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

Merge pull request #409 from nkpz/bnb8bit

Browse Source 
Configurable quantization level, fix for broken toggles in model settings
This commit is contained in:
henk717
2023-07-19 14:22:44 +02:00
committed by GitHub
parent 22e7baec52 9581e51476
commit dc4404f29c
2 changed files with 19 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
modeling/inference_models/generic_hf_torch/class.py
View File

@@ -36,13 +36,14 @@ class model_backend(HFTorchInferenceModel):
else:
temp = {}
requested_parameters.append({
"uitype": "toggle",
"unit": "bool",
"label": "Use 4-bit",
"id": "use_4_bit",
"default": temp['use_4_bit'] if 'use_4_bit' in temp else False,
"tooltip": "Whether or not to use BnB's 4-bit mode",
"uitype": "dropdown",
"unit": "text",
"label": "Quantization",
"id": "quantization",
"default": temp['quantization'] if 'quantization' in temp else 'none',
"tooltip": "Whether or not to use BnB's 4-bit or 8-bit mode",
"menu_path": "Layers",
"children": [{'text': 'None', 'value':'none'},{'text': '4-bit', 'value': '4bit'}, {'text': '8-bit', 'value': '8bit'}],
"extra_classes": "",
"refresh_model_inputs": False
})
@@ -52,7 +53,7 @@ class model_backend(HFTorchInferenceModel):
def set_input_parameters(self, parameters):
super().set_input_parameters(parameters)
self.use_4_bit = parameters['use_4_bit'] if 'use_4_bit' in parameters else False
self.quantization = parameters['quantization'] if 'quantization' in parameters else False
def _load(self, save_model: bool, initial_load: bool) -> None:
utils.koboldai_vars.allowsp = True
@@ -82,7 +83,15 @@ class model_backend(HFTorchInferenceModel):
"low_cpu_mem_usage": True,
}
if self.use_4_bit or utils.koboldai_vars.colab_arg:
if self.quantization == "8bit":
tf_kwargs.update({
"quantization_config":BitsAndBytesConfig(
load_in_8bit=True,
llm_int8_enable_fp32_cpu_offload=True
),
})
if self.quantization == "4bit" or utils.koboldai_vars.colab_arg:
tf_kwargs.update({
"quantization_config":BitsAndBytesConfig(
load_in_4bit=True,
@@ -297,7 +306,7 @@ class model_backend(HFTorchInferenceModel):
"disk_layers": self.disk_layers
if "disk_layers" in vars(self)
else 0,
"use_4_bit": self.use_4_bit,
"quantization": self.quantization,
},
f,
indent="",

4
static/koboldai.js
View File

@@ -2011,7 +2011,7 @@ function load_model() {
data = {}
if (settings_area) {
for (const element of settings_area.querySelectorAll(".model_settings_input:not(.hidden)")) {
var element_data = element.value;
var element_data = element.getAttribute("data_type") === "bool" ? element.checked : element.value;
if ((element.tagName == "SELECT") && (element.multiple)) {
element_data = [];
for (var i=0, iLen=element.options.length; i<iLen; i++) {
@@ -2024,8 +2024,6 @@ function load_model() {
element_data = parseInt(element_data);
} else if (element.getAttribute("data_type") == "float") {
element_data = parseFloat(element_data);
} else if (element.getAttribute("data_type") == "bool") {
element_data = (element_data == 'on');
}
}
data[element.id.split("|")[1].replace("_value", "")] = element_data;