From 961bc95d96f8dc8268cac42390cbb9a15fd77e68 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Stanis=C5=82aw=20Szymczyk?= Date: Sat, 14 Mar 2026 20:17:21 +0100 Subject: [PATCH] convert : DSA indexer weights are bf16 in the original fp8 model, so I think it's best not to quantize them. --- convert_hf_to_gguf.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 3fdeb27794..212836398b 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -618,6 +618,8 @@ class ModelBase: gguf.MODEL_TENSOR.SSM_CONV1D_Q, gguf.MODEL_TENSOR.SSM_CONV1D_K, gguf.MODEL_TENSOR.SSM_CONV1D_V, + # DSA indexer weights should be F32 + gguf.MODEL_TENSOR.INDEXER_PROJ, ) ) or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")