ai-edge-torch-nightly 0.3.0.dev20241002__py3-none-any.whl → 0.3.0.dev20241005__py3-none-any.whl

ai_edge_torch/generative/examples/qwen/qwen.py

@@ -15,28 +15,10 @@
 
 """Example of building Qwen 2.5 models."""
 
-import copy
-
-from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
 import ai_edge_torch.generative.layers.model_config as cfg
-import ai_edge_torch.generative.utilities.loader as loading_utils
-from torch import nn
-
-TENSOR_NAMES = copy.copy(tiny_llama.TENSOR_NAMES)
-# Qwen re-uses the embedding as the head projection layer.
-TENSOR_NAMES.lm_head = None
-
-
-class Qwen(tiny_llama.TinyLlama):
-  """A Qwen model built from the Edge Generative API layers.
-
-  Qwen 2.5 shares the same architecture as TinyLlama.
-  """
+from ai_edge_torch.generative.utilities import model_builder
 
-  def __init__(self, config: cfg.ModelConfig):
-    super().__init__(config)
-    # Qwen re-uses the embedding as the head projection layer.
-    self.lm_head.weight.data = self.tok_embedding.weight.data
+TENSOR_NAMES = model_builder.TENSOR_NAMES
 
 
 def get_3b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
@@ -119,23 +101,31 @@ def get_fake_model_config(**kwargs) -> cfg.ModelConfig:
   return config
 
 
-def _build_model(checkpoint_path: str, config: cfg.ModelConfig) -> nn.Module:
-  model = Qwen(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 build_3b_model(checkpoint_path: str, **kwargs) -> nn.Module:
-  return _build_model(checkpoint_path, get_3b_model_config(**kwargs))
+def build_3b_model(
+    checkpoint_path: str, **kwargs
+) -> model_builder.DecoderOnlyModel:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_3b_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+  )
 
 
-def build_1_5b_model(checkpoint_path: str, **kwargs) -> nn.Module:
-  return _build_model(checkpoint_path, get_1_5b_model_config(**kwargs))
+def build_1_5b_model(
+    checkpoint_path: str, **kwargs
+) -> model_builder.DecoderOnlyModel:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_1_5b_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+  )
 
 
-def build_0_5b_model(checkpoint_path: str, **kwargs) -> nn.Module:
-  return _build_model(checkpoint_path, get_0_5b_model_config(**kwargs))
+def build_0_5b_model(
+    checkpoint_path: str, **kwargs
+) -> model_builder.DecoderOnlyModel:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_0_5b_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+  )

ai_edge_torch/generative/examples/smollm/smollm.py

@@ -15,29 +15,10 @@
 
 """Example of building a SmolLM model."""
 
-import copy
-
-from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
 import ai_edge_torch.generative.layers.model_config as cfg
-import ai_edge_torch.generative.utilities.loader as loading_utils
-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.
+from ai_edge_torch.generative.utilities import model_builder
 
-  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
+TENSOR_NAMES = model_builder.TENSOR_NAMES
 
 
 def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
@@ -91,12 +72,11 @@ def get_fake_model_config(**kwargs) -> cfg.ModelConfig:
   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 build_model(
+    checkpoint_path: str, **kwargs
+) -> model_builder.DecoderOnlyModel:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+  )

ai_edge_torch/generative/examples/stable_diffusion/clip.py

@@ -75,9 +75,7 @@ class CLIP(nn.Module):
     )
 
   @torch.inference_mode
-  def forward(self, tokens: torch.LongTensor) -> torch.FloatTensor:
-    tokens = tokens.type(torch.int)
-
+  def forward(self, tokens: torch.IntTensor) -> torch.FloatTensor:
     state = self.tok_embedding(tokens) + self.tok_embedding_position
     for layer in self.transformer_blocks:
       state = layer(state, mask=self.mask_cache)

ai_edge_torch/generative/examples/stable_diffusion/convert_to_tflite.py

@@ -13,47 +13,54 @@
 # limitations under the License.
 # ==============================================================================
 
-import argparse
 import os
-from pathlib import Path
-from typing import Optional
+import pathlib
 
