[quant] Use Int8DynActInt4WeightQuantizer in torchao

Summary:
att

Test Plan:
python3 -m examples.models.llama2.export_llama -c stories110M.pt -p params.json -qmode 8da4w -X -d fp32

Reviewers:

Subscribers:

Tasks:

Tags:

ghstack-source-id: b66176d
Pull Request resolved: #2551

Author: jerryzh168

examples/models/llama2/export_llama_lib.py

@@ -38,11 +38,7 @@
 
 from .builder import DType, LlamaEdgeManager, load_llama_model, WeightType
 
-from .quantize import (
-    EmbeddingOnlyInt8QuantHandler,
-    Int8DynActInt4WeightQuantHandler,
-    WeightOnlyInt8QuantHandler,
-)
+from .quantize import EmbeddingOnlyInt8QuantHandler, WeightOnlyInt8QuantHandler
 
 
 IS_FBCODE = True  #  os.environ.get("FBCODE_PLATFORM", False)
@@ -241,10 +237,11 @@ def quantize(
         # Add quantization mode options here: group size, bit width, etc.
         return WeightOnlyInt8QuantHandler(model).quantized_model()
     elif qmode == "8da4w":
-        model = Int8DynActInt4WeightQuantHandler(
-            model,
-            precision=torch_dtype,
-        ).quantized_model()
+        from torchao.quantization.quant_api import Int8DynActInt4WeightQuantizer
+
+        model = Int8DynActInt4WeightQuantizer(precision=torch_dtype).quantize(
+            model
+        )
         print("quantized model:", model)
         return model
     elif qmode == "8da4w-gptq":