vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Improve help text, expand warning

Browse source
This commit is contained in:
KerfuffleV2 2023-08-20 13:15:01 -06:00
parent ff25134390
commit 80912f0741
1 changed files with 4 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

8
convert-llama-ggmlv3-to-gguf.py
View file

@ -227,19 +227,19 @@ def handle_args():
parser.add_argument('--input', '-i', help = 'Input GGMLv3 filename') parser.add_argument('--input', '-i', help = 'Input GGMLv3 filename')
parser.add_argument('--output', '-o', help ='Output GGUF filename') parser.add_argument('--output', '-o', help ='Output GGUF filename')
parser.add_argument('--gqa', type = int, default = 1, help = 'grouped-query attention factor (use 8 for LLaMA2 70B)') parser.add_argument('--gqa', type = int, default = 1, help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
parser.add_argument('--eps', default = '5.0e-06', help = 'RMS norm eps (use 1e-5 for LLaMA2)') parser.add_argument('--eps', default = '5.0e-06', help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
parser.add_argument('--context-length', '-c', type=int, default = 2048, help = 'Default max context length') parser.add_argument('--context-length', '-c', type=int, default = 2048, help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
return parser.parse_args() return parser.parse_args()
def main(): def main():
cfg = handle_args() cfg = handle_args()
print(f'* Using config: {cfg}') print(f'* Using config: {cfg}')
print('\n=== WARNING === Be aware that this conversion script is best-effort. Use a native GGUF model if possible. === WARNING ===\n') print('\n=== WARNING === Be aware that this conversion script is best-effort. Special tokens may not be converted correctly. Use a native GGUF model if possible. === WARNING ===\n')
data = np.memmap(cfg.input, mode = 'r') data = np.memmap(cfg.input, mode = 'r')
model = GGMLV3Model() model = GGMLV3Model()
print('* Scanning GGML input file') print('* Scanning GGML input file')
offset = model.load(data, 0) offset = model.load(data, 0)
print(model.hyperparameters) print(f'* GGML model hyperparameters: {model.hyperparameters}')
converter = GGMLToGGUF(model, data, cfg) converter = GGMLToGGUF(model, data, cfg)
converter.save() converter.save()
print(f'* Successful completion. Output saved to: {cfg.output}') print(f'* Successful completion. Output saved to: {cfg.output}')