Merge nucleic/sleek-ember-seal-uady into dev

README.md

@@ -239,7 +239,10 @@ and removes each package immediately after prediction; this avoids Core ML Tools
 one full weight copy per calibration step until process exit. It prints progress while it
 runs. The candidate uses per-tensor asymmetric uint8 activations,
 per-channel symmetric int8 linear weights, and per-tensor asymmetric uint8 embedding
-weights. It fails the command if the resulting package exceeds 25 MiB.
+weights. Activation quantization is limited to floating-point linear operations; applying
+Core ML Tools' global policy also selects integer embedding-index additions and produces
+an invalid quantize operation. It fails the command if the resulting package exceeds
+25 MiB.
 
 Run the frozen gate with CPU+Neural Engine placement and compare labels directly with the
 accepted int8 ONNX artifact. Gated Core ML evaluation fails closed without

quantize_coreml.py

@@ -86,7 +86,12 @@ def _optimization_configs(optimize: Any) -> tuple[Any, Any]:
         granularity="per_tensor",
     )
     activation_config = optimize.coreml.OptimizationConfig(
-        global_config=activation,
+        # A global activation policy also selects integer `add` operations in the
+        # embedding/index path. Core ML's quantize op requires its floating-point scale
+        # to match a floating-point input, so quantize only accelerator-supported linear
+        # activations. Attention matmul activation quantization is not supported by this
+        # Core ML graph pass.
+        op_type_configs={"linear": activation},
     )
 
     linear_weight = optimize.coreml.OpLinearQuantizerConfig(

tests/test_quantize_coreml.py

@@ -37,12 +37,15 @@ class FakeOptimize:
 class CoreMLQuantizationConfigTests(unittest.TestCase):
     def test_matches_accepted_qdq_policy(self):
         activation, weights = quantize_coreml._optimization_configs(FakeOptimize)
-        self.assertEqual("linear", activation.global_config.values["mode"])
+        self.assertIsNone(activation.global_config)
-        self.assertIs(np.uint8, activation.global_config.values["dtype"])
+        activation_linear = activation.op_type_configs["linear"].values
+        self.assertEqual("linear", activation_linear["mode"])
+        self.assertIs(np.uint8, activation_linear["dtype"])
         self.assertEqual(
             "per_tensor",
-            activation.global_config.values["granularity"],
+            activation_linear["granularity"],
         )
+        self.assertEqual({"linear"}, set(activation.op_type_configs))
 
         linear = weights.op_type_configs["linear"].values
         self.assertEqual("linear_symmetric", linear["mode"])