BREAKING_CHANGE: switch aimet_torch default API from v1 to v2 (#3689)

Switched the default API of aimet_torch from v1 to v2.

From this commit onwards, any package or module named ``aimet_torch.<X>`` will point to ``aimet_torch.v2.<X>`` (if any).
For example,  ``aimet_torch.quantsim`` will now point to ``aimet_torch.v2.quantsim``, not ``aimet_torch.v1.quantsim``
Also, any v1-specific modules/packages will be moved under ``aimet_torch.v1`` subpackage, such as ``aimet_torch.v1.qc_quantize_op`` or ``aimet_torch.v1.tensor_quantizer``

To keep using ``aimet_torch.v1`` as the default API for backwards-compatibility with the legacy code, set environment variable "AIMET_DEFAULT_API=v1"

Signed-off-by: Kyunggeun Lee <[email protected]>

TrainingExtensions/torch/src/python/aimet_torch/utils.py

@@ -1307,7 +1307,7 @@ def _get_metadata_and_state_dict(safetensor_file_path: str) -> [dict, dict]:
 
 
 def _get_default_api() -> Union[Literal["v1"], Literal["v2"]]:
-    default_api = os.getenv("AIMET_DEFAULT_API", "v1").lower()
+    default_api = os.getenv("AIMET_DEFAULT_API", "v2").lower()
 
     if default_api not in ("v1", "v2"):
         raise RuntimeError("Invalid value specified for environment variable AIMET_DEFAULT_API. "

TrainingExtensions/torch/test/python/test_import.py

@@ -43,104 +43,124 @@
 import pytest
 
 import aimet_torch
+from aimet_torch.utils import _get_default_api
 
 
 def test_default_import():
+    if _get_default_api() == "v1":
+        from aimet_torch.v1 import quantsim as default_quantsim
+        from aimet_torch.v1.quantsim import QuantizationSimModel as default_QuantizationSimModel
+        from aimet_torch.v1.adaround import adaround_weight as default_adaround_weight
+        from aimet_torch.v1.adaround.adaround_weight import Adaround as default_Adaround
+        from aimet_torch.v1 import seq_mse as default_seq_mse
+        from aimet_torch.v1.seq_mse import apply_seq_mse as default_apply_seq_mse
+        from aimet_torch.v1.nn.modules import custom as default_custom
+        from aimet_torch.v1.nn.modules.custom import Add as default_Add
+        from aimet_torch.v1 import auto_quant as default_auto_quant
+        from aimet_torch.v1.auto_quant import AutoQuant as default_AutoQuant
+        from aimet_torch.v1 import quant_analyzer as default_quant_analyzer
+        from aimet_torch.v1.quant_analyzer import QuantAnalyzer as default_QuantAnalyzer
+        from aimet_torch.v1 import batch_norm_fold as default_batch_norm_fold
+        from aimet_torch.v1.batch_norm_fold import fold_all_batch_norms_to_scale as default_fold_all_batch_norms_to_scale
+        from aimet_torch.v1 import mixed_precision as default_mixed_precision
+        from aimet_torch.v1.mixed_precision import choose_mixed_precision as default_choose_mixed_precision
+    else:
+        from aimet_torch.v2 import quantsim as default_quantsim
+        from aimet_torch.v2.quantsim import QuantizationSimModel as default_QuantizationSimModel
+        from aimet_torch.v2.adaround import adaround_weight as default_adaround_weight
+        from aimet_torch.v2.adaround.adaround_weight import Adaround as default_Adaround
+        from aimet_torch.v2 import seq_mse as default_seq_mse
+        from aimet_torch.v2.seq_mse import apply_seq_mse as default_apply_seq_mse
+        from aimet_torch.v2.nn.modules import custom as default_custom
+        from aimet_torch.v2.nn.modules.custom import Add as default_Add
+        from aimet_torch.v2 import auto_quant as default_auto_quant
+        from aimet_torch.v2.auto_quant import AutoQuant as default_AutoQuant
+        from aimet_torch.v2 import quant_analyzer as default_quant_analyzer
+        from aimet_torch.v2.quant_analyzer import QuantAnalyzer as default_QuantAnalyzer
+        from aimet_torch.v2 import batch_norm_fold as default_batch_norm_fold
+        from aimet_torch.v2.batch_norm_fold import fold_all_batch_norms_to_scale as default_fold_all_batch_norms_to_scale
+        from aimet_torch.v2 import mixed_precision as default_mixed_precision
+        from aimet_torch.v2.mixed_precision import choose_mixed_precision as default_choose_mixed_precision
+
     """
     When: Import from aimet_torch.quantsim
-    Then: Import should be redirected to aimet_torch.v1.quantsim
+    Then: Import should be redirected to aimet_torch.v1 or aimet_torch.v2.quantsim
     """
-    from aimet_torch    import quantsim
-    from aimet_torch.v1 import quantsim as v1_quantsim
-    assert quantsim.QuantizationSimModel is v1_quantsim.QuantizationSimModel
+    from aimet_torch import quantsim
+    assert quantsim.QuantizationSimModel is default_quantsim.QuantizationSimModel
 
