mmarch/text-generation-webui
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Add llama-2-70b GGML support (#3285)

Browse source
This commit is contained in:
oobabooga 2023-07-24 16:37:03 -03:00 • committed by GitHub
parent 6f4830b4d3
commit a07d070b6c
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
8 changed files with 20 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

2
modules/llamacpp_hf.py
View file

@ -106,6 +106,8 @@ class LlamacppHF(PreTrainedModel):
'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,
'n_gqa': shared.args.n_gqa or None,
'rms_norm_eps': shared.args.rms_norm_eps or None,
'logits_all': True,
}

2
modules/llamacpp_model.py
View file

@ -53,6 +53,8 @@ class LlamaCppModel:
'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,
'n_gqa': shared.args.n_gqa or None,
'rms_norm_eps': shared.args.rms_norm_eps or None,
}
result.model = Llama(**params)

4
modules/loaders.py
View file

@ -30,6 +30,8 @@ loaders_and_params = {
],
'llama.cpp': [
'n_ctx',
'n_gqa',
'rms_norm_eps',
'n_gpu_layers',
'n_batch',
'threads',
@ -42,6 +44,8 @@ loaders_and_params = {
],
'llamacpp_HF': [
'n_ctx',
'n_gqa',
'rms_norm_eps',
'n_gpu_layers',
'n_batch',
'threads',

2
modules/shared.py
View file

@ -127,6 +127,8 @@ parser.add_argument('--cache-capacity', type=str, help='Maximum cache capacity.
parser.add_argument('--n-gpu-layers', type=int, default=0, help='Number of layers to offload to the GPU.')
parser.add_argument('--n_ctx', type=int, default=2048, help='Size of the prompt context.')
parser.add_argument('--llama_cpp_seed', type=int, default=0, help='Seed for llama-cpp models. Default 0 (random)')
parser.add_argument('--n_gqa', type=int, default=0, help='grouped-query attention. Must be 8 for llama2 70b.')
parser.add_argument('--rms_norm_eps', type=float, default=0, help='Must be 1e-5 for llama2 70b.')
# GPTQ
parser.add_argument('--wbits', type=int, default=0, help='Load a pre-quantized model with specified precision in bits. 2, 3, 4 and 8 are supported.')

2
modules/ui.py
View file

@ -61,6 +61,8 @@ def list_model_elements():
'mlock',
'n_gpu_layers',
'n_ctx',
'n_gqa',
'rms_norm_eps',
'llama_cpp_seed',
'gpu_split',
'max_seq_len',