!16059 伪量化npu_weight_quant_batchmatmul接口增加inner_precise参数

Merge pull request !16059 from chenxu/master

torch_npu/contrib/module/linear_weight_quant.py

@@ -63,6 +63,7 @@ def __init__(
         quant_scale: bool = False,
         quant_offset: bool = False,
         antiquant_group_size: int = 0,
+        inner_precise: int = 0,
     ) -> None:
         super(LinearWeightQuant, self).__init__()
         self.weight = Parameter(torch.empty((out_features, in_features), device=device), False)
@@ -89,6 +90,7 @@ def __init__(
             self.register_parameter('bias', None)
 
         self.antiquant_group_size = antiquant_group_size
+        self.inner_precise = inner_precise
 
     def forward(self, x: Tensor) -> Tensor:
         antiquant_scale = self.antiquant_scale
@@ -100,4 +102,4 @@ def forward(self, x: Tensor) -> Tensor:
                 antiquant_offset = self.antiquant_offset.transpose(-1, -2)
         return torch_npu.npu_weight_quant_batchmatmul(x, self.weight.transpose(-1, -2), antiquant_scale,
                                                       antiquant_offset, self.quant_scale, self.quant_offset,
-                                                      self.bias, self.antiquant_group_size)
+                                                      self.bias, self.antiquant_group_size, self.inner_precise)

torch_npu/meta/_meta_registrations.py

@@ -340,7 +340,7 @@ def npu_mm_all_reduce_base_forward(x1, x2, hcom, reduce_op='sum', bias=None, ant
 
 
 @impl(m, "npu_weight_quant_batchmatmul")
-def npu_weight_quant_batchmatmul_meta(x, weight, antiquant_scale, antiquant_offset=None, quant_scale=None, quant_offset=None, bias=None, antiquant_group_size=0):
+def npu_weight_quant_batchmatmul_meta(x, weight, antiquant_scale, antiquant_offset=None, quant_scale=None, quant_offset=None, bias=None, antiquant_group_size=0, inner_precise=0):
     dim_m = x.size(0)
     if weight.dtype == torch.int32 and weight.is_contiguous():
         dim_n = weight.size(1) * 8