Remove group_size arg in Float8DynamicActivationInt4WeightConfig (#2779)

Summary:
Fixes: #2763

Test Plan:
python test/quantization/quantize_/workflows/int4/test_int4_preshuffled_tensor.py

Reviewers:

Subscribers:

Tasks:

Tags:

Author: jerryzh168

test/quantization/quantize_/workflows/int4/test_int4_preshuffled_tensor.py

@@ -33,8 +33,8 @@
     version=2,
 )
 
+# only 128 group_size is supported
 FP8_ACT_CONFIG = Float8DynamicActivationInt4WeightConfig(
-    group_size=128,
     packing_format="preshuffled",
 )
 

test/quantization/test_qat.py

@@ -1927,7 +1927,7 @@ def test_quantize_api_fp8_int4(self):
             quantize_(model, QATConfig(Float8DynamicActivationInt4WeightConfig(), step="convert"))
         """
         self._test_quantize_api_against_ptq(
-            Float8DynamicActivationInt4WeightConfig(group_size=128),
+            Float8DynamicActivationInt4WeightConfig(),
             target_prepare_sqnr=15,
             target_convert_sqnr=float("inf"),
         )

torchao/quantization/quant_api.py

@@ -1156,13 +1156,13 @@ def _int4_weight_only_transform(
 class Float8DynamicActivationInt4WeightConfig(AOBaseConfig):
     """Configuration for apply float8 dynamic per row quantization and int4
     per group weight quantization to linear
+    (only group_size 128 is supported right now since underlying kernel used only supports 128
+    and above and no benefits of making it bigger)
 
     Args:
-        `group_size`: group size for groupwise quantization for weight
         `packing_format`: how the weight is packed, only preshuffled is supported
     """
 
-    group_size: int = 128
     packing_format: PackingFormat = "preshuffled"
 
 
@@ -1174,13 +1174,13 @@ def _float8_dynamic_activation_int4_weight_transform(
         "applying int8 weight only quant requires module to have weight attribute"
         + " but {module} does not have one"
     )
-    group_size = config.group_size
     packing_format = config.packing_format
 
     assert packing_format == "preshuffled", (
         f"only preshuffled packing_format supported right now, got: {packing_format}"
     )
     weight = module.weight
+    group_size = 128
     block_size = tuple([1 for _ in range(weight.ndim - 1)] + [group_size])
     new_weight = Int4PreshuffledTensor.from_hp(
         module.weight,