rusticluftig's picture
Add a custom subtensor endpoint
26ec5d2
raw
history blame
No virus
16.3 kB
import functools
import traceback
import gradio as gr
import bittensor as bt
from typing import Dict, List, Any, Optional, Tuple
from bittensor.extrinsics.serving import get_metadata
from dataclasses import dataclass
import wandb
import math
import os
import datetime
import time
import json
import pandas as pd
from dotenv import load_dotenv
from huggingface_hub import HfApi
from apscheduler.schedulers.background import BackgroundScheduler
import pandas as pd
load_dotenv()
FONT = (
"""<link href="https://fonts.cdnfonts.com/css/jmh-typewriter" rel="stylesheet">"""
)
TITLE = """<h1 align="center" id="space-title" class="typewriter">Subnet 9 Leaderboard</h1>"""
HEADER = """<h2 align="center" class="typewriter"><a href="https://github.com/RaoFoundation/pretraining" target="_blank">Subnet 9</a> is a <a href="https://bittensor.com/" target="_blank">Bittensor</a> subnet that rewards miners for producing pretrained Foundation-Models on the <a href="https://huggingface.co/datasets/tiiuae/falcon-refinedweb" target="_blank">Falcon Refined Web dataset</a>. It acts like a continuous benchmark whereby miners are rewarded for attaining the best losses on randomly sampled pages of Falcon.<br/>The models with the best head-to-head loss on the evaluation data receive a steady emission of TAO.</h3>"""
EVALUATION_DETAILS = """<ul><li><b>Name:</b> the 🤗 Hugging Face model name (click to go to the model card)</li><li><b>Rewards / Day:</b> the expected rewards per day based on current ranking.</li><li><b>Last Average Loss:</b> the last loss value on the evaluation data for the model as calculated by a validator (lower is better)</li><li><b>UID:</b> the Bittensor UID of the miner</li><li><b>Block:</b> the Bittensor block that the model was submitted in</li></ul><br/>More stats on <a href="https://taostats.io/subnets/netuid-9/" target="_blank">taostats</a>."""
EVALUATION_HEADER = """<h3 align="center">Shows the latest internal evaluation statistics as calculated by the Opentensor validator</h3>"""
VALIDATOR_WANDB_PROJECT = "opentensor-dev/pretraining-subnet"
BENCHMARK_WANDB_PROJECT = "raofoundation/pretraining-leaderboard-data"
H4_TOKEN = os.environ.get("H4_TOKEN", None)
API = HfApi(token=H4_TOKEN)
WANDB_TOKEN = os.environ.get("WANDB_API_KEY", None)
SUBTENSOR_ENDPOINT=os.environ.get("SUBTENSOR_ENDPOINT", None)
REPO_ID = "RaoFoundation/pretraining-leaderboard"
MAX_AVG_LOSS_POINTS = 1
RETRIES = 5
DELAY_SECS = 3
NETUID = 9
SECONDS_PER_BLOCK = 12
@dataclass
class ModelData:
uid: int
hotkey: str
namespace: str
name: str
commit: str
hash: str
block: int
incentive: float
emission: float
@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 None,
hash=tokens[3] if tokens[3] != "None" else None,
block=block,
incentive=incentive,
emission=emission,
)
def run_with_retries(func, *args, **kwargs):
for i in range(0, RETRIES):
try:
return func(*args, **kwargs)
except (Exception, RuntimeError):
if i == RETRIES - 1:
raise
time.sleep(DELAY_SECS)
raise RuntimeError("Should never happen")
def get_subtensor_and_metagraph() -> Tuple[bt.subtensor, bt.metagraph]:
def _internal() -> Tuple[bt.subtensor, bt.metagraph]:
subtensor = bt.subtensor("finney")
metagraph = bt.metagraph(NETUID, lite=False)
return subtensor, metagraph
return run_with_retries(_internal)
def get_validator_weights(
metagraph: bt.metagraph,
) -> Dict[int, Tuple[float, int, Dict[int, float]]]:
"""Returns a dictionary of validator UIDs to (vtrust, stake, {uid: weight})."""
