]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commitdiff
convert: add error message for mistral3 quantized weight (#17686)
authorXuan-Son Nguyen <redacted>
Tue, 2 Dec 2025 10:48:31 +0000 (11:48 +0100)
committerGitHub <redacted>
Tue, 2 Dec 2025 10:48:31 +0000 (11:48 +0100)
convert_hf_to_gguf.py

index a54cce887bb747de41c7cf18ccaf7e4ff00f877d..8ddb6d04cd992a473acfba4700918b715ac0f5e2 100755 (executable)
@@ -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 []