ai-edge-torch-nightly 0.3.0.dev20240926__py3-none-any.whl → 0.3.0.dev20240929__py3-none-any.whl

ai_edge_torch/generative/examples/test_models/toy_model.py

@@ -44,13 +44,10 @@ class ToySingleLayerModel(torch.nn.Module):
     self.rope_cache = attn_utils.build_rope_cache(
         size=config.max_seq_len,
         dim=int(attn_config.rotary_percentage * attn_config.head_dim),
-        base=10_000,
-        condense_ratio=1,
-        dtype=torch.float32,
-        device=torch.device('cpu'),
+        base=attn_config.rotary_base,
     )
     self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.max_seq_len, dtype=torch.float32, device=torch.device('cpu')
+        size=config.max_seq_len,
     )
     self.config = config
 
@@ -93,13 +90,10 @@ class ToySingleLayerModelWeightSharing(torch.nn.Module):
     self.rope_cache = attn_utils.build_rope_cache(
         size=config.max_seq_len,
         dim=int(attn_config.rotary_percentage * attn_config.head_dim),
-        base=10_000,
-        condense_ratio=1,
-        dtype=torch.float32,
-        device=torch.device('cpu'),
+        base=attn_config.rotary_base,
     )
     self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.max_seq_len, dtype=torch.float32, device=torch.device('cpu')
+        size=config.max_seq_len,
     )
     self.config = config
 
@@ -124,6 +118,7 @@ def get_model_config() -> cfg.ModelConfig:
       num_heads=32,
       head_dim=4,
       num_query_groups=4,
+      rotary_base=10000,
       rotary_percentage=1.0,
       enable_kv_cache=False,
   )

ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py

@@ -51,13 +51,10 @@ class ToyModelWithKVCache(torch.nn.Module):
     self.rope_cache = attn_utils.build_rope_cache(
         size=config.max_seq_len,
         dim=int(attn_config.rotary_percentage * attn_config.head_dim),
-        base=10_000,
-        condense_ratio=1,
-        dtype=torch.float32,
-        device=torch.device('cpu'),
+        base=attn_config.rotary_base,
     )
     self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.max_seq_len, dtype=torch.float32, device=torch.device('cpu')
+        size=config.max_seq_len,
     )
     self.config = config
 
@@ -91,6 +88,7 @@ def get_model_config() -> cfg.ModelConfig:
       num_heads=32,
       head_dim=4,
       num_query_groups=4,
+      rotary_base=10000,
       rotary_percentage=1.0,
   )
   ff_config = cfg.FeedForwardConfig(

ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py

@@ -67,15 +67,10 @@ class TinyLlama(nn.Module):
     self.rope_cache = attn_utils.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"),
+        base=attn_config.rotary_base,
     )
     self.mask_cache = attn_utils.build_causal_mask_cache(
         size=config.kv_cache_max,
-        dtype=torch.float32,
-        device=torch.device("cpu"),
     )
     self.config = config
 
@@ -132,6 +127,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       num_heads=32,
       head_dim=64,
       num_query_groups=4,
+      rotary_base=10000,
       rotary_percentage=1.0,
   )
   ff_config = cfg.FeedForwardConfig(

ai_edge_torch/generative/examples/tiny_llama/verify.py

@@ -21,6 +21,7 @@ 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
 
@@ -30,16 +31,20 @@ _PROMPTS = flags.DEFINE_multi_string(
     "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)
-  wrapper_model = verifier.ModelWrapper(
-      model=transformers.AutoModelForCausalLM.from_pretrained(
-          checkpoint, trust_remote_code=True
-      ),
+  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
@@ -52,10 +57,13 @@ def main(_):
   tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint)
 
   verifier.verify_reauthored_model(
-      original_model=wrapper_model,
-      reauthored_model=reauthored_model,
-      tokenizer=tokenizer,
+      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,
   )
 

ai_edge_torch/generative/layers/model_config.py

@@ -83,6 +83,8 @@ class AttentionConfig:
   # Used to determine number of groups in grouped query attention (GQA)
   # https://arxiv.org/pdf/2305.13245.pdf
   num_query_groups: Optional[int]
+  # Base of rotary positional embedding.
+  rotary_base: int = 10_000
   # Percentage of Rotary Positional Embedding added Q and K projections.
   rotary_percentage: Optional[float] = None
   # Whether to transpose the query groups of qkv bundled tensor before