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
) -> List[ModelData]:
result = []
for uid in metagraph.uids.tolist():
hotkey = metagraph.hotkeys[uid]
metadata = None
try:
metadata = run_with_retries(
functools.partial(get_metadata, subtensor, metagraph.netuid, hotkey)
)
except:
print(f"Failed to get metadata for UID {uid}: {traceback.format_exc()}")
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 is_floatable(x) -> bool:
return (
isinstance(x, float) and not math.isnan(x) and not math.isinf(x)
) or isinstance(x, int)
def get_wandb_runs(project: str, filters: Dict[str, Any]) -> List:
"""Get the latest runs from Wandb, retrying infinitely until we get them."""
while True:
api = wandb.Api(api_key=WANDB_TOKEN)
runs = list(
api.runs(
project,
filters=filters,
)
)
if len(runs) > 0:
return runs
# WandDB API is quite unreliable. Wait another minute and try again.
print("Failed to get runs from Wandb. Trying again in 60 seconds.")
time.sleep(60)
def get_scores(
uids: List[int],
wandb_runs: List,
) -> Dict[int, Dict[str, Optional[float]]]:
result = {}
previous_timestamp = None
# Iterate through the runs until we've processed all the uids.
for i, run in enumerate(wandb_runs):
if not "original_format_json" in run.summary:
continue
data = json.loads(run.summary["original_format_json"])
all_uid_data = data["uid_data"]
timestamp = data["timestamp"]
# Make sure runs are indeed in descending time order.
assert (
previous_timestamp is None or timestamp < previous_timestamp
), f"Timestamps are not in descending order: {timestamp} >= {previous_timestamp}"
previous_timestamp = timestamp
for uid in uids:
if uid in result:
continue
if str(uid) in all_uid_data:
uid_data = all_uid_data[str(uid)]
# Only the most recent run is fresh.
is_fresh = i == 0
result[uid] = {
"avg_loss": uid_data.get("average_loss", None),
"win_rate": uid_data.get("win_rate", None),
"win_total": uid_data.get("win_total", None),
"weight": uid_data.get("weight", None),
"fresh": is_fresh,
}
if len(result) == len(uids):
break
return result
def get_losses_over_time(wandb_runs: List) -> pd.DataFrame:
"""Returns a dataframe of the best average model loss over time."""
timestamps = []
best_losses = []
for run in wandb_runs:
if "original_format_json" not in run.summary:
continue
data = json.loads(run.summary["original_format_json"])
all_uid_data = data["uid_data"]
timestamp = datetime.datetime.fromtimestamp(data["timestamp"])
best_loss = math.inf
for _, uid_data in all_uid_data.items():
loss = uid_data.get("average_loss", math.inf)
# Filter out the numbers from the exploit.
if loss < best_loss and (loss > 2.5 or timestamp > datetime.datetime(2024,2,8)):
best_loss = uid_data["average_loss"]
if best_loss != math.inf:
timestamps.append(timestamp)
best_losses.append(best_loss)
return pd.DataFrame({"timestamp": timestamps, "best_loss": best_losses})
def format_score(uid: int, scores, key) -> Optional[float]:
if uid in scores:
if key in scores[uid]:
point = scores[uid][key]
if is_floatable(point):
return round(scores[uid][key], 4)
return None
def next_epoch(subtensor: bt.subtensor, block: int) -> int:
return (
block
+ subtensor.get_subnet_hyperparameters(NETUID).tempo
- subtensor.blocks_since_epoch(NETUID, block)
)
def get_next_update_div(current_block: int, next_update_block: int) -> str:
now = datetime.datetime.now()
blocks_to_go = next_update_block - current_block
next_update_time = now + datetime.timedelta(
seconds=blocks_to_go * SECONDS_PER_BLOCK
)
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 get_last_updated_div() -> str:
return f"""<div>Last Updated: {datetime.datetime.utcnow().strftime("%Y-%m-%d %H:%M:%S")} (UTC)</div>"""
def leaderboard_data(
leaderboard: List[ModelData],
scores: Dict[int, Dict[str, Optional[float]]],
show_stale: bool,
) -> List[List[Any]]:
"""Returns the leaderboard data, based on models data and UID scores."""
