Update app.py
Browse files
app.py
CHANGED
@@ -1,367 +1,129 @@
|
|
1 |
-
"""
|
2 |
-
This module is used to launch Axolotl with user defined configurations.
|
3 |
-
"""
|
4 |
-
|
5 |
import gradio as gr
|
6 |
-
import
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
model_type
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
adapter
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
""
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
with gr.Column():
|
133 |
-
with gr.Row():
|
134 |
-
gptq = gr.Checkbox(label="GPTQ", value=False)
|
135 |
-
gptq_groupsize = gr.Number(label="GPTQ Groupsize", value=128)
|
136 |
-
gptq_model_v1 = gr.Checkbox(label="GPTQ Model V1", value=False)
|
137 |
-
load_in_8bit = gr.Checkbox(label="Load in 8-bit", value=False)
|
138 |
-
load_in_4bit = gr.Checkbox(label="Load in 4-bit", value=False)
|
139 |
-
with gr.Row():
|
140 |
-
bf16 = gr.Checkbox(label="BF16", value=False)
|
141 |
-
fp16 = gr.Checkbox(label="FP16", value=False)
|
142 |
-
tf32 = gr.Checkbox(label="TF32", value=False)
|
143 |
-
bfloat16 = gr.Checkbox(label="BFloat16", value=False)
|
144 |
-
float16 = gr.Checkbox(label="Float16", value=False)
|
145 |
-
|
146 |
-
with gr.Accordion("GPU & LoRA Settings", open=False):
|
147 |
-
gpu_memory_limit = gr.Textbox(label="GPU Memory Limit")
|
148 |
-
lora_on_cpu = gr.Checkbox(label="LoRA on CPU", value=False)
|
149 |
-
datasets = gr.TextArea(label="Datasets",
|
150 |
-
placeholder="YAML or JSON format for datasets")
|
151 |
-
test_datasets = gr.TextArea(
|
152 |
-
label="Test Datasets",
|
153 |
-
placeholder="YAML or JSON format for test datasets")
|
154 |
-
rl = gr.Textbox(label="RL")
|
155 |
-
chat_template = gr.Textbox(label="Chat Template")
|
156 |
-
default_system_message = gr.Textbox(label="Default System Message")
|
157 |
-
dataset_prepared_path = gr.Textbox(label="Dataset Prepared Path")
|
158 |
-
push_dataset_to_hub = gr.Textbox(label="Push Dataset to Hub")
|
159 |
-
dataset_processes = gr.Number(label="Dataset Processes", value=1)
|
160 |
-
dataset_keep_in_memory = gr.Checkbox(label="Dataset Keep in Memory",
|
161 |
-
value=False)
|
162 |
-
with gr.Row():
|
163 |
-
hub_model_id = gr.Textbox(label="Hub Model ID")
|
164 |
-
hub_strategy = gr.Textbox(label="Hub Strategy")
|
165 |
-
hf_use_auth_token = gr.Checkbox(label="HF Use Auth Token",
|
166 |
-
value=False)
|
167 |
-
with gr.Row():
|
168 |
-
val_set_size = gr.Number(label="Validation Set Size",
|
169 |
-
value=0.04,
|
170 |
-
step=0.01)
|
171 |
-
dataset_shard_num = gr.Number(label="Dataset Shard Num")
|
172 |
-
dataset_shard_idx = gr.Number(label="Dataset Shard Index")
|
173 |
-
|
174 |
-
with gr.Accordion("Training & Evaluation", open=False):
|
175 |
-
with gr.Row():
|
176 |
-
sequence_len = gr.Number(label="Sequence Length", value=2048)
|
177 |
-
pad_to_sequence_len = gr.Checkbox(label="Pad to Sequence Length",
|
178 |
-
value=False)
|
179 |
-
with gr.Row():
|
180 |
-
sample_packing = gr.Checkbox(label="Sample Packing", value=False)
|
181 |
-
eval_sample_packing = gr.Checkbox(label="Eval Sample Packing",
