readme detais
Browse files- README.md +15 -1
- app-old.py +0 -78
- finetuning.ipynb +0 -0
README.md
CHANGED
@@ -9,4 +9,18 @@ app_file: app.py
|
|
9 |
pinned: false
|
10 |
---
|
11 |
|
12 |
-
An example chatbot using [Gradio](https://gradio.app), [`huggingface_hub`](https://huggingface.co/docs/huggingface_hub/v0.22.2/en/index), and the [Hugging Face Inference API](https://huggingface.co/docs/api-inference/index).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
pinned: false
|
10 |
---
|
11 |
|
12 |
+
An example chatbot using [Gradio](https://gradio.app), [`huggingface_hub`](https://huggingface.co/docs/huggingface_hub/v0.22.2/en/index), and the [Hugging Face Inference API](https://huggingface.co/docs/api-inference/index).
|
13 |
+
|
14 |
+
### [HuggingFace Space with Quantized LLMs](https://huggingface.co/spaces/Robzy/llm)
|
15 |
+
|
16 |
+
**Baseline model**: Llama-3.2-1B-Instruct with 4-bit quantization
|
17 |
+
|
18 |
+
**Training infrastracture**:
|
19 |
+
* Google Colab with NVIDIA Tesla T4 GPU
|
20 |
+
* Finetuning with parameter-effecient finetuning (PEFT) by low-rank adaption (LORA) using Unsloth and HuggingFace's supervised finetuning libraries.
|
21 |
+
* Weight & Biases for model training monitoring and model checkpointing. Checkpointing every 10 steps.
|
22 |
+
|
23 |
+
**Finetuning details**
|
24 |
+
|
25 |
+
**Datasets**:
|
26 |
+
* [Code instructions Alpaca 120k](https://huggingface.co/datasets/iamtarun/code_instructions_120k_alpaca)
|
app-old.py
DELETED
@@ -1,78 +0,0 @@
|
|
1 |
-
from llama_cpp import Llama
|
2 |
-
import gradio as gr
|
3 |
-
|
4 |
-
llm = Llama.from_pretrained(
|
5 |
-
repo_id="Robzy/Llama-3.2-1B-Instruct-Finetuned-q4_k_m",
|
6 |
-
filename="unsloth.Q4_K_M.gguf",
|
7 |
-
)
|
8 |
-
|
9 |
-
llm2 = Llama.from_pretrained(
|
10 |
-
repo_id="Robzy/Llama-3.2-1B-Instruct-Finetuned-16bit",
|
11 |
-
filename="unsloth.F16.gguf",
|
12 |
-
)
|
13 |
-
|
14 |
-
def predict(message, history):
|
15 |
-
messages = [{"role": "system", "content": "You are a helpful assistant."}]
|
16 |
-
for user_message, bot_message in history:
|
17 |
-
if user_message:
|
18 |
-
messages.append({"role": "user", "content": user_message})
|
19 |
-
if bot_message:
|
20 |
-
messages.append({"role": "assistant", "content": bot_message})
|
21 |
-
messages.append({"role": "user", "content": message})
|
22 |
-
|
23 |
-
response = ""
|
24 |
-
for chunk in llm.create_chat_completion(
|
25 |
-
stream=True,
|
26 |
-
messages=messages,
|
27 |
-
):
|
28 |
-
part = chunk["choices"][0]["delta"].get("content", None)
|
29 |
-
if part:
|
30 |
-
response += part
|
31 |
-
yield response
|
32 |
-
|
33 |
-
|
34 |
-
def predict2(message, history):
|
35 |
-
messages = [{"role": "system", "content": "You are a helpful assistant."}]
|
36 |
-
for user_message, bot_message in history:
|
37 |
-
if user_message:
|
38 |
-
messages.append({"role": "user", "content": user_message})
|
39 |
-
if bot_message:
|
40 |
-
messages.append({"role": "assistant", "content": bot_message})
|
41 |
-
messages.append({"role": "user", "content": message})
|
42 |
-
|
43 |
-
response = ""
|
44 |
-
for chunk in llm2.create_chat_completion(
|
45 |
-
stream=True,
|
46 |
-
messages=messages,
|
47 |
-
):
|
48 |
-
part = chunk["choices"][0]["delta"].get("content", None)
|
49 |
-
if part:
|
50 |
-
response += part
|
51 |
-
yield response
|
52 |
-
|
53 |
-
|
54 |
-
chat1 = gr.ChatInterface(predict, title="4-bit")
|
55 |
-
chat2 = gr.ChatInterface(predict2, title="16-bit")
|
56 |
-
chat3 = gr.ChatInterface(predict2, title="16-bit")
|
57 |
-
|
58 |
-
def update_chat(value):
|
59 |
-
if value == "4-bit":
|
60 |
-
chat1.render(visible=True)
|
61 |
-
chat2.render(visible=False)
|
62 |
-
chat3.render(visible=False)
|
63 |
-
elif value == "16-bit":
|
64 |
-
chat1.render(visible=False)
|
65 |
-
chat2.render(visible=True)
|
66 |
-
chat3.render(visible=False)
|
67 |
-
else:
|
68 |
-
chat1.render(visible=False)
|
69 |
-
chat2.render(visible=False)
|
70 |
-
chat3.render(visible=True)
|
71 |
-
|
72 |
-
with gr.Blocks() as demo:
|
73 |
-
|
74 |
-
gr.Markdown("# Quantized Llama Comparison for Code Generation")
|
75 |
-
dropdown = gr.Dropdown(["4-bit", "16-bit", "32-bit"], label="Choose model version", value="4-bit")
|
76 |
-
dropdown.change(fn=update_chat, inputs=dropdown, outputs=[chat1, chat2, chat3])
|
77 |
-
|
78 |
-
demo.launch()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
finetuning.ipynb
ADDED
The diff for this file is too large to render.
See raw diff
|
|