|
import gradio as gr |
|
import requests |
|
from PIL import Image |
|
import io |
|
from typing import Any, Tuple |
|
import os |
|
|
|
|
|
class Client: |
|
def __init__(self, server_url: str): |
|
self.server_url = server_url |
|
|
|
def send_request(self, task_name: str, model_name: str, text: str, normalization_type: str) -> Tuple[Any, str]: |
|
response = requests.post( |
|
self.server_url, |
|
json={ |
|
"task_name": task_name, |
|
"model_name": model_name, |
|
"text": text, |
|
"normalization_type": normalization_type |
|
}, |
|
timeout=60 |
|
) |
|
if response.status_code == 200: |
|
response_data = response.json() |
|
img_data = bytes.fromhex(response_data["image"]) |
|
img = Image.open(io.BytesIO(img_data)) |
|
return img, "OK" |
|
else: |
|
return "Error, please retry", "Error: Could not get response from server" |
|
|
|
client = Client(f"http://{os.environ['SERVER']}/predict") |
|
|
|
def get_layerwise_nonlinearity(task_name: str, model_name: str, text: str, normalization_type: str) -> Tuple[Any, str]: |
|
return client.send_request(task_name, model_name, text, normalization_type) |
|
|
|
def update_output(task_name: str, model_name: str, text: str, normalization_type: str) -> Tuple[Any]: |
|
img, _ = get_layerwise_nonlinearity(task_name, model_name, text, normalization_type) |
|
return img |
|
|
|
def set_default(task_name: str) -> str: |
|
if task_name in ["Layer wise non-linearity", "Next-token prediction from intermediate representations", "Tokenwise loss without i-th layer"]: |
|
return "token-wise" |
|
return "global" |
|
|
|
def check_normalization(task_name: str, normalization_name) -> Tuple[str]: |
|
if task_name == "Contextualization measurement" and normalization_name == "token-wise": |
|
return "global" |
|
return normalization_name |
|
|
|
def update_description(task_name: str) -> str: |
|
descriptions = { |
|
"Layer wise non-linearity": "Non-linearity per layer: shows how complex each layer's transformation is. Red = more nonlinear.", |
|
"Next-token prediction from intermediate representations": "Layerwise token prediction: when does the model start guessing correctly?", |
|
"Contextualization measurement": "Context stored in each token: how well can the model reconstruct the previous context?", |
|
"Layerwise predictions (logit lens)": "Logit lens: what does each layer believe the next token should be?", |
|
"Tokenwise loss without i-th layer": "Layer ablation: how much does performance drop if a layer is removed?" |
|
} |
|
return descriptions.get(task_name, "βΉοΈ No description available.") |
|
|
|
with gr.Blocks() as demo: |
|
|
|
gr.Markdown("# π¬ LLM-Microscope β A Look Inside the Black Box") |
|
gr.Markdown("Select a model, analysis mode, and input β then peek inside the black box of an LLM to see which layers matter most, which tokens carry the most memory, and how predictions evolve.") |
|
|
|
with gr.Row(): |
|
model_selector = gr.Dropdown( |
|
choices=[ |
|
"facebook/opt-1.3b", |
|
"TheBloke/Llama-2-7B-fp16", |
|
"Qwen/Qwen3-8B" |
|
], |
|
value="facebook/opt-1.3b", |
|
label="Select Model" |
|
) |
|
task_selector = gr.Dropdown( |
|
choices=[ |
|
"Layer wise non-linearity", |
|
"Next-token prediction from intermediate representations", |
|
"Contextualization measurement", |
|
"Layerwise predictions (logit lens)", |
|
"Tokenwise loss without i-th layer" |
|
], |
|
value="Layer wise non-linearity", |
|
label="Select Mode" |
|
) |
|
normalization_selector = gr.Dropdown( |
|
choices=["global", "token-wise"], |
|
value="token-wise", |
|
label="Select Normalization" |
|
) |
|
|
|
task_description = gr.Markdown("βΉοΈ Choose a mode to see what it does.") |
|
|
|
with gr.Column(): |
|
text_message = gr.Textbox(label="Enter your input text:", value="I love to live my life") |
|
submit = gr.Button("Submit") |
|
box_for_plot = gr.Image(label="Visualization", type="pil") |
|
|
|
with gr.Accordion("π More Info and Explanation", open=False): |
|
gr.Markdown(""" |
|
This heatmap shows **how each token is processed** across layers of a language model. Here's how to read it: |
|
|
|
- **Rows**: layers of the model (bottom = deeper) |
|
- **Columns**: input tokens |
|
- **Colors**: intensity of effect (depends on the selected metric) |
|
|
|
--- |
|
|
|
### Metrics explained: |
|
|
|
- `Layer wise non-linearity`: how nonlinear the transformation is at each layer (red = more nonlinear). |
|
- `Next-token prediction from intermediate representations`: shows which layers begin to make good predictions. |
|
- `Contextualization measurement`: tokens with more context info get lower scores (green = more context). |
|
- `Layerwise predictions (logit lens)`: tracks how the modelβs guesses evolve at each layer. |
|
- `Tokenwise loss without i-th layer`: shows how much each token depends on a specific layer. Red means performance drops if we skip this layer. |
|
|
|
Use this tool to **peek inside the black box** β it reveals which layers matter most, which tokens carry the most memory, and how LLMs evolve their predictions. |
|
|
|
--- |
|
|
|
You can also use `llm-microscope` as a Python library to run these analyses on **your own models and data**. |
|
|
|
Just install it with: `pip install llm-microscope` |
|
|
|
More details provided in [GitHub repo](https://github.com/AIRI-Institute/LLM-Microscope). |
|
""") |
|
|
|
task_selector.change(fn=update_description, inputs=[task_selector], outputs=[task_description]) |
|
task_selector.select(set_default, [task_selector], [normalization_selector]) |
|
normalization_selector.select(check_normalization, [task_selector, normalization_selector], [normalization_selector]) |
|
submit.click( |
|
fn=update_output, |
|
inputs=[task_selector, model_selector, text_message, normalization_selector], |
|
outputs=[box_for_plot] |
|
) |
|
|
|
if __name__ == "__main__": |
|
demo.launch(share=True, server_port=7860, server_name="0.0.0.0") |
|
|