ai-edge-quantizer-nightly 0.3.0.dev20250612__py3-none-any.whl → 0.3.0.dev20250614__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- ai_edge_quantizer/algorithm_manager.py +2 -0
- ai_edge_quantizer/algorithms/uniform_quantize/common_quantize.py +17 -0
- ai_edge_quantizer/calibrator.py +1 -1
- ai_edge_quantizer/default_policy.py +10 -8
- ai_edge_quantizer/params_generator.py +1 -1
- ai_edge_quantizer/qtyping.py +1 -0
- ai_edge_quantizer/utils/tfl_flatbuffer_utils.py +1 -0
- {ai_edge_quantizer_nightly-0.3.0.dev20250612.dist-info → ai_edge_quantizer_nightly-0.3.0.dev20250614.dist-info}/METADATA +1 -1
- {ai_edge_quantizer_nightly-0.3.0.dev20250612.dist-info → ai_edge_quantizer_nightly-0.3.0.dev20250614.dist-info}/RECORD +12 -12
- {ai_edge_quantizer_nightly-0.3.0.dev20250612.dist-info → ai_edge_quantizer_nightly-0.3.0.dev20250614.dist-info}/LICENSE +0 -0
- {ai_edge_quantizer_nightly-0.3.0.dev20250612.dist-info → ai_edge_quantizer_nightly-0.3.0.dev20250614.dist-info}/WHEEL +0 -0
- {ai_edge_quantizer_nightly-0.3.0.dev20250612.dist-info → ai_edge_quantizer_nightly-0.3.0.dev20250614.dist-info}/top_level.txt +0 -0
@@ -112,6 +112,7 @@ MIN_MAX_OP_NAME_MATERIALIZE_FUNC_DICT = {
|
|
112
112
|
common_quantize.materialize_squared_difference
|
113
113
|
),
|
114
114
|
_TFLOpName.MAX_POOL_2D: common_quantize.materialize_max_pool_2d,
|
115
|
+
_TFLOpName.RESIZE_BILINEAR: common_quantize.materialize_resize_bilinear,
|
115
116
|
}
|
116
117
|
for op_name, materialize_func in MIN_MAX_OP_NAME_MATERIALIZE_FUNC_DICT.items():
|
117
118
|
register_quantized_op(
|
@@ -250,6 +251,7 @@ _OCTAV_OP_NAME_MATERIALIZE_FUNC_DICT = immutabledict({
|
|
250
251
|
common_quantize.materialize_squared_difference
|
251
252
|
),
|
252
253
|
_TFLOpName.MAX_POOL_2D: common_quantize.materialize_max_pool_2d,
|
254
|
+
_TFLOpName.RESIZE_BILINEAR: common_quantize.materialize_resize_bilinear,
|
253
255
|
})
|
254
256
|
|
255
257
|
for op_name, materialize_func in _OCTAV_OP_NAME_MATERIALIZE_FUNC_DICT.items():
|
@@ -728,6 +728,23 @@ def materialize_max_pool_2d(
|
|
728
728
|
)
|
729
729
|
|
730
730
|
|
731
|
+
def materialize_resize_bilinear(
|
732
|
+
get_tensor_quant_params_fn: qtyping.GetTensorQuantParamsFuncSignature,
|
733
|
+
op_info: qtyping.OpInfo,
|
734
|
+
graph_info: qtyping.GraphInfo,
|
735
|
+
tensor_name_to_qsv: dict[str, Any],
|
736
|
+
) -> list[qtyping.TensorTransformationParams]:
|
737
|
+
"""Materialize tensors in tfl.resize_bilinear."""
|
738
|
+
return common_utils.materialize_standard_op(
|
739
|
+
op_info,
|
740
|
+
graph_info,
|
741
|
+
tensor_name_to_qsv,
|
742
|
+
get_tensor_quant_params_fn,
|
743
|
+
constraint=_OpQuantConstraint.SAME_AS_INPUT_SCALE,
|
744
|
+
inputs_to_ignore=[1], # Resize size does not need to be quantized.
|
745
|
+
)
|
746
|
+
|
747
|
+
|
731
748
|
def _get_tensor_shape_for_blockwise(
|
732
749
|
tensor_shape: Sequence[int], quantized_dim: int, block_size: int
|
733
750
|
) -> list[int]:
|
ai_edge_quantizer/calibrator.py
CHANGED
@@ -165,7 +165,7 @@ class Calibrator:
|
|
165
165
|
)
|
166
166
|
if algorithm_name == algorithm_manager.AlgorithmName.NO_QUANTIZE:
|
167
167
|
continue
|
168
|
-
if policy.