ai_edge_torch/generative/test/test_model_conversion_large.py

@@ -19,9 +19,11 @@ import ai_edge_torch
 from ai_edge_torch import config as ai_edge_config
 from ai_edge_torch.generative.examples.gemma import gemma1
 from ai_edge_torch.generative.examples.gemma import gemma2
+from ai_edge_torch.generative.examples.llama import llama
 from ai_edge_torch.generative.examples.openelm import openelm
 from ai_edge_torch.generative.examples.phi import phi2
 from ai_edge_torch.generative.examples.phi import phi3
+from ai_edge_torch.generative.examples.qwen import qwen
 from ai_edge_torch.generative.examples.smollm import smollm
 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
@@ -102,6 +104,15 @@ class TestModelConversion(googletest.TestCase):
     pytorch_model = gemma2.Gemma2(config).eval()
     self._test_model(config, pytorch_model, "prefill", atol=1e-4, rtol=1e-5)
 
+  @googletest.skipIf(
+      ai_edge_config.Config.use_torch_xla,
+      reason="tests with custom ops are not supported on oss",
+  )
+  def test_llama(self):
+    config = llama.get_fake_model_config()
+    pytorch_model = llama.Llama(config).eval()
+    self._test_model(config, pytorch_model, "prefill", atol=1e-3, rtol=1e-5)
+
   @googletest.skipIf(
       ai_edge_config.Config.use_torch_xla,
       reason="tests with custom ops are not supported on oss",
@@ -142,6 +153,15 @@ class TestModelConversion(googletest.TestCase):
     pytorch_model = openelm.OpenELM(config).eval()
     self._test_model(config, pytorch_model, "prefill", atol=1e-4, rtol=1e-5)
 
+  @googletest.skipIf(
+      ai_edge_config.Config.use_torch_xla,
+      reason="tests with custom ops are not supported on oss",
+  )
+  def test_qwen(self):
+    config = qwen.get_fake_model_config()
+    pytorch_model = qwen.Qwen(config).eval()
+    self._test_model(config, pytorch_model, "prefill", atol=1e-3, rtol=1e-5)
+
   @googletest.skipIf(
       ai_edge_config.Config.use_torch_xla,
       reason="tests with custom ops are not supported on oss",

ai_edge_torch/generative/utilities/transformers_verifier.py

@@ -0,0 +1,42 @@
+# 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.
+# ==============================================================================
+
+"""Utilities for the models predefined in HuggingFace transformers."""
+
+from typing import cast
+
+from ai_edge_torch.generative.utilities import verifier
+import torch
+import transformers
+
+
+class TransformersModelWrapper(verifier.ModelWrapper):
+  """A wrapper for the model predefined in HuggingFace transformers.
+
+  Verifier expects forward() to return logits while Transformers models return
+  an object with `logits` field.
+
+  Transformers models get `max_new_tokens` settings for generate() via
+  GenerationConfig.
+  """
+
+  def forward(self, tokens: torch.Tensor) -> torch.Tensor:
+    return self.model.forward(tokens).logits
+
+  def generate(
+      self, inputs: torch.Tensor, max_new_tokens: int
+  ) -> torch.IntTensor:
+    gen_config = transformers.GenerationConfig(max_new_tokens=max_new_tokens)
+    return self.model.generate(inputs=inputs, generation_config=gen_config)

ai_edge_torch/generative/utilities/verifier.py

@@ -16,111 +16,129 @@
 """Common utility functions to verify the reauthored models."""
 
 import logging
-from typing import List, Optional, Union
+from typing import List
 
 from ai_edge_torch.generative.layers import kv_cache as kv_utils
 import torch
-import transformers
 
 
 class ModelWrapper(torch.nn.Module):
-  """A wrapper for the model to be verified, this could be a HuggingFace model
+  """A wrapper for the model to be verified.
 
-  or a regular PyTorch model.
+  It unifies the interface of forward() and generate() of models for the
+  verification to call.
   """
 
-  def __init__(
-      self,
-      model: torch.nn.Module,
-      model_format: str = "huggingface",
-      hf_generation_config: Optional[transformers.GenerationConfig] = None,
-  ):
+  def __init__(self, model: torch.nn.Module):
     """Initializes the wrapper.
 
