Tiago Caldeira
commited on
Commit
Β·
9f37a6e
1
Parent(s):
44e14ac
different approach using unsloth
Browse files- _app.py +75 -0
- app.py +29 -36
- requirements.txt +5 -4
_app.py
ADDED
|
@@ -0,0 +1,75 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
from transformers import AutoProcessor, Gemma3nForConditionalGeneration
|
| 3 |
+
import torch
|
| 4 |
+
import textwrap
|
| 5 |
+
from huggingface_hub import login
|
| 6 |
+
import os
|
| 7 |
+
|
| 8 |
+
# Log in using the HF token (automatically read from secret)
|
| 9 |
+
hf_token = os.getenv("HUGGINGFACE_HUB_TOKEN")
|
| 10 |
+
login(token=hf_token)
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
# π Load model and processor
|
| 15 |
+
model_id = "google/gemma-3n-e2b-it"
|
| 16 |
+
model_id = "google/gemma-3n-E2B"
|
| 17 |
+
model_id = "lmstudio-community/gemma-3n-E2B-it-MLX-4bit"
|
| 18 |
+
model_id = "unsloth/gemma-3n-E2B-it-unsloth-bnb-4bit"
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
processor = AutoProcessor.from_pretrained(model_id)
|
| 22 |
+
model = Gemma3nForConditionalGeneration.from_pretrained(
|
| 23 |
+
model_id,
|
| 24 |
+
torch_dtype=torch.float32,
|
| 25 |
+
device_map="cpu"
|
| 26 |
+
).eval()
|
| 27 |
+
|
| 28 |
+
# π οΈ Helper to format output
|
| 29 |
+
def print_response(text: str) -> str:
|
| 30 |
+
return "\n".join(textwrap.fill(line, 100) for line in text.split("\n"))
|
| 31 |
+
|
| 32 |
+
# π Inference function for text-only input
|
| 33 |
+
def predict_text(system_prompt: str, user_prompt: str) -> str:
|
| 34 |
+
messages = [
|
| 35 |
+
{"role": "system", "content": [{"type": "text", "text": system_prompt.strip()}]},
|
| 36 |
+
{"role": "user", "content": [{"type": "text", "text": user_prompt.strip()}]},
|
| 37 |
+
]
|
| 38 |
+
|
| 39 |
+
inputs = processor.apply_chat_template(
|
| 40 |
+
messages,
|
| 41 |
+
add_generation_prompt=True,
|
| 42 |
+
tokenize=True,
|
| 43 |
+
return_dict=True,
|
| 44 |
+
return_tensors="pt"
|
| 45 |
+
).to(model.device)
|
| 46 |
+
|
| 47 |
+
input_len = inputs["input_ids"].shape[-1]
|
| 48 |
+
|
| 49 |
+
with torch.inference_mode():
|
| 50 |
+
output = model.generate(
|
| 51 |
+
**inputs,
|
| 52 |
+
max_new_tokens=500,
|
| 53 |
+
do_sample=False,
|
| 54 |
+
use_cache=False # π₯ Fixes CPU bug
|
| 55 |
+
)
|
| 56 |
+
|
| 57 |
+
gen = output[0][input_len:]
|
| 58 |
+
decoded = processor.decode(gen, skip_special_tokens=True)
|
| 59 |
+
return print_response(decoded)
|
| 60 |
+
|
| 61 |
+
# ποΈ Gradio Interface
|
| 62 |
+
demo = gr.Interface(
|
| 63 |
+
fn=predict_text,
|
| 64 |
+
inputs=[
|
| 65 |
+
gr.Textbox(lines=2, label="System Prompt", value="You are a helpful assistant."),
|
| 66 |
+
gr.Textbox(lines=4, label="User Prompt", placeholder="Ask something..."),
|
| 67 |
+
],
|
| 68 |
+
outputs=gr.Textbox(label="Gemma 3n Response"),
|
| 69 |
+
title="Gemma 3n Text-Only Chat",
|
| 70 |
+
description="Interact with the Gemma 3n language model using plain text. Image input not required.",
|
| 71 |
+
)
|
| 72 |
+
|
| 73 |
+
if __name__ == "__main__":
|
| 74 |
+
demo.launch()
|
| 75 |
+
|
app.py
CHANGED
|
@@ -1,63 +1,56 @@
|
|
| 1 |
-
import gradio as gr
|
| 2 |
-
from transformers import AutoProcessor, Gemma3nForConditionalGeneration
|
| 3 |
import torch
|
|
|
|
|
|
|
|
|
|
| 4 |
import textwrap
|
| 5 |
-
from huggingface_hub import login
|
| 6 |
-
import os
|
| 7 |
-
|
| 8 |
-
# Log in using the HF token (automatically read from secret)
|
| 9 |
-
hf_token = os.getenv("HUGGINGFACE_HUB_TOKEN")
|
| 10 |
-
login(token=hf_token)
|
| 11 |
-
|
| 12 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
model_id = "lmstudio-community/gemma-3n-E2B-it-MLX-4bit"
|
| 18 |
-
model_id = "unsloth/gemma-3n-E2B-it-unsloth-bnb-4bit"
|
| 19 |
-
|
| 20 |
-
processor = AutoProcessor.from_pretrained(model_id)
|
| 21 |
-
model = Gemma3nForConditionalGeneration.from_pretrained(
|
| 22 |
-
model_id,
|
| 23 |
-
torch_dtype=torch.float32,
|
| 24 |
-
device_map="cpu"
|
| 25 |
-
).eval()
|
| 26 |
|
| 27 |
-
