Merge pull request #372 from jondea/fix-perf-drop-after-torchao-api-change

pytorch: fix perf issue from torchao API change

Author: nSircombe

ML-Frameworks/pytorch-aarch64/CHANGELOG.md

@@ -20,6 +20,7 @@ where `YY` is the year, and `MM` the month of the increment.
   - KLEIDIAI_HASH to bd2e6ae060014035e25bf4986be682762c446c2d, v1.14 from main.
 - Update torchvision from 0.23.0 to a nightly build, 0.25.0.dev20250923
 - Change of flag name in `./build.sh` from `--force` to `--fresh`
+- Add `intx_packing_format="opaque_aten_kleidiai"` to `Int8DynamicActivationIntxWeightConfig` due to torchao API change
 
 ### Removed
 - Removes WIP ComputeLibrary patch https://review.mlplatform.org/c/ml/ComputeLibrary/+/12818/1.

ML-Frameworks/pytorch-aarch64/examples/llama_vision_instruct.py

@@ -52,6 +52,7 @@ def main(args):
                 weight_mapping_type=MappingType.SYMMETRIC_NO_CLIPPING_ERR, # MappingType.SYMMETRIC can also be used but increases error
                 layout=layout,
                 weight_dtype=torch.int4,
+                intx_packing_format="opaque_aten_kleidiai",
             ),
         )
 

ML-Frameworks/pytorch-aarch64/examples/transformers_llm_text_gen.py

@@ -113,6 +113,7 @@ def get_quantized_model(args):
         weight_mapping_type=mapping_type,
         layout=layout,
         weight_dtype=torch.int4,
+        intx_packing_format="opaque_aten_kleidiai",
     )
 
     print("Quantization config:")