ai-edge-torch-nightly 0.3.0.dev20240926__py3-none-any.whl → 0.3.0.dev20240928__py3-none-any.whl

ai_edge_torch/generative/examples/openelm/verify.py

@@ -20,6 +20,7 @@ import pathlib
 from absl import app
 from absl import flags
 from ai_edge_torch.generative.examples.openelm import openelm
+from ai_edge_torch.generative.utilities import transformers_verifier
 from ai_edge_torch.generative.utilities import verifier
 import transformers
 
@@ -29,15 +30,18 @@ _PROMPTS = flags.DEFINE_multi_string(
     "What is the meaning of life?",
     "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 = "apple/OpenELM-3B"
   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.
@@ -53,10 +57,13 @@ def main(_):
   tokenizer = transformers.AutoTokenizer.from_pretrained(tokenizer_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,
   )
 
 

ai_edge_torch/generative/examples/phi/phi2.py

@@ -65,15 +65,10 @@ class Phi2(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
 
@@ -129,6 +124,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       num_heads=32,
       head_dim=80,
       num_query_groups=32,
+      rotary_base=10000,
       rotary_percentage=0.4,
       qkv_use_bias=True,
       output_proj_use_bias=True,

ai_edge_torch/generative/examples/phi/phi3.py

@@ -97,15 +97,15 @@ ROPE_SHORT_FACTOR = [
 ]
 
 
-def build_rope_cache(
+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,
+    base: int,
+    condense_ratio: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    theta_factors: torch.Tensor,
+    scale: float,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
   """Precomputes Rotary Positional Embeddings for Phi-3.5 model.
 
@@ -116,26 +116,20 @@ def build_rope_cache(
   Args:
       size (int): The size of the built cache.
       dim (int): Each sequence's dimmension.
-      base (int, optional): Rope base value. Defaults to 10000.
+      base (int, optional): Rope base value.
       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.
+        condensed.
+      dtype (torch.dtype, optional): Output tensor's data type.
+      device (torch.device, optional): Output tensor's data type.
       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.
+        scale the theta values.
+      scale (float, optional): A float used to scale the rope values.
 
   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
+  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
@@ -167,10 +161,10 @@ class Phi3_5Mini(nn.Module):
         config.final_norm_config,
     )
     attn_config = block_config.attn_config
-    self.rope_cache = build_rope_cache(
+    self.rope_cache = _build_rope_cache(
         size=config.kv_cache_max,
         dim=int(attn_config.rotary_percentage * attn_config.head_dim),
-        base=10_000,
+        base=attn_config.rotary_base,
         condense_ratio=1,
         dtype=torch.float32,
         device=torch.device("cpu"),
@@ -181,8 +175,6 @@ class Phi3_5Mini(nn.Module):
     )
     self.mask_cache = attn_utils.build_causal_mask_cache(
         size=config.kv_cache_max,
-        dtype=torch.float32,
-        device=torch.device("cpu"),
     )
     self.config = config
 
@@ -238,6 +230,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       num_heads=32,
       head_dim=96,
       num_query_groups=32,
+      rotary_base=10000,
       rotary_percentage=1.0,
       qkv_transpose_before_split=True,
   )

ai_edge_torch/generative/examples/phi/verify.py

@@ -19,6 +19,7 @@ import logging
 from absl import app
 from absl import flags
 from ai_edge_torch.generative.examples.phi import phi2
+from ai_edge_torch.generative.utilities import transformers_verifier
 from ai_edge_torch.generative.utilities import verifier
 import kagglehub
 import transformers
@@ -39,12 +40,7 @@ _MAX_NEW_TOKENS = flags.DEFINE_integer(
 def main(_):
   checkpoint = kagglehub.model_download("Microsoft/phi/transformers/2")
   logging.info("Loading the original model from: %s", checkpoint)
-  generation_config = transformers.GenerationConfig.from_pretrained(checkpoint)
-  generation_config.max_new_tokens = _MAX_NEW_TOKENS.value
-  wrapper_model = verifier.ModelWrapper(
-      model=transformers.AutoModelForCausalLM.from_pretrained(checkpoint),
-      hf_generation_config=generation_config,
-  )
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint)
 
   logging.info("Building the reauthored model from: %s", checkpoint)
   reauthored_model = phi2.build_model(checkpoint)
@@ -53,10 +49,13 @@ 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-03,
   )
 

ai_edge_torch/generative/examples/phi/verify_phi3.py

@@ -21,6 +21,7 @@ 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 transformers_verifier
 from ai_edge_torch.generative.utilities import verifier
 import transformers
 
