[SW-199696] Added PatchedLinearBase (#192)

* Added PatchedLinearBase

* Fixed PatchedLinear forward_qdq

* Changed quant strategy - scale to fix ci

* Renamed QuantStrategy to QuantWrapper

* Removed instance member from QuantWrapper

* [SW-224403] Added ticket and throwing error when using row_parallel_linear_allreduce_quantization

* Changed QuantWrapper to a simple method that stores scale

* [SW-224538] Added ticket to TODO comment for init_linear

* Pushed requires_grad to the tensor creation

* Update neural_compressor/torch/algorithms/fp8_quant/_quant_common/helper_modules.py

* Update neural_compressor/torch/algorithms/fp8_quant/_quant_common/helper_modules.py

* Moved copy_scale functions inside PatchedLinearBase

* Update helper_modules.py

* Update helper_modules.py

* Update neural_compressor/torch/algorithms/fp8_quant/_quant_common/helper_modules.py

* Update helper_modules.py

* Update helper_modules.py

* Update helper_modules.py copy scale

* Update neural_compressor/torch/algorithms/fp8_quant/_quant_common/helper_modules.py

Author: HolyFalafel

neural_compressor/torch/algorithms/fp8_quant/_core/quant_dequant.py

@@ -130,9 +130,6 @@ def forward(self, x):
         scale = self.input_scales_creator.calc_scales(x, QuantTensorType.DYNAMIC)
         scale_inv = self.input_scales_creator.calc_invert_scales()
 
-        scale = create_scale_tensor(scale, self.scale_format)
-        scale_inv = create_scale_tensor(scale_inv, self.scale_format)
-
         ret = self.cast_to_op(x, scale_inv, False, False, self.lp_dtype)
 
         return ret, scale

neural_compressor/torch/algorithms/fp8_quant/_core/scale_handler.py

@@ -39,9 +39,9 @@ def create_scale_tensor(orig_tensor, scale_format):
         return orig_tensor.to("cpu").to(torch.float)
     if scale_format == ScaleFormat.CONST:
         if isinstance(orig_tensor, torch.Tensor):
-            return torch.nn.Parameter(orig_tensor)
+            return torch.nn.Parameter(orig_tensor, requires_grad=False)
         elif isinstance(orig_tensor, list):
-            return [torch.nn.Parameter(x) for x in orig_tensor]
+            return [torch.nn.Parameter(x, requires_grad=False) for x in orig_tensor]
     elif scale_format == ScaleFormat.SCALAR:
         if isinstance(orig_tensor, (torch.Tensor, float)):
             return scale_to_scalar(orig_tensor)