Sebastien De Greef
commited on
Commit
•
6baccb3
1
Parent(s):
4af8a78
adds "gradio" to the requirements.txt and handle buttons up to training
Browse files- .gitignore +1 -0
- app.py +278 -18
- requirements.txt +4 -2
- unsloth.png +0 -0
.gitignore
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
.venv/*
|
app.py
CHANGED
@@ -1,32 +1,292 @@
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
|
3 |
-
|
4 |
-
|
5 |
-
|
6 |
-
|
|
|
|
|
7 |
# Dropdown options
|
8 |
-
model_options = [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
|
10 |
# Create the Gradio interface
|
11 |
with gr.Blocks() as demo:
|
12 |
-
gr.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
|
14 |
-
|
15 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
|
17 |
-
# Checkboxes
|
18 |
-
checkbox1 = gr.Checkbox(label="Checkbox 1")
|
19 |
-
checkbox2 = gr.Checkbox(label="Checkbox 2")
|
20 |
|
21 |
-
|
22 |
-
|
23 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
|
|
25 |
# Output
|
26 |
-
output = gr.Textbox(label="Output")
|
27 |
|
28 |
-
|
29 |
-
|
30 |
-
submit_btn.click(process_input, inputs=[model_name, checkbox1, checkbox2, text1, text2], outputs=output)
|
31 |
|
32 |
demo.launch()
|
|
|
1 |
import gradio as gr
|
2 |
+
from huggingface_hub import HfApi
|
3 |
+
from unsloth import FastLanguageModel
|
4 |
+
from trl import SFTTrainer
|
5 |
+
from transformers import TrainingArguments, TrainerCallback
|
6 |
+
from unsloth import is_bfloat16_supported
|
7 |
+
import torch
|
8 |
+
from datasets import load_dataset
|
9 |
+
import logging
|
10 |
+
from io import StringIO
|
11 |
+
import time
|
12 |
+
import asyncio
|
13 |
+
# Create a string stream to capture log messages
|
14 |
+
log_stream = StringIO()
|
15 |
|
16 |
+
# Configure logging to use the string stream
|
17 |
+
logging.basicConfig(stream=log_stream, level=logging.DEBUG, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
|
18 |
+
logger = logging.getLogger(__name__)
|
19 |
+
log_contents = log_stream.getvalue()
|
20 |
+
print(log_contents)
|
21 |
+
logger.debug('This is a debug message')
|
22 |
# Dropdown options
|
23 |
+
model_options = [
|
24 |
+
"unsloth/mistral-7b-v0.3-bnb-4bit", # New Mistral v3 2x faster!
|
25 |
+
"unsloth/mistral-7b-instruct-v0.3-bnb-4bit",
|
26 |
+
"unsloth/llama-3-8b-bnb-4bit", # Llama-3 15 trillion tokens model 2x faster!
|
27 |
+
"unsloth/llama-3-8b-Instruct-bnb-4bit",
|
28 |
+
"unsloth/llama-3-70b-bnb-4bit",
|
29 |
+
"unsloth/Phi-3-mini-4k-instruct", # Phi-3 2x faster!
|
30 |
+
"unsloth/Phi-3-medium-4k-instruct",
|
31 |
+
"unsloth/mistral-7b-bnb-4bit",
|
32 |
+
"unsloth/gemma-2-9b-bnb-4bit",
|
33 |
+
"unsloth/gemma-2-27b-bnb-4bit", # Gemma 2x faster!