-    from aimet_torch.quantsim    import QuantizationSimModel
-    from aimet_torch.v1.quantsim import QuantizationSimModel as v1_QuantizationSimModel
-    assert QuantizationSimModel is v1_QuantizationSimModel
+    from aimet_torch.quantsim import QuantizationSimModel
+    assert QuantizationSimModel is default_QuantizationSimModel
 
     """
     When: Import from aimet_torch.adaround
-    Then: Import should be redirected to aimet_torch.v1.adaround
+    Then: Import should be redirected to aimet_torch.v1 or aimet_torch.v2.adaround
     """
-    from aimet_torch.adaround    import adaround_weight
-    from aimet_torch.v1.adaround import adaround_weight as v1_adaround_weight
-    assert adaround_weight.Adaround is v1_adaround_weight.Adaround
+    from aimet_torch.adaround import adaround_weight
+    assert adaround_weight.Adaround is default_adaround_weight.Adaround
 
-    from aimet_torch.adaround.adaround_weight    import Adaround
-    from aimet_torch.v1.adaround.adaround_weight import Adaround as v1_Adaround
-    assert Adaround is v1_Adaround
+    from aimet_torch.adaround.adaround_weight import Adaround
+    assert Adaround is default_Adaround
 
     """
     When: Import from aimet_torch.seq_mse
-    Then: Import should be redirected to aimet_torch.v1.seq_mse
+    Then: Import should be redirected to aimet_torch.v1 or aimet_torch.v2.seq_mse
     """
-    from aimet_torch    import seq_mse
-    from aimet_torch.v1 import seq_mse as v1_seq_mse
-    assert seq_mse.apply_seq_mse is v1_seq_mse.apply_seq_mse
+    from aimet_torch import seq_mse
+    assert seq_mse.apply_seq_mse is default_seq_mse.apply_seq_mse
 
-    from aimet_torch.seq_mse    import apply_seq_mse
-    from aimet_torch.v1.seq_mse import apply_seq_mse as v1_apply_seq_mse
-    assert apply_seq_mse is v1_apply_seq_mse
+    from aimet_torch.seq_mse import apply_seq_mse
+    assert apply_seq_mse is default_apply_seq_mse
 
     """
     When: Import from aimet_torch.nn
-    Then: Import should be redirected to aimet_torch.v1.nn
+    Then: Import should be redirected to aimet_torch.v1 or aimet_torch.v2.nn
     """
-    from aimet_torch.nn.modules    import custom
-    from aimet_torch.v1.nn.modules import custom as v1_custom
-    assert custom.Add is v1_custom.Add
+    from aimet_torch.nn.modules import custom
+    assert custom.Add is default_custom.Add
 
-    from aimet_torch.nn.modules.custom    import Add
-    from aimet_torch.v1.nn.modules.custom import Add as v1_Add
-    assert Add is v1_Add
+    from aimet_torch.nn.modules.custom import Add
+    assert Add is default_Add
 
     """
     When: Import from aimet_torch.auto_quant
-    Then: Import should be redirected to aimet_torch.v1.auto_quant
+    Then: Import should be redirected to aimet_torch.v1 or aimet_torch.v2.auto_quant
     """
-    from aimet_torch    import auto_quant
-    from aimet_torch.v1 import auto_quant as v1_auto_quant
-    assert auto_quant.AutoQuant is v1_auto_quant.AutoQuant
+    from aimet_torch import auto_quant
+    assert auto_quant.AutoQuant is default_auto_quant.AutoQuant
 
-    from aimet_torch.auto_quant    import AutoQuant
-    from aimet_torch.v1.auto_quant import AutoQuant as v1_AutoQuant
-    assert AutoQuant is v1_AutoQuant
+    from aimet_torch.auto_quant import AutoQuant
+    assert AutoQuant is default_AutoQuant
 
