remove dequantize_symmetric_with_bias

modules/sdnq/dequantizer.py

@@ -85,19 +85,6 @@ def dequantize_symmetric(
     return result
 
 
-@devices.inference_context()
-def dequantize_symmetric_with_bias(weight: torch.Tensor, scale: torch.FloatTensor, bias: torch.FloatTensor, hadamard: torch.FloatTensor | None = None, dtype: torch.dtype = None, result_shape: torch.Size = None) -> torch.FloatTensor:
-    if hadamard is not None:
-        result = rotate_hadamard(weight.to(dtype=scale.dtype).mul_(scale), hadamard=hadamard).add_(bias)
-    else:
-        result = torch.addcmul(bias, weight.to(dtype=scale.dtype), scale)
-    if dtype is not None:
-        result = result.to(dtype=dtype)
-    if result_shape is not None:
-        result = result.view(result_shape)
-    return result
-
-
 @devices.inference_context()
 def dequantize_packed_int_asymmetric(weight: torch.Tensor, scale: torch.FloatTensor, zero_point: torch.FloatTensor, shape: torch.Size, weights_dtype: str, svd_up: torch.FloatTensor | None = None, svd_down: torch.FloatTensor | None = None, hadamard: torch.FloatTensor | None = None, dtype: torch.dtype = None, result_shape: torch.Size = None, skip_quantized_matmul: bool = False, re_quantize_for_matmul: bool = False) -> torch.FloatTensor:
     return dequantize_asymmetric(unpack_int(weight, weights_dtype, shape), scale, zero_point, svd_up=svd_up, svd_down=svd_down, hadamard=hadamard, dtype=dtype, result_shape=result_shape, skip_quantized_matmul=skip_quantized_matmul, re_quantize_for_matmul=re_quantize_for_matmul)

modules/sdnq/layers/conv/conv_fp16.py

@@ -3,7 +3,7 @@
 import torch
 
 from ...common import compile_func, fp_mm_func
-from ...dequantizer import dequantize_symmetric, dequantize_symmetric_with_bias
+from ...dequantizer import dequantize_symmetric, dequantize_asymmetric
 from ...quant_utils import rotate_hadamard, get_hadamard
 from ...packed_float import unpack_float
 
@@ -59,7 +59,7 @@ def conv_fp16_matmul(
             result.append(fp_mm_func(input[:, i], weight[:, i]))
         result = torch.cat(result, dim=-1).to(dtype=input_scale.dtype).mul_(input_scale)
     if bias is not None:
-        dequantize_symmetric_with_bias(result, scale, bias, dtype=return_dtype, result_shape=mm_output_shape)
+        dequantize_asymmetric(result, scale, bias, dtype=return_dtype, result_shape=mm_output_shape)
     else:
         dequantize_symmetric(result, scale, dtype=return_dtype, result_shape=mm_output_shape)
 

modules/sdnq/layers/conv/conv_fp8.py

@@ -3,7 +3,7 @@
 import torch
 
 from ...common import compile_func
-from ...dequantizer import dequantize_symmetric, dequantize_symmetric_with_bias
+from ...dequantizer import dequantize_symmetric, dequantize_asymmetric
 from ...quant_utils import rotate_hadamard, get_hadamard
 from ...packed_float import unpack_float
 
@@ -58,7 +58,7 @@ def conv_fp8_matmul(
             result.append(torch._scaled_mm(input[:, i], weight[:, i], scale_a=dummy_input_scale, scale_b=dummy_input_scale, bias=None, out_dtype=input_scale.dtype))
         result = torch.cat(result, dim=-1).mul_(input_scale)
     if bias is not None:
-        dequantize_symmetric_with_bias(result, scale, bias, dtype=return_dtype, result_shape=mm_output_shape)
+        dequantize_asymmetric(result, scale, bias, dtype=return_dtype, result_shape=mm_output_shape)
     else:
         dequantize_symmetric(result, scale, dtype=return_dtype, result_shape=mm_output_shape)
 

modules/sdnq/layers/conv/conv_int8.py

@@ -3,7 +3,7 @@
 import torch
 
 from ...common import compile_func, int_mm_func
-from ...dequantizer import dequantize_symmetric, dequantize_symmetric_with_bias
+from ...dequantizer import dequantize_symmetric, dequantize_asymmetric
 from ...quant_utils import rotate_hadamard, get_hadamard
 from ...packed_int import unpack_int
 
@@ -73,7 +73,7 @@ def conv_int8_matmul(
             result.append(int_mm_func(input[:, i], weight[:, i]))
         result = torch.cat(result, dim=-1).to(dtype=input_scale.dtype).mul_(input_scale)
     if bias is not None:
-        result = dequantize_symmetric_with_bias(result, scale, bias, dtype=return_dtype, result_shape=mm_output_shape)
+        result = dequantize_asymmetric(result, scale, bias, dtype=return_dtype, result_shape=mm_output_shape)
     else:
         result = dequantize_symmetric(result, scale, dtype=return_dtype, result_shape=mm_output_shape)
 

modules/sdnq/layers/conv/conv_uint8.py

@@ -3,7 +3,7 @@
 import torch
 
 from ...common import compile_func, int_mm_func
-from ...dequantizer import dequantize_symmetric, dequantize_symmetric_with_bias
+from ...dequantizer import dequantize_asymmetric
 from ...quant_utils import rotate_hadamard, get_hadamard
 from ...packed_int import unpack_int
 
