Mirrors/text-generation-webui
1
0
Fork 0
mirror of https://github.com/oobabooga/text-generation-webui.git synced 2024-11-24 17:06:53 +01:00
Code Issues Packages Projects Releases Wiki Activity

Read more GGUF metadata (scale_linear and freq_base) (#3877)

Browse Source
This commit is contained in:
Gennadij 2023-09-12 23:02:42 +03:00 committed by GitHub
parent 90fca6a77d
commit 460c40d8ab
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 9 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
models/config.yaml
View File

@ -226,4 +226,4 @@ llama-65b-gptq-3bit:
.*codellama: .*codellama:
rope_freq_base: 1000000 rope_freq_base: 1000000
.*codellama.*instruct: .*codellama.*instruct:
instruction_template: 'Llama-v2' instruction_template: 'Llama-v2'

3
modules/metadata_gguf.py
View File

@ -70,6 +70,9 @@ def load_metadata(fname):
GGUF_VERSION = struct.unpack("<I", file.read(4))[0] GGUF_VERSION = struct.unpack("<I", file.read(4))[0]
ti_data_count = struct.unpack("<Q", file.read(8))[0] ti_data_count = struct.unpack("<Q", file.read(8))[0]
kv_data_count = struct.unpack("<Q", file.read(8))[0] kv_data_count = struct.unpack("<Q", file.read(8))[0]
if GGUF_VERSION == 1:
raise Exception('You are using an outdated GGUF, please download a new one.')
for i in range(kv_data_count): for i in range(kv_data_count):
key_length = struct.unpack("<Q", file.read(8))[0] key_length = struct.unpack("<Q", file.read(8))[0]

5
modules/models_settings.py
View File

@ -17,6 +17,7 @@ def get_fallback_settings():
'truncation_length': shared.settings['truncation_length'], 'truncation_length': shared.settings['truncation_length'],
'n_ctx': 2048, 'n_ctx': 2048,
'rope_freq_base': 0, 'rope_freq_base': 0,
'compress_pos_emb': 1,
} }
@ -48,6 +49,10 @@ def get_model_metadata(model):
metadata = metadata_gguf.load_metadata(model_file) metadata = metadata_gguf.load_metadata(model_file)
if 'llama.context_length' in metadata: if 'llama.context_length' in metadata:
model_settings['n_ctx'] = metadata['llama.context_length'] model_settings['n_ctx'] = metadata['llama.context_length']
if 'llama.rope.scale_linear' in metadata:
model_settings['compress_pos_emb'] = metadata['llama.rope.scale_linear']
if 'llama.rope.freq_base' in metadata:
model_settings['rope_freq_base'] = metadata['llama.rope.freq_base']
return model_settings return model_settings