|
182 |
-
value=False)
|
183 |
-
sample_packing_eff_est = gr.Number(label="Sample Packing Eff Est")
|
184 |
-
with gr.Row():
|
185 |
-
total_num_tokens = gr.Number(label="Total Num Tokens")
|
186 |
-
device_map = gr.Textbox(label="Device Map")
|
187 |
-
max_memory = gr.Textbox(label="Max Memory")
|
188 |
-
adapter = gr.Textbox(label="Adapter")
|
189 |
-
with gr.Column():
|
190 |
-
lora_model_dir = gr.Textbox(label="LoRA Model Dir")
|
191 |
-
lora_r = gr.Number(label="LoRA R", value=8)
|
192 |
-
lora_alpha = gr.Number(label="LoRA Alpha", value=16)
|
193 |
-
lora_dropout = gr.Number(label="LoRA Dropout", value=0.05, step=0.01)
|
194 |
-
lora_target_modules = gr.TextArea(label="LoRA Target Modules")
|
195 |
-
lora_target_linear = gr.Checkbox(label="LoRA Target Linear",
|
196 |
-
value=False)
|
197 |
-
lora_modules_to_save = gr.TextArea(label="LoRA Modules to Save")
|
198 |
-
lora_fan_in_fan_out = gr.Checkbox(label="LoRA Fan In Fan Out",
|
199 |
-
value=False)
|
200 |
-
peft = gr.Textbox(label="PEFT")
|
201 |
-
with gr.Row():
|
202 |
-
relora_steps = gr.Number(label="ReLoRA Steps")
|
203 |
-
relora_warmup_steps = gr.Number(label="ReLoRA Warmup Steps")
|
204 |
-
relora_anneal_steps = gr.Number(label="ReLoRA Anneal Steps")
|
205 |
-
relora_prune_ratio = gr.Number(label="ReLoRA Prune Ratio")
|
206 |
-
relora_cpu_offload = gr.Checkbox(label="ReLoRA CPU Offload",
|
207 |
-
value=False)
|
208 |
-
with gr.Row():
|
209 |
-
wandb_mode = gr.Textbox(label="WandB Mode")
|
210 |
-
wandb_project = gr.Textbox(label="WandB Project")
|
211 |
-
wandb_entity = gr.Textbox(label="WandB Entity")
|
212 |
-
wandb_watch = gr.Checkbox(label="WandB Watch", value=False)
|
213 |
-
wandb_name = gr.Textbox(label="WandB Name")
|
214 |
-
wandb_run_id = gr.Textbox(label="WandB Run ID")
|
215 |
-
wandb_log_model = gr.Checkbox(label="WandB Log Model", value=False)
|
216 |
-
with gr.Column():
|
217 |
-
mlflow_tracking_uri = gr.Textbox(label="MLFlow Tracking URI")
|
218 |
-
mlflow_experiment_name = gr.Textbox(label="MLFlow Experiment Name")
|
219 |
-
output_dir = gr.Textbox(label="Output Dir")
|
220 |
-
torch_compile = gr.Checkbox(label="Torch Compile", value=False)
|
221 |
-
torch_compile_backend = gr.Textbox(label="Torch Compile Backend")
|
222 |
-
gradient_accumulation_steps = gr.Number(
|
223 |
-
label="Gradient Accumulation Steps", value=1)
|
224 |
-
micro_batch_size = gr.Number(label="Micro Batch Size", value=2)
|
225 |
-
eval_batch_size = gr.Number(label="Eval Batch Size", value=2)
|
226 |
-
num_epochs = gr.Number(label="Number of Epochs", value=4)
|
227 |
-
warmup_steps = gr.Number(label="Warmup Steps", value=100)
|
228 |
-
warmup_ratio = gr.Number(label="Warmup Ratio")
|
229 |
-
learning_rate = gr.Number(label="Learning Rate",
|
230 |
-
value=0.00003,
|
231 |
-
step=1e-5)
|
232 |
-
lr_quadratic_warmup = gr.Checkbox(label="LR Quadratic Warmup",
|
233 |
-
value=False)
|
234 |
-
logging_steps = gr.Number(label="Logging Steps", value=1)
|
235 |
-
eval_steps = gr.Textbox(label="Eval Steps")
|
236 |
-
evals_per_epoch = gr.Number(label="Evals per Epoch", value=4)
|
237 |
-
save_strategy = gr.Textbox(label="Save Strategy")
|
238 |
-
save_steps = gr.Textbox(label="Save Steps")
|
239 |
-
saves_per_epoch = gr.Number(label="Saves per Epoch", value=1)
|
240 |
-
save_total_limit = gr.Number(label="Save Total Limit")
|
241 |
-
max_steps = gr.Number(label="Max Steps")
|
242 |
-
eval_table_size = gr.Number(label="Eval Table Size")
|
243 |
-
eval_max_new_tokens = gr.Number(label="Eval Max New Tokens",
|
244 |
-
value=128)
|
245 |
-
eval_causal_lm_metrics = gr.TextArea(label="Eval Causal LM Metrics")
|
246 |
-
loss_watchdog_threshold = gr.Number(label="Loss Watchdog Threshold")
|
247 |
-
loss_watchdog_patience = gr.Number(label="Loss Watchdog Patience",
|
248 |
-
value=3)
|
249 |
-
save_safetensors = gr.Checkbox(label="Save SafeTensors", value=False)
|
250 |
-
train_on_inputs = gr.Checkbox(label="Train on Inputs", value=False)
|
251 |
-
group_by_length = gr.Checkbox(label="Group by Length", value=False)
|
252 |
-
gradient_checkpointing = gr.Checkbox(label="Gradient Checkpointing",
|
253 |
-
value=False)
|
254 |
-
early_stopping_patience = gr.Number(label="Early Stopping Patience",
|
255 |
-
value=3)
|
256 |
-
lr_scheduler = gr.Textbox(label="LR Scheduler")
|
257 |
-
lr_scheduler_kwargs = gr.TextArea(label="LR Scheduler KWArgs")
|
258 |
-
cosine_min_lr_ratio = gr.Number(label="Cosine Min LR Ratio")
|
259 |
-
cosine_constant_lr_ratio = gr.Number(
|
260 |
-
label="Cosine Constant LR Ratio")
|
261 |
-
lr_div_factor = gr.Number(label="LR Div Factor")
|
262 |
-
log_sweep_min_lr = gr.Number(label="Log Sweep Min LR")
|
263 |
-
log_sweep_max_lr = gr.Number(label="Log Sweep Max LR")
|
264 |
-
optimizer = gr.Textbox(label="Optimizer")
|
265 |
-
weight_decay = gr.Number(label="Weight Decay", value=0.0, step=0.01)
|
266 |
-
adam_beta1 = gr.Number(label="Adam Beta1", value=0.9, step=0.01)
|
267 |
-
adam_beta2 = gr.Number(label="Adam Beta2", value=0.999, step=0.001)
|
268 |
-
adam_epsilon = gr.Number(label="Adam Epsilon", value=1e-8, step=1e-9)
|
269 |
-
max_grad_norm = gr.Number(label="Max Grad Norm")
|
270 |
-
neftune_noise_alpha = gr.Number(label="NEFTune Noise Alpha")
|
271 |
-
flash_optimum = gr.Checkbox(label="Flash Optimum", value=False)
|
272 |
-
xformers_attention = gr.Checkbox(label="XFormers Attention",
|
273 |
-
value=False)
|
274 |
-
flash_attention = gr.Checkbox(label="Flash Attention", value=False)
|
275 |
-
flash_attn_cross_entropy = gr.Checkbox(
|
276 |
-
label="Flash Attn Cross Entropy", value=False)
|
277 |
-
flash_attn_rms_norm = gr.Checkbox(label="Flash Attn RMS Norm",
|
278 |
-
value=False)
|
279 |
-
flash_attn_fuse_qkv = gr.Checkbox(label="Flash Attn Fuse QKV",
|
280 |
-
value=False)
|
281 |
-
flash_attn_fuse_mlp = gr.Checkbox(label="Flash Attn Fuse MLP",
|
282 |
-
value=False)
|
283 |
-
sdp_attention = gr.Checkbox(label="SDP Attention", value=False)
|
284 |
-
s2_attention = gr.Checkbox(label="S2 Attention", value=False)
|
285 |
-
resume_from_checkpoint = gr.Textbox(label="Resume From Checkpoint")
