ai-edge-torch-nightly 0.6.0.dev20250521__py3-none-any.whl → 0.6.0.dev20250522__py3-none-any.whl

ai_edge_torch/generative/examples/qwen/convert_v3_to_tflite.py

@@ -0,0 +1,61 @@
+# Copyright 2025 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 Qwen 3.0 models to multi-signature tflite model."""
+
+from absl import app
+from ai_edge_torch.generative.examples.qwen import qwen3
+from ai_edge_torch.generative.utilities import converter
+from ai_edge_torch.generative.utilities import export_config
+from ai_edge_torch.generative.utilities import loader
+
+flags = converter.define_conversion_flags('qwen')
+
+_MODEL_SIZE = flags.DEFINE_enum(
+    'model_size',
+    '1.7b',
+    ['0.6b', '1.7b', '4b'],
+    'The size of the model to convert.',
+)
+
+_BUILDER = {
+    '0.6b': qwen3.build_0_6b_model,
+    '1.7b': qwen3.build_1_7b_model,
+    '4b': qwen3.build_4b_model,
+}
+
+
+def main(_):
+  checkpoint_path = flags.FLAGS.checkpoint_path
+  pytorch_model = _BUILDER[_MODEL_SIZE.value](
+      checkpoint_path,
+      custom_loader=loader.maybe_get_custom_loader(
+          checkpoint_path, flags.FLAGS.custom_checkpoint_loader
+      ),
+      kv_cache_max_len=flags.FLAGS.kv_cache_max_len,
+  )
+  converter.convert_to_tflite(
+      pytorch_model,
+      output_path=flags.FLAGS.output_path,
+      output_name_prefix=flags.FLAGS.output_name_prefix,
+      prefill_seq_len=flags.FLAGS.prefill_seq_lens,
+      quantize=flags.FLAGS.quantize,
+      lora_ranks=flags.FLAGS.lora_ranks,
+      export_config=export_config.get_from_flags(),
+  )
+
+
+if __name__ == '__main__':
+  app.run(main)

ai_edge_torch/generative/examples/qwen/qwen3.py

@@ -0,0 +1,171 @@
+# Copyright 2025 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 Qwen 3.0 models."""
+
+from typing import Callable, Dict
+import ai_edge_torch.generative.layers.model_config as cfg
+from ai_edge_torch.generative.utilities import loader as loading_utils
+from ai_edge_torch.generative.utilities import model_builder
+import torch
+from torch import nn
+
+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_query_norm="model.layers.{}.self_attn.q_norm",
+    attn_key_norm="model.layers.{}.self_attn.k_norm",
+    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 Qwen3(model_builder.DecoderOnlyModel):
+  """A Qwen3 model built from the Edge Generative API layers."""
+
+  pass
+
+
+def get_4b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Qwen 3.0 4B 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.
+  """
+  norm_config = cfg.NormalizationConfig(
+      type=cfg.NormalizationType.RMS_NORM, epsilon=1e-06
+  )
+  attn_config = cfg.AttentionConfig(
+      num_heads=32,
+      head_dim=128,
+      num_query_groups=8,
+      query_norm_config=norm_config,
+      key_norm_config=norm_config,
+      rotary_base=1000000,
+      rotary_percentage=1.0,
+      qkv_use_bias=False,
+      qkv_transpose_before_split=True,
+      qkv_fused_interleaved=False,  # No interleaved qkv projection.
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.GATED,
+      activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
+      intermediate_size=9728,
+  )
+  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=151936,
+      num_layers=36,
+      max_seq_len=40960,
+      embedding_dim=2560,
+      kv_cache_max_len=kv_cache_max_len,
+      block_configs=block_config,
+      final_norm_config=norm_config,
+  )
+  return config
+
+
+def get_1_7b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Qwen 3.0 1.7B model."""
+  config = get_4b_model_config(kv_cache_max_len)
+  # Qwen has only one block config.
+  block_config = config.block_config(0)
+  block_config.attn_config.num_heads = 16
+  block_config.attn_config.head_dim = 128
+  block_config.ff_config.intermediate_size = 6144
+  config.num_layers = 28
+  config.embedding_dim = 2048
+  return config
+
+
+def get_0_6b_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Qwen 3.0 0.6B model."""
+  config = get_4b_model_config(kv_cache_max_len)
+  # Qwen has only one block config.
+  block_config = config.block_config(0)
+  block_config.attn_config.num_heads = 16
+  block_config.attn_config.head_dim = 128
+  block_config.ff_config.intermediate_size = 3072
+  config.num_layers = 28
+  config.embedding_dim = 1024
+  return config
+
+
+def get_fake_model_config(**kwargs) -> cfg.ModelConfig:
+  config = get_4b_model_config(**kwargs)
+  config.vocab_size = 128
+  config.num_layers = 2
+  # Qwen has only one block config.
+  config.block_config(0).ff_config.intermediate_size = 64
+  return config
+
+
+def build_4b_model(
+    checkpoint_path: str,
+    custom_loader: Callable[[str], Dict[str, torch.Tensor]] = None,
+    **kwargs
+) -> nn.Module:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_4b_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+      model_class=Qwen3,
+      custom_loader=custom_loader,
+  )
+
+
+def build_1_7b_model(
+    checkpoint_path: str,
+    custom_loader: Callable[[str], Dict[str, torch.Tensor]] = None,
+    **kwargs
+) -> nn.Module:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_1_7b_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+      model_class=Qwen3,
+      custom_loader=custom_loader,
+  )
+
+
+def build_0_6b_model(
+    checkpoint_path: str,
+    custom_loader: Callable[[str], Dict[str, torch.Tensor]] = None,
+    **kwargs
+) -> nn.Module:
+  return model_builder.build_decoder_only_model(
+      checkpoint_path=checkpoint_path,
+      config=get_0_6b_model_config(**kwargs),
+      tensor_names=TENSOR_NAMES,
+      model_class=Qwen3,
+      custom_loader=custom_loader,
+  )

