=torch.long, device="cpu")
+  tokens = torch.full((1, kv_cache_max_len), 0, dtype=torch.int, 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))
+  input_pos = torch.arange(0, kv_cache_max_len, dtype=torch.int)
+  kv = kv_utils.KVCache.from_model_config(model.config)
+  output = model.forward(tokens, input_pos, kv)
+  print("comparing with goldens..")
+  assert torch.allclose(
+      phi2_goldens, output["logits"][0, idx.shape[1] - 1, :], atol=1e-02
+  )
 
 
 if __name__ == "__main__":
-  input_checkpoint_path = os.path.join(Path.home(), "Downloads/phi2")
+  input_checkpoint_path = os.path.join(
+      pathlib.Path.home(), "Downloads/llm_data/phi2"
+  )
   define_and_run(input_checkpoint_path)

ai_edge_torch/generative/examples/{experimental/tiny_llama → smollm}/convert_to_tflite.py

@@ -12,28 +12,26 @@
 # 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.
 
+"""Example of converting SmolLM model to multi-signature tflite model."""
 
 import os
-from pathlib import Path
+import pathlib
 
 import ai_edge_torch
-from ai_edge_torch.generative.examples.experimental.tiny_llama import tiny_llama
-from ai_edge_torch.generative.layers.experimental import ekv_cache as kv_utils
+from ai_edge_torch.generative.examples.smollm import smollm
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
 from ai_edge_torch.generative.quantize import quant_recipes
 import torch
 
 
-def convert_tiny_llama_to_tflite(
+def convert_smollm_to_tflite(
     checkpoint_path: str,
     prefill_seq_len: int = 512,
     kv_cache_max_len: int = 1024,
     quantize: bool = True,
 ):
-  """An example for converting TinyLlama model to multi-signature tflite model.
+  """Converts SmolLM model to multi-signature tflite model.
 
   Args:
       checkpoint_path (str): The filepath to the model checkpoint, or directory
@@ -45,15 +43,15 @@ def convert_tiny_llama_to_tflite(
       quantize (bool, optional): Whether the model should be quanized. Defaults
         to True.
   """
-  pytorch_model = tiny_llama.build_model(
+  pytorch_model = smollm.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)
+  prefill_tokens = torch.full((1, prefill_seq_len), 0, dtype=torch.int)
+  prefill_input_pos = torch.arange(0, prefill_seq_len, dtype=torch.int)
+  decode_token = torch.tensor([[0]], dtype=torch.int)
+  decode_input_pos = torch.tensor([0], dtype=torch.int)
+  kv = kv_utils.KVCache.from_model_config(pytorch_model.config)
 
   quant_config = quant_recipes.full_int8_dynamic_recipe() if quantize else None
   edge_model = (
@@ -77,11 +75,12 @@ def convert_tiny_llama_to_tflite(
       )
       .convert(quant_config=quant_config)
   )
+  quant_suffix = 'q8' if quantize else 'f32'
   edge_model.export(
-      f'/tmp/tiny_llama_seq{prefill_seq_len}_ekv{kv_cache_max_len}.tflite'
+      f'/tmp/smollm_{quant_suffix}_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)
+  path = os.path.join(pathlib.Path.home(), 'Downloads/llm_data/smollm')
+  convert_smollm_to_tflite(path)

