app.py

import gradio as gr
import bittensor as bt
import typing
from bittensor.extrinsics.serving import get_metadata
from dataclasses import dataclass
import requests
import wandb
import math
import os
import datetime
import time
import functools
import multiprocessing
from dotenv import load_dotenv
from huggingface_hub import HfApi
from apscheduler.schedulers.background import BackgroundScheduler
from tqdm import tqdm

load_dotenv()

FONT = """<link href="https://fonts.cdnfonts.com/css/jmh-typewriter" rel="stylesheet">"""
TITLE = """<h1 align="center" id="space-title" class="typewriter">Subnet 6 Leaderboard</h1>"""
IMAGE = """<a href="https://discord.gg/jqVphNsB4H" target="_blank"><img src="https://i.ibb.co/88wyVQ7/nousgirl.png" alt="nousgirl" style="margin: auto; width: 20%; border: 0;" /></a>"""
HEADER = """<h2 align="center" class="typewriter"><a href="https://github.com/NousResearch/finetuning-subnet" target="_blank">Subnet 6</a> is a <a href="https://bittensor.com/" target="_blank">Bittensor</a> subnet that incentivizes the creation of the best open models by evaluating submissions on a constant stream of newly generated synthetic GPT-4 data. The models with the best <a href="https://github.com/NousResearch/finetuning-subnet/blob/master/docs/validator.md" target="_blank">head-to-head loss</a> on the evaluation data receive a steady emission of TAO.</h3>"""
EVALUATION_DETAILS = """<b>Name</b> is the 🤗 Hugging Face model name (click to go to the model card). <b>Rewards / Day</b> are the expected rewards per day for each model. <b>Perplexity</b> is represents the loss on all of the evaluation data for the model as calculated by the validator (lower is better). <b>UID</b> is the Bittensor user id of the submitter. <b>Block</b> is the Bittensor block that the model was submitted in. More stats on <a href="https://taostats.io/subnets/netuid-6/" target="_blank">taostats</a>."""
EVALUATION_HEADER = """<h3 align="center">Shows the latest internal evaluation statistics as calculated by a validator run by Nous Research</h3>"""
VALIDATOR_WANDB_PROJECT = os.environ["VALIDATOR_WANDB_PROJECT"]
H4_TOKEN = os.environ.get("H4_TOKEN", None)
API = HfApi(token=H4_TOKEN)
REPO_ID = "NousResearch/finetuning_subnet_leaderboard"
METAGRAPH_RETRIES = 10
METAGRAPH_DELAY_SECS = 30
METADATA_TTL = 10
NETUID = 6
SUBNET_START_BLOCK = 2225782
SECONDS_PER_BLOCK = 12
SUBTENSOR = os.environ.get("SUBTENSOR", "finney")

@dataclass
class Competition:
    id: str
    name: str

COMPETITIONS = [Competition(id="m1", name="mistral-7b"), Competition(id="g1", name="gemma-2b")]
DEFAULT_COMPETITION_ID = "m1"

def run_in_subprocess(func: functools.partial, ttl: int) -> typing.Any:
    """Runs the provided function on a subprocess with 'ttl' seconds to complete.

    Args:
        func (functools.partial): Function to be run.
        ttl (int): How long to try for in seconds.

    Returns:
        Any: The value returned by 'func'
    """

    def wrapped_func(func: functools.partial, queue: multiprocessing.Queue):
        try:
            result = func()
            queue.put(result)
        except (Exception, BaseException) as e:
            # Catch exceptions here to add them to the queue.
            queue.put(e)

    # Use "fork" (the default on all POSIX except macOS), because pickling doesn't seem
    # to work on "spawn".
    ctx = multiprocessing.get_context("fork")
    queue = ctx.Queue()
    process = ctx.Process(target=wrapped_func, args=[func, queue])

    process.start()

    process.join(timeout=ttl)

    if process.is_alive():
        process.terminate()
        process.join()
        raise TimeoutError(f"Failed to {func.func.__name__} after {ttl} seconds")

    # Raises an error if the queue is empty. This is fine. It means our subprocess timed out.
    result = queue.get(block=False)