+from absl import app
+from absl import flags
 import ai_edge_torch
-import ai_edge_torch.generative.examples.stable_diffusion.clip as clip
-import ai_edge_torch.generative.examples.stable_diffusion.decoder as decoder
-import ai_edge_torch.generative.examples.stable_diffusion.diffusion as diffusion
-from ai_edge_torch.generative.examples.stable_diffusion.encoder import Encoder
-import ai_edge_torch.generative.examples.stable_diffusion.util as util
+from ai_edge_torch.generative.examples.stable_diffusion import clip
+from ai_edge_torch.generative.examples.stable_diffusion import decoder
+from ai_edge_torch.generative.examples.stable_diffusion import diffusion
+from ai_edge_torch.generative.examples.stable_diffusion import util
 from ai_edge_torch.generative.quantize import quant_recipes
-import ai_edge_torch.generative.utilities.stable_diffusion_loader as stable_diffusion_loader
+from ai_edge_torch.generative.utilities import stable_diffusion_loader
 import torch
 
-arg_parser = argparse.ArgumentParser()
-arg_parser.add_argument(
-    '--clip_ckpt',
-    type=str,
+_CLIP_CKPT = flags.DEFINE_string(
+    'clip_ckpt',
+    None,
     help='Path to source CLIP model checkpoint',
     required=True,
 )
-arg_parser.add_argument(
-    '--diffusion_ckpt',
-    type=str,
+
+_DIFFUSION_CKPT = flags.DEFINE_string(
+    'diffusion_ckpt',
+    None,
     help='Path to source diffusion model checkpoint',
     required=True,
 )
-arg_parser.add_argument(
-    '--decoder_ckpt',
-    type=str,
+
+_DECODER_CKPT = flags.DEFINE_string(
+    'decoder_ckpt',
+    None,
     help='Path to source image decoder model checkpoint',
     required=True,
 )
-arg_parser.add_argument(
-    '--output_dir',
-    type=str,
+
+_OUTPUT_DIR = flags.DEFINE_string(
+    'output_dir',
+    None,
     help='Path to the converted TF Lite directory.',
     required=True,
 )
 
+_QUANTIZE = flags.DEFINE_bool(
+    'quantize',
+    help='Whether to quantize the model during conversion.',
+    default=True,
+)
+
 
 @torch.inference_mode
 def convert_stable_diffusion_to_tflite(
@@ -111,7 +118,7 @@ def convert_stable_diffusion_to_tflite(
   time_embedding = util.get_time_embedding(timestamp)
 
   if not os.path.exists(output_dir):
-    Path(output_dir).mkdir(parents=True, exist_ok=True)
+    pathlib.Path(output_dir).mkdir(parents=True, exist_ok=True)
 
   quant_config = (
       quant_recipes.full_int8_weight_only_recipe() if quantize else None
@@ -142,14 +149,15 @@ def convert_stable_diffusion_to_tflite(
   ).export(f'{output_dir}/decoder.tflite')
 
 
-if __name__ == '__main__':
-  args = arg_parser.parse_args()
+def main(_):
   convert_stable_diffusion_to_tflite(
-      output_dir=args.output_dir,
-      clip_ckpt_path=args.clip_ckpt,
-      diffusion_ckpt_path=args.diffusion_ckpt,
-      decoder_ckpt_path=args.decoder_ckpt,
-      image_height=512,
-      image_width=512,
-      quantize=True,
+      output_dir=_OUTPUT_DIR.value,
+      clip_ckpt_path=_CLIP_CKPT.value,
+      diffusion_ckpt_path=_DIFFUSION_CKPT.value,
+      decoder_ckpt_path=_DECODER_CKPT.value,
+      quantize=_QUANTIZE.value,
   )
+
+
+if __name__ == '__main__':
+  app.run(main)

ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py

@@ -15,102 +15,10 @@
 
 """Example of building a TinyLlama model."""
 
-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
+from ai_edge_torch.generative.utilities import model_builder
 
-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",
-    post_attn_norm="model.layers.{}.post_attention_layernorm",
-    embedding="model.embed_tokens",
-    final_norm="model.norm",
-    lm_head="lm_head",
-)
-
-
-class TinyLlama(nn.Module):
-  """A TinyLlama 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
-    )
-    # TinyLlama 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 = attn_utils.build_rope_cache(
-        size=config.kv_cache_max,
-        dim=int(attn_config.rotary_percentage * attn_config.head_dim),
-        base=attn_config.rotary_base,
-    )
-    self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.kv_cache_max,
-    )
-    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]
-
-    # 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.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}
+TENSOR_NAMES = model_builder.TENSOR_NAMES_WITH_SEPARATE_LM_HEAD
 
 
 def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
@@ -150,6 +58,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       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
@@ -164,10 +73,11 @@ def get_fake_model_config(**kwargs) -> cfg.ModelConfig:
   return config
 
