From 5f74b2c64c1001a3f0a5f032583616bb11d1c93c Mon Sep 17 00:00:00 2001 From: Dani Date: Sat, 26 Apr 2025 22:34:05 -0400 Subject: [PATCH] Fixing CUDA errors --- .gitignore | 3 ++- dashboard/dashboard.py | 8 +++++-- dashboard/templates/brainmap.html | 1 + dashboard/templates/index.html | 37 ++++++++++++++++++++++--------- main.py | 25 ++++++++++++++++----- model/dream_replay.py | 3 +++ model/dynamic_expand.py | 9 ++++---- model/rehearsal.py | 11 ++++++--- model/scheduler.py | 19 ++++++++++++++++ model/trainer.py | 21 ++++++++++++------ reader/reader.py | 3 +++ 11 files changed, 105 insertions(+), 35 deletions(-) create mode 100644 model/scheduler.py diff --git a/.gitignore b/.gitignore index 79d8774..4f5689d 100644 --- a/.gitignore +++ b/.gitignore @@ -174,4 +174,5 @@ cython_debug/ /data/memory/context.json /data/memory/dreams.json data/memory/brainmap.json -/data/memory/vocab.json \ No newline at end of file +/data/memory/vocab.json +data/memory/book_progress.json diff --git a/dashboard/dashboard.py b/dashboard/dashboard.py index 8884fe0..6ede7f0 100644 --- a/dashboard/dashboard.py +++ b/dashboard/dashboard.py @@ -5,6 +5,7 @@ from model.memory import load_dreams from model.tokenizer import Tokenizer from model.abstraction import cluster_vocab from model.memory import load_dreams +from model.scheduler import get_time_until_next_action, get_next_action_label from context.context import load_context import json import os @@ -37,13 +38,16 @@ def index(): top_dreams = dreams[:5] memory_size = len(load_context()) loss_data = load_loss_data() - remaining = max(0, int(next_cycle_time - time.time())) + next_cycle = get_time_until_next_action() + next_action_label = get_next_action_label() + return render_template("index.html", vocab_size=len(tokenizer.vocab), top_dreams=top_dreams, memory_size=memory_size, loss_data=loss_data, - next_cycle=remaining) + next_cycle=next_cycle, + next_action_label=next_action_label) @app.route("/growth") diff --git a/dashboard/templates/brainmap.html b/dashboard/templates/brainmap.html index 5c6fba7..381ef1d 100644 --- a/dashboard/templates/brainmap.html +++ b/dashboard/templates/brainmap.html @@ -2,6 +2,7 @@ + Ruby's Brain Map - +
+ 🏠 Home + 📓 Journal + 🧠 Concepts + 🕸️ Brain Map + 📈 Growth + 💬 Dreams +
-
- 🏠 Home - 📓 Journal - 🧠 Concepts - 🕸️ Brain Map - 📈 Growth - 💬 Dreams -
-

