]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
convert-*.py: GGUF Naming Convention Refactor and Metadata Override Refactor (#7499)
authorBrian <redacted>
Thu, 18 Jul 2024 10:40:15 +0000 (20:40 +1000)
committerGitHub <redacted>
Thu, 18 Jul 2024 10:40:15 +0000 (20:40 +1000)
commit672a6f101839a150180fc28891ebed379c8f2353
tree0fe1866907aae3695e938839f383826b6f7f36ba
parent3807c3de04cde853418033c95e96642876545f3e
convert-*.py: GGUF Naming Convention Refactor and Metadata Override Refactor (#7499)

Main thing is that the default output filename will take this form

{name}{parameters}{finetune}{version}{encoding}{kind}

In addition this add and remove some entries in the KV store and adds a metadata class with automatic heuristics capability to derive some values based on model card content

* No Change:
  - Internal GGUF Spec
    - `general.architecture`
    - `general.quantization_version`
    - `general.alignment`
    - `general.file_type`
  - General Model Details
    - `general.name`
    - `general.author`
    - `general.version`
    - `general.description`
  - Licensing details
    - `general.license`
  - Typically represents the converted GGUF repo (Unless made from scratch)
    - `general.url`
  - Model Source during conversion
    - `general.source.url`

* Removed:
  - Model Source during conversion
    - `general.source.huggingface.repository`

* Added:
  - General Model Details
    - `general.organization`
    - `general.finetune`
    - `general.basename`
    - `general.quantized_by`
    - `general.size_label`
  - Licensing details
    - `general.license.name`
    - `general.license.link`
  - Typically represents the converted GGUF repo (Unless made from scratch)
    - `general.doi`
    - `general.uuid`
    - `general.repo_url`
  - Model Source during conversion
    - `general.source.doi`
    - `general.source.uuid`
    - `general.source.repo_url`
  - Base Model Source
    - `general.base_model.count`
    - `general.base_model.{id}.name`
    - `general.base_model.{id}.author`
    - `general.base_model.{id}.version`
    - `general.base_model.{id}.organization`
    - `general.base_model.{id}.url` (Model Website/Paper)
    - `general.base_model.{id}.doi`
    - `general.base_model.{id}.uuid`
    - `general.base_model.{id}.repo_url` (Model Source Repository (git/svn/etc...))
  - Array based KV stores
    - `general.tags`
    - `general.languages`
    - `general.datasets`

---------

Co-authored-by: compilade <redacted>
Co-authored-by: Xuan Son Nguyen <redacted>
13 files changed:
convert_hf_to_gguf.py
convert_lora_to_gguf.py
examples/convert_legacy_llama.py
gguf-py/README.md
gguf-py/gguf/__init__.py
gguf-py/gguf/constants.py
gguf-py/gguf/gguf_writer.py
gguf-py/gguf/metadata.py [new file with mode: 0644]
gguf-py/gguf/utility.py [new file with mode: 0644]
gguf-py/pyproject.toml
gguf-py/tests/__init__.py [new file with mode: 0644]
gguf-py/tests/test_gguf.py [deleted file]
gguf-py/tests/test_metadata.py [new file with mode: 0755]