Heuristic:Ggml org Llama cpp Warning Deprecated Legacy Converters
| Knowledge Sources | |
|---|---|
| Domains | Model_Conversion, Migration |
| Last Updated | 2026-02-15 00:00 GMT |
Overview
Deprecation warning for legacy model conversion tools that convert from obsolete formats (GGML, llama2.c) to GGUF.
Description
Several model conversion scripts in the llama.cpp repository handle legacy formats that are being phased out. The primary conversion path is now convert_hf_to_gguf.py (HuggingFace to GGUF). The following tools exist primarily for backward compatibility:
- convert_llama_ggml_to_gguf.py — Converts old GGML/GGJT format files to GGUF. The script itself warns: "Be aware that this conversion script is best-effort. Use a native GGUF model if possible."
- convert_legacy_llama.py — Legacy conversion logic now superseded by the modern convert_hf_to_gguf.py pipeline
- convert-llama2c-to-ggml — Converts from Karpathy's llama2.c format, an external legacy format
- convert_hf_to_gguf_update.py — Developer-only maintenance tool, not for end users
Additionally, the pocs/vdot/ directory contains proof-of-concept benchmarks (vdot, q8dot) that are experimental code, not production tools.
Usage
Consult this heuristic when encountering any of the legacy conversion tools. Prefer convert_hf_to_gguf.py for all new model conversions. Only use legacy converters when working with models that are exclusively available in old formats.
The Insight (Rule of Thumb)
- Action: Always prefer convert_hf_to_gguf.py for new conversions
- Value: Legacy converters should only be used as a last resort for old-format models
- Trade-off: Legacy converters may produce lower-quality results or miss features supported by the modern pipeline
Reasoning
The llama.cpp project has standardized on GGUF as the model format. Older formats (GGML, GGJT, llama2.c) are no longer actively developed. The legacy conversion scripts remain for backward compatibility but receive limited maintenance. Users should convert models from their original HuggingFace format using the primary conversion pipeline whenever possible.