@@ -75,7 +75,7 @@ def conv_uint8_matmul(
         for i in range(groups):
             result.append(int_mm_func(input[:, i], weight[:, i]))
         result = torch.cat(result, dim=-1).to(dtype=input_scale.dtype).mul_(input_scale)
-    result = dequantize_symmetric_with_bias(result, scale, zero_bias, dtype=return_dtype, result_shape=mm_output_shape)
+    result = dequantize_asymmetric(result, scale, zero_bias, dtype=return_dtype, result_shape=mm_output_shape)
 
     if conv_type == 1:
         result = result.transpose_(1,2)

modules/sdnq/layers/linear/linear_fp16.py

@@ -3,7 +3,7 @@
 import torch
 
 from ...common import compile_func, fp_mm_func
-from ...dequantizer import dequantize_symmetric, dequantize_symmetric_with_bias
+from ...dequantizer import dequantize_symmetric, dequantize_asymmetric
 from ...quant_utils import rotate_hadamard, get_hadamard
 from ...packed_float import unpack_float
 
@@ -40,7 +40,7 @@ def fp16_matmul(
     input, input_scale = quantize_fp_mm_input(input, dtype=scale.dtype, matmul_dtype="float16")
     input, weight = check_mats(input, weight)
     if bias is not None:
-        return dequantize_symmetric_with_bias(fp_mm_func(input, weight).to(dtype=input_scale.dtype).mul_(input_scale), scale, bias, dtype=return_dtype, result_shape=output_shape)
+        return dequantize_asymmetric(fp_mm_func(input, weight).to(dtype=input_scale.dtype).mul_(input_scale), scale, bias, dtype=return_dtype, result_shape=output_shape)
     else:
         return dequantize_symmetric(fp_mm_func(input, weight).to(dtype=input_scale.dtype).mul_(input_scale), scale, dtype=return_dtype, result_shape=output_shape)
 

modules/sdnq/layers/linear/linear_fp8.py

@@ -3,7 +3,7 @@
 import torch
 
 from ...common import compile_func
-from ...dequantizer import dequantize_symmetric, dequantize_symmetric_with_bias
+from ...dequantizer import dequantize_symmetric, dequantize_asymmetric
 from ...quant_utils import quantize_fp_mm, rotate_hadamard, get_hadamard
 from ...packed_float import unpack_float
 
@@ -48,7 +48,7 @@ def fp8_matmul(
     input, input_scale = quantize_fp_mm_input(input, dtype=scale.dtype)
     input, weight = check_mats(input, weight, allow_contiguous_mm=False)
     if bias is not None:
-        return dequantize_symmetric_with_bias(torch._scaled_mm(input, weight, scale_a=dummy_input_scale, scale_b=dummy_input_scale, bias=None, out_dtype=input_scale.dtype).mul_(input_scale), scale, bias, dtype=return_dtype, result_shape=output_shape)
+        return dequantize_asymmetric(torch._scaled_mm(input, weight, scale_a=dummy_input_scale, scale_b=dummy_input_scale, bias=None, out_dtype=input_scale.dtype).mul_(input_scale), scale, bias, dtype=return_dtype, result_shape=output_shape)
     else:
         return dequantize_symmetric(torch._scaled_mm(input, weight, scale_a=dummy_input_scale, scale_b=dummy_input_scale, bias=None, out_dtype=input_scale.dtype).mul_(input_scale), scale, dtype=return_dtype, result_shape=output_shape)
 

modules/sdnq/layers/linear/linear_int8.py

@@ -3,7 +3,7 @@
 import torch
 
 from ...common import compile_func, int_mm_func
-from ...dequantizer import dequantize_symmetric, dequantize_symmetric_with_bias
+from ...dequantizer import dequantize_symmetric, dequantize_asymmetric
 from ...quant_utils import quantize_int_mm, rotate_hadamard, get_hadamard
 from ...packed_int import unpack_int
 
@@ -67,7 +67,7 @@ def int8_matmul(
     input, weight = check_mats(input, weight)
 
     if bias is not None:
-        return dequantize_symmetric_with_bias(int_mm_func(input, weight).to(dtype=input_scale.dtype).mul_(input_scale), scale, bias, dtype=return_dtype, result_shape=output_shape)
+        return dequantize_asymmetric(int_mm_func(input, weight).to(dtype=input_scale.dtype).mul_(input_scale), scale, bias, dtype=return_dtype, result_shape=output_shape)
     else:
         return dequantize_symmetric(int_mm_func(input, weight).to(dtype=input_scale.dtype).mul_(input_scale), scale, dtype=return_dtype, result_shape=output_shape)
 

modules/sdnq/layers/linear/linear_uint8.py

@@ -3,7 +3,7 @@
 import torch
 
 from ...common import compile_func, int_mm_func
-from ...dequantizer import dequantize_symmetric, dequantize_symmetric_with_bias
+from ...dequantizer import dequantize_asymmetric
 from ...quant_utils import quantize_uint_mm, rotate_hadamard, get_hadamard
 from ...packed_int import unpack_int
 
@@ -70,7 +70,7 @@ def uint8_matmul(
         zero_bias.add_(bias)
 
     input, weight = check_mats(input, weight)
-    return dequantize_symmetric_with_bias(int_mm_func(input, weight).to(dtype=input_scale.dtype).mul_(input_scale), scale, zero_bias, dtype=return_dtype, result_shape=output_shape)
+    return dequantize_asymmetric(int_mm_func(input, weight).to(dtype=input_scale.dtype).mul_(input_scale), scale, zero_bias, dtype=return_dtype, result_shape=output_shape)
 
 
 def quantized_linear_forward_uint8_matmul(self, input: torch.FloatTensor) -> torch.FloatTensor: