Update app.py
Browse files
app.py
CHANGED
@@ -1,73 +1,64 @@
|
|
1 |
import gradio as gr
|
2 |
import spaces
|
3 |
-
import
|
4 |
-
from transformers import Trainer, TrainingArguments
|
5 |
-
from datasets import load_dataset
|
6 |
-
from transformers import (
|
7 |
-
AutoConfig,
|
8 |
-
AutoTokenizer,
|
9 |
-
AutoModelForCausalLM,
|
10 |
-
DataCollatorForLanguageModeling,
|
11 |
-
)
|
12 |
|
13 |
-
|
14 |
-
def run_finetuning():
|
15 |
-
# Load dataset
|
16 |
-
ds = load_dataset("Magpie-Align/Magpie-Reasoning-V2-250K-CoT-Deepseek-R1-Llama-70B")
|
17 |
-
# maybe select a small subset (like 1000 rows) or you'll likely time out
|
18 |
-
ds_small = ds["train"].select(range(1000))
|
19 |
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
config = AutoConfig.from_pretrained(
|
27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
model = AutoModelForCausalLM.from_pretrained(
|
29 |
"wuhp/myr1",
|
30 |
subfolder="myr1",
|
31 |
config=config,
|
32 |
-
torch_dtype=
|
33 |
-
device_map="auto",
|
34 |
trust_remote_code=True
|
35 |
)
|
|
|
|
|
36 |
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
|
|
|
|
|
|
|
|
|
|
41 |
|
42 |
-
|
43 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
44 |
|
45 |
-
|
46 |
-
args = TrainingArguments(
|
47 |
-
output_dir="finetuned_model",
|
48 |
-
num_train_epochs=1,
|
49 |
-
per_device_train_batch_size=1,
|
50 |
-
logging_steps=5,
|
51 |
-
fp16=True,
|
52 |
-
save_strategy="no",
|
53 |
-
)
|
54 |
-
trainer = Trainer(
|
55 |
-
model=model,
|
56 |
-
args=args,
|
57 |
-
train_dataset=ds_small,
|
58 |
-
data_collator=collator,
|
59 |
-
)
|
60 |
-
trainer.train()
|
61 |
-
|
62 |
-
# Save
|
63 |
-
trainer.save_model("finetuned_model")
|
64 |
-
tokenizer.save_pretrained("finetuned_model")
|
65 |
-
return "Finetuning done!"
|
66 |
-
|
67 |
-
# Then define a Gradio UI that calls run_finetuning
|
68 |
with gr.Blocks() as demo:
|
69 |
-
|
70 |
-
|
71 |
-
|
|
|
|
|
|
|
|
|
72 |
|
73 |
demo.launch()
|
|
|
1 |
import gradio as gr
|
2 |
import spaces
|
3 |
+
from transformers import AutoConfig, AutoTokenizer, AutoModelForCausalLM, pipeline
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
|
5 |
+
text_pipeline = None # global var to hold our pipeline once loaded
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
+
@spaces.GPU(duration=120) # request up to 120s GPU time to load the model
|
8 |
+
def load_model():
|
9 |
+
"""
|
10 |
+
This function will run in a *child* process that has GPU allocated.
|
11 |
+
We can safely do device_map="auto" or .to("cuda") here.
|
12 |
+
"""
|
13 |
+
config = AutoConfig.from_pretrained(
|
14 |
+
"wuhp/myr1",
|
15 |
+
subfolder="myr1",
|
16 |
+
trust_remote_code=True
|
17 |
+
)
|
18 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
19 |
+
"wuhp/myr1",
|
20 |
+
subfolder="myr1",
|
21 |
+
trust_remote_code=True
|
22 |
+
)
|
23 |
model = AutoModelForCausalLM.from_pretrained(
|
24 |
"wuhp/myr1",
|
25 |
subfolder="myr1",
|
26 |
config=config,
|
27 |
+
torch_dtype="auto", # triggers GPU usage
|
28 |
+
device_map="auto", # triggers GPU usage
|
29 |
trust_remote_code=True
|
30 |
)
|
31 |
+
text_pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
32 |
+
return text_pipe
|
33 |
|
34 |
+
def ensure_pipeline():
|
35 |
+
"""
|
36 |
+
If we've never loaded the pipeline, call load_model() now.
|
37 |
+
If ZeroGPU has deallocated it, we might need to reload again.
|
38 |
+
"""
|
39 |
+
global text_pipeline
|
40 |
+
if text_pipeline is None:
|
41 |
+
text_pipeline = load_model() # <-- calls the GPU-wrapped function
|
42 |
+
return text_pipeline
|
43 |
|
44 |
+
@spaces.GPU(duration=60) # up to 60s for each generate call
|
45 |
+
def predict(prompt, max_new_tokens=64):
|
46 |
+
"""
|
47 |
+
Called when the user clicks 'Generate'; ensures the model is loaded,
|
48 |
+
then runs inference on GPU.
|
49 |
+
"""
|
50 |
+
pipe = ensure_pipeline()
|
51 |
+
outputs = pipe(prompt, max_new_tokens=int(max_new_tokens))
|
52 |
+
return outputs[0]["generated_text"]
|
53 |
|
54 |
+
# Build the Gradio UI
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
with gr.Blocks() as demo:
|
56 |
+
gr.Markdown("# ZeroGPU Inference Demo")
|
57 |
+
prompt = gr.Textbox(label="Prompt")
|
58 |
+
max_tok = gr.Slider(1, 256, value=64, step=1, label="Max New Tokens")
|
59 |
+
output = gr.Textbox(label="Generated Text")
|
60 |
+
|
61 |
+
generate_btn = gr.Button("Generate")
|
62 |
+
generate_btn.click(fn=predict, inputs=[prompt, max_tok], outputs=output)
|
63 |
|
64 |
demo.launch()
|