ai_edge_torch/generative/examples/qwen/{verify.py → verify_qwen2.py}

@@ -48,6 +48,7 @@ _CHECKPOINT = {
 def main(_):
   verify_util.verify_qwen(
       model_size=_MODEL_SIZE.value,
+      model_version="v2",
       checkpoint_dir=_CHECKPOINT[_MODEL_SIZE.value],
       max_new_tokens=_MAX_NEW_TOKENS.value,
       prompts=_PROMPTS.value,

ai_edge_torch/generative/examples/qwen/verify_qwen3.py

@@ -0,0 +1,59 @@
+# Copyright 2025 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.
+# ==============================================================================
+
+"""Verifies the reauthored Qwen 3.0 0.6B, 1.7B, and 4B models."""
+
+
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.qwen import verify_util
+
+
+_MODEL_SIZE = flags.DEFINE_enum(
+    "model_size",
+    "0.6b",
+    ["0.6b", "1.7b", "4b"],
+    "The size of the model to verify.",
+)
+_PROMPTS = flags.DEFINE_multi_string(
+    "prompts",
+    "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.",
+)
+
+_CHECKPOINT = {
+    "0.6b": "Qwen/Qwen3-0.6B",
+    "1.7b": "Qwen/Qwen3-1.7B",
+    "4b": "Qwen/Qwen3-4B",
+}
+
+
+def main(_):
+  verify_util.verify_qwen(
+      model_size=_MODEL_SIZE.value,
+      model_version="v3",
+      checkpoint_dir=_CHECKPOINT[_MODEL_SIZE.value],
+      max_new_tokens=_MAX_NEW_TOKENS.value,
+      prompts=_PROMPTS.value,
+  )
+
+
+if __name__ == "__main__":
+  app.run(main)

ai_edge_torch/generative/examples/qwen/verify_util.py

@@ -17,24 +17,36 @@ import logging
 import os
 import pathlib
 
-from ai_edge_torch.generative.examples.qwen import qwen
+from ai_edge_torch.generative.examples.qwen import qwen, qwen3
 from ai_edge_torch.generative.utilities import loader
 from ai_edge_torch.generative.utilities import transformers_verifier
 from ai_edge_torch.generative.utilities import verifier
 import transformers
 
 
-_BUILDER = {
+_BUILDER_V2 = {
     "0.5b": qwen.build_0_5b_model,
     "1.5b": qwen.build_1_5b_model,
     "3b": qwen.build_3b_model,
 }
 
+_BUILDER_V3 = {
+    "0.6b": qwen3.build_0_6b_model,
+    "1.7b": qwen3.build_1_7b_model,
+    "4b": qwen3.build_4b_model,
+}
+
+_BUILDER = {
+    "v2": _BUILDER_V2,
+    "v3": _BUILDER_V3,
+}
+
 DEFAULT_PROMPTS = ["What is the meaning of life?"]
 
 
 def verify_qwen(
     model_size: str,
+    model_version: str,
     checkpoint_dir: str,
     weight_filename: str = "model.safetensors",
     max_new_tokens: int = 30,
@@ -64,7 +76,7 @@ def verify_qwen(
     reauthored_checkpoint = os.path.join(checkpoint_dir, weight_filename)
 
   logging.info("Building the reauthored model from: %s", reauthored_checkpoint)
-  reauthored_model = _BUILDER[model_size](
+  reauthored_model = _BUILDER[model_version][model_size](
       checkpoint_path=reauthored_checkpoint,
       custom_loader=custom_loader,
   )

ai_edge_torch/generative/utilities/transformers_verifier.py

@@ -15,8 +15,6 @@
 
 """Utilities for the models predefined in HuggingFace transformers."""
 
-from typing import cast
-
 from ai_edge_torch.generative.utilities import verifier
 import torch
 import transformers
@@ -39,4 +37,8 @@ class TransformersModelWrapper(verifier.ModelWrapper):
       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)
+    # Do not override GenerationConfig with model defaults. Always keep greedy
+    # sampling.
+    return self.model.generate(
+        inputs=inputs, generation_config=gen_config, use_model_defaults=False
+    )

ai_edge_torch/version.py

@@ -13,4 +13,4 @@
 # limitations under the License.
 # ==============================================================================
 
-__version__ = "0.6.0.dev20250521"
+__version__ = "0.6.0.dev20250522"

{ai_edge_torch_nightly-0.6.0.dev20250521.dist-info → ai_edge_torch_nightly-0.6.0.dev20250522.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.6.0.dev20250521
+Version: 0.6.0.dev20250522
 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

{ai_edge_torch_nightly-0.6.0.dev20250521.dist-info → ai_edge_torch_nightly-0.6.0.dev20250522.dist-info}/RECORD

@@ -2,7 +2,7 @@ ai_edge_torch/__init__.py,sha256=lemyLCNoGYRnJsmDuGZu7qOqLbLqG6CGDFtu3ue1syU,129
 ai_edge_torch/_config.py,sha256=AiqhbcheF7j_ozIGDLC89k1we95aVgFDa-tR6h7UI0s,2529
 ai_edge_torch/conftest.py,sha256=r0GTrhMRhlmOGrrkvumHN8hkmyug6WvF60vWq8wRIBI,758
 ai_edge_torch/model.py,sha256=wxjSFq_rBSxSqbUE8E8EJTCkgvgaRLjq_ZuAM-IZpCU,5606
-ai_edge_torch/version.py,sha256=lmyCstaeVZjTAbBP4s9Z02tpX00ynyLPsymBY2tCe4A,706
+ai_edge_torch/version.py,sha256=khTRYaBBm9J5iM-bJK8PgD2F4rAYy3w4cGEKRbQ498M,706
 ai_edge_torch/_convert/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/_convert/conversion.py,sha256=iQk3R-pLq4c1nfLqPB4xTRj78gghxPGzJCJtILLdg5o,6123
 ai_edge_torch/_convert/conversion_utils.py,sha256=Sr8qXVcTwc-ZnZmK7yxVrIOOp1S_vNrwzC0zUvLTI2o,2160
@@ -118,9 +118,12 @@ ai_edge_torch/generative/examples/phi/verify_phi4.py,sha256=2MlgQrfRkhE7Dya8MIix
 ai_edge_torch/generative/examples/phi/verify_util.py,sha256=kRREOMSikn_BRbTDkQiXBllPZwmWHa9KUk-kK5lCkbU,2945
 ai_edge_torch/generative/examples/qwen/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/qwen/convert_to_tflite.py,sha256=TnzyARHQgmWeOdYsV9WpRj5vhKGBH0kAbp3tMj8ZCYw,1998
+ai_edge_torch/generative/examples/qwen/convert_v3_to_tflite.py,sha256=GVV8CVj3rdgt_ZTOlpLSa6AD1pMMpMnZEuowzN2AIGM,2004
 ai_edge_torch/generative/examples/qwen/qwen.py,sha256=EcIHVeBcJLc290TiPkPfE7jdG_VXZYKlVGf0XQXzqo8,4554
-ai_edge_torch/generative/examples/qwen/verify.py,sha256=mP1SIAX2B1vFO02vRkAZC0UCyvBBxeWxK_456gG5a1s,1633
-ai_edge_torch/generative/examples/qwen/verify_util.py,sha256=jEmqYnOkOcQhOmHJrHsX0vdLq7JSahROvEBrG6n7tqg,2919
+ai_edge_torch/generative/examples/qwen/qwen3.py,sha256=g6aVHjnlPo4YhLjSdXxONaDcKT3fZOh8cewlvf3cfoQ,5554
+ai_edge_torch/generative/examples/qwen/verify_qwen2.py,sha256=ry-c2QesH-0KnrSQygfjUFs6d4kOFvJz2ts_8mP156I,1659
+ai_edge_torch/generative/examples/qwen/verify_qwen3.py,sha256=hmE0gdyzgcDpEDcWiwOzKQcxt4XeAe9DPRspy_I-lc8,1628
+ai_edge_torch/generative/examples/qwen/verify_util.py,sha256=vPROwLRABTChMGo5yWJkZURXP6TKWgh5FJj1Z3Zs6HU,3153
 ai_edge_torch/generative/examples/qwen_vl/__init__.py,sha256=JaAnrFoXTl3RJX97XspklkTyqOHVyAgRJsZtzNDd10c,671
 ai_edge_torch/generative/examples/qwen_vl/convert_to_tflite.py,sha256=BM-ed7KrmPwzI3MvDs2R7P-kJgE1SK_cNVqIfXhtJjs,2411
 ai_edge_torch/generative/examples/qwen_vl/decoder.py,sha256=plOi-3LltxReW_HVxhxwee_rYCQq-gsOwbGZtRsM8N8,4443
@@ -212,7 +215,7 @@ ai_edge_torch/generative/utilities/model_builder.py,sha256=tBfOcsI_NcneggHqkCSyd
 ai_edge_torch/generative/utilities/moonshine_loader.py,sha256=_RpFabSqtGH5PHiP3_1f6QfO14qMADUxr_HGRlVDFB0,4891
 ai_edge_torch/generative/utilities/stable_diffusion_loader.py,sha256=dqPD9qRXEWtU3ombslOC-BE2l_dMwHoCNu7NsIJhsso,36158
 ai_edge_torch/generative/utilities/t5_loader.py,sha256=tEsfy8-ymzbbjOIc-oesXF3yGyyWtJgFXn2s7VOavt8,16961
-ai_edge_torch/generative/utilities/transformers_verifier.py,sha256=8sp9m_FMcXn7nqOrochtu2jIANkJKhnhIBUmH0ZTDR4,1549
+ai_edge_torch/generative/utilities/transformers_verifier.py,sha256=l54bmmhj613eB2oCoONIAKEHhf8TQOhC9Gwjp6lxHAE,1659
 ai_edge_torch/generative/utilities/types.py,sha256=gZI9hIPB3XAo4oecKIIoVDfiyibLaSNFhecPFx4VDTM,2913
 ai_edge_torch/generative/utilities/verifier.py,sha256=ETO2ShU5KXG7MLP8eVOWuzuRLCUtapafYHcZ6TZHIkw,13061
 ai_edge_torch/hlfb/__init__.py,sha256=sH4um75na-O8tzxN6chFyp6Y4xnexsE7kUQpZySv6dE,735
@@ -264,8 +267,8 @@ ai_edge_torch/testing/__init__.py,sha256=_yGgvnBZWb7T3IN3mc4x1sS4vM96HZwM8pwIcPG
 ai_edge_torch/testing/export.py,sha256=k5mGDGzwc23Z4zaIVDs8CNh-oOt64gsf9MS9NjhbPy4,3293
 ai_edge_torch/testing/model_coverage/__init__.py,sha256=5P8J6Zk5YYtDvTBucFvB9NGSRI7Gw_24WnrbhXgycEE,765
 ai_edge_torch/testing/model_coverage/model_coverage.py,sha256=UPB448aMDUyC0HNYVqio2rcJPnDN0tBQMP08J6vPYew,4718
-ai_edge_torch_nightly-0.6.0.dev20250521.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
-ai_edge_torch_nightly-0.6.0.dev20250521.dist-info/METADATA,sha256=_UC8q7Xe3xMUCwKKbF4CJ5hewK9PLIJ26ksKCAeWjik,2074
-ai_edge_torch_nightly-0.6.0.dev20250521.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
-ai_edge_torch_nightly-0.6.0.dev20250521.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
-ai_edge_torch_nightly-0.6.0.dev20250521.dist-info/RECORD,,
+ai_edge_torch_nightly-0.6.0.dev20250522.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.6.0.dev20250522.dist-info/METADATA,sha256=4AO6S0O6AyT4GmWBjgWBjxRvbwD34vv4IoRR9kZW7V8,2074
+ai_edge_torch_nightly-0.6.0.dev20250522.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
+ai_edge_torch_nightly-0.6.0.dev20250522.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.6.0.dev20250522.dist-info/RECORD,,