     Args:
-      model (torch.nn.Module): The original model. This could be a model built
-        from HuggingFace transformers, or a regular PyTorch model.
-      model_format (str): The format of the model. It should be either
-        "huggingface" or "pytorch".
-      hf_generation_config (transformers.GenerationConfig): The HuggingFace
-        generation config. This config will only be used if the underlying model
-        is built from HuggingFace transformers.
+      model (torch.nn.Module): The model which might have different interfaces
+        of forward() and generate(). It could be a model built from HuggingFace
+        transformers, a regular PyTorch model, or a model re-authored with
+        ai_edge_torch Generative API.
     """
     super().__init__()
     self.model = model
-    self.model_format = model_format
-    self.hf_generation_config = hf_generation_config
+
+  def forward(self, tokens: torch.Tensor) -> torch.Tensor:
+    """Gets output logits by forwarding the input tokens.
+
+    Args:
+      tokens (torch.Tensor): The input tokens to forward. Its dimension is
+        expected to be (batch_size=1, kv_cache_max_len).
+
+    Returns:
+      The output logits.
+    """
+    raise NotImplementedError("forward() is not implemented.")
 
   def generate(
-      self, inputs: torch.Tensor
-  ) -> Union[transformers.utils.ModelOutput, torch.LongTensor]:
-    if self.model_format == "huggingface":
-      return self.model.generate(
-          inputs=inputs, generation_config=self.hf_generation_config
-      )
-    else:
-      raise NotImplementedError(
-          "generate() is not implemented for model format: %s"
-          % self.model_format
-      )
+      self, prompts: torch.Tensor, max_new_tokens: int
+  ) -> torch.IntTensor:
+    """Returns the response token IDs to the given prompts tensor.
 
-  def forward(
-      self,
-      inputs: torch.Tensor,
-  ):
-    return self.model.forward(inputs)
+    The maximum number of tokens to generate might be set by subclasses.
 
+    Args:
+      prompts (torch.Tensor): The input token IDs to generate with. Its shape is
+        expected to be (batch_size=1, input_ids_len).
+      max_new_tokens (int): The maximum number of response token IDs to
+        generate.
+
+    Returns:
+      The tensor of response token IDs with shape of (batch_size=1,
+      response_ids_len).
+    """
+    raise NotImplementedError("generate() is not implemented.")
 
-def forward(
-    model: torch.nn.Module,
-    tokens: torch.Tensor,
-    kv_cache: kv_utils.KVCache,
-) -> tuple[torch.Tensor, kv_utils.KVCache]:
-  """Forwards the model reauthored with ai_edge_torch Generative API.
 
-  Args:
-    model (torch.nn.Module): The model to forward. It should be a model built
-      with ai_edge_torch Generative API.
-    tokens (torch.Tensor): The input tokens to forward.
-    kv_cache (KVCache): The KV cache to forward.
+class ReauthoredModelWrapper(ModelWrapper):
+  """A wrapper for the model reauthored with ai_edge_torch Generative API."""
 
-  Returns:
-    The output logits and the updated KV cache.
-  """
-  input_pos = torch.arange(0, tokens.shape[1], dtype=torch.int)
-  output = model.forward(tokens, input_pos, kv_cache)
-  return output["logits"], output["kv_cache"]
+  def _init_kv_cache(self):
+    """Returns an initialized KV cache."""
+    return kv_utils.KVCache.from_model_config(self.model.config)
 
+  def _forward_with_kv_cache(
+      self,
+      tokens: torch.Tensor,
+      kv_cache: kv_utils.KVCache,
+  ) -> tuple[torch.Tensor, kv_utils.KVCache]:
+    """Forwards the model and updates an external KV cache.
 
-def generate(
-    model: torch.nn.Module, prompts: torch.Tensor, response_len: int
-) -> torch.Tensor:
-  """Generates the response to the prompts.
+    Args:
+      tokens (torch.Tensor): The input tokens to forward.
+      kv_cache (KVCache): The KV cache to forward.
 
-  It appends tokens output by the model to the prompts and feeds them back to
-  the model up to decode_len.
+    Returns:
+      The output logits and the updated KV cache.
+    """
+    input_pos = torch.arange(0, tokens.shape[1], dtype=torch.int)
+    output = self.model.forward(tokens, input_pos, kv_cache)
+    return output["logits"], output["kv_cache"]
 
-  Args:
-    model (torch.nn.Module): The model to generate. It should be a model built
-      with ai_edge_torch Generative API.
-    prompts (torch.Tensor): The prompts to generate.
-    response_len (int): The number of tokens to generate.
+  def forward(self, tokens: torch.Tensor) -> torch.Tensor:
+    logits, _ = self._forward_with_kv_cache(tokens, self._init_kv_cache())
+    return logits
 
-  Returns:
-    The generated tokens.
-  """
-  input_ids = prompts[0].int().tolist()
-  kv_cache = kv_utils.KVCache.from_model_config(model.config)
-  for _ in range(response_len - len(input_ids)):
-    logits, kv_cache = forward(model, torch.tensor([input_ids]), kv_cache)
-    generated_token = logits[0][-1].argmax().item()
-    input_ids.append(generated_token)
-  return torch.tensor([input_ids])
+  def generate(
+      self, prompts: torch.Tensor, max_new_tokens: int
+  ) -> torch.IntTensor:
+    input_ids = prompts[0].int().tolist()
+    kv_cache = self._init_kv_cache()
+    for _ in range(max_new_tokens):
+      tokens = torch.tensor([input_ids])
+      logits, kv_cache = self._forward_with_kv_cache(tokens, kv_cache)
+      generated_token = logits[0][-1].argmax().item()
+      input_ids.append(generated_token)
+    return torch.tensor([input_ids])
+
+
+class TokenizerWrapper(torch.nn.Module):
+  """A wrapper for the tokenizer used for verification."""
+
+  def __init__(self, tokenizer: torch.nn.Module):
+    """Initializes the wrapper.
+
+    Args:
+      tokenizer (torch.nn.Module): The tokenizer to wrap.
+    """
+    super().__init__()
+    self.tokenizer = tokenizer
+
+  def encode(self, prompts: str) -> torch.Tensor:
+    """Encodes the prompts to token IDs."""
+    return self.tokenizer.encode(prompts, return_tensors="pt")
+
+  def decode(self, token_ids: torch.Tensor) -> str:
+    """Decodes the token IDs to a string."""
+    return self.tokenizer.decode(token_ids)
 
 
 def verify_with_input_ids(
     original_model: ModelWrapper,
-    reauthored_model: torch.nn.Module,
+    reauthored_model: ReauthoredModelWrapper,
     input_ids: List[int],
     kv_cache_max_len: int = 1024,
     rtol: float = 1e-05,
@@ -132,8 +150,8 @@ def verify_with_input_ids(
 
   Args:
     original_model (ModelWrapper): The original model.
-    reauthored_model (torch.nn.Module): The model reauthored with ai_edge_torch
-      Generative API.
+    reauthored_model (ReauthoredModelWrapper): The model reauthored with
+      ai_edge_torch Generative API.
     input_ids (List[int]): The input token IDs to forward with.
     kv_cache_max_len (int): The maximum sequence length of the KV cache.
     rtol (float): The relative tolerance for the comparison.
@@ -147,13 +165,12 @@ def verify_with_input_ids(
 
   logging.info("Forwarding the original model...")
   outputs_original = original_model.forward(tokens)
-  logits_original = outputs_original.logits[0, len(input_ids) - 1, :]
+  logits_original = outputs_original[0, len(input_ids) - 1, :]
   logging.info("logits_original: %s", logits_original)
 
   logging.info("Forwarding the reauthored model...")
-  kv_cache = kv_utils.KVCache.from_model_config(reauthored_model.config)
-  outputs_reauthored = forward(reauthored_model, tokens, kv_cache)
-  logits_reauthored = outputs_reauthored[0][0, len(input_ids) - 1, :]
+  outputs_reauthored = reauthored_model.forward(tokens)
+  logits_reauthored = outputs_reauthored[0, len(input_ids) - 1, :]
   logging.info("logits_reauthored: %s", logits_reauthored)
 
   return torch.allclose(
@@ -163,9 +180,10 @@ def verify_with_input_ids(
 
 def verify_model_with_prompts(
     original_model: ModelWrapper,
-    reauthored_model: torch.nn.Module,
-    tokenizer: torch.nn.Module,
+    reauthored_model: ReauthoredModelWrapper,
+    tokenizer: TokenizerWrapper,
     prompts: str,
+    max_new_tokens: int,
 ) -> bool:
   """Verifies if the model reauthored generates the same answer of the oringal.
 
@@ -174,24 +192,24 @@ def verify_model_with_prompts(
 
   Args:
     original_model (ModelWrapper): The original model.
-    reauthored_model (torch.nn.Module): The model reauthored with ai_edge_torch
-      Generative API.
-    tokenizer (torch.nn.Module): The tokenizer.
+    reauthored_model (ReauthoredModelWrapper): The model reauthored with
+      ai_edge_torch Generative API.
+    tokenizer (TokenizerWrapper): The tokenizer.
     prompts (str): The input prompts to generate answers.
+    max_new_tokens (int): The maximum number of new tokens to generate.
 
   Returns:
     True if the model reauthored generates the same answer of the original.
   """
-  prompt_tokens = tokenizer.encode(prompts, return_tensors="pt")
+  prompt_tokens = tokenizer.encode(prompts)
 
   logging.info("Generating answer with the original model...")
-  outputs_original = original_model.generate(prompt_tokens)
+  outputs_original = original_model.generate(prompt_tokens, max_new_tokens)
   response_original = tokenizer.decode(outputs_original[0])
   logging.info("outputs_from_original_model: [[%s]]", response_original)
 
   logging.info("Generating answer with the reauthored model...")
-  generate_len = len(outputs_original[0])
-  outputs_reauthored = generate(reauthored_model, prompt_tokens, generate_len)
+  outputs_reauthored = reauthored_model.generate(prompt_tokens, max_new_tokens)
   response_reauthored = tokenizer.decode(outputs_reauthored[0])
   logging.info("outputs from reauthored model: [[%s]]", response_reauthored)
 
@@ -200,9 +218,10 @@ def verify_model_with_prompts(
 
 def verify_reauthored_model(
     original_model: ModelWrapper,
-    reauthored_model: torch.nn.Module,
-    tokenizer: torch.nn.Module,
+    reauthored_model: ReauthoredModelWrapper,
+    tokenizer: TokenizerWrapper,
     generate_prompts: List[str],
+    max_new_tokens: int = 30,
     forward_input_ids: List[List[int]] = [[1, 2, 3, 4]],
     rtol: float = 1e-05,
     atol: float = 1e-05,
@@ -219,10 +238,11 @@ def verify_reauthored_model(
 
   Args:
     original_model (ModelWrapper): The original model.
-    reauthored_model (torch.nn.Module): The model reauthored with ai_edge_torch
-      Generative API.
-    tokenizer (torch.nn.Module): The tokenizer.
+    reauthored_model (ReauthoredModelWrapper): The model reauthored with
+      ai_edge_torch Generative API.
+    tokenizer (TokenizerWrapper): The tokenizer.
     generate_prompts (List[str]): List of the input prompts to generate answers.
+    max_new_tokens (int): The maximum number of new tokens to generate.
     forward_input_ids (List[torch.Tensor]): List if ihe input token IDs to
       forward with.
     rtol (float): The relative tolerance for the comparison.
@@ -235,13 +255,13 @@ def verify_reauthored_model(
     ):
       logging.info("PASS")
     else:
-      logging.info("FAILED")
+      logging.error("FAILED")
 
   for prompts in generate_prompts:
     logging.info("Verifying the reauthored model with prompts:%s", prompts)
     if verify_model_with_prompts(
-        original_model, reauthored_model, tokenizer, prompts
+        original_model, reauthored_model, tokenizer, prompts, max_new_tokens
     ):
       logging.info("PASS")
     else:
-      logging.info("FAILED")
+      logging.error("FAILED")

ai_edge_torch/version.py

@@ -13,4 +13,4 @@
 # limitations under the License.
 # ==============================================================================
 
-__version__ = "0.3.0.dev20240926"
+__version__ = "0.3.0.dev20240929"

{ai_edge_torch_nightly-0.3.0.dev20240926.dist-info → ai_edge_torch_nightly-0.3.0.dev20240929.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.3.0.dev20240926
+Version: 0.3.0.dev20240929
 Summary: Supporting PyTorch models with the Google AI Edge TFLite runtime.
 Home-page: https://github.com/google-ai-edge/ai-edge-torch
 Keywords: On-Device ML,AI,Google,TFLite,PyTorch,LLMs,GenAI