return [
[
f"[{c.namespace}/{c.name} ({c.commit[0:8]})](https://huggingface.co/{c.namespace}/{c.name}/commit/{c.commit})",
format_score(c.uid, scores, "win_rate"),
format_score(c.uid, scores, "avg_loss"),
format_score(c.uid, scores, "weight"),
c.uid,
c.block,
]
for c in leaderboard
if (c.uid in scores and scores[c.uid]["fresh"]) or show_stale
]
def get_benchmarks() -> Tuple[pd.DataFrame, datetime.datetime]:
"""Returns the latest benchmarks and the time they were run."""
runs = get_wandb_runs(project=BENCHMARK_WANDB_PROJECT, filters=None)
for run in runs:
artifacts = list(run.logged_artifacts())
if artifacts:
table = artifacts[-1].get("benchmarks")
if table:
return table.get_dataframe(), datetime.datetime.strptime(run.metadata["startedAt"], "%Y-%m-%dT%H:%M:%S.%f")
bt.logging.error("Failed to get benchmarks from Wandb.")
return None, None
def restart_space():
API.restart_space(repo_id=REPO_ID, token=H4_TOKEN)
def main():
# To avoid leaderboard failures, infinitely try until we get all data
# needed to populate the dashboard
while True:
try:
subtensor, metagraph = get_subtensor_and_metagraph()
model_data: List[ModelData] = get_subnet_data(subtensor, metagraph)
model_data.sort(key=lambda x: x.incentive, reverse=True)
vali_runs = get_wandb_runs(project=VALIDATOR_WANDB_PROJECT, filters={"config.type": "validator", "config.uid": 238})
scores = get_scores([x.uid for x in model_data], vali_runs)
# TODO: Re-enable once ""SubtensorModule.BlocksSinceEpoch" not found" issue is resolved.
# current_block = metagraph.block.item()
# next_epoch_block = next_epoch(subtensor, current_block)
validator_df = get_validator_weights(metagraph)
weight_keys = set()
for uid, stats in validator_df.items():
weight_keys.update(stats[-1].keys())
benchmarks, benchmark_timestamp = get_benchmarks()
break
except Exception as e:
print(f"Failed to get data: {e}")
time.sleep(30)
demo = gr.Blocks(css=".typewriter {font-family: 'JMH Typewriter', sans-serif;}")
with demo:
gr.HTML(FONT)
gr.HTML(TITLE)
gr.HTML(HEADER)
# TODO: Re-enable once ""SubtensorModule.BlocksSinceEpoch" not found" issue is resolved.
# gr.HTML(value=get_next_update_div(current_block, next_epoch_block))
gr.Label(
value={
f"{c.namespace}/{c.name} ({c.commit[0:8]}) · (τ{round(c.emission, 2):,})": c.incentive
for c in model_data
if c.incentive
},
num_top_classes=10,
)
if benchmarks is not None:
with gr.Accordion("Top Model Benchmarks"):
gr.components.Dataframe(benchmarks)
gr.HTML("""<div>PPL computed using a stride of 512. See <a href='https://github.com/RaoFoundation/pretraining/blob/dev/scripts/run_benchmarks.py'>here</a> for the full code.</div>""")
gr.HTML(f"""<div>Last Updated: {benchmark_timestamp.strftime("%Y-%m-%d %H:%M:%S")} (UTC)</div>""")
with gr.Accordion("Evaluation Stats"):
gr.HTML(EVALUATION_HEADER)
show_stale = gr.Checkbox(label="Show Stale", interactive=True)
leaderboard_table = gr.components.Dataframe(
value=leaderboard_data(model_data, scores, show_stale.value),
headers=["Name", "Win Rate", "Average Loss", "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 stale: leaderboard_data(model_data, scores, stale),
inputs=[show_stale],
outputs=leaderboard_table,
)
gr.LinePlot(
get_losses_over_time(vali_runs),
x="timestamp",
x_title="Date",
y="best_loss",
y_title="Average Loss",
tooltip="best_loss",
interactive=True,
visible=True,
width=1024,
title="Best Average Loss Over Time",
)
with gr.Accordion("Validator Stats"):
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 model_data
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]})"
for c in model_data
if c.incentive
],
datatype=["number", "number", "number"]
+ ["number" for c in model_data if c.incentive],
interactive=False,
visible=True,
)
gr.HTML(value=get_last_updated_div())
scheduler = BackgroundScheduler()
scheduler.add_job(
restart_space, "interval", seconds=60 * 30
) # restart every 15 minutes
scheduler.start()
demo.launch()
main()