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

2026-07-30 21:47:07 -07:00
parent a374fa40b6
commit 82545b833b
3 changed files with 16 additions and 5 deletions
@@ -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(