|
168
|
+
if policy.is_non_quantizable_composite_op(op):
|
169
169
|
continue
|
170
170
|
|
171
171
|
# Step2.2: query algorithm_manager to get/call the related
|
@@ -185,7 +185,8 @@ DEFAULT_JSON_POLICY = """
|
|
185
185
|
"SELECT_V2",
|
186
186
|
"STABLEHLO_COMPOSITE",
|
187
187
|
"PAD",
|
188
|
-
"MAX_POOL_2D"
|
188
|
+
"MAX_POOL_2D",
|
189
|
+
"RESIZE_BILINEAR"
|
189
190
|
],
|
190
191
|
"static_wi8_ai8": [
|
191
192
|
"ADD",
|
@@ -219,7 +220,8 @@ DEFAULT_JSON_POLICY = """
|
|
219
220
|
"STABLEHLO_COMPOSITE",
|
220
221
|
"PAD",
|
221
222
|
"SQUARED_DIFFERENCE",
|
222
|
-
"MAX_POOL_2D"
|
223
|
+
"MAX_POOL_2D",
|
224
|
+
"RESIZE_BILINEAR"
|
223
225
|
],
|
224
226
|
"static_wi4_ai8": ["FULLY_CONNECTED", "CONV_2D", "INPUT", "OUTPUT", "EMBEDDING_LOOKUP"],
|
225
227
|
"static_wi4_ai16": ["FULLY_CONNECTED", "CONV_2D", "INPUT", "OUTPUT", "EMBEDDING_LOOKUP"],
|
@@ -245,6 +247,7 @@ DEFAULT_JSON_POLICY = """
|
|
245
247
|
}
|
246
248
|
}
|
247
249
|
"""
|
250
|
+
QUANTIZABLE_COMPOSITES = ["od" + "ml.npu_call", "od" + "ml.rms_norm"]
|
248
251
|
|
249
252
|
|
250
253
|
def _unroll_json_config(
|
@@ -322,10 +325,10 @@ def _unroll_json_config(
|
|
322
325
|
|
323
326
|
|
324
327
|
# TODO: b/401024954 - Have a better way to specify recipes based on op options.
|
325
|
-
def
|
328
|
+
def is_non_quantizable_composite_op(
|
326
329
|
op: Union[schema.Operator, schema.OperatorT],
|
327
330
|
) -> bool:
|
328
|
-
"""Checks if the operator is
|
331
|
+
"""Checks if the operator is a non-quantizable composite op.
|
329
332
|
|
330
333
|
We may want to quantize an op only when its has certain options.
|
331
334
|
Policies/recipes
|
@@ -340,10 +343,9 @@ def is_conditionally_unquantized(
|
|
340
343
|
if opts := flatbuffer_utils.get_options_as(
|
341
344
|
op, schema.StableHLOCompositeOptionsT
|
342
345
|
):
|
343
|
-
name
|
344
|
-
|
345
|
-
|
346
|
-
return ("od" + "ml.npu_call") not in name.decode("utf-8")
|
346
|
+
name = opts.name.decode("utf-8")
|
347
|
+
if name not in QUANTIZABLE_COMPOSITES:
|
348
|
+
return True
|
347
349
|
|
348
350
|
return False
|
349
351
|
|
@@ -109,7 +109,7 @@ class ParamsGenerator:
|
|
109
109
|
algorithm_name, op_quant_config = (
|
110
110
|
model_recipe_manager.get_quantization_configs(op_key, op_scope)
|
111
111
|
)
|
112
|
-
if policy.
|
112
|
+
if policy.is_non_quantizable_composite_op(op):
|
113
113
|
algorithm_name = algorithm_manager.AlgorithmName.NO_QUANTIZE
|
114
114
|
|
115
115
|
if algorithm_name == algorithm_manager.AlgorithmName.NO_QUANTIZE:
|
ai_edge_quantizer/qtyping.py
CHANGED
@@ -59,6 +59,7 @@ TFL_OP_NAME_TO_CODE = immutabledict.immutabledict({
|
|
59
59
|
_TFLOpName.PAD: schema.BuiltinOperator.PAD,
|
60
60
|
_TFLOpName.SQUARED_DIFFERENCE: schema.BuiltinOperator.SQUARED_DIFFERENCE,
|
61
61
|
_TFLOpName.MAX_POOL_2D: schema.BuiltinOperator.MAX_POOL_2D,
|
62
|
+
_TFLOpName.RESIZE_BILINEAR: schema.BuiltinOperator.RESIZE_BILINEAR,
|
62
63
|
})
|
63
64
|
|
64
65
|
TFL_OP_CODE_TO_NAME = immutabledict.immutabledict(
|
@@ -1,6 +1,6 @@
|
|
1
1
|
Metadata-Version: 2.1
|
2
2
|
Name: ai-edge-quantizer-nightly
|
3
|
-
Version: 0.3.0.
