[GGML] Support for customizable RoPE (#3083)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
This commit is contained in:
randoentity 2023-07-18 03:32:37 +02:00 committed by GitHub
parent 89e0d15cf5
commit a69955377a
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 9 additions and 1 deletions

View File

@ -97,6 +97,8 @@ class LlamacppHF(PreTrainedModel):
'use_mlock': shared.args.mlock, 'use_mlock': shared.args.mlock,
'low_vram': shared.args.low_vram, 'low_vram': shared.args.low_vram,
'n_gpu_layers': shared.args.n_gpu_layers, 'n_gpu_layers': shared.args.n_gpu_layers,
'rope_freq_base': 10000 * shared.args.alpha_value ** (64/63.),
'rope_freq_scale': 1.0 / shared.args.compress_pos_emb,
'logits_all': True, 'logits_all': True,
} }

View File

@ -50,7 +50,9 @@ class LlamaCppModel:
'use_mmap': not shared.args.no_mmap, 'use_mmap': not shared.args.no_mmap,
'use_mlock': shared.args.mlock, 'use_mlock': shared.args.mlock,
'low_vram': shared.args.low_vram, 'low_vram': shared.args.low_vram,
'n_gpu_layers': shared.args.n_gpu_layers 'n_gpu_layers': shared.args.n_gpu_layers,
'rope_freq_base': 10000 * shared.args.alpha_value ** (64/63.),
'rope_freq_scale': 1.0 / shared.args.compress_pos_emb,
} }
result.model = Llama(**params) result.model = Llama(**params)

View File

@ -37,6 +37,8 @@ loaders_and_params = {
'low_vram', 'low_vram',
'mlock', 'mlock',
'llama_cpp_seed', 'llama_cpp_seed',
'compress_pos_emb',
'alpha_value',
], ],
'llamacpp_HF': [ 'llamacpp_HF': [
'n_ctx', 'n_ctx',
@ -47,6 +49,8 @@ loaders_and_params = {
'low_vram', 'low_vram',
'mlock', 'mlock',
'llama_cpp_seed', 'llama_cpp_seed',
'compress_pos_emb',
'alpha_value',
'llamacpp_HF_info', 'llamacpp_HF_info',
], ],
'Transformers': [ 'Transformers': [