pytorch · vkuzo · Sep 29, 2025 · Sep 25, 2025 · Sep 25, 2025 · Sep 25, 2025
diff --git a/torchao/prototype/mx_formats/nvfp4_tensor.py b/torchao/prototype/mx_formats/nvfp4_tensor.py
@@ -133,6 +133,9 @@ def __new__(
     def __repr__(self):
         return f"NVFP4Tensor: blockwise_scales: {self._scale_e4m3}, per_tensor_scale: {self._per_tensor_scale}, d: {self.qdata}, d_hp: {self.to_dtype(self._orig_dtype)}"
 
+    def _quantization_type(self):
+        return f"{self._is_swizzled_scales=}, {self.use_triton_kernel=}, {self.act_quant_kwargs=}"
+
     @classmethod
     def __torch_dispatch__(cls, func, types, args, kwargs=None):
         # Use NVFP4-specific ops table