|
3
|
+
Version: 0.3.0.dev20250614
|
4
4
|
Summary: A quantizer for advanced developers to quantize converted AI Edge models.
|
5
5
|
Home-page: https://github.com/google-ai-edge/ai-edge-quantizer
|
6
6
|
Keywords: On-Device ML,AI,Google,TFLite,Quantization,LLMs,GenAI
|
@@ -1,18 +1,18 @@
|
|
1
1
|
ai_edge_quantizer/__init__.py,sha256=4pFSkukSwahYyzwqia0yPRyz8TnFQfGRthVJhYpMWas,793
|
2
|
-
ai_edge_quantizer/algorithm_manager.py,sha256=
|
2
|
+
ai_edge_quantizer/algorithm_manager.py,sha256=rMTM89YDPkmLKlUQV_Rjr7B2KpcvldAHzfpgUqaOqdU,12216
|
3
3
|
ai_edge_quantizer/algorithm_manager_api.py,sha256=u903TG0s1uIDhJqfeJne3CFl8A93phZrwgV2-hwdcXU,9247
|
4
4
|
ai_edge_quantizer/algorithm_manager_api_test.py,sha256=w6bSONvXkX6bzXAGc0-7b6gNDt9oz9ieq97KP8Sg_JU,7666
|
5
|
-
ai_edge_quantizer/calibrator.py,sha256
|
5
|
+
ai_edge_quantizer/calibrator.py,sha256=Sms7_AIHPH9G5xFaz5Ef3a5gPhxuIWQI8d2LUM8C96I,12071
|
6
6
|
ai_edge_quantizer/calibrator_test.py,sha256=C_oWOaRugPKYX74jF-eRFH-k6nGOdA8I9_uPiocaOuE,11900
|
7
7
|
ai_edge_quantizer/conftest.py,sha256=SxCz-5LlRD_lQm4hQc4c6IGG7DS8d7IyEWY9gnscPN0,794
|
8
|
-
ai_edge_quantizer/default_policy.py,sha256=
|
8
|
+
ai_edge_quantizer/default_policy.py,sha256=zghBh9dTB-ouPFumV-0siBSnEbp0WxF6tGOsn3TLirg,11242
|
9
9
|
ai_edge_quantizer/model_modifier.py,sha256=teGa8I6kGvn6TQY6Xv53YFIc_pQEhNvM9Zb4bvhezyw,7110
|
10
10
|
ai_edge_quantizer/model_modifier_test.py,sha256=cJd04SLOG-fQZZNZPcisoBLx3cLtWEwGqUBbLb-pif4,4751
|
11
11
|
ai_edge_quantizer/model_validator.py,sha256=Hj0_5o-Oa3dSlJ3ryVjRhvsyelHNyek1GrtG9buMczg,13153
|
12
12
|
ai_edge_quantizer/model_validator_test.py,sha256=EeqOP_mrZsnZ3rug756s0ryDDqd2KgIDld5Lm_gDuWY,13020
|
13
|
-
ai_edge_quantizer/params_generator.py,sha256=
|
13
|
+
ai_edge_quantizer/params_generator.py,sha256=gC7G6Ne4Fumc8RSmIAbx96ZBhszZlHqBKSmE9p6RPTo,20099
|
14
14
|
ai_edge_quantizer/params_generator_test.py,sha256=RDYoRZDJfEZRtjlTAU2kZ_4t3JHOqEHxfJX9V4ETAhg,40597
|
15
|
-
ai_edge_quantizer/qtyping.py,sha256=
|
15
|
+
ai_edge_quantizer/qtyping.py,sha256=kX1AoD-YlHYbDI1RfGVXIbPn-CYT7HUF2x77-hPtKBM,16565
|
16
16
|
ai_edge_quantizer/quantizer.py,sha256=g3DMqFMrMpt9jQttCE0WcdNbMtk0JZnmN5MmCHrNdyM,13202
|
17
17
|
ai_edge_quantizer/quantizer_test.py,sha256=K_HBA56JkFI3HL8VLWCqGEfC0ISh5ldMKoNyBdGRAJg,20368
|
18
18
|
ai_edge_quantizer/recipe.py,sha256=FR0uJceumZrnle2VRSOQZ1uXup4S1cTYKRH-N53mWRo,2919
|
@@ -28,7 +28,7 @@ ai_edge_quantizer/algorithms/nonlinear_quantize/__init__.py,sha256=lpq1g2ayg3lCP
|
|
28
28
|
ai_edge_quantizer/algorithms/nonlinear_quantize/float_casting.py,sha256=Bs9CK7wZAw6jNaZ8xEtbwO2vM34VYXNZSMVWvxJo9nw,9297
|
29
29
|
ai_edge_quantizer/algorithms/nonlinear_quantize/float_casting_test.py,sha256=EqIHGEZ1LgUrTN7zf880RuAzEv3Qy7kgh5ivObJGHSo,22646
|
30