    # If we put an exception on the queue then raise instead of returning.
    if isinstance(result, Exception):
        raise result
    if isinstance(result, BaseException):
        raise Exception(f"BaseException raised in subprocess: {str(result)}")

    return result


def get_subtensor_and_metagraph() -> typing.Tuple[bt.subtensor, bt.metagraph]:
    for i in range(0, METAGRAPH_RETRIES):
        try:
            print("Connecting to subtensor...")
            subtensor: bt.subtensor = bt.subtensor(SUBTENSOR)
            print("Pulling metagraph...")
            metagraph: bt.metagraph = subtensor.metagraph(NETUID, lite=False)
            return subtensor, metagraph
        except:
            if i == METAGRAPH_RETRIES - 1:
                raise
            print(f"Error connecting to subtensor or pulling metagraph, retry {i + 1} of {METAGRAPH_RETRIES} in {METAGRAPH_DELAY_SECS} seconds...")
            time.sleep(METAGRAPH_DELAY_SECS)
    raise RuntimeError()

@dataclass
class ModelData:
    uid: int
    hotkey: str
    namespace: str
    name: str
    commit: str
    hash: str
    block: int
    incentive: float
    emission: float
    competition: str

    @classmethod
    def from_compressed_str(cls, uid: int, hotkey: str, cs: str, block: int, incentive: float, emission: float):
        """Returns an instance of this class from a compressed string representation"""
        tokens = cs.split(":")
        return ModelData(
            uid=uid,
            hotkey=hotkey,
            namespace=tokens[0],
            name=tokens[1],
            commit=tokens[2] if tokens[2] != "None" else "",
            hash=tokens[3] if tokens[3] != "None" else "",
            competition=tokens[4] if len(tokens) > 4 and tokens[4] != "None" else DEFAULT_COMPETITION_ID,
            block=block,
            incentive=incentive,
            emission=emission
        )

def get_tao_price() -> float:
    for i in range(0, METAGRAPH_RETRIES):
        try:
            return float(requests.get("https://api.kucoin.com/api/v1/market/stats?symbol=TAO-USDT").json()["data"]["last"])
        except:
                if i == METAGRAPH_RETRIES - 1:
                    raise
                time.sleep(METAGRAPH_DELAY_SECS)
        raise RuntimeError()

def get_validator_weights(metagraph: bt.metagraph) -> typing.Dict[int, typing.Tuple[float, int, typing.Dict[int, float]]]:
    ret = {}
    for uid in metagraph.uids.tolist():
        vtrust = metagraph.validator_trust[uid].item()
        if vtrust > 0:
            ret[uid] = (vtrust, metagraph.S[uid].item(), {})
            for ouid in metagraph.uids.tolist():
                if ouid == uid:
                    continue
                weight = round(metagraph.weights[uid][ouid].item(), 4)
                if weight > 0:
                    ret[uid][-1][ouid] = weight
    return ret

def get_subnet_data(subtensor: bt.subtensor, metagraph: bt.metagraph) -> typing.List[ModelData]:
    result = []
    for uid in tqdm(metagraph.uids.tolist(), desc="Metadata for hotkeys"):
        hotkey = metagraph.hotkeys[uid]
        try:
            # Wrap calls to the subtensor in a subprocess with a timeout to handle potential hangs.
            partial = functools.partial(get_metadata, subtensor, metagraph.netuid, hotkey)
            metadata = run_in_subprocess(partial, METADATA_TTL)
        except KeyboardInterrupt:
            raise
        except:
            metadata = None
        if not metadata:
            continue

        commitment = metadata["info"]["fields"][0]
        hex_data = commitment[list(commitment.keys())[0]][2:]
        chain_str = bytes.fromhex(hex_data).decode()
        block = metadata["block"]
        incentive = metagraph.incentive[uid].nan_to_num().item()
        emission = metagraph.emission[uid].nan_to_num().item() * 20 # convert to daily TAO

        model_data = None
        try:
            model_data = ModelData.from_compressed_str(uid, hotkey, chain_str, block, incentive, emission)    
        except:
            continue

        result.append(model_data)
    return result

def floatable(x) -> bool:
    return (isinstance(x, float) and not math.isnan(x) and not math.isinf(x)) or isinstance(x, int)
    
