From: Sigbjørn Skjæret <redacted>
Date: Sun, 12 Apr 2026 17:22:29 +0000 (+0200)
Subject: convert : force f16 or f32 on step3-vl conv weights (#21646)
X-Git-Tag: upstream/0.0.10438~1670
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=1e9d771e2c2f1113a5ebdd0dc15bafe57dce64be;p=pkg%2Fggml%2Fsources%2Fllama.cpp

convert : force f16 or f32 on step3-vl conv weights (#21646)
---

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index c96afc78b..374a55fb1 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -4992,6 +4992,8 @@ class Step3VLVisionModel(MmprojModel):
     def tensor_force_quant(self, name, new_name, bid, n_dims):
         if ".position_embd." in new_name:
             return gguf.GGMLQuantizationType.F32
+        if ("mm.0." in new_name or "mm.1." in new_name) and new_name.endswith(".weight"):
+            return gguf.GGMLQuantizationType.F16 if self.ftype == gguf.LlamaFileType.MOSTLY_F16 else gguf.GGMLQuantizationType.F32
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: