From 5f78ec32a5e8c4bf16286702e223ab15ed50b1f3 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Dani=C3=ABl=20de=20Kok?= Date: Thu, 16 Jan 2025 13:44:32 +0100 Subject: [PATCH] Do not convert weight scale to e4m3fnuz on CUDA (#2917) --- .../text_generation_server/layers/compressed_tensors/w8an_fp.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/server/text_generation_server/layers/compressed_tensors/w8an_fp.py b/server/text_generation_server/layers/compressed_tensors/w8an_fp.py index 15bdce08..ebcc06d6 100644 --- a/server/text_generation_server/layers/compressed_tensors/w8an_fp.py +++ b/server/text_generation_server/layers/compressed_tensors/w8an_fp.py @@ -147,7 +147,7 @@ class W8ANFpLoader(WeightsLoader): else None ) - if self.load_weight_scale or SYSTEM == "rocm": + if self.load_weight_scale and SYSTEM == "rocm": w, weight_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz( w, weight_scale, input_scale )