def get_float_score(key: str, history, competition_id: str) -> typing.Tuple[typing.Optional[float], bool]:
    if key in history and "competition_id" in history:
        data = list(history[key])
        if len(data) > 0:
            competitions = list(history["competition_id"])
            while True:
                if competitions.pop() != competition_id:
                    data.pop()
                    continue
                if floatable(data[-1]):
                    return float(data[-1]), True
                else:
                    data = [float(x) for x in data if floatable(x)]
                    if len(data) > 0:
                        return float(data[-1]), False
                break
    return None, False

def get_sample(uid, history, competition_id: str) -> typing.Optional[typing.Tuple[str, str, str]]:
    prompt_key = f"sample_prompt_data.{uid}"
    response_key = f"sample_response_data.{uid}"
    truth_key = f"sample_truth_data.{uid}"
    if prompt_key in history and response_key in history and truth_key in history and "competition_id" in history:
        competitions = list(history["competition_id"])
        prompts = list(history[prompt_key])
        responses = list(history[response_key])
        truths = list(history[truth_key])
        while True:
            prompt = prompts.pop()
            response = responses.pop()
            truth = truths.pop()
            if competitions.pop() != competition_id:
                continue
            if isinstance(prompt, str) and isinstance(response, str) and isinstance(truth, str):
                return prompt, response, truth
            break
    return None

def get_scores(uids: typing.List[int], competition_id: str) -> typing.Dict[int, typing.Dict[str, typing.Optional[float | str]]]:
    api = wandb.Api()
    runs = list(api.runs(VALIDATOR_WANDB_PROJECT))

    result = {}
    for run in runs:
        history = run.history()
        for uid in uids:
            if uid in result.keys():
                continue
            perplexity, perplexity_fresh = get_float_score(f"perplexity_data.{uid}", history, competition_id)
            win_rate, win_rate_fresh = get_float_score(f"win_rate_data.{uid}", history, competition_id)
            win_total, win_total_fresh = get_float_score(f"win_total_data.{uid}", history, competition_id)
            weight, weight_fresh = get_float_score(f"weight_data.{uid}", history, competition_id)
            sample = get_sample(uid, history, competition_id)
            result[uid] = {
                "perplexity": perplexity,
                "win_rate": win_rate,
                "win_total": win_total,
                "weight": weight,
                "sample": sample,
                "fresh": perplexity_fresh and win_rate_fresh and win_total_fresh
            }
        if len(result.keys()) == len(uids):
            break
    return result

def format_score(uid, scores, key) -> typing.Optional[float]:
    if uid in scores:
        if key in scores[uid]:
            point = scores[uid][key]
            if floatable(point):
                return round(scores[uid][key], 4)
    return None

def next_tempo(start_block, tempo, block):
    start_num = start_block + tempo
    intervals = (block - start_num) // tempo
    nearest_num = start_num + ((intervals + 1) * tempo)
    return nearest_num

subtensor, metagraph = get_subtensor_and_metagraph()

tao_price = get_tao_price()

leaderboard_df = get_subnet_data(subtensor, metagraph)
leaderboard_df.sort(key=lambda x: x.incentive, reverse=True)

competition_scores = {
    y.id: get_scores([x.uid for x in leaderboard_df if x.competition == y.id], y.id)
    for y in COMPETITIONS
}
    
current_block = metagraph.block.item()
next_update = next_tempo(
    SUBNET_START_BLOCK,
    subtensor.get_subnet_hyperparameters(NETUID).tempo,
    current_block
)
blocks_to_go = next_update - current_block
current_time = datetime.datetime.now()
next_update_time = current_time + datetime.timedelta(seconds=blocks_to_go * SECONDS_PER_BLOCK)

validator_df = get_validator_weights(metagraph)
weight_keys = set()
for uid, stats in validator_df.items():
    weight_keys.update(stats[-1].keys())

def get_next_update():
    now = datetime.datetime.now()
    delta = next_update_time - now
    return f"""<div align="center" style="font-size: larger;">Next reward update: <b>{blocks_to_go}</b> blocks (~{int(delta.total_seconds() // 60)} minutes)</div>"""