 
-def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
-  config = get_model_config(**kwargs)
-  model = TinyLlama(config)
-  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
-  loader.load(model)
-  model.eval()
-  return model
+def build_model(
+    checkpoint_path: str, **kwargs
+) -> model_builder.DecoderOnlyModel:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+  )

ai_edge_torch/generative/layers/model_config.py

@@ -184,8 +184,14 @@ class ModelConfig:
       default_factory=NormalizationConfig
   )
 
+  # Scale factor of the embedding.
+  embedding_scale: Optional[float] = None
+
   # Use bias term within LLM's HEAD.
   lm_head_use_bias: bool = False
+  # Whether LLM's HEAD shares the weight of the embedding.
+  lm_head_share_weight_with_embedding: bool = True
+
   # Whether to turn on high-level function boundary.
   enable_hlfb: bool = False
 

ai_edge_torch/generative/test/test_loader.py

@@ -19,6 +19,7 @@ import tempfile
 
 from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
 from ai_edge_torch.generative.utilities import loader as loading_utils
+from ai_edge_torch.generative.utilities import model_builder
 import safetensors.torch
 import torch
 
@@ -71,7 +72,7 @@ class TestLoader(googletest.TestCase):
       safetensors.torch.save_file(test_weights, file_path)
       cfg = tiny_llama.get_model_config()
       cfg.num_layers = 1
-      model = tiny_llama.TinyLlama(cfg)
+      model = model_builder.DecoderOnlyModel(cfg)
 
       loader = loading_utils.ModelLoader(file_path, tiny_llama.TENSOR_NAMES)
       # if returns successfully, it means all the tensors were initiallized.

ai_edge_torch/generative/test/test_model_conversion.py

@@ -21,6 +21,7 @@ from ai_edge_torch.generative.examples.test_models import toy_model_with_kv_cach
 from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
 from ai_edge_torch.generative.layers import kv_cache
 from ai_edge_torch.generative.test import utils as test_utils
+from ai_edge_torch.generative.utilities import model_builder
 import numpy as np
 import torch
 
@@ -42,31 +43,40 @@ class TestModelConversion(googletest.TestCase):
         )
     )
 
-  def _test_model_with_kv_cache(self, config, pytorch_model):
+  def _get_params(self, enable_hlfb: bool):
+    """Returns a model, edge model and the kwargs to use for testing."""
+    config = toy_model_with_kv_cache.get_model_config()
+    config.enable_hlfb = enable_hlfb
+    pytorch_model = toy_model_with_kv_cache.ToyModelWithKVCache(config).eval()
     tokens, input_pos = torch.tensor([[1]], dtype=torch.int), torch.tensor(
         [10], dtype=torch.int
     )
     kv = kv_cache.KVCache.from_model_config(config)
+    kwargs = {
+        "tokens": tokens,
+        "input_pos": input_pos,
+        "kv_cache": kv,
+    }
 
     edge_model = ai_edge_torch.convert(
         pytorch_model,
-        sample_kwargs={
-            "tokens": tokens,
-            "input_pos": input_pos,
-            "kv_cache": kv,
-        },
+        sample_kwargs=kwargs,
     )
     edge_model.set_interpreter_builder(
         self._interpreter_builder(edge_model.tflite_model())
     )
+    return pytorch_model, edge_model, kwargs
+
+  def _test_model_with_kv_cache(self, enable_hlfb: bool):
+    pytorch_model, edge_model, kwargs = self._get_params(enable_hlfb)
 
     self.assertTrue(
         test_utils.compare_tflite_torch(
             edge_model,
             pytorch_model,
-            tokens,
-            input_pos,
-            kv,
+            kwargs["tokens"],
+            kwargs["input_pos"],
+            kwargs["kv_cache"],
             signature_name="serving_default",
             atol=1e-5,
             rtol=1e-5,
@@ -78,19 +88,31 @@ class TestModelConversion(googletest.TestCase):
       reason="tests with custom ops are not supported on oss",
   )
   def test_toy_model_with_kv_cache(self):
-    config = toy_model_with_kv_cache.get_model_config()
-    pytorch_model = toy_model_with_kv_cache.ToyModelWithKVCache(config).eval()
-    self._test_model_with_kv_cache(config, pytorch_model)
+    self._test_model_with_kv_cache(enable_hlfb=False)
 
   @googletest.skipIf(
       ai_edge_config.Config.use_torch_xla,
       reason="tests with custom ops are not supported on oss",
   )
   def test_toy_model_with_kv_cache_with_hlfb(self):
-    config = toy_model_with_kv_cache.get_model_config()
-    config.enable_hlfb = True
-    pytorch_model = toy_model_with_kv_cache.ToyModelWithKVCache(config).eval()
-    self._test_model_with_kv_cache(config, pytorch_model)
+    self._test_model_with_kv_cache(enable_hlfb=True)
+
+  @googletest.skipIf(
+      ai_edge_config.Config.use_torch_xla,
+      reason="tests with custom ops are not supported on oss",
+  )
+  def test_toy_model_has_ekv_op(self):
+    """Tests that the model has the external kv cache op."""
+    _, edge_model, _ = self._get_params(enable_hlfb=True)
+    interpreter_ = interpreter.InterpreterWithCustomOps(
+        custom_op_registerers=["GenAIOpsRegisterer"],
+        model_content=edge_model.tflite_model(),
+        experimental_default_delegate_latest_features=True,
+    )
+
+    # pylint: disable=protected-access
+    op_names = [op["op_name"] for op in interpreter_._get_ops_details()]
+    self.assertIn("odml.update_external_kv_cache", op_names)
 
   def _test_multisig_model(self, config, pytorch_model, atol, rtol):
     # prefill
@@ -163,7 +185,7 @@ class TestModelConversion(googletest.TestCase):
   )
   def test_tiny_llama_multisig(self):
     config = tiny_llama.get_fake_model_config()
-    pytorch_model = tiny_llama.TinyLlama(config).eval()
+    pytorch_model = model_builder.DecoderOnlyModel(config).eval()
     self._test_multisig_model(config, pytorch_model, atol=1e-5, rtol=1e-5)
 
 

ai_edge_torch/generative/test/test_model_conversion_large.py

@@ -29,6 +29,7 @@ from ai_edge_torch.generative.examples.stable_diffusion import clip as sd_clip
 from ai_edge_torch.generative.examples.stable_diffusion import decoder as sd_decoder
 from ai_edge_torch.generative.examples.stable_diffusion import diffusion as sd_diffusion
 from ai_edge_torch.generative.layers import kv_cache
+from ai_edge_torch.generative.utilities import model_builder
 from ai_edge_torch.generative.test import utils as test_utils
 import numpy as np
 import torch
@@ -90,7 +91,7 @@ class TestModelConversion(googletest.TestCase):
   )
   def test_gemma1(self):
     config = gemma1.get_fake_model_config()
-    pytorch_model = gemma1.Gemma(config).eval()
+    pytorch_model = model_builder.DecoderOnlyModel(config).eval()
     self._test_model(
         config, pytorch_model, "serving_default", atol=1e-2, rtol=1e-5
     )
@@ -119,7 +120,7 @@ class TestModelConversion(googletest.TestCase):
   )
   def test_phi2(self):
     config = phi2.get_fake_model_config()
-    pytorch_model = phi2.Phi2(config).eval()
+    pytorch_model = model_builder.DecoderOnlyModel(config).eval()
     self._test_model(
         config, pytorch_model, "serving_default", atol=1e-3, rtol=1e-3
     )
@@ -139,7 +140,7 @@ class TestModelConversion(googletest.TestCase):
   )
   def test_smollm(self):
     config = smollm.get_fake_model_config()
-    pytorch_model = smollm.SmolLM(config).eval()
+    pytorch_model = model_builder.DecoderOnlyModel(config).eval()
     self._test_model(config, pytorch_model, "prefill", atol=1e-4, rtol=1e-5)
 
   @googletest.skipIf(
@@ -148,7 +149,7 @@ class TestModelConversion(googletest.TestCase):
   )
   def test_openelm(self):
     config = openelm.get_fake_model_config()
-    pytorch_model = openelm.OpenELM(config).eval()
+    pytorch_model = model_builder.DecoderOnlyModel(config).eval()
     self._test_model(config, pytorch_model, "prefill", atol=1e-4, rtol=1e-5)
 
   @googletest.skipIf(
@@ -157,7 +158,7 @@ class TestModelConversion(googletest.TestCase):
   )
   def test_qwen(self):
     config = qwen.get_fake_model_config()
-    pytorch_model = qwen.Qwen(config).eval()
+    pytorch_model = model_builder.DecoderOnlyModel(config).eval()
     self._test_model(config, pytorch_model, "prefill", atol=1e-3, rtol=1e-5)
 
   @googletest.skipIf(