|
286 |
-
auto_resume_from_checkpoints = gr.Checkbox(
|
287 |
-
label="Auto Resume From Checkpoints", value=False)
|
288 |
-
local_rank = gr.Number(label="Local Rank")
|
289 |
-
special_tokens = gr.TextArea(label="Special Tokens")
|
290 |
-
tokens = gr.TextArea(label="Tokens")
|
291 |
-
fsdp = gr.Checkbox(label="FSDP", value=False)
|
292 |
-
fsdp_config = gr.TextArea(label="FSDP Config")
|
293 |
-
deepspeed = gr.Textbox(label="Deepspeed")
|
294 |
-
ddp_timeout = gr.Number(label="DDP Timeout")
|
295 |
-
ddp_bucket_cap_mb = gr.Number(label="DDP Bucket Cap MB")
|
296 |
-
ddp_broadcast_buffers = gr.Checkbox(label="DDP Broadcast Buffers",
|
297 |
-
value=False)
|
298 |
-
torchdistx_path = gr.Textbox(label="TorchDistX Path")
|
299 |
-
pretraining_dataset = gr.Textbox(label="Pretraining Dataset")
|
300 |
-
debug = gr.Checkbox(label="Debug", value=False)
|
301 |
-
seed = gr.Number(label="Seed", value=42)
|
302 |
-
strict = gr.Checkbox(label="Strict", value=False)
|
303 |
-
|
304 |
-
submit_button = gr.Button("Launch Configuration")
|
305 |
-
output_area = gr.TextArea(label="Configuration Output")
|
306 |
|
307 |
-
submit_button.click(
|
308 |
-
config,
|
309 |
-
inputs=[
|
310 |
-
base_model, base_model_ignore_patterns, base_model_config,
|
311 |
-
model_revision, tokenizer_config, model_type, tokenizer_type,
|
312 |
-
trust_remote_code, tokenizer_use_fast, tokenizer_legacy,
|
313 |
-
resize_token_embeddings_to_32x, is_falcon_derived_model,
|
314 |
-
is_llama_derived_model, is_mistral_derived_model,
|
315 |
-
is_qwen_derived_model, model_config, bnb_config_kwargs, gptq,
|
316 |
-
gptq_groupsize, gptq_model_v1, load_in_8bit, load_in_4bit, bf16,
|
317 |
-
fp16, tf32, bfloat16, float16, gpu_memory_limit, lora_on_cpu,
|
318 |
-
datasets, test_datasets, rl, chat_template, default_system_message,
|
319 |
-
dataset_prepared_path, push_dataset_to_hub, dataset_processes,
|
320 |
-
dataset_keep_in_memory, hub_model_id, hub_strategy,
|
321 |
-
hf_use_auth_token, val_set_size, dataset_shard_num,
|
322 |
-
dataset_shard_idx, sequence_len, pad_to_sequence_len, sample_packing,
|
323 |
-
eval_sample_packing, sample_packing_eff_est, total_num_tokens,
|
324 |
-
device_map, max_memory, adapter, lora_model_dir, lora_r, lora_alpha,
|
325 |
-
lora_dropout, lora_target_modules, lora_target_linear,
|
326 |
-
lora_modules_to_save, lora_fan_in_fan_out, peft, relora_steps,
|
327 |
-
relora_warmup_steps, relora_anneal_steps, relora_prune_ratio,
|
328 |
-
relora_cpu_offload, wandb_mode, wandb_project, wandb_entity,
|
329 |
-
wandb_watch, wandb_name, wandb_run_id, wandb_log_model,
|
330 |
-
mlflow_tracking_uri, mlflow_experiment_name, output_dir,
|
331 |
-
torch_compile, torch_compile_backend, gradient_accumulation_steps,
|
332 |
-
micro_batch_size, eval_batch_size, num_epochs, warmup_steps,
|
333 |
-
warmup_ratio, learning_rate, lr_quadratic_warmup, logging_steps,
|
334 |
-
eval_steps, evals_per_epoch, save_strategy, save_steps,
|
335 |
-
saves_per_epoch, save_total_limit, max_steps, eval_table_size,
|
336 |
-
