Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-18 16:54:33 +02:00
Code Issues Packages Projects Releases Wiki Activity

SDNQ use torch.bitwise ops instead of python

Browse Source
This commit is contained in:
Disty0
2025-05-27 03:02:36 +03:00
parent 280be31883
commit dec460e665
1 changed files with 2 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/model_quant_sdnq.py
+2 -3
View File
@@ -433,15 +433,14 @@ def pack_uint4(tensor: torch.Tensor) -> torch.Tensor:
if tensor.dtype != torch.uint8:
raise RuntimeError(f"Invalid tensor dtype {tensor.type}. torch.uint8 type is supported.")
packed_tensor = tensor.contiguous().reshape(-1, 2)
packed_tensor = torch.bitwise_and(packed_tensor[..., ::2], 15) | packed_tensor[..., 1::2] << 4
packed_tensor = torch.bitwise_or(torch.bitwise_and(packed_tensor[:, 0], 15), torch.bitwise_left_shift(packed_tensor[:, 1], 4))
return packed_tensor
def pack_int4(tensor: torch.Tensor) -> torch.Tensor:
if tensor.dtype != torch.int8:
raise RuntimeError(f"Invalid tensor dtype {tensor.type}. torch.int8 type is supported.")
tensor = tensor + 8
return pack_uint4(tensor.to(dtype=torch.uint8))
return pack_uint4((tensor + 8).to(dtype=torch.uint8))
def unpack_uint4(packed_tensor: torch.Tensor, shape: torch.Size) -> torch.Tensor: