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

Override lumina2 to use diffusers lora loading

Browse Source
This commit is contained in:
Disty0
2025-07-02 06:40:27 +03:00
parent adcee933da
commit 3406083d14
2 changed files with 5 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/lora/lora_overrides.py
+1
View File
@@ -32,6 +32,7 @@ force_models_diffusers = [ # forced always
'kandinsky',
'hunyuandit',
'auraflow',
'lumina2',
# video models
'hunyuanvideo',
'cogvideo',
modules/sdnq/forward.py
+4 -4
View File
@@ -44,15 +44,15 @@ def get_forward_func(layer_class_name: str, use_quantized_matmul: bool, is_integ
def quantize_fp8_matmul_input(input: torch.FloatTensor) -> Tuple[torch.Tensor, torch.FloatTensor]:
input = input.flatten(0,-2).contiguous()
input_scale = torch.amax(input.abs(), dim=-1, keepdims=True).div_(448)
input = torch.div(input, input_scale).clamp_(-448, 448).to(torch.float8_e4m3fn)
input_scale = input_scale.to(torch.float32)
input = torch.div(input, input_scale).clamp_(-448, 448).to(dtype=torch.float8_e4m3fn)
input_scale = input_scale.to(dtype=torch.float32)
return input, input_scale
def quantize_fp8_matmul_input_tensorwise(input: torch.FloatTensor, scale: torch.FloatTensor) -> Tuple[torch.Tensor, torch.FloatTensor]:
input = input.flatten(0,-2).contiguous()
input_scale = torch.amax(input.abs(), dim=-1, keepdims=True).div_(448)
input = torch.div(input, input_scale).clamp_(-448, 448).to(torch.float8_e4m3fn)
input = torch.div(input, input_scale).clamp_(-448, 448).to(dtype=torch.float8_e4m3fn)
scale = torch.mul(input_scale, scale)
if scale.dtype == torch.float16: # fp16 will overflow
scale = scale.to(dtype=torch.float32)
@@ -62,7 +62,7 @@ def quantize_fp8_matmul_input_tensorwise(input: torch.FloatTensor, scale: torch.
def quantize_int8_matmul_input(input: torch.FloatTensor, scale: torch.FloatTensor) -> Tuple[torch.CharTensor, torch.FloatTensor]:
input = input.flatten(0,-2).contiguous()
input_scale = torch.amax(input.abs(), dim=-1, keepdims=True).div_(127)
input = torch.div(input, input_scale).round_().clamp_(-128, 127).to(torch.int8)
input = torch.div(input, input_scale).round_().clamp_(-128, 127).to(dtype=torch.int8)
scale = torch.mul(input_scale, scale)
if scale.dtype == torch.float16: # fp16 will overflow
scale = scale.to(dtype=torch.float32)