ai-edge-quantizer-nightly 0.1.0.dev20250327__py3-none-any.whl → 0.1.0.dev20250328__py3-none-any.whl

ai_edge_quantizer/params_generator.py

@@ -394,14 +394,6 @@ def _compatible_tensor_params(
   ]
   if _same_tensor_params_except_id(params1, params2):
     return True
-  if (
-      params1.transformations[0] != _QuantTrans.NO_QUANTIZE
-      and params2.transformations[0] != _QuantTrans.NO_QUANTIZE
-  ):
-    # NO_QUANTIZE has no parameters. So only if both params aren't NO_QUANTIZE
-    # do we expect the parameters to be the same.
-    if params1.parameters != params2.parameters:
-      return False
   # We only need to check the first transformation because transformations are
   # applied in order, and as long as the one that's immediately after the tensor
   # is the same, it's compatible.
@@ -413,6 +405,7 @@ def _compatible_tensor_params(
   if (
       params1.transformations[0] in quantized_source_transformations
       and params2.transformations[0] in quantized_source_transformations
+      and params1.parameters == params2.parameters
   ):
     return True
   return False

ai_edge_quantizer/params_generator_test.py

@@ -914,6 +914,58 @@ class ParamsGeneratorTest(parameterized.TestCase):
           ),
           expected=True,
       ),
+      dict(
+          testcase_name='compatible_no_numeric_check',
+          param1=qtyping.TensorTransformationParams(
+              tensor_name='tfl.quantize',
+              producer=None,
+              consumers=[
+                  qtyping.OpToTensorParams(
+                      subgraph_op_id=4,
+                      transformations=[
+                          qtyping.QuantTransformation.ADD_QUANTIZE,
+                      ],
+                      parameters=qtyping.UniformQuantParams(
+                          8, None, np.array([0.00028806]), np.array([0])
+                      ),
+                  ),
+                  qtyping.OpToTensorParams(
+                      subgraph_op_id=5,
+                      transformations=[
+                          qtyping.QuantTransformation.ADD_QUANTIZE,
+                      ],
+                      parameters=qtyping.UniformQuantParams(
+                          8, None, np.array([0.00027501]), np.array([0])
+                      ),
+                  ),
+              ],
+          ),
+          param2=qtyping.TensorTransformationParams(
+              tensor_name='tfl.quantize',
+              producer=None,
+              consumers=[
+                  qtyping.OpToTensorParams(
+                      subgraph_op_id=4,
+                      transformations=[
+                          qtyping.QuantTransformation.ADD_QUANTIZE,
+                      ],
+                      parameters=qtyping.UniformQuantParams(
+                          8, None, np.array([0.00028806]), np.array([0])
+                      ),
+                  ),
+                  qtyping.OpToTensorParams(
+                      subgraph_op_id=5,
+                      transformations=[
+                          qtyping.QuantTransformation.ADD_QUANTIZE,
+                      ],
+                      parameters=qtyping.UniformQuantParams(
+                          8, None, np.array([0.00027501]), np.array([0])
+                      ),
+                  ),
+              ],
+          ),
+          expected=True,
+      ),
   )
   def test_params_compatible(self, param1, param2, expected):
     # adding a test to make production coverage happy.

ai_edge_quantizer/transformations/quantize_tensor.py

@@ -121,6 +121,26 @@ def _perform_channelwise_quantization(
   return flatbuffer_quantization
 
 
+def _downcast_and_truncate_scale(input_scale: np.ndarray) -> np.ndarray:
+  """Given a fp32 scale, downcast it to fp16 and truncate mantissa to 7 bits.
+
+  CPU kernel can only utilize 7 bits of mantissa for fp16, so we want to produce
+  scale this way to unify behaviours across different platforms.
+
+  Args:
+    input_scale: The input scale in fp32.
+
+  Returns:
+    The downcasted & truncated scale in fp16.
+  """
+
+  # A regular fp16 has 10 bits of mantissa, so we need to zero out the 3 least
+  # significant bits.
+  return (
+      input_scale.astype(np.float16).view(dtype=np.uint16) & np.uint16(0xFFF8)
+  ).view(dtype=np.float16)
+
+
 def _perform_blockwise_quantization(
     transformation_input: transformation_utils.TransformationInput,
 ) -> schema_py_generated.QuantizationParametersT():
@@ -142,9 +162,13 @@ def _perform_blockwise_quantization(
   )
   tensor = transformation_input.subgraph.tensors[transformation_input.tensor_id]
   blockwise_details = schema_py_generated.BlockwiseQuantizationT()
+  # Downcast and truncate the scale to fp16.
+  downcasted_scale = _downcast_and_truncate_scale(
+      transformation_input.quant_params.scale
+  )
   scale_tensor_id = transformation_utils.add_new_constant_tensor(
       tensor.name + b"_scales",
-      transformation_input.quant_params.scale.astype(np.float16),
+      downcasted_scale,
       schema_py_generated.TensorType.FLOAT16,
       transformation_input.subgraph,
       transformation_input.buffers,

{ai_edge_quantizer_nightly-0.1.0.dev20250327.dist-info → ai_edge_quantizer_nightly-0.1.0.dev20250328.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-quantizer-nightly
-Version: 0.1.0.dev20250327
+Version: 0.1.0.dev20250328
 Summary: A quantizer for advanced developers to quantize converted AI Edge models.
 Home-page: https://github.com/google-ai-edge/ai-edge-quantizer
 Keywords: On-Device ML,AI,Google,TFLite,Quantization,LLMs,GenAI

{ai_edge_quantizer_nightly-0.1.0.dev20250327.dist-info → ai_edge_quantizer_nightly-0.1.0.dev20250328.dist-info}/RECORD