|
34 |
+
]
|
35 |
+
gpu_stats = torch.cuda.get_device_properties(0)
|
36 |
+
start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
|
37 |
+
max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
|
38 |
+
|
39 |
+
model=None
|
40 |
+
tokenizer = None
|
41 |
+
dataset = None
|
42 |
+
max_seq_length = 2048
|
43 |
+
|
44 |
+
class PrinterCallback(TrainerCallback):
|
45 |
+
step = 0
|
46 |
+
def __init__(self, progress):
|
47 |
+
self.progress = progress
|
48 |
+
def on_log(self, args, state, control, logs=None, **kwargs):
|
49 |
+
_ = logs.pop("total_flos", None)
|
50 |
+
if state.is_local_process_zero:
|
51 |
+
print(logs)
|
52 |
+
def on_step_end(self, args, state, control, **kwargs):
|
53 |
+
if state.is_local_process_zero:
|
54 |
+
self.step = state.global_step
|
55 |
+
self.progress.update(self.step)
|
56 |
+
print("**Step ", state.global_step)
|
57 |
+
|
58 |
+
|
59 |
+
|
60 |
+
def formatting_prompts_func(examples, prompt):
|
61 |
+
EOS_TOKEN = tokenizer.eos_token # Must add EOS_TOKEN
|
62 |
+
instructions = examples["instruction"]
|
63 |
+
inputs = examples["input"]
|
64 |
+
outputs = examples["output"]
|
65 |
+
texts = []
|
66 |
+
for instruction, input, output in zip(instructions, inputs, outputs):
|
67 |
+
# Must add EOS_TOKEN, otherwise your generation will go on forever!
|
68 |
+
text = prompt.format(instruction, input, output) + EOS_TOKEN
|
69 |
+
texts.append(text)
|
70 |
+
return { "text" : texts, }
|
71 |
+
pass
|
72 |
+
|
73 |
+
|
74 |
+
def load_model(initial_model_name, load_in_4bit, max_sequence_length):
|
75 |
+
global model, tokenizer, max_seq_length
|
76 |
+
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
|
77 |
+
max_seq_length = max_sequence_length
|
78 |
+
model, tokenizer = FastLanguageModel.from_pretrained(
|
79 |
+
model_name = initial_model_name,
|
80 |
+
max_seq_length = max_sequence_length,
|
81 |
+
dtype = dtype,
|
82 |
+
load_in_4bit = load_in_4bit,
|
83 |
+
# token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
|
84 |
+
)
|
85 |
+
log_contents = log_stream.getvalue()
|
86 |
+
print(log_contents)
|
87 |
+
return f"Model {initial_model_name} loaded, using {max_sequence_length} as max sequence length.", gr.update(visible=True, interactive=True), gr.update(interactive=True),gr.update(interactive=False), gr.update(interactive=False), gr.update(interactive=False)
|
88 |
+
|
89 |
+
def load_data(dataset_name, data_template_style, data_template):
|
90 |
+
global dataset
|
91 |
+
dataset = load_dataset(dataset_name, split = "train")
|
92 |
+
dataset = dataset.map(lambda examples: formatting_prompts_func(examples, data_template), batched=True)
|
93 |
+
return f"Data loaded {len(dataset)} records loaded.", gr.update(visible=True, interactive=True), gr.update(visible=True, interactive=True)
|
94 |
+
|
95 |
+
|
96 |
+
|
97 |
+
|
98 |
+
async def train_model(model_name: str, lora_r: int, lora_alpha: int, lora_dropout: float, per_device_train_batch_size: int, warmup_steps: int, max_steps: int,
|
99 |
+
gradient_accumulation_steps: int, logging_steps: int, log_to_tensorboard: bool, optim, learning_rate, weight_decay, lr_scheduler_type, seed: int, output_dir, progress= gr.Progress()):
|
100 |
+
global model, tokenizer
|
101 |
+
print(f"$$$ Training model {model_name} with {lora_r} R, {lora_alpha} alpha, {lora_dropout} dropout, {per_device_train_batch_size} per device train batch size, {warmup_steps} warmup steps, {max_steps} max steps, {gradient_accumulation_steps} gradient accumulation steps, {logging_steps} logging steps, {log_to_tensorboard} log to tensorboard, {optim} optimizer, {learning_rate} learning rate, {weight_decay} weight decay, {lr_scheduler_type} lr scheduler type, {seed} seed, {output_dir} output dir")
|
102 |
+
iseed = seed
|
103 |
+
model = FastLanguageModel.get_peft_model(
|
104 |
+
model,
|
105 |
+
r = lora_r,
|
106 |
+
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
|
107 |
+
"gate_proj", "up_proj", "down_proj",],
|
108 |
+
lora_alpha = lora_alpha,
|
109 |
+
lora_dropout = lora_dropout,
|
110 |
+
bias = "none",
|
111 |
+
use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
|
112 |
+
random_state=iseed,
|
113 |
+
use_rslora = False, # We support rank stabilized LoRA
|
114 |
+
loftq_config = None, # And LoftQ
|
115 |
+
)
|
116 |
+
|
117 |
+
trainer = SFTTrainer(
|
118 |
+
model = model,
|
119 |
+
tokenizer = tokenizer,
|
120 |
+
train_dataset = dataset,
|
121 |
+
dataset_text_field = "text",
|
122 |
+
max_seq_length = max_seq_length,
|
123 |
+
dataset_num_proc = 2,
|
124 |
+
packing = False, # Can make training 5x faster for short sequences.
