ai-edge-torch-nightly 0.7.0.dev20251006__py3-none-any.whl → 0.7.0.dev20251008__py3-none-any.whl

ai_edge_torch/generative/utilities/converter.py

@@ -143,9 +143,23 @@ def define_conversion_flags(
       '`prefill_seq_lens` as the maximum of kv_cache size and prefill lengths '
       'in the graph.',
   )
+  flags.DEFINE_bool(
+      'export_gpu_dynamic_shape_verifications',
+      False,
+      'If true, the conversion script will export signatures used only for '
+      'verification of GPU dynamic shapes.',
+  )
   return flags
 
 
+# Context length for verifying GPU dynamic shapes.
+_CONTEXT_LENGTH_TO_VERIFY_MAGIC_NUMBERS = 1280
+# Long prefill length for verifying GPU dynamic shapes.
+_LONG_PREFILL_LENGTH_TO_VERIFY_MAGIC_NUMBERS = 1024
+# Short prefill length for verifying GPU dynamic shapes.
+_SHORT_PREFILL_LENGTH_TO_VERIFY_MAGIC_NUMBERS = 64
+
+
 def is_magic_number_(num: int) -> bool:
   """Returns true if the number is a magic number, i.e. prime number > 10."""
   if num < 10:
@@ -263,6 +277,10 @@ def convert_to_tflite(
     config: cfg.ModelConfig = None,
     lora_ranks: Optional[list[int]] = None,
     export_config: ExportConfig = None,
+    extra_model: torch.nn.Module = None,
+    extra_prefill_seq_lens: list[int] = None,
+    extra_kv_cache_max_len: int = 0,
+    extra_signature_prefix: str = '',
 ):
   """Converts a nn.Module model to multi-signature tflite model.
 
@@ -315,6 +333,15 @@ def convert_to_tflite(
         no LoRA signatures will be added.
       export_config (ExportConfig, optional): The export configuration. If None,
         it uses the default export configuration.
+      extra_model (torch.nn.Module, optional): PyTorch model to export in
+        addition to the pytorch_model. This model can have different
+        prefill_seq_lens and kv_cache_max_len.
+      extra_prefill_seq_lens (list[int], optional): The prefill sequence
+        lengths for extra_model. Meaningful only when extra_model is not None.
+      extra_kv_cache_max_len (int, optional): The maximum size of KV cache
+        buffer for extra_model. Meaningful only when extra_model is not None.
+      extra_signature_prefix (str, optional): The prefix of the extra model
+        signatures. Meaningful only when extra_model is not None.
   """
   # pylint: disable=protected-access
   torch._dynamo.config.cache_size_limit = 64
@@ -353,32 +380,51 @@ def convert_to_tflite(
   )
   output_file = os.path.join(output_path, output_filename)
 
-  _export_helper(
+  converter = converter_utils.Converter()
+  _add_signatures(
+      converter,
       pytorch_model,
-      output_file,
       prefill_seq_lens,
       kv_cache_max_len,
       pixel_values_size,
       pixel_seq_len,
-      quantize,
       config,
       loras,
       export_config,
   )
+
+  if extra_model is not None and extra_prefill_seq_lens:
+    _add_signatures(
+        converter,
+        extra_model,
+        extra_prefill_seq_lens,
+        extra_kv_cache_max_len,
+        pixel_values_size,
+        pixel_seq_len,
+        config,
+        loras,
+        export_config,
+        signature_prefix=extra_signature_prefix,
+    )
+
+  edge_model = converter.convert(
+      quant_config=get_quant_recipe_from_flag(quantize, config),
+  )
+  edge_model.export(output_file)
   return output_file
 
