update

Author: dsikka

src/compressed_tensors/quantization/quant_scheme.py

@@ -163,46 +163,59 @@ def is_preset_scheme(name: str) -> bool:
     )
 )
 
-MXFP4 = dict(
+
+NVFP4 = dict(
     weights=QuantizationArgs(
         num_bits=4,
         type=QuantizationType.FLOAT,
-        strategy=QuantizationStrategy.GROUP,
+        strategy=QuantizationStrategy.TENSOR_GROUP,
         symmetric=True,
         dynamic=False,
-        group_size=32,
+        group_size=16,
+        observer="static_minmax",
     ),
     input_activations=QuantizationArgs(
+        num_bits=4,
+        type=QuantizationType.FLOAT,
+        strategy=QuantizationStrategy.TENSOR_GROUP,
+        symmetric=True,
+        dynamic=DynamicType.LOCAL,
+        group_size=16,
+        observer="static_minmax",
+    ),
+)
+
+MXFP4A16 = dict(
+    weights=QuantizationArgs(
         num_bits=4,
         type=QuantizationType.FLOAT,
         strategy=QuantizationStrategy.GROUP,
-        dynamic=True,
         symmetric=True,
+        dynamic=False,
         group_size=32,
-    ),
+    )
 )
 
-NVFP4 = dict(
+MXFP4 = dict(
     weights=QuantizationArgs(
         num_bits=4,
         type=QuantizationType.FLOAT,
-        strategy=QuantizationStrategy.TENSOR_GROUP,
+        strategy=QuantizationStrategy.GROUP,
         symmetric=True,
         dynamic=False,
-        group_size=16,
-        observer="static_minmax",
+        group_size=32,
     ),
     input_activations=QuantizationArgs(
         num_bits=4,
         type=QuantizationType.FLOAT,
-        strategy=QuantizationStrategy.TENSOR_GROUP,
+        strategy=QuantizationStrategy.GROUP,
+        dynamic=True,
         symmetric=True,
-        dynamic=DynamicType.LOCAL,
-        group_size=16,
-        observer="static_minmax",
+        group_size=32,
     ),
 )
 
+
 # 8 bit integer weights and 8 bit activations quantization
 INT8_W8A8 = dict(
     weights=QuantizationArgs(