# π οΈ
|
| 28 |
def print_response(text: str) -> str:
|
| 29 |
return "\n".join(textwrap.fill(line, 100) for line in text.split("\n"))
|
| 30 |
|
| 31 |
-
# π Inference function for
|
| 32 |
def predict_text(system_prompt: str, user_prompt: str) -> str:
|
| 33 |
messages = [
|
| 34 |
{"role": "system", "content": [{"type": "text", "text": system_prompt.strip()}]},
|
| 35 |
{"role": "user", "content": [{"type": "text", "text": user_prompt.strip()}]},
|
| 36 |
]
|
| 37 |
|
| 38 |
-
inputs =
|
| 39 |
messages,
|
| 40 |
add_generation_prompt=True,
|
| 41 |
tokenize=True,
|
| 42 |
return_dict=True,
|
| 43 |
-
return_tensors="pt"
|
| 44 |
-
).to(
|
| 45 |
-
|
| 46 |
-
input_len = inputs["input_ids"].shape[-1]
|
| 47 |
|
| 48 |
with torch.inference_mode():
|
| 49 |
-
|
| 50 |
**inputs,
|
| 51 |
-
max_new_tokens=
|
| 52 |
-
|
| 53 |
-
|
|
|
|
| 54 |
)
|
| 55 |
|
| 56 |
-
|
| 57 |
-
decoded =
|
| 58 |
return print_response(decoded)
|
| 59 |
|
| 60 |
-
# ποΈ Gradio
|
| 61 |
demo = gr.Interface(
|
| 62 |
fn=predict_text,
|
| 63 |
inputs=[
|
|
@@ -66,7 +59,7 @@ demo = gr.Interface(
|
|
| 66 |
],
|
| 67 |
outputs=gr.Textbox(label="Gemma 3n Response"),
|
| 68 |
title="Gemma 3n Text-Only Chat",
|
| 69 |
-
description="Interact with the Gemma 3n language model using plain text.
|
| 70 |
)
|
| 71 |
|
| 72 |
if __name__ == "__main__":
|
|
|
|
|
|
|
|
|
|
| 1 |
import torch
|
| 2 |
+
import gradio as gr
|
| 3 |
+
from unsloth import FastModel
|
| 4 |
+
from transformers import TextStreamer, AutoTokenizer
|
| 5 |
import textwrap
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
|
| 7 |
+
# Load model (4-bit quantized)
|
| 8 |
+
model, tokenizer = FastModel.from_pretrained(
|
| 9 |
+
model_name = "unsloth/gemma-3n-E4B-it",
|
| 10 |
+
dtype = None, # Auto-detect FP16/32
|
| 11 |
+
max_seq_length = 1024,
|
| 12 |
+
load_in_4bit = True,
|
| 13 |
+
full_finetuning = False,
|
| 14 |
+
# token = "hf_..." # Uncomment if model is gated
|
| 15 |
+
)
|
| 16 |
|
| 17 |
+
model.eval()
|
| 18 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 19 |
+
model.to(device)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
+
# π οΈ Format output
|
| 22 |
def print_response(text: str) -> str:
|
| 23 |
return "\n".join(textwrap.fill(line, 100) for line in text.split("\n"))
|
| 24 |
|
| 25 |
+
# π Inference function for Gradio
|
| 26 |
def predict_text(system_prompt: str, user_prompt: str) -> str:
|
| 27 |
messages = [
|
| 28 |
{"role": "system", "content": [{"type": "text", "text": system_prompt.strip()}]},
|
| 29 |
{"role": "user", "content": [{"type": "text", "text": user_prompt.strip()}]},
|
| 30 |
]
|
| 31 |
|
| 32 |
+
inputs = tokenizer.apply_chat_template(
|
| 33 |
messages,
|
| 34 |
add_generation_prompt=True,
|
| 35 |
tokenize=True,
|
| 36 |
return_dict=True,
|
| 37 |
+
return_tensors="pt",
|
| 38 |
+
).to(device)
|
|
|
|
|
|
|
| 39 |
|
| 40 |
with torch.inference_mode():
|
| 41 |
+
outputs = model.generate(
|
| 42 |
**inputs,
|
| 43 |
+
max_new_tokens=256,
|
| 44 |
+
temperature=1.0,
|
| 45 |
+
top_p=0.95,
|
| 46 |
+
top_k=64,
|
| 47 |
)
|
| 48 |
|
| 49 |
+
generated = outputs[0][inputs["input_ids"].shape[-1]:]
|
| 50 |
+
decoded = tokenizer.decode(generated, skip_special_tokens=True)
|
| 51 |
return print_response(decoded)
|
| 52 |
|
| 53 |
+
# ποΈ Gradio UI
|
| 54 |
demo = gr.Interface(
|
| 55 |
fn=predict_text,
|
| 56 |
inputs=[
|
|
|
|
| 59 |
],
|
| 60 |
outputs=gr.Textbox(label="Gemma 3n Response"),
|
| 61 |
title="Gemma 3n Text-Only Chat",
|
| 62 |
+
description="Interact with the Gemma 3n language model using plain text. 4-bit quantized for efficiency.",
|
| 63 |
)
|
| 64 |
|
| 65 |
if __name__ == "__main__":
|
requirements.txt
CHANGED
|
@@ -1,6 +1,7 @@
|
|
| 1 |
-
transformers>=4.42.0
|
| 2 |
-
torch
|
| 3 |
gradio
|
| 4 |
-
accelerate
|
| 5 |
timm
|
| 6 |
-
bitsandbytes
|
|
|
|
|
|
| 1 |
+
#transformers>=4.42.0
|
| 2 |
+
#torch
|
| 3 |
gradio
|
| 4 |
+
#accelerate
|
| 5 |
timm
|
| 6 |
+
#bitsandbytes
|
| 7 |
+
unsloth
|