 
-def _export_helper(
+def _add_signatures(
+    converter: converter_utils.Converter,
     pytorch_model: torch.nn.Module,
-    output_file: str,
     prefill_seq_lens: list[int],
     kv_cache_max_len: int,
     pixel_values_size: torch.Size,
     pixel_seq_len: int,
-    quantize: str,
     config: cfg.ModelConfig,
     loras: list[None | lora_utils.LoRA],
     export_config: ExportConfig,
+    signature_prefix: str = '',
 ):
   """Helper function to export a model to tflite."""
   prefill_tokens_list = []
@@ -423,17 +469,14 @@ def _export_helper(
       kv_layout=export_config.kvcache_layout,
   )
 
-  quant_config = get_quant_recipe_from_flag(quantize, config)
-
   # For export, we create a module that captures any non-exportable,
   # arugments, e.g. the generation config object.
   mod = ExportableModule(pytorch_model, export_config=export_config).eval()
 
-  converter = converter_utils.Converter()
   for lora in loras:
     for i in range(len(prefill_seq_lens)):
       prefill_seq_len = prefill_seq_lens[i]
-      prefill_signature_name = f'prefill_{prefill_seq_len}'
+      prefill_signature_name = f'{signature_prefix}prefill_{prefill_seq_len}'
 
       sample_kwargs = {
           'tokens': prefill_tokens_list[i],
@@ -488,17 +531,15 @@ def _export_helper(
     if lora is not None:
       sample_kwargs['lora'] = lora
 
+    decode_signature_name = f'{signature_prefix}decode'
+    if lora is not None:
+      decode_signature_name += f'_lora_r{lora.get_rank()}'
     converter.add_signature(
-        'decode' if lora is None else f'decode_lora_r{lora.get_rank()}',
+        decode_signature_name,
         mod,
         sample_kwargs=sample_kwargs,
     )
 
-  edge_model = converter.convert(
-      quant_config=quant_config,
-  )
-  edge_model.export(output_file)
-
 
 def build_and_convert_to_tflite_from_flags(
     model_builder: Callable[
@@ -521,11 +562,36 @@ def build_and_convert_to_tflite_from_flags(
       get_mask_cache_size_from_flags(),
   )
 
+  # Extra model for GPU dynamic shape verification if needed.
+  extra_model = None
+  extra_prefill_seq_lens = None
+  extra_kv_cache_max_len = 0
   if flags.FLAGS.gpu_dynamic_shapes:
     prefill_seq_lens = [
         get_magic_number_for(l) for l in flags.FLAGS.prefill_seq_lens
     ]
     kv_cache_max_len = get_magic_number_for(flags.FLAGS.kv_cache_max_len)
+
+    if flags.FLAGS.export_gpu_dynamic_shape_verifications:
+      extra_kv_cache_max_len = _CONTEXT_LENGTH_TO_VERIFY_MAGIC_NUMBERS
+      if extra_kv_cache_max_len > flags.FLAGS.kv_cache_max_len:
+        extra_kv_cache_max_len = flags.FLAGS.kv_cache_max_len
+      extra_model = model_builder(
+          checkpoint_path,
+          loader.maybe_get_custom_loader(
+              checkpoint_path, flags.FLAGS.custom_checkpoint_loader
+          ),
+          extra_kv_cache_max_len,
+      )
+      extra_prefill_seq_lens = []
+      if extra_kv_cache_max_len > _SHORT_PREFILL_LENGTH_TO_VERIFY_MAGIC_NUMBERS:
+        extra_prefill_seq_lens.append(
+            _SHORT_PREFILL_LENGTH_TO_VERIFY_MAGIC_NUMBERS
+        )
+      if extra_kv_cache_max_len > _LONG_PREFILL_LENGTH_TO_VERIFY_MAGIC_NUMBERS:
+        extra_prefill_seq_lens.append(
+            _LONG_PREFILL_LENGTH_TO_VERIFY_MAGIC_NUMBERS
+        )
   else:
     prefill_seq_lens = flags.FLAGS.prefill_seq_lens
     kv_cache_max_len = flags.FLAGS.kv_cache_max_len
@@ -539,6 +605,10 @@ def build_and_convert_to_tflite_from_flags(
       quantize=flags.FLAGS.quantize,
       lora_ranks=flags.FLAGS.lora_ranks,
       export_config=export_config_lib.get_from_flags(),
+      extra_model=extra_model,
+      extra_prefill_seq_lens=extra_prefill_seq_lens,
+      extra_kv_cache_max_len=extra_kv_cache_max_len,
+      extra_signature_prefix='test_' if extra_model is not None else '',
   )
 
 

ai_edge_torch/version.py

@@ -15,4 +15,4 @@
 
 # The next version of ai-edge-torch.
 # The minor version code should be bumped after every release.
-__version__ = "0.7.0.dev20251006"
+__version__ = "0.7.0.dev20251008"

{ai_edge_torch_nightly-0.7.0.dev20251006.dist-info → ai_edge_torch_nightly-0.7.0.dev20251008.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.7.0.dev20251006
+Version: 0.7.0.dev20251008
 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-0.7.0.dev20251006.dist-info → ai_edge_torch_nightly-0.7.0.dev20251008.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=A7loFu8jE9CsXsfMmHYZ-KDFJiaD8Kkqwm_9d3IVzk0,5638
-ai_edge_torch/version.py,sha256=NWXVGUPY6DtGiDOywYxBG91TAP5aYrDt8Ayzv2kwLhs,806
+ai_edge_torch/version.py,sha256=Jd2ZmbryaZTSc314Yj8KLDdZImrRPNAWsBVxJ18z8dk,806
 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
@@ -208,7 +208,7 @@ ai_edge_torch/generative/test/test_model_conversion_large.py,sha256=NkEwrjO8vIcd
 ai_edge_torch/generative/test/test_quantize.py,sha256=kKJ01wscTC2t_Ylr7huO5gNKES01gm3dT1gx52z15PA,7356
 ai_edge_torch/generative/test/utils.py,sha256=tF6aCfAGJnc9dmzCnZCEOuKNVimfWOqscv9og0DDLHU,2656
 ai_edge_torch/generative/utilities/__init__.py,sha256=-_jxnnFnCgnTU4oTm4MnRsvL5lqhomBNdFBbqfmfHPo,720
-ai_edge_torch/generative/utilities/converter.py,sha256=Bt-48O1wf-7YcGVof53eVKI7wJwNvrc1Bv5zE3JuFdk,20093
+ai_edge_torch/generative/utilities/converter.py,sha256=d8pehTq6EzEdVR8ioL2b1ECGTR4G1K1fczc9amu_Oyk,23106
 ai_edge_torch/generative/utilities/export_config.py,sha256=5B15nYyqf96kjjYlHfPctUfsIdsBsh1f8rxKitJpwKQ,2384
 ai_edge_torch/generative/utilities/litertlm_builder.py,sha256=0cNuaqhc7cQcAa4NRalUXyoPQUQC9O3-aHAJEDV1Mps,4265
 ai_edge_torch/generative/utilities/loader.py,sha256=drgKBmNibuc3PCdc0kU0pVcp2Nt1_mjLYh67RyXOn7U,15952
@@ -270,8 +270,8 @@ 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-0.7.0.dev20251006.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
-ai_edge_torch_nightly-0.7.0.dev20251006.dist-info/METADATA,sha256=9-6TbSzTZaYh69AVVx-R794sgQjC01l9C0jLlWmTNOg,2074
-ai_edge_torch_nightly-0.7.0.dev20251006.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
-ai_edge_torch_nightly-0.7.0.dev20251006.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
-ai_edge_torch_nightly-0.7.0.dev20251006.dist-info/RECORD,,
+ai_edge_torch_nightly-0.7.0.dev20251008.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.7.0.dev20251008.dist-info/METADATA,sha256=Xg4GCLMHL1FhKweyTtY2OAUdPGHFwpGlpZw-bUvs3FY,2074
+ai_edge_torch_nightly-0.7.0.dev20251008.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
+ai_edge_torch_nightly-0.7.0.dev20251008.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.7.0.dev20251008.dist-info/RECORD,,