Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 09:14:35 +02:00
Code Issues Packages Projects Releases Wiki Activity

SDNQ remove unnecessary .contiguous()

Browse Source
This commit is contained in:
Disty0
2025-08-21 02:21:05 +03:00
parent c79471b2cd
commit f324b7c0e5
4 changed files with 5 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sdnq/__init__.py
+2 -2
View File
@@ -163,8 +163,8 @@ def sdnq_quantize_layer(layer, weights_dtype="int8", torch_dtype=None, group_siz
scale.transpose_(0,1)
layer.weight.transpose_(0,1)
if not dtype_dict[weights_dtype]["is_integer"]:
stride = layer.weight.stride()
if stride[0] > stride[1] and stride[1] == 1:
weight_stride = layer.weight.stride()
if not (weight_stride[0] == 1 and weight_stride[1] > 1):
layer.weight.data = layer.weight.t().contiguous().t()
if not use_tensorwise_fp8_matmul:
scale = scale.to(torch.float32)
modules/sdnq/layers/linear/linear_fp8.py
+1 -1
View File
@@ -8,7 +8,7 @@ from ...common import use_torch_compile # noqa: TID252
def quantize_fp8_matmul_input(input: torch.FloatTensor) -> Tuple[torch.Tensor, torch.FloatTensor]:
input = input.flatten(0,-2).contiguous().to(dtype=torch.float32)
input = input.flatten(0,-2).to(dtype=torch.float32)
input_scale = torch.amax(input.abs(), dim=-1, keepdims=True).div_(448)
input = torch.div(input, input_scale).clamp_(-448, 448).to(dtype=torch.float8_e4m3fn)
return input, input_scale
modules/sdnq/layers/linear/linear_fp8_tensorwise.py
+1 -1
View File
@@ -9,7 +9,7 @@ from ...dequantizer import dequantize_symmetric, dequantize_symmetric_with_bias
def quantize_fp8_matmul_input_tensorwise(input: torch.FloatTensor, scale: torch.FloatTensor) -> Tuple[torch.Tensor, torch.FloatTensor]:
input = input.flatten(0,-2).contiguous().to(dtype=scale.dtype)
input = input.flatten(0,-2).to(dtype=scale.dtype)
input_scale = torch.amax(input.abs(), dim=-1, keepdims=True).div_(448)
input = torch.div(input, input_scale).clamp_(-448, 448).to(dtype=torch.float8_e4m3fn)
scale = torch.mul(input_scale, scale)
modules/sdnq/layers/linear/linear_int8.py
+1 -1
View File
@@ -10,7 +10,7 @@ from ...dequantizer import dequantize_symmetric, dequantize_symmetric_with_bias
def quantize_int8_matmul_input(input: torch.FloatTensor, scale: torch.FloatTensor) -> Tuple[torch.CharTensor, torch.FloatTensor]:
input = input.flatten(0,-2).contiguous().to(dtype=scale.dtype)
input = input.flatten(0,-2).to(dtype=scale.dtype)
input_scale = torch.amax(input.abs(), dim=-1, keepdims=True).div_(127)
input = torch.div(input, input_scale).round_().clamp_(-128, 127).to(dtype=torch.int8)
scale = torch.mul(input_scale, scale)