eval_max_new_tokens, eval_causal_lm_metrics, loss_watchdog_threshold,
|
337 |
-
loss_watchdog_patience, save_safetensors, train_on_inputs,
|
338 |
-
group_by_length, gradient_checkpointing, early_stopping_patience,
|
339 |
-
lr_scheduler, lr_scheduler_kwargs, cosine_min_lr_ratio,
|
340 |
-
cosine_constant_lr_ratio, lr_div_factor, log_sweep_min_lr,
|
341 |
-
log_sweep_max_lr, optimizer, weight_decay, adam_beta1, adam_beta2,
|
342 |
-
adam_epsilon, max_grad_norm, neftune_noise_alpha, flash_optimum,
|
343 |
-
xformers_attention, flash_attention, flash_attn_cross_entropy,
|
344 |
-
flash_attn_rms_norm, flash_attn_fuse_qkv, flash_attn_fuse_mlp,
|
345 |
-
sdp_attention, s2_attention, resume_from_checkpoint,
|
346 |
-
auto_resume_from_checkpoints, local_rank, special_tokens, tokens,
|
347 |
-
fsdp, fsdp_config, deepspeed, ddp_timeout, ddp_bucket_cap_mb,
|
348 |
-
ddp_broadcast_buffers, torchdistx_path, pretraining_dataset, debug,
|
349 |
-
seed, strict
|
350 |
-
],
|
351 |
-
outputs=output_area)
|
352 |
-
"""
|
353 |
-
This section is used to create a configuration file from user text.
|
354 |
-
"""
|
355 |
-
with gr.Tab(label="YML text"):
|
356 |
-
yml_config_text = gr.TextArea(label='YML Config',
|
357 |
-
lines=50,
|
358 |
-
value=example_yml)
|
359 |
-
create_config = gr.Button("Create config")
|
360 |
-
output = gr.TextArea(label="Generated config")
|
361 |
-
create_config.click(
|
362 |
-
yml_config,
|
363 |
-
inputs=[yml_config_text],
|
364 |
-
outputs=output,
|
365 |
-
)
|
366 |
|
367 |
-
demo.launch(share=True)
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
+
import os
|
3 |
+
|
4 |
+
|
5 |
+
class Main:
|
6 |
+
|
7 |
+
async def train_model(self,max_steps, base_model, model_type, tokenizer_type, is_llama_derived_model,
|
8 |
+
strict, datasets_path, dataset_format, shards,
|
9 |
+
val_set_size, output_dir, adapter, lora_model_dir, sequence_len, sample_packing,
|
10 |
+
pad_to_sequence_len, lora_r, lora_alpha, lora_dropout,
|
11 |
+
lora_target_modules, lora_target_linear, lora_fan_in_fan_out, gradient_accumulation_steps,
|
12 |
+
micro_batch_size, num_epochs, optimizer, lr_scheduler, learning_rate, train_on_inputs,
|
13 |
+
group_by_length, bf16, fp16, tf32, gradient_checkpointing,
|
14 |
+
resume_from_checkpoint, local_rank, logging_steps, xformers_attention, flash_attention,
|
15 |
+
load_best_model_at_end, warmup_steps, evals_per_epoch, eval_table_size, saves_per_epoch,
|
16 |
+
debug, weight_decay, wandb_project, wandb_entity, wandb_watch,
|
17 |
+
wandb_name, wandb_log_model,last_tab,progress=gr.Progress(track_tqdm=True)):
|
18 |
+
|
19 |
+
a = [base_model, model_type, tokenizer_type, is_llama_derived_model,
|
20 |
+
strict, datasets_path, dataset_format, shards,
|
21 |
+
val_set_size, output_dir, adapter, lora_model_dir, sequence_len, sample_packing,
|
22 |
+
pad_to_sequence_len, lora_r, lora_alpha, lora_dropout,
|
23 |
+
lora_target_modules, lora_target_linear, lora_fan_in_fan_out, gradient_accumulation_steps,
|
24 |
+
micro_batch_size, num_epochs, optimizer, lr_scheduler, learning_rate, train_on_inputs,