     """
     When: Import from aimet_torch.quant_analyzer
-    Then: Import should be redirected to aimet_torch.v1.quant_analyzer
+    Then: Import should be redirected to aimet_torch.v1 or aimet_torch.v2.quant_analyzer
     """
-    from aimet_torch    import quant_analyzer
-    from aimet_torch.v1 import quant_analyzer as v1_auto_quant
-    assert quant_analyzer.QuantAnalyzer is v1_auto_quant.QuantAnalyzer
+    from aimet_torch import quant_analyzer
+    assert quant_analyzer.QuantAnalyzer is default_quant_analyzer.QuantAnalyzer
 
-    from aimet_torch.quant_analyzer    import QuantAnalyzer
-    from aimet_torch.v1.quant_analyzer import QuantAnalyzer as v1_QuantAnalyzer
-    assert QuantAnalyzer is v1_QuantAnalyzer
+    from aimet_torch.quant_analyzer import QuantAnalyzer
+    assert QuantAnalyzer is default_QuantAnalyzer
 
     """
     When: Import from aimet_torch.batch_norm_fold
-    Then: Import should be redirected to aimet_torch.v1.batch_norm_fold
+    Then: Import should be redirected to aimet_torch.v1 or aimet_torch.v2.batch_norm_fold
     """
-    from aimet_torch    import batch_norm_fold
-    from aimet_torch.v1 import batch_norm_fold as v1_batch_norm_fold
-    assert batch_norm_fold.fold_all_batch_norms_to_scale is v1_batch_norm_fold.fold_all_batch_norms_to_scale
+    from aimet_torch import batch_norm_fold
+    assert batch_norm_fold.fold_all_batch_norms_to_scale is default_batch_norm_fold.fold_all_batch_norms_to_scale
 
-    from aimet_torch.batch_norm_fold    import fold_all_batch_norms_to_scale
-    from aimet_torch.v1.batch_norm_fold import fold_all_batch_norms_to_scale as v1_fold_all_batch_norms_to_scale
-    assert fold_all_batch_norms_to_scale is v1_fold_all_batch_norms_to_scale
+    from aimet_torch.batch_norm_fold import fold_all_batch_norms_to_scale
+    assert fold_all_batch_norms_to_scale is default_fold_all_batch_norms_to_scale
 
     """
     When: Import from aimet_torch.mixed_precision
-    Then: Import should be redirected to aimet_torch.v1.mixed_precision
+    Then: Import should be redirected to aimet_torch.v1 or aimet_torch.v2.mixed_precision
     """
-    from aimet_torch    import mixed_precision
-    from aimet_torch.v1 import mixed_precision as v1_mixed_precision
-    assert mixed_precision.choose_mixed_precision is v1_mixed_precision.choose_mixed_precision
+    from aimet_torch import mixed_precision
+    assert mixed_precision.choose_mixed_precision is default_mixed_precision.choose_mixed_precision
 
-    from aimet_torch.mixed_precision    import choose_mixed_precision
-    from aimet_torch.v1.mixed_precision import choose_mixed_precision as v1_choose_mixed_precision
-    assert choose_mixed_precision is v1_choose_mixed_precision
+    from aimet_torch.mixed_precision import choose_mixed_precision
+    assert choose_mixed_precision is default_choose_mixed_precision
 
 
 def _get_all_modules():

TrainingExtensions/torch/test/python/v1/test_auto_quant_with_amp.py

@@ -54,7 +54,7 @@
 from aimet_torch.amp.mixed_precision_algo import GreedyMixedPrecisionAlgo
 from aimet_common.defs import QuantizationDataType
 from aimet_torch import utils
-from aimet_torch.quantsim import QuantizationSimModel
+from aimet_torch.v1.quantsim import QuantizationSimModel
 from aimet_torch.save_utils import SaveUtils
 
 

TrainingExtensions/torch/test/python/v1/test_bn_fold.py

@@ -43,7 +43,7 @@
 import torch
 from torchvision import models
 
-from aimet_torch.batch_norm_fold import fold_given_batch_norms
+from aimet_torch.v1.batch_norm_fold import fold_given_batch_norms
 from ..models.test_models import TransposedConvModel
 from aimet_torch.model_preparer import prepare_model
 from aimet_common.defs import QuantScheme