rixty/KoboldAI-Client
1
1
Fork 0
mirror of https://github.com/KoboldAI/KoboldAI-Client.git synced 2025-06-05 21:59:24 +02:00
Code Issues Projects Releases Wiki Activity

Add gpt_bigcode support, fix 8-bit GPTQ incoherence

Browse Source
This commit is contained in:
0cc4m
2023-05-23 06:59:28 +02:00
parent 47b371b9d3
commit 12df8220fb
3 changed files with 7 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
docs/gptq-whl-links.html
View File

@@ -11,3 +11,6 @@
<a href="https://github.com/0cc4m/GPTQ-for-LLaMa/releases/download/2023-05-19-2/gptq_koboldai-0.0.4-cp38-cp38-linux_x86_64.whl">gptq_koboldai-0.0.4-cp38-cp38-linux_x86_64.whl</a> <a href="https://github.com/0cc4m/GPTQ-for-LLaMa/releases/download/2023-05-19-2/gptq_koboldai-0.0.4-cp38-cp38-linux_x86_64.whl">gptq_koboldai-0.0.4-cp38-cp38-linux_x86_64.whl</a>
<a href="https://github.com/0cc4m/GPTQ-for-LLaMa/releases/download/2023-05-19-2/gptq_koboldai-0.0.4-cp38-cp38-win_amd64.whl">gptq_koboldai-0.0.4-cp38-cp38-win_amd64.whl</a> <a href="https://github.com/0cc4m/GPTQ-for-LLaMa/releases/download/2023-05-19-2/gptq_koboldai-0.0.4-cp38-cp38-win_amd64.whl">gptq_koboldai-0.0.4-cp38-cp38-win_amd64.whl</a>
<a href="https://github.com/0cc4m/GPTQ-for-LLaMa/releases/download/2023-05-23/gptq_koboldai-0.0.5-cp38-cp38-linux_x86_64.whl">gptq_koboldai-0.0.5-cp38-cp38-linux_x86_64.whl</a>
<a href="https://github.com/0cc4m/GPTQ-for-LLaMa/releases/download/2023-05-23/gptq_koboldai-0.0.5-cp38-cp38-win_amd64.whl">gptq_koboldai-0.0.5-cp38-cp38-win_amd64.whl</a>

2
environments/huggingface.yml
View File

@@ -47,6 +47,6 @@ dependencies:
- diffusers - diffusers
- git+https://github.com/0cc4m/hf_bleeding_edge/ - git+https://github.com/0cc4m/hf_bleeding_edge/
- --find-links=https://0cc4m.github.io/KoboldAI/gptq-whl-links.html - --find-links=https://0cc4m.github.io/KoboldAI/gptq-whl-links.html
- gptq_koboldai==0.0.4 - gptq_koboldai==0.0.5
- einops - einops
- peft==0.3.0 - peft==0.3.0

3
modeling/inference_models/gptq_hf_torch/class.py
View File

@@ -35,6 +35,7 @@ from gptq.gptj import load_quant as gptj_load_quant
from gptq.gptneox import load_quant as gptneox_load_quant from gptq.gptneox import load_quant as gptneox_load_quant
from gptq.llama import load_quant as llama_load_quant from gptq.llama import load_quant as llama_load_quant
from gptq.opt import load_quant as opt_load_quant from gptq.opt import load_quant as opt_load_quant
from gptq.bigcode import load_quant as bigcode_load_quant
from gptq.mpt import load_quant as mpt_load_quant from gptq.mpt import load_quant as mpt_load_quant
from gptq.offload import load_quant_offload from gptq.offload import load_quant_offload
@@ -220,6 +221,8 @@ class model_backend(HFTorchInferenceModel):
model = load_quant_offload(opt_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, self.gpu_layers_list) model = load_quant_offload(opt_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, self.gpu_layers_list)
elif model_type == "mpt": elif model_type == "mpt":
model = load_quant_offload(mpt_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, self.gpu_layers_list) model = load_quant_offload(mpt_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, self.gpu_layers_list)
elif model_type == "gpt_bigcode":
model = load_quant_offload(bigcode_load_quant, location, path_4bit, utils.koboldai_vars.gptq_bits, groupsize, self.gpu_layers_list).half()
elif autogptq_support: elif autogptq_support:
# Monkey patch in hf_bleeding_edge to avoid having to trust remote code # Monkey patch in hf_bleeding_edge to avoid having to trust remote code
auto_gptq.modeling._utils.AutoConfig = hf_bleeding_edge.AutoConfig auto_gptq.modeling._utils.AutoConfig = hf_bleeding_edge.AutoConfig