def leaderboard_data(show_stale: bool, scores: typing.Dict[int, typing.Dict[str, typing.Optional[float | str]]], competition_id: str):
    value = [
        [
            f'[{c.namespace}/{c.name} ({c.commit[0:8]}, UID={c.uid})](https://huggingface.co/{c.namespace}/{c.name}/commit/{c.commit})',
            format_score(c.uid, scores, "win_rate"),
            format_score(c.uid, scores, "perplexity"),
            format_score(c.uid, scores, "weight"),
            c.uid,
            c.block
        ] for c in leaderboard_df if c.competition == competition_id and (scores[c.uid]["fresh"] or show_stale)
    ]
    return value

demo = gr.Blocks(css=".typewriter {font-family: 'JMH Typewriter', sans-serif;}")
with demo:
    gr.HTML(FONT)
    gr.HTML(TITLE)
    gr.HTML(IMAGE)
    gr.HTML(HEADER)

    gr.HTML(value=get_next_update())

    with gr.Tabs():
        for competition in COMPETITIONS:
            with gr.Tab(competition.name):
                scores = competition_scores[competition.id]
                print(scores)

                class_denominator = sum(leaderboard_df[i].incentive for i in range(0, 10) if leaderboard_df[i].incentive and leaderboard_df[i].competition == competition.id)

                class_values = {
                    f"{leaderboard_df[i].namespace}/{leaderboard_df[i].name} ({leaderboard_df[i].commit[0:8]}, UID={leaderboard_df[i].uid}) · ${round(leaderboard_df[i].emission * tao_price, 2):,} (τ{round(leaderboard_df[i].emission, 2):,})": \
                        leaderboard_df[i].incentive / class_denominator for i in range(0, 10) if leaderboard_df[i].incentive and leaderboard_df[i].competition == competition.id
                }

                gr.Label(
                    value=class_values,
                    num_top_classes=10,
                )
                
                with gr.Accordion("Evaluation Stats"):
                    gr.HTML(EVALUATION_HEADER)

                    with gr.Tabs():
                        for entry in leaderboard_df:
                            if entry.competition == competition.id:
                                sample = scores[entry.uid]["sample"]
                                if sample is not None:
                                    name = f"{entry.namespace}/{entry.name} ({entry.commit[0:8]}, UID={entry.uid})"
                                    with gr.Tab(name):
                                        gr.Chatbot([(sample[0], sample[1])])
                                        # gr.Chatbot([(sample[0], f"*{name}*: {sample[1]}"), (None, f"*GPT-4*: {sample[2]}")])

                    show_stale = gr.Checkbox(label="Show Stale", interactive=True)
                    leaderboard_table = gr.components.Dataframe(
                        value=leaderboard_data(show_stale.value, scores, competition.id),
                        headers=["Name", "Win Rate", "Perplexity", "Weight", "UID", "Block"],
                        datatype=["markdown", "number", "number", "number", "number", "number"],
                        elem_id="leaderboard-table",
                        interactive=False,
                        visible=True,
                    )
                    gr.HTML(EVALUATION_DETAILS)
                    show_stale.change(lambda x: leaderboard_data(x, scores, competition.id), [show_stale], leaderboard_table)

    with gr.Accordion("Validator Stats"):
        validator_table = gr.components.Dataframe(
            value=[
                [uid, int(validator_df[uid][1]), round(validator_df[uid][0], 4)] + [validator_df[uid][-1].get(c.uid) for c in leaderboard_df if c.incentive]
                for uid, _ in sorted(
                        zip(validator_df.keys(), [validator_df[x][1] for x in validator_df.keys()]),
                        key=lambda x: x[1],
                        reverse=True
                    )
            ],
            headers=["UID", "Stake (τ)", "V-Trust"] + [f"{c.namespace}/{c.name} ({c.commit[0:8]}, UID={c.uid})" for c in leaderboard_df if c.incentive],
            datatype=["number", "number", "number"] + ["number" for c in leaderboard_df if c.incentive],
            interactive=False,
            visible=True,
        )

def restart_space():
    API.restart_space(repo_id=REPO_ID, token=H4_TOKEN)

scheduler = BackgroundScheduler()
scheduler.add_job(restart_space, "interval", seconds=60 * 15) # restart every 15 minutes
scheduler.start()

demo.launch()