|
25 |
+
group_by_length, bf16, fp16, tf32, gradient_checkpointing,
|
26 |
+
resume_from_checkpoint, local_rank, logging_steps, xformers_attention, flash_attention,
|
27 |
+
load_best_model_at_end, warmup_steps, evals_per_epoch, eval_table_size, saves_per_epoch,
|
28 |
+
debug, weight_decay, wandb_project, wandb_entity, wandb_watch,
|
29 |
+
wandb_name, wandb_log_model,last_tab]
|
30 |
+
|
31 |
+
return a
|
32 |
+
|
33 |
+
|
34 |
+
|
35 |
+
|
36 |
+
|
37 |
+
def initiate_userInterface(self):
|
38 |
+
with gr.Blocks() as self.app:
|
39 |
+
gr.Markdown("### Axolotl UI")
|
40 |
+
|
41 |
+
# Finetuning Tab
|
42 |
+
with gr.Tab("FineTuning UI"):
|
43 |
+
base_model = gr.Dropdown(choices=["NousResearch/Llama-2-7b-hf", "mistralai/Mistral-7B-Instruct-v0.2"], label="Select Model", value="NousResearch/Llama-2-7b-hf")
|
44 |
+
datasets_path = gr.Textbox(label="datasets_path", value="mhenrichsen/alpaca_2k_test")
|
45 |
+
dataset_format = gr.Radio(choices=['Alpaca'], label="Dataset Format", value='Alpaca')
|
46 |
+
shards = gr.Slider(minimum=0, maximum=20, step=1, label="shards", value=10)
|
47 |
+
last_tab = gr.Checkbox(label='last_tab',value=False,visible=False)
|
48 |
+
|
49 |
+
with gr.Accordion("Advanced Settings",open=False):
|
50 |
+
with gr.Tab("YAML Configuration"):
|
51 |
+
model_type = gr.Radio(label="model_type", choices=['MistralForCausalLM','LlamaForCausalLM'],info="",value="LlamaForCausalLM")
|
52 |
+
tokenizer_type = gr.Textbox(label="tokenizer_type", value="LlamaTokenizer",visible=False)
|
53 |
+
is_llama_derived_model = gr.Checkbox(label="is_llama_derived_model", value=True,info="Determines the padding strategy based on the parent type of the model")
|
54 |
+
strict = gr.Checkbox(label="strict", value=False,visible=False)
|
55 |
+
val_set_size = gr.Slider(minimum=0, maximum=1, step=0.1, label="val_set_size", value=0.05,info="Percentage of training data to be used for validation")
|
56 |
+
output_dir = gr.Textbox(label="output_dir", value="./finetune-out",info="Output directory of the finetuned model")
|
57 |
+
adapter = gr.Radio(choices=["qlora", "lora"], label="adapter",value='qlora',info="Parameter efficient training strategy")
|
58 |
+
lora_model_dir = gr.Textbox(label="lora_model_dir",info="Directory of a custom adapter can be provided",visible=False)
|
59 |
+
sequence_len = gr.Slider(minimum=512, maximum=4096, step=10,label="sequence_len", value=1024,info="The maximum length input allowed to train")
|
60 |
+
sample_packing = gr.Checkbox(label="sample_packing", value=True,info="Speeds up data preparation but recommended false for small datasets")
|
61 |
+
pad_to_sequence_len = gr.Checkbox(label="pad_to_sequence_len", value=True, info="Pads the input to match sequence length to avoid memory fragmentation and out of memory issues. Recommended true")
|
62 |
+
# eval_sample_packing = gr.Checkbox(label="eval_sample_packing", value=False)
|
63 |
+
lora_r = gr.Slider(minimum=8, maximum=64, step=2,label="lora_r", value=32,info="The number of parameters in adaptation layers.")
