Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-20 01:31:13 +02:00
Code Issues Packages Projects Releases Wiki Activity

update sdnq

Browse Source
This commit is contained in:
Disty0
2026-01-14 16:23:26 +03:00
parent 4dbdee10e3
commit 784cda80aa
13 changed files with 121 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sdnq/dequantizer.py
+10 -7
View File
@@ -95,7 +95,7 @@ def quantize_int_mm(input: torch.FloatTensor, dim: int = -1, matmul_dtype: str =
@devices.inference_context()
def quantize_int_mm_sr(input: torch.FloatTensor, dim: int = -1, matmul_dtype: str = "int8") -> Tuple[torch.Tensor, torch.FloatTensor]:
scale = torch.amax(input.abs(), dim=dim, keepdims=True).div_(dtype_dict[matmul_dtype]["max"])
input = torch.div(input, scale).add_(torch.randn_like(input), alpha=0.1).round_().clamp_(dtype_dict[matmul_dtype]["min"], dtype_dict[matmul_dtype]["max"]).to(dtype=dtype_dict[matmul_dtype]["torch_dtype"])
input = torch.div(input, scale).add_(torch.rand_like(input), alpha=0.1).round_().clamp_(dtype_dict[matmul_dtype]["min"], dtype_dict[matmul_dtype]["max"]).to(dtype=dtype_dict[matmul_dtype]["torch_dtype"])
return input, scale
@@ -308,6 +308,7 @@ class SDNQDequantizer:
def __call__(self, weight, scale, zero_point, svd_up, svd_down, skip_quantized_matmul: bool = False, skip_compile: bool = False, dtype: torch.dtype = None): # pylint: disable=unused-argument
if dtype is None:
dtype = self.result_dtype
re_quantize_for_matmul = self.re_quantize_for_matmul or self.is_packed
if self.is_packed:
if self.is_integer:
if self.is_unsigned:
@@ -317,9 +318,9 @@ class SDNQDequantizer:
return dequantize_packed_int_asymmetric_compiled(weight, scale, zero_point, self.quantized_weight_shape, self.weights_dtype, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul)
else:
if skip_compile:
return dequantize_packed_int_symmetric(weight, scale, self.quantized_weight_shape, self.weights_dtype, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul, re_quantize_for_matmul=self.re_quantize_for_matmul)
return dequantize_packed_int_symmetric(weight, scale, self.quantized_weight_shape, self.weights_dtype, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul, re_quantize_for_matmul=re_quantize_for_matmul)
else:
return dequantize_packed_int_symmetric_compiled(weight, scale, self.quantized_weight_shape, self.weights_dtype, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul, re_quantize_for_matmul=self.re_quantize_for_matmul)
return dequantize_packed_int_symmetric_compiled(weight, scale, self.quantized_weight_shape, self.weights_dtype, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul, re_quantize_for_matmul=re_quantize_for_matmul)
else:
if self.is_unsigned:
if skip_compile: # compiled training needs to be traced with the original function
@@ -328,9 +329,9 @@ class SDNQDequantizer:
return dequantize_packed_float_asymmetric_compiled(weight, scale, zero_point, self.quantized_weight_shape, self.weights_dtype, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul)
else:
if skip_compile:
return dequantize_packed_float_symmetric(weight, scale, self.quantized_weight_shape, self.weights_dtype, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul, re_quantize_for_matmul=self.re_quantize_for_matmul)
return dequantize_packed_float_symmetric(weight, scale, self.quantized_weight_shape, self.weights_dtype, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul, re_quantize_for_matmul=re_quantize_for_matmul)
else:
return dequantize_packed_float_symmetric_compiled(weight, scale, self.quantized_weight_shape, self.weights_dtype, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul, re_quantize_for_matmul=self.re_quantize_for_matmul)
return dequantize_packed_float_symmetric_compiled(weight, scale, self.quantized_weight_shape, self.weights_dtype, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul, re_quantize_for_matmul=re_quantize_for_matmul)
else:
if self.is_unsigned:
if skip_compile:
@@ -339,9 +340,9 @@ class SDNQDequantizer:
return dequantize_asymmetric_compiled(weight, scale, zero_point, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul)
else:
if skip_compile:
return dequantize_symmetric(weight, scale, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul, re_quantize_for_matmul=self.re_quantize_for_matmul)
return dequantize_symmetric(weight, scale, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul, re_quantize_for_matmul=re_quantize_for_matmul)
else:
return dequantize_symmetric_compiled(weight, scale, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul, re_quantize_for_matmul=self.re_quantize_for_matmul)
return dequantize_symmetric_compiled(weight, scale, svd_up=svd_up, svd_down=svd_down, dtype=dtype, result_shape=self.result_shape, skip_quantized_matmul=skip_quantized_matmul, re_quantize_for_matmul=re_quantize_for_matmul)
dequantize_asymmetric_compiled = compile_func(dequantize_asymmetric)
@@ -356,3 +357,5 @@ re_quantize_matmul_packed_int_asymmetric_compiled = compile_func(re_quantize_mat
re_quantize_matmul_packed_int_symmetric_compiled = compile_func(re_quantize_matmul_packed_int_symmetric)
re_quantize_matmul_packed_float_asymmetric_compiled = compile_func(re_quantize_matmul_packed_float_asymmetric)
re_quantize_matmul_packed_float_symmetric_compiled = compile_func(re_quantize_matmul_packed_float_symmetric)
torch.serialization.add_safe_globals([SDNQDequantizer])