ai_edge_torch/generative/examples/smollm/smollm.py

@@ -0,0 +1,131 @@
+# 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 SmolLM model."""
+
+import copy
+import os
+import pathlib
+
+from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
+import ai_edge_torch.generative.layers.model_config as cfg
+import ai_edge_torch.generative.utilities.loader as loading_utils
+import numpy as np
+import torch
+from torch import nn
+
+TENSOR_NAMES = copy.copy(tiny_llama.TENSOR_NAMES)
+# SmolLM re-uses the embedding as the head projection layer.
+TENSOR_NAMES.lm_head = None
+
+
+class SmolLM(tiny_llama.TinyLlama):
+  """A SmolLM model built from the Edge Generative API layers.
+
+  SmolLM shares the same architecture as TinyLlama, but with different model
+  sizes.
+  """
+
+  def __init__(self, config: cfg.ModelConfig):
+    super().__init__(config)
+    # SmolLM re-uses the embedding as the head projection layer.
+    self.lm_head.weight.data = self.tok_embedding.weight.data
+
+
+def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a SmolLM 135M model.
+
+  Args:
+    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
+      is 1024.
+
+  Returns:
+    The model config for a SmolLM model.
+  """
+  attn_config = cfg.AttentionConfig(
+      num_heads=9,
+      head_dim=64,
+      num_query_groups=3,
+      rotary_percentage=1.0,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.GATED,
+      activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
+      intermediate_size=1536,
+  )
+  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)
+  block_config = cfg.TransformerBlockConfig(
+      attn_config=attn_config,
+      ff_config=ff_config,
+      pre_attention_norm_config=norm_config,
+      post_attention_norm_config=norm_config,
+  )
+  config = cfg.ModelConfig(
+      vocab_size=49152,
+      num_layers=30,
+      max_seq_len=2048,
+      embedding_dim=576,
+      kv_cache_max_len=kv_cache_max_len,
+      block_configs=block_config,
+      final_norm_config=norm_config,
+      enable_hlfb=True,
+  )
+  return config
+
+
+def get_fake_model_config(**kwargs) -> cfg.ModelConfig:
+  config = get_model_config(**kwargs)
+  config.vocab_size = 128
+  config.num_layers = 2
+  # SmolLM 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:
+  config = get_model_config(**kwargs)
+  model = SmolLM(config)
+  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
+  # Since embedding and lm-head use the same weight, we need to set strict
+  # to False.
+  loader.load(model, strict=False)
+  model.eval()
+  return model
+
+
+def define_and_run(checkpoint_path: str) -> None:
+  """Instantiates and runs a SmolLM model."""
+
+  current_dir = pathlib.Path(__file__).parent.resolve()
+  smollm_goldens = torch.load(current_dir / "smollm_lm_logits.pt")
+  kv_cache_max_len = 1024
+  model = build_model(checkpoint_path, kv_cache_max_len=kv_cache_max_len)
+  idx = torch.from_numpy(np.array([[1, 2, 3, 4]]))
+  tokens = torch.full((1, kv_cache_max_len), 0, dtype=torch.int, device="cpu")
+  tokens[0, :4] = idx
+  input_pos = torch.arange(0, kv_cache_max_len, dtype=torch.int)
+  kv = kv_utils.KVCache.from_model_config(model.config)
+  output = model.forward(tokens, input_pos, kv)
+  assert torch.allclose(
+      smollm_goldens, output["logits"][0, idx.shape[1] - 1, :], atol=1e-05
+  )
+
+
+if __name__ == "__main__":
+  input_checkpoint_path = os.path.join(
+      pathlib.Path.home(), "Downloads/llm_data/smollm"
+  )
+  define_and_run(input_checkpoint_path)

ai_edge_torch/generative/examples/stable_diffusion/clip.py

@@ -61,8 +61,10 @@ class CLIP(nn.Module):
     )
 
     self.config = config
+    # CLIP has only one block config.
+    block_config = config.block_config(0)
     self.transformer_blocks = nn.ModuleList(
-        TransformerBlock(config) for _ in range(config.num_layers)
+        TransformerBlock(block_config, config) for _ in range(config.num_layers)
     )
     self.final_norm = builder.build_norm(
         config.embedding_dim, config.final_norm_config
@@ -74,7 +76,7 @@ class CLIP(nn.Module):
 
   @torch.inference_mode
   def forward(self, tokens: torch.LongTensor) -> torch.FloatTensor:
-    tokens = tokens.type(torch.long)
+    tokens = tokens.type(torch.int)
 
     state = self.tok_embedding(tokens) + self.tok_embedding_position
     for layer in self.transformer_blocks:
@@ -112,15 +114,19 @@ def get_model_config() -> cfg.ModelConfig:
 
   norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.LAYER_NORM)
 
+  block_config = cfg.TransformerBlockConfig(
+      attn_config=attn_config,
+      ff_config=ff_config,
+      pre_attention_norm_config=norm_config,
+      post_attention_norm_config=norm_config,
+  )
+
   config = cfg.ModelConfig(
       vocab_size=vocab_size,
       num_layers=num_layers,
       max_seq_len=max_seq_len,
       embedding_dim=embedding_dim,
-      attn_config=attn_config,
-      ff_config=ff_config,
-      pre_attention_norm_config=norm_config,
-      post_attention_norm_config=norm_config,
+      block_configs=block_config,
       final_norm_config=norm_config,
       enable_hlfb=True,
   )

ai_edge_torch/generative/examples/stable_diffusion/convert_to_tflite.py

@@ -94,7 +94,7 @@ def convert_stable_diffusion_to_tflite(
   n_tokens = 77
   timestamp = 0
   len_prompt = 1
-  prompt_tokens = torch.full((1, n_tokens), 0, dtype=torch.long)
+  prompt_tokens = torch.full((1, n_tokens), 0, dtype=torch.int)
   input_image = torch.full(
       (1, 3, image_height, image_width), 0, dtype=torch.float32
   )

ai_edge_torch/generative/examples/t5/convert_to_tflite.py

@@ -29,24 +29,24 @@ def convert_t5_to_tflite_singlesig(checkpoint_path: str):
 
   # encoder
   seq_len = 512
-  prefill_e_tokens = torch.full((1, seq_len), 0, dtype=torch.long)
+  prefill_e_tokens = torch.full((1, seq_len), 0, dtype=torch.int)
   prompt_e_token = [1, 2, 3, 4, 5, 6]
   prefill_e_tokens[0, : len(prompt_e_token)] = torch.tensor(
-      prompt_e_token, dtype=torch.long
+      prompt_e_token, dtype=torch.int
   )
-  prefill_e_input_pos = torch.arange(0, seq_len)
-  prefill_d_tokens = torch.full((1, seq_len), 0, dtype=torch.long)
+  prefill_e_input_pos = torch.arange(0, seq_len, dtype=torch.int)
+  prefill_d_tokens = torch.full((1, seq_len), 0, dtype=torch.int)
   prompt_d_token = [1, 2, 3, 4, 5, 6]
   prefill_d_tokens[0, : len(prompt_d_token)] = torch.tensor(
-      prompt_d_token, dtype=torch.long
+      prompt_d_token, dtype=torch.int
   )
-  prefill_d_input_pos = torch.arange(0, seq_len)
+  prefill_d_input_pos = torch.arange(0, seq_len, dtype=torch.int)
 
   # decoder
-  decode_token = torch.tensor([[1]], dtype=torch.long)
-  decode_input_pos = torch.tensor([0], dtype=torch.int64)
-  decode_d_token = torch.tensor([[1]], dtype=torch.long)
-  decode_d_input_pos = torch.tensor([0], dtype=torch.int64)
+  decode_token = torch.tensor([[1]], dtype=torch.int)
+  decode_input_pos = torch.tensor([0], dtype=torch.int)
+  decode_d_token = torch.tensor([[1]], dtype=torch.int)
+  decode_d_input_pos = torch.tensor([0], dtype=torch.int)
 
   # Pad mask for self attention only on "real" tokens.
   # Pad with `-inf` for any tokens indices that aren't desired.
@@ -81,24 +81,24 @@ def convert_t5_to_tflite_multisig(checkpoint_path: str):
 
   # encoder
   seq_len = 512
-  prefill_e_tokens = torch.full((1, seq_len), 0, dtype=torch.long)
+  prefill_e_tokens = torch.full((1, seq_len), 0, dtype=torch.int)
   prompt_e_token = [1, 2, 3, 4, 5, 6]
   prefill_e_tokens[0, : len(prompt_e_token)] = torch.tensor(
-      prompt_e_token, dtype=torch.long
+      prompt_e_token, dtype=torch.int
   )
-  prefill_e_input_pos = torch.arange(0, seq_len)
-  prefill_d_tokens = torch.full((1, seq_len), 0, dtype=torch.long)
+  prefill_e_input_pos = torch.arange(0, seq_len, dtype=torch.int)
+  prefill_d_tokens = torch.full((1, seq_len), 0, dtype=torch.int)
   prompt_d_token = [1, 2, 3, 4, 5, 6]
   prefill_d_tokens[0, : len(prompt_d_token)] = torch.tensor(
-      prompt_d_token, dtype=torch.long
+      prompt_d_token, dtype=torch.int
   )
-  prefill_d_input_pos = torch.arange(0, seq_len)
+  prefill_d_input_pos = torch.arange(0, seq_len, dtype=torch.int)
 
   # decoder
-  decode_token = torch.tensor([[1]], dtype=torch.long)
-  decode_input_pos = torch.tensor([0], dtype=torch.int64)
-  decode_d_token = torch.tensor([[1]], dtype=torch.long)
-  decode_d_input_pos = torch.tensor([0], dtype=torch.int64)
+  decode_token = torch.tensor([[1]], dtype=torch.int)
+  decode_input_pos = torch.tensor([0], dtype=torch.int)
+  decode_d_token = torch.tensor([[1]], dtype=torch.int)
+  decode_d_input_pos = torch.tensor([0], dtype=torch.int)
 
   # Pad mask for self attention only on "real" tokens.
   # Pad with `-inf` for any tokens indices that aren't desired.