|
64 |
+
lora_alpha = gr.Slider(minimum=8, maximum=64, step=0.1,label="lora_alpha", value=16,info="How much adapted weights affect base model's")
|
65 |
+
lora_dropout = gr.Slider(minimum=0, maximum=1, label="lora_dropout", value=0.05, step=0.01,info="The ratio of weights ignored randomly within adapted weights")
|
66 |
+
lora_target_modules = gr.Textbox(label="lora_target_modules", value="q_proj, v_proj, k_proj",info="All dense layers can be targeted using parameter efficient tuning")
|
67 |
+
lora_target_linear = gr.Checkbox(label="lora_target_linear", value=True,info="Lora Target Modules will be ignored and all linear layers will be used")
|
68 |
+
lora_fan_in_fan_out = gr.Textbox(label="lora_fan_in_fan_out",visible=False)
|
69 |
+
|
70 |
+
gradient_accumulation_steps = gr.Slider(minimum=4, maximum=64, step=1,label="gradient_accumulation_steps", value=4,info="Number of steps required to update the weights with cumulative gradients")
|
71 |
+
micro_batch_size = gr.Slider(minimum=1, maximum=64, step=2,label="micro_batch_size", value=2,info="Number of samples sent to each gpu")
|
72 |
+
num_epochs = gr.Slider(minimum=1, maximum=4, step=1,label="num_epochs", value=1)
|
73 |
+
max_steps = gr.Textbox(label="max_steps",value='1',info="Maximum number of steps to be trained. Overwrites the number of epochs",visible=False)
|
74 |
+
optimizer = gr.Radio(choices=["adamw_hf",'adamw_torch','adamw_torch_fused','adamw_torch_xla','adamw_apex_fused','adafactor','adamw_anyprecision','sgd','adagrad','adamw_bnb_8bit','lion_8bit','lion_32bit','paged_adamw_32bit','paged_adamw_8bit','paged_lion_32bit','paged_lion_8bit'], value="paged_adamw_32bit",label='optimizer',info="Use an optimizer which aligns with the quantization of model")
|
75 |
+
lr_scheduler = gr.Radio(label="lr_scheduler", choices=['one_cycle', 'log_sweep', 'cosine'],value="cosine",info="Determines dynamic learning rate based on current step")
|
76 |
+
learning_rate = gr.Textbox(label="max_learning_rate", value="2e-5",info="")
|
77 |
+
train_on_inputs = gr.Checkbox(label="train_on_inputs", value=False,visible=False)
|
78 |
+
group_by_length = gr.Checkbox(label="group_by_length", value=False,visible=False)
|
79 |
+
bf16 = gr.Checkbox(label="bfloat16", value=False,info="Enable bfloat16 precision for tensors; supported only on Ampere or newer GPUs.")
|
80 |
+
fp16 = gr.Checkbox(label="Half Precision", value=True,info="Enable half precision (FP16) for tensor processing.")
|
81 |
+
tf32 = gr.Checkbox(label="TensorFloat32", value=False,info="Enable TensorFloat32 precision for tensors; supported only on Ampere or newer GPUs.")
