From: Xuan-Son Nguyen <redacted>
Date: Tue, 2 Dec 2025 10:48:31 +0000 (+0100)
Subject: convert: add error message for mistral3 quantized weight (#17686)
X-Git-Tag: upstream/0.0.7446~214
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=2c453c6c7786df267b3ae6fd5019eee126a35a29;p=pkg%2Fggml%2Fsources%2Fllama.cpp

convert: add error message for mistral3 quantized weight (#17686)
---

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index a54cce88..8ddb6d04 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -2842,6 +2842,10 @@ class Mistral3Model(LlamaModel):
             self.gguf_writer.add_attn_temperature_scale(rope_params["llama_4_scaling_beta"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        # TODO: probably not worth supporting quantized weight, as official BF16 is also available
+        if name.endswith("weight_scale_inv"):
+            raise ValueError("This is a quantized weight, please use BF16 weight instead")
+
         name = name.replace("language_model.", "")
         if "multi_modal_projector" in name or "vision_tower" in name:
             return []