30
|
ai_edge_quantizer/algorithms/uniform_quantize/__init__.py,sha256=lpq1g2ayg3lCPLy79t2VicYcnGKw64FfYIj1V7J-4m8,676
|
31
|
-
ai_edge_quantizer/algorithms/uniform_quantize/common_quantize.py,sha256=
|
31
|
+
ai_edge_quantizer/algorithms/uniform_quantize/common_quantize.py,sha256=rImKK2ax7LrRx6XurSdvRTk0h6WtFGtQn9sYNJcn-uw,30222
|
32
32
|
ai_edge_quantizer/algorithms/uniform_quantize/common_quantize_test.py,sha256=GGf_n3wIeg3GB_eGsmyNJ0fTcxgpeMMbugTMRONK6TQ,3553
|
33
33
|
ai_edge_quantizer/algorithms/uniform_quantize/dequantized_weight_recovery.py,sha256=BDdn_uBZakfHyzdMJPKadsOqxqyC-s6W2ZzFH99L4fE,8652
|
34
34
|
ai_edge_quantizer/algorithms/uniform_quantize/dequantized_weight_recovery_test.py,sha256=sT5eX5TLZEHTtPfnSkCPDlS0sQxlTFWbCsbvOuj--yY,8889
|
@@ -64,14 +64,14 @@ ai_edge_quantizer/utils/__init__.py,sha256=lpq1g2ayg3lCPLy79t2VicYcnGKw64FfYIj1V
|
|
64
64
|
ai_edge_quantizer/utils/calibration_utils.py,sha256=1Fj9MIO6aLZIRgyd4axvZN4S_O64nB_-Miu1WP664js,2536
|
65
65
|
ai_edge_quantizer/utils/calibration_utils_test.py,sha256=Z-AcdTieesWFKyKBb08ZXm4Mgu6cvJ4bg2-MJ7hLD10,2856
|
66
66
|
ai_edge_quantizer/utils/test_utils.py,sha256=Y2pdMvn1k4gmqDo3noJfzx3fJcDHX_1hcsP6oiIz65Y,8240
|
67
|
-
ai_edge_quantizer/utils/tfl_flatbuffer_utils.py,sha256=
|
67
|
+
ai_edge_quantizer/utils/tfl_flatbuffer_utils.py,sha256=pZv8FMWyjBSLN5MGJ2K_dZ6oqkJGbp9RI4CfnlPuPII,10830
|
68
68
|
ai_edge_quantizer/utils/tfl_flatbuffer_utils_test.py,sha256=K1SbK8q92qYVtiVj0I0GtugsPTkpIpEKv9zakvFV_Sc,8555
|
69
69
|
ai_edge_quantizer/utils/tfl_interpreter_utils.py,sha256=EtOv6cpKM_F0uv2bWuSXylYmTeXT6zUc182pw4sdYSI,13889
|
70
70
|
ai_edge_quantizer/utils/tfl_interpreter_utils_test.py,sha256=6fjkM-rycZ95L4yfvlr0TN6RlrhfPzxNUYrZaYO_F0A,12013
|
71
71
|
ai_edge_quantizer/utils/validation_utils.py,sha256=oYw33Sg547AqtGw-choPUJmp9SAKkV46J_ddqSsum2Q,3950
|
72
72
|
ai_edge_quantizer/utils/validation_utils_test.py,sha256=V_qNDikPD4OPB-siOLQCWNVWTAu87h2IgNYt7teFd-o,2934
|
73
|
-
ai_edge_quantizer_nightly-0.3.0.
|
74
|
-
ai_edge_quantizer_nightly-0.3.0.
|
75
|
-
ai_edge_quantizer_nightly-0.3.0.
|
76
|
-
ai_edge_quantizer_nightly-0.3.0.
|
77
|
-
ai_edge_quantizer_nightly-0.3.0.
|
73
|
+
ai_edge_quantizer_nightly-0.3.0.dev20250614.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
|
74
|
+
ai_edge_quantizer_nightly-0.3.0.dev20250614.dist-info/METADATA,sha256=5ZPSscczc1tLmVN4sCf-xtX2qvmabAWOAkIjZVCb_7U,1528
|
75
|
+
ai_edge_quantizer_nightly-0.3.0.dev20250614.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
|
76
|
+
ai_edge_quantizer_nightly-0.3.0.dev20250614.dist-info/top_level.txt,sha256=8QTfPnFXNVUhScFLaa-NWZMFWMn72M50DVPubpwWB1g,18
|
77
|
+
ai_edge_quantizer_nightly-0.3.0.dev20250614.dist-info/RECORD,,
|
File without changes
|
File without changes
|