From: Sigbjørn Skjæret <redacted>
Date: Wed, 17 Dec 2025 21:15:53 +0000 (+0100)
Subject: convert : force patch_merger tensors to f16/f32 (#18124)
X-Git-Tag: upstream/0.0.7599~136
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=5166aaf86898c5d5268b2c4f770a35cb7ca1a0f6;p=pkg%2Fggml%2Fsources%2Fllama.cpp

convert : force patch_merger tensors to f16/f32 (#18124)
---

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index bd16ba31..806b3d7b 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -1838,7 +1838,7 @@ class MmprojModel(ModelBase):
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
         del bid, name, n_dims  # unused
-        if ".patch_embd.weight" in new_name:
+        if ".patch_embd.weight" in new_name or ".patch_merger.weight" in new_name:
             return gguf.GGMLQuantizationType.F16 if self.ftype == gguf.LlamaFileType.MOSTLY_F16 else gguf.GGMLQuantizationType.F32
         return False