|
125 |
+
callbacks = [PrinterCallback(progress)],
|
126 |
+
args = TrainingArguments(
|
127 |
+
per_device_train_batch_size = per_device_train_batch_size,
|
128 |
+
gradient_accumulation_steps = gradient_accumulation_steps,
|
129 |
+
warmup_steps = warmup_steps,
|
130 |
+
max_steps = 60, # Set num_train_epochs = 1 for full training runs
|
131 |
+
learning_rate = learning_rate,
|
132 |
+
fp16 = not is_bfloat16_supported(),
|
133 |
+
bf16 = is_bfloat16_supported(),
|
134 |
+
logging_steps = logging_steps,
|
135 |
+
optim = "adamw_8bit",
|
136 |
+
weight_decay = weight_decay,
|
137 |
+
lr_scheduler_type = "linear",
|
138 |
+
seed = iseed,
|
139 |
+
report_to="tensorboard" if log_to_tensorboard else None,
|
140 |
+
output_dir = output_dir
|
141 |
+
),
|
142 |
+
)
|
143 |
+
trainer.train()
|
144 |
+
return "Model training",gr.update(visible=True, interactive=False), gr.update(visible=True, interactive=True), gr.update(interactive=True)
|
145 |
+
|
146 |
+
def save_model():
|
147 |
+
return "Model saved", gr.update(visible=True, interactive=True), gr.update(visible=True, interactive=False), gr.update(interactive=False)
|
148 |
+
|
149 |
|
150 |
# Create the Gradio interface
|
151 |
with gr.Blocks() as demo:
|
152 |
+
with gr.Column():
|
153 |
+
gr.Image("unsloth.png", width="300px", interactive=False, show_download_button=False, show_label=False)
|
154 |
+
with gr.Column():
|
155 |
+
gr.Markdown(f"**GPU Information:** {gpu_stats.name} ({max_memory} GB)")
|
156 |
+
with gr.Tab("Base Model Parameters"):
|
157 |
+
|
158 |
+
with gr.Row():
|
159 |
+
initial_model_name = gr.Dropdown(choices=model_options, label="Select Base Model", allow_custom_value=True)
|
160 |
+
load_in_4bit = gr.Checkbox(label="Load 4bit model", value=True)
|
161 |
+
|
162 |
+
gr.Markdown("### Target Model Parameters")
|
163 |
+
with gr.Row():
|
164 |
+
max_sequence_length = gr.Slider(minimum=128, value=512, step=64, maximum=128*1024, interactive=True, label="Max Sequence Length")
|
165 |
+
load_btn = gr.Button("Load")
|
166 |
+
output = gr.Textbox(label="Model Load Status", value="Model not loaded", interactive=False)
|
167 |
+
gr.Markdown("---")
|
168 |
+
|
169 |
+
with gr.Tab("Data Preparation"):
|
170 |
+
with gr.Row():
|
171 |
+
dataset_name = gr.Textbox(label="Dataset Name", value="yahma/alpaca-cleaned")
|
172 |
+
data_template_style = gr.Dropdown(label="Template", choices=["alpaca","custom"], value="alpaca", allow_custom_value=True)
|
173 |
+
with gr.Row():
|
174 |
+
data_tempalte = gr.TextArea(label="Data Template", value="""Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
|
175 |
+
|
176 |
+
### Instruction:
|
177 |
+
{}
|
178 |
+
|
179 |
+
### Input:
|
180 |
+
{}
|
181 |
+
|
182 |
+
### Response:
|
183 |
+
{}""")
|
184 |
+
gr.Markdown("---")
|
185 |
+
output_load_data = gr.Textbox(label="Data Load Status", value="Data not loaded", interactive=False)
|
186 |
+
load_data_btn = gr.Button("Load Dataset", interactive=True)
|
187 |
+
