README.md

metadata

license: mit
tags:
  - merge

Quantizations of BigCodeLLama LFG 🚀

Experimental CodeLlaMA frankenstein of 70b instruct, python and base to see how it benchmarks

Models Merged

The following models were included in the merge:

• ../CodeLlama-70b-hf
• ../CodeLlama-70b-Instruct-hf
• ../CodeLlama-70b-Python-hf

Configuration

The following YAML configuration was used to produce this model:

dtype: bfloat16
merge_method: passthrough
slices:
- sources:
  - layer_range: [0, 69]
    model:
      model:
        path: ../CodeLlama-70b-hf
- sources:
  - layer_range: [66, 76]
    model:
      model:
        path: ../CodeLlama-70b-Instruct-hf
- sources:
  - layer_range: [42, 66]
    model:
      model:
        path: ../CodeLlama-70b-hf
- sources:
  - layer_range: [13, 37]
    model:
      model:
        path: ../CodeLlama-70b-Python-hf
- sources:
  - layer_range: [10, 80]
    model:
      model:
        path: ../CodeLlama-70b-Instruct-hf

To reunite each file after downloading do

cat BigCodeLlama-169b-q2k.gguf.part0 BigCodeLlama-169b-q2k.gguf.part1 > BigCodeLlama-169b-q2k.gguf

cat BigCodeLlama-169b-q3km.gguf.part0 BigCodeLlama-169b-q3km.gguf.part1 > BigCodeLlama-169b-q3km.gguf

cat BigCodeLlama-169b-q4ks.gguf.part0 BigCodeLlama-169b-q4ks.gguf.part1 > BigCodeLlama-169b-q4ks.gguf

cat BigCodeLlama-169b-q5km.gguf.part0 BigCodeLlama-169b-q5km.gguf.part1 BigCodeLlama-169b-q5km.gguf.part2 > BigCodeLlama-169b-q5km.gguf

cat BigCodeLlama-169b-q8.gguf.part0 BigCodeLlama-169b-q8.gguf.part1 BigCodeLlama-169b-q8.gguf.part2 BigCodeLlama-169b-q8.gguf.part3 > BigCodeLlama-169b-q8.gguf