|
82 |
+
gradient_checkpointing = gr.Checkbox(label="gradient_checkpointing", value=True,info='',visible=False)
|
83 |
+
resume_from_checkpoint = gr.Textbox(label="resume_from_checkpoint",visible=False)
|
84 |
+
local_rank = gr.Textbox(label="local_rank",visible=False)
|
85 |
+
logging_steps = gr.Slider(minimum=1, maximum=100, step=1,label="logging_steps", value=1,info='',visible=False)
|
86 |
+
xformers_attention = gr.Checkbox(label="xformers_attention", value=False,visible=False)
|
87 |
+
flash_attention = gr.Checkbox(label="flash_attention", value=False,info='',visible=False)
|
88 |
+
load_best_model_at_end = gr.Checkbox(label="load_best_model_at_end", value=False,visible=False)
|
89 |
+
warmup_steps = gr.Slider(minimum=1, maximum=100, step=1,label="warmup_steps", value=10,visible=False)
|
90 |
+
evals_per_epoch = gr.Slider(minimum=1, maximum=100, step=1,label="evals_per_epoch", value=4,info='No. of Evaluation Per Epoch',visible=False)
|
91 |
+
eval_table_size = gr.Textbox(label="eval_table_size",visible=False)
|
92 |
+
saves_per_epoch = gr.Slider(minimum=1, maximum=100, step=1,label="saves_per_epoch", value=1,info='No. of checkpoints to be saved')
|
93 |
+
|
94 |
+
debug = gr.Checkbox(label="debug", value=False,visible=False)
|
95 |
+
|
96 |
+
weight_decay = gr.Number(label="weight_decay", value=0.0,visible=False)
|
97 |
+
wandb_watch = gr.Checkbox(label="wandb_watch", value=False,visible=False)
|
98 |
+
wandb_log_model = gr.Checkbox(label="wandb_log_model", value=False,visible=False)
|
99 |
+
wandb_project = gr.Textbox(label="wandb_project",visible=False)
|
100 |
+
wandb_entity = gr.Textbox(label="wandb_entity",visible=False)
|
101 |
+
wandb_name = gr.Textbox(label="wandb_name",visible=False)
|
102 |
+
|
103 |
+
|
104 |
+
train_btn = gr.Button("Start Training")
|
105 |
+
train_btn.click(
|
106 |
+
self.train_model,
|
107 |
+
inputs=[max_steps, base_model, model_type, tokenizer_type, is_llama_derived_model,
|
108 |
+
strict, datasets_path, dataset_format, shards,
|
109 |
+
val_set_size, output_dir, adapter, lora_model_dir, sequence_len, sample_packing,
|
110 |
+
pad_to_sequence_len, lora_r, lora_alpha, lora_dropout,
|
111 |
+
lora_target_modules, lora_target_linear, lora_fan_in_fan_out, gradient_accumulation_steps,
|
112 |
+
micro_batch_size, num_epochs, optimizer, lr_scheduler, learning_rate, train_on_inputs,
|
113 |
+
group_by_length, bf16, fp16, tf32, gradient_checkpointing,
|
114 |
+
resume_from_checkpoint, local_rank, logging_steps, xformers_attention, flash_attention,
|
115 |
+
load_best_model_at_end, warmup_steps, evals_per_epoch, eval_table_size, saves_per_epoch,
|
116 |
+
debug, weight_decay, wandb_project, wandb_entity, wandb_watch,
|
117 |
+
wandb_name, wandb_log_model,last_tab],
|
118 |
+
outputs=[gr.Textbox(label="Training Output",interactive=False)]
|
119 |
+
)
|
120 |
+
|
121 |
+
return self.app
|
122 |
+
|
123 |
+
|
124 |
+
if __name__ == "__main__":
|
125 |
+
main = Main()
|
126 |
+
app = main.initiate_userInterface()
|
127 |
+
app.queue().launch(share=True,server_name='0.0.0.0')
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
128 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
129 |
|
|