llama.cpp: error loading model — bad magic / unsupported GGUF

Environment: llama.cpp and downstream tools (Ollama, LM Studio, koboldcpp) loading a GGUF file built by an incompatible version.

Severity: medium — file is unusable until re-quantized.

• GGUF v1/v2 file loaded by a llama.cpp build that only supports v3+
• Newer GGUF v3 file with new metadata loaded by old llama.cpp
• File is actually a different format renamed to .gguf (e.g. AWQ safetensors)
• Partial download — first bytes are HTML from a CDN error page
• Custom-quantized file with non-standard tensor types your build wasn't compiled with

1. Verify the file isn't corrupted — first 4 bytes should be GGUF:

xxd model.gguf | head -1
# 00000000: 4747 5546 ...   (i.e. "GGUF")

If you see HTML or zeros, redownload.

2. Update llama.cpp — most "bad magic" errors are stale builds:

cd llama.cpp
git pull
make clean
GGML_CUDA=1 make -j

3. Re-quantize from the original safetensors with a current llama-quantize:

# Convert HF model to GGUF F16
python convert_hf_to_gguf.py /path/to/model --outfile model-f16.gguf
# Quantize to your target
./llama-quantize model-f16.gguf model-Q4_K_M.gguf Q4_K_M

4. Pull a known-good GGUF from a maintained repo:

hf download bartowski/Meta-Llama-3.1-8B-Instruct-GGUF \
  Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf

5. If the file came from a vendor build, check their llama.cpp commit pin in the model card and match it:

git checkout <commit-sha>
make clean && make -j