From 67e0f4d833eb4bcb07f4e64c77be57f7fe4265af Mon Sep 17 00:00:00 2001 From: Disty0 Date: Thu, 29 May 2025 03:22:40 +0300 Subject: [PATCH] Cleanup --- modules/model_quant_sdnq.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/modules/model_quant_sdnq.py b/modules/model_quant_sdnq.py index 256299f4e..eb73c0973 100644 --- a/modules/model_quant_sdnq.py +++ b/modules/model_quant_sdnq.py @@ -279,8 +279,9 @@ def unpack_int4(packed_tensor: torch.Tensor, shape: torch.Size, dtype: Optional[ def quantize_fp8_matmul_input(input: torch.FloatTensor) -> Tuple[torch.FloatTensor, torch.FloatTensor]: input = input.flatten(0,-2).contiguous() - input_scale = torch.div(input.abs().amax(dim=-1), 448).unsqueeze(-1).to(torch.float32) + input_scale = torch.div(input.abs().amax(dim=-1), 448).unsqueeze(-1) input = torch.div(input, input_scale).clamp_(-448, 448).to(torch.float8_e4m3fn) + input_scale = input_scale.to(torch.float32) return input, input_scale