@@ -40,12 +41,7 @@ _MAX_NEW_TOKENS = flags.DEFINE_integer(
 def main(_):
   checkpoint = "microsoft/Phi-3.5-mini-instruct"
   logging.info("Loading the original model from: %s", checkpoint)
-  generation_config = transformers.GenerationConfig.from_pretrained(checkpoint)
-  generation_config.max_new_tokens = _MAX_NEW_TOKENS.value
-  wrapper_model = verifier.ModelWrapper(
-      model=transformers.AutoModelForCausalLM.from_pretrained(checkpoint),
-      hf_generation_config=generation_config,
-  )
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint)
 
   # Locate the cached dir.
   cached_config_file = transformers.utils.cached_file(
@@ -59,10 +55,13 @@ 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,
   )
 
 

ai_edge_torch/generative/examples/smollm/smollm.py

@@ -54,6 +54,7 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       num_heads=9,
       head_dim=64,
       num_query_groups=3,
+      rotary_base=10000,
       rotary_percentage=1.0,
   )
   ff_config = cfg.FeedForwardConfig(

ai_edge_torch/generative/examples/smollm/verify.py

@@ -21,6 +21,7 @@ import pathlib
 from absl import app
 from absl import flags
 from ai_edge_torch.generative.examples.smollm import smollm
+from ai_edge_torch.generative.utilities import transformers_verifier
 from ai_edge_torch.generative.utilities import verifier
 import transformers
 
@@ -30,14 +31,18 @@ _PROMPTS = flags.DEFINE_multi_string(
     "What is the meaning of life?",
     "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 = "HuggingFaceTB/SmolLM-135M"
   logging.info("Loading the original model from: %s", checkpoint)
-  wrapper_model = verifier.ModelWrapper(
-      model=transformers.AutoModelForCausalLM.from_pretrained(checkpoint),
-  )
+  original_model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint)
+
   # Locate the cached dir.
   cached_config_file = transformers.utils.cached_file(
       checkpoint, transformers.utils.CONFIG_NAME
@@ -50,10 +55,13 @@ 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/examples/stable_diffusion/clip.py

@@ -98,6 +98,7 @@ def get_model_config() -> cfg.ModelConfig:
       num_heads=num_heads,
       head_dim=embedding_dim // num_heads,
       num_query_groups=num_query_groups,
+      rotary_base=0,
       rotary_percentage=0.0,
       qkv_use_bias=True,
       qkv_transpose_before_split=True,
@@ -148,6 +149,7 @@ def get_fake_model_config() -> cfg.ModelConfig:
       num_heads=num_heads,
       head_dim=embedding_dim // num_heads,
       num_query_groups=num_query_groups,
+      rotary_base=0,
       rotary_percentage=0.0,
       qkv_use_bias=True,
       qkv_transpose_before_split=True,

ai_edge_torch/generative/examples/stable_diffusion/decoder.py

@@ -295,6 +295,7 @@ def get_model_config() -> unet_cfg.AutoEncoderConfig:
           enable_kv_cache=False,
           qkv_transpose_before_split=True,
           qkv_fused_interleaved=False,
+          rotary_base=0,
           rotary_percentage=0.0,
       ),
       enable_hlfb=False,
@@ -351,6 +352,7 @@ def get_fake_model_config() -> unet_cfg.AutoEncoderConfig:
           enable_kv_cache=False,
           qkv_transpose_before_split=True,
           qkv_fused_interleaved=False,
+          rotary_base=0,
           rotary_percentage=0.0,
       ),
       enable_hlfb=False,

ai_edge_torch/generative/examples/stable_diffusion/diffusion.py

@@ -199,6 +199,7 @@ def build_attention_config(
     num_heads,
     dim,
     num_query_groups,
+    rotary_base=0,
     rotary_percentage=0.0,
     qkv_transpose_before_split=True,
     qkv_use_bias=False,
@@ -211,6 +212,7 @@ def build_attention_config(
       num_heads=num_heads,
       head_dim=dim // num_heads,
       num_query_groups=num_query_groups,
+      rotary_base=rotary_base,
       rotary_percentage=rotary_percentage,
       qkv_transpose_before_split=qkv_transpose_before_split,
       qkv_use_bias=qkv_use_bias,

ai_edge_torch/generative/examples/t5/t5.py

@@ -335,8 +335,6 @@ class T5Decoder(nn.Module):
 
     self.dec_attn_mask_cache = attn_utils.build_causal_mask_cache(
         size=config.kv_cache_max,
-        dtype=torch.float32,
-        device=torch.device("cpu"),
     )
 
   @torch.inference_mode

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

@@ -19,6 +19,7 @@ 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
@@ -102,6 +103,15 @@ 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",

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)