load_data_btn.click(load_data, inputs=[dataset_name, data_template_style, data_tempalte], outputs=[output_load_data, load_data_btn])
|
188 |
+
|
189 |
+
with gr.Tab("Fine-Tuning"):
|
190 |
+
gr.Markdown("""### Fine-Tuned Model Parameters""")
|
191 |
+
with gr.Row():
|
192 |
+
model_name = gr.Textbox(label="Model Name", value=initial_model_name.value, interactive=True)
|
193 |
+
|
194 |
+
gr.Markdown("""### Lora Parameters""")
|
195 |
+
|
196 |
+
with gr.Row():
|
197 |
+
lora_r = gr.Number(label="R", value=16, interactive=True)
|
198 |
+
lora_alpha = gr.Number(label="Lora Alpha", value=16, interactive=True)
|
199 |
+
lora_dropout = gr.Number(label="Lora Dropout", value=0.1, interactive=True)
|
200 |
+
|
201 |
+
gr.Markdown("---")
|
202 |
+
gr.Markdown("""### Training Parameters""")
|
203 |
+
with gr.Row():
|
204 |
+
with gr.Column():
|
205 |
+
with gr.Row():
|
206 |
+
per_device_train_batch_size = gr.Number(label="Per Device Train Batch Size", value=2, interactive=True)
|
207 |
+
warmup_steps = gr.Number(label="Warmup Steps", value=5, interactive=True)
|
208 |
+
max_steps = gr.Number(label="Max Steps", value=60, interactive=True)
|
209 |
+
gradient_accumulation_steps = gr.Number(label="Gradient Accumulation Steps", value=4, interactive=True)
|
210 |
+
with gr.Row():
|
211 |
+
logging_steps = gr.Number(label="Logging Steps", value=1, interactive=True)
|
212 |
+
log_to_tensorboard = gr.Checkbox(label="Log to Tensorboard", value=True, interactive=True)
|
213 |
+
|
214 |
+
with gr.Row():
|
215 |
+
optim = gr.Dropdown(choices=["adamw_8bit", "adamw", "sgd"], label="Optimizer", value="adamw_8bit")
|
216 |
+
learning_rate = gr.Number(label="Learning Rate", value=2e-4, interactive=True)
|
217 |
+
|
218 |
+
with gr.Row():
|
219 |
+
weight_decay = gr.Number(label="Weight Decay", value=0.01, interactive=True)
|
220 |
+
lr_scheduler_type = gr.Dropdown(choices=["linear", "cosine", "constant"], label="LR Scheduler Type", value="linear")
|
221 |
+
gr.Markdown("---")
|
222 |
+
|
223 |
+
with gr.Row():
|
224 |
+
seed = gr.Number(label="Seed", value=3407, interactive=True)
|
225 |
+
output_dir = gr.Textbox(label="Output Directory", value="outputs", interactive=True)
|
226 |
+
gr.Markdown("---")
|
227 |
+
|
228 |
+
train_output = gr.Textbox(label="Training Status", value="Model not trained", interactive=False)
|
229 |
+
train_btn = gr.Button("Train", visible=True)
|
230 |
+
train_btn.click(train_model, inputs=[model_name, lora_r, lora_alpha, lora_dropout, per_device_train_batch_size, warmup_steps, max_steps, gradient_accumulation_steps, logging_steps, log_to_tensorboard, optim, learning_rate, weight_decay, lr_scheduler_type, seed, output_dir], outputs=[train_output, train_btn])
|
231 |
+
|
232 |
+
with gr.Tab("Save & Push Options"):
|
233 |
+
|
234 |
+
|
235 |
+
|
236 |
+
with gr.Row():
|
237 |
+
gr.Markdown("### Merging Options")
|
238 |
+
with gr.Column():
|
239 |
+
merge_16bit = gr.Checkbox(label="Merge to 16bit", value=False, interactive=True)
|
240 |
+