@@ -10,8 +10,8 @@ ai_edge_quantizer/model_modifier.py,sha256=SPt9X-xBzRvcd4xIS24zLHt3aUS2QwsNDqweF
 ai_edge_quantizer/model_modifier_test.py,sha256=cJd04SLOG-fQZZNZPcisoBLx3cLtWEwGqUBbLb-pif4,4751
 ai_edge_quantizer/model_validator.py,sha256=fRNz0jO54cthPTibsCuViUXUuFRHl_fbvEiCukIVy20,13030
 ai_edge_quantizer/model_validator_test.py,sha256=EeqOP_mrZsnZ3rug756s0ryDDqd2KgIDld5Lm_gDuWY,13020
-ai_edge_quantizer/params_generator.py,sha256=f-KhJMFdRv2oHxfM8tAANPOtfBMw8vD7Vjv0rYQbnF4,16062
-ai_edge_quantizer/params_generator_test.py,sha256=zmDS6jG5zKhHL_hzJw2wlMTx1LLcNCK6S5WlwogWF-A,41122
+ai_edge_quantizer/params_generator.py,sha256=46XDjnP4R3m4xsoXNp7brv0sNQPdQMg217_CbEl-Wgg,15780
+ai_edge_quantizer/params_generator_test.py,sha256=9WTUl87XqbM4NruX5ypLuVRtuhcw-CmxndsMOUzZ92Q,43171
 ai_edge_quantizer/qtyping.py,sha256=UBZ3HgO8IDLY6VJmO05rGtFv_idMD3Os3WWsnriA0NA,15235
 ai_edge_quantizer/quantizer.py,sha256=g3DMqFMrMpt9jQttCE0WcdNbMtk0JZnmN5MmCHrNdyM,13202
 ai_edge_quantizer/quantizer_test.py,sha256=K_HBA56JkFI3HL8VLWCqGEfC0ISh5ldMKoNyBdGRAJg,20368
@@ -50,7 +50,7 @@ ai_edge_quantizer/transformations/emulated_subchannel.py,sha256=HVaRxoC8PCAvy3xe
 ai_edge_quantizer/transformations/emulated_subchannel_test.py,sha256=gZP6u9NdPXl7s19qB_Un8evou9ZZV6I9Gy0E1rdobHM,7722
 ai_edge_quantizer/transformations/quant_insert.py,sha256=jn6HsJaV-sqBiFPY-Aqbd64t8zgcYVkEkZI375x_FWY,3958
 ai_edge_quantizer/transformations/quant_insert_test.py,sha256=X9ptPDvJCFkR5tejKnD1SlHFGPazQTW-wNNMV9MEAuw,10107
-ai_edge_quantizer/transformations/quantize_tensor.py,sha256=vzKtrXILqVsr1NGlribhdtKEIsXA93o37embLRe9TwQ,7493
+ai_edge_quantizer/transformations/quantize_tensor.py,sha256=y6As38mTzhva50YvNQ7p0SFpuWet3LPqFwE3qIO0gEQ,8231
 ai_edge_quantizer/transformations/quantize_tensor_test.py,sha256=mHLO3_MRt36A8-ZN8ADn5tBBJlqjTWa7ZUN8Mmu5Rcw,9116
 ai_edge_quantizer/transformations/transformation_utils.py,sha256=R42OIbzwQ7JYJ-Qt46jsqwb6u4MfDGiIPCRZCUGLVCw,4664
 ai_edge_quantizer/transformations/transformation_utils_test.py,sha256=xH64SF3UHDh84vYbt-WvmXNjM-Jg-mefES1ACO1tkqw,6269
@@ -64,8 +64,8 @@ ai_edge_quantizer/utils/tfl_interpreter_utils.py,sha256=x2xA2CFPpe_2trcV8v5xGaBE
 ai_edge_quantizer/utils/tfl_interpreter_utils_test.py,sha256=Op3JxtOqlrjzmYF18jnnstL1k9xiY9kKJ8S2vklKGkc,11327
 ai_edge_quantizer/utils/validation_utils.py,sha256=oYw33Sg547AqtGw-choPUJmp9SAKkV46J_ddqSsum2Q,3950
 ai_edge_quantizer/utils/validation_utils_test.py,sha256=V_qNDikPD4OPB-siOLQCWNVWTAu87h2IgNYt7teFd-o,2934
-ai_edge_quantizer_nightly-0.1.0.dev20250327.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-ai_edge_quantizer_nightly-0.1.0.dev20250327.dist-info/METADATA,sha256=KES2W7tXAwTOpymOVvoQXovvZ2eaObKUUUT5L06a2gw,1527
-ai_edge_quantizer_nightly-0.1.0.dev20250327.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
-ai_edge_quantizer_nightly-0.1.0.dev20250327.dist-info/top_level.txt,sha256=8QTfPnFXNVUhScFLaa-NWZMFWMn72M50DVPubpwWB1g,18
-ai_edge_quantizer_nightly-0.1.0.dev20250327.dist-info/RECORD,,
+ai_edge_quantizer_nightly-0.1.0.dev20250328.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+ai_edge_quantizer_nightly-0.1.0.dev20250328.dist-info/METADATA,sha256=ajE8NjYUjSmr3VZQWBzARMc9MLqLkeHooAmK-fCg7Ms,1527
+ai_edge_quantizer_nightly-0.1.0.dev20250328.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
+ai_edge_quantizer_nightly-0.1.0.dev20250328.dist-info/top_level.txt,sha256=8QTfPnFXNVUhScFLaa-NWZMFWMn72M50DVPubpwWB1g,18
+ai_edge_quantizer_nightly-0.1.0.dev20250328.dist-info/RECORD,,