merge_4bit = gr.Checkbox(label="Merge to 4bit", value=False, interactive=True)
|
241 |
+
just_lora = gr.Checkbox(label="Just LoRA Adapter", value=False, interactive=True)
|
242 |
+
gr.Markdown("---")
|
243 |
+
|
244 |
+
with gr.Row():
|
245 |
+
gr.Markdown("### GGUF Options")
|
246 |
+
with gr.Column():
|
247 |
+
merge_16bit = gr.Checkbox(label="Quantize to f16", value=False, interactive=True)
|
248 |
+
merge_16bit = gr.Checkbox(label="Quantize to 8bit (Q8_0)", value=False, interactive=True)
|
249 |
+
merge_16bit = gr.Checkbox(label="Quantize to 4bit (q4_k_m)", value=False, interactive=True)
|
250 |
+
with gr.Column():
|
251 |
+
merge_custom = gr.Checkbox(label="Custom", value=False, interactive=True)
|
252 |
+
merge_custom_value = gr.Textbox(label="", value="Q5_K", interactive=True)
|
253 |
+
gr.Markdown("---")
|
254 |
|
255 |
+
with gr.Row():
|
256 |
+
gr.Markdown("### Hugging Face Hub Options")
|
257 |
+
push_to_hub = gr.Checkbox(label="Push to Hub", value=False, interactive=True)
|
258 |
+
with gr.Column():
|
259 |
+
hub_model_name = gr.Textbox(label="Hub Model Name", value=f"username/model_name", interactive=True)
|
260 |
+
hub_token = gr.Textbox(label="Hub Token", interactive=True, type="password")
|
261 |
+
ollama_pub_key = gr.Button("HuggingFace Access Token")
|
262 |
+
gr.Markdown("---")
|
263 |
+
|
264 |
+
with gr.Row():
|
265 |
+
gr.Markdown("### Ollama options")
|
266 |
+
with gr.Column():
|
267 |
+
ollama_create_local = gr.Checkbox(label="Create in Ollama (local)", value=False, interactive=True)
|
268 |
+
ollama_push_to_hub = gr.Checkbox(label="Push to Ollama", value=False, interactive=True)
|
269 |
+
with gr.Column():
|
270 |
+
ollama_model_name = gr.Textbox(label="Ollama Model Name", value="user/model_name")
|
271 |
+
ollama_pub_key = gr.Button("Ollama Pub Key")
|
272 |
+
gr.Markdown("---")
|
273 |
|
|
|
|
|
|
|
274 |
|
275 |
+
with gr.Tab("Inference"):
|
276 |
+
with gr.Row():
|
277 |
+
gr.Textbox(label="Input Text", lines=4, value="""\
|
278 |
+
Continue the fibonnaci sequence.
|
279 |
+
# instruction
|
280 |
+
1, 1, 2, 3, 5, 8
|
281 |
+
# input
|
282 |
+
""", interactive=True)
|
283 |
+
gr.Textbox(label="Output Text", lines=4, value="""\
|
284 |
+
""", interactive=False)
|
285 |
|
286 |
+
inference_button = gr.Button("Inference", visible=False, interactive=False)
|
287 |
# Output
|
|
|
288 |
|
289 |
+
# Button click events
|
290 |
+
load_btn.click(load_model, inputs=[initial_model_name, load_in_4bit, max_sequence_length], outputs=[output, load_btn, train_btn, initial_model_name, load_in_4bit, max_sequence_length])
|
|
|
291 |
|
292 |
demo.launch()
|
requirements.txt
CHANGED
@@ -1,6 +1,8 @@
|
|
1 |
unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git
|
2 |
-
xformers
|
3 |
trl<0.9.0
|
4 |
peft
|
5 |
accelerate
|
6 |
-
bitsandbytes
|
|
|
|
|
|
1 |
unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git
|
2 |
+
xformers<0.0.27
|
3 |
trl<0.9.0
|
4 |
peft
|
5 |
accelerate
|
6 |
+
bitsandbytes
|
7 |
+
gradio
|
8 |
+
tensorboard
|
unsloth.png
ADDED