Ruby is Running 🧠

-

⏳ Next Cycle Countdown

+

⏳ Next Cycle

+

Next: {{ next_action_label }}

{{ next_cycle }} seconds

diff --git a/main.py b/main.py index 76e11f1..69ed472 100644 --- a/main.py +++ b/main.py @@ -8,8 +8,10 @@ from model.brain import generate_response from model.cleanup import full_cleanup from model.dream_replay import replay_dreams from model.rehearsal import simulate_conversation +from model.scheduler import set_next_action from reader.reader import read_books_forever from dashboard.dashboard import run_dashboard +import threading load_dotenv() TOKEN = os.getenv("DISCORD_TOKEN") @@ -43,27 +45,38 @@ threading.Thread(target=run_dashboard, daemon=True).start() async def background_cleanup_loop(): while True: full_cleanup() + set_next_action(300, "Cleaning up") await asyncio.sleep(300) # 5 minutes async def dream_replay_loop(): while True: replay_dreams() + set_next_action(900, "Dreaming new dreams") await asyncio.sleep(900) # Replay every 15 minutes async def rehearsal_loop(): while True: simulate_conversation() + set_next_action(1200, "Practicing Conversations") await asyncio.sleep(1200) # Every 20 minutes -# Launch background tasks -loop = asyncio.get_event_loop() -loop.create_task(read_books_forever()) # Book reader task -loop.create_task(background_cleanup_loop()) -loop.create_task(dream_replay_loop()) -loop.create_task(rehearsal_loop()) +# Start Ruby's Brain Loops in a separate thread +def start_brain_loops(): + loop = asyncio.new_event_loop() + asyncio.set_event_loop(loop) + + loop.create_task(read_books_forever()) + loop.create_task(dream_replay_loop()) + loop.create_task(background_cleanup_loop()) + loop.create_task(rehearsal_loop()) + + loop.run_forever() + + +threading.Thread(target=start_brain_loops, daemon=True).start() # Launch Discord bot (blocking) client.run(TOKEN) diff --git a/model/dream_replay.py b/model/dream_replay.py index 34a3e76..4f93380 100644 --- a/model/dream_replay.py +++ b/model/dream_replay.py @@ -1,9 +1,12 @@ import random from model.memory import load_dreams from model.trainer import train_on_message +from model.dynamic_expand import expand_model_if_needed def replay_dreams(): + expand_model_if_needed() + dreams = load_dreams() if not dreams: return diff --git a/model/dynamic_expand.py b/model/dynamic_expand.py index f6f8e92..5b69c68 100644 --- a/model/dynamic_expand.py +++ b/model/dynamic_expand.py @@ -6,18 +6,17 @@ optimizer = torch.optim.Adam(model.parameters(), lr=1e-4) def get_optimizer(): - global optimizer return optimizer def expand_model_if_needed(): global model, optimizer - current_vocab_size = len(tokenizer.vocab) + 10 # Buffer + current_vocab_size = len(tokenizer.vocab) + 10 old_vocab_size = model.head.out_features if current_vocab_size <= old_vocab_size: - return + return # No expansion needed print(f"Expanding model from {old_vocab_size} -> {current_vocab_size}") @@ -31,6 +30,6 @@ def expand_model_if_needed(): param.copy_(old_state[name]) model = new_model - opt = get_optimizer() + optimizer = torch.optim.Adam(model.parameters(), lr=1e-4) - print("Model expanded and optimizer rebuilt.") + print("Expansion complete.") diff --git a/model/rehearsal.py b/model/rehearsal.py index 088b0f1..0515ef5 100644 --- a/model/rehearsal.py +++ b/model/rehearsal.py @@ -1,13 +1,18 @@ import torch from model.brain import model, tokenizer, DEVICE from model.trainer import train_on_message +from model.dynamic_expand import expand_model_if_needed def simulate_conversation(): - seed = torch.randint(0, tokenizer.next_id, (1, 5), device=DEVICE) - output = model(seed) - preds = torch.argmax(output, dim=-1).squeeze().tolist() + expand_model_if_needed() + model.eval() + seed = torch.randint(0, tokenizer.next_id, (1, 5), device=DEVICE) + seed = seed[:, -128:] # Safety clamp + output = model(seed) + + preds = torch.argmax(output, dim=-1).squeeze().tolist() if isinstance(preds, int): preds = [preds] diff --git a/model/scheduler.py b/model/scheduler.py new file mode 100644 index 0000000..269b5b5 --- /dev/null +++ b/model/scheduler.py @@ -0,0 +1,19 @@ +import time + +_next_action_time = time.time() + 60 # default 1 minute from now +_next_action_label = "Waiting" + + +def set_next_action(seconds_from_now: int, label: str = "Thinking"): + global _next_action_time, _next_action_label + _next_action_time = time.time() + seconds_from_now + _next_action_label = label + + +def get_time_until_next_action() -> int: + remaining = int(_next_action_time - time.time()) + return max(0, remaining) + + +def get_next_action_label() -> str: + return _next_action_label diff --git a/model/trainer.py b/model/trainer.py index 22e6623..155298f 100644 --- a/model/trainer.py +++ b/model/trainer.py @@ -1,9 +1,8 @@ import torch import time -from model.brain_state import model, tokenizer, DEVICE,loss_fn -from context.context import add_to_context, get_recent_context from model.dynamic_expand import expand_model_if_needed, get_optimizer -from model.brainmap import update_brainmap +from model.brain_state import model, tokenizer, DEVICE, loss_fn +from context.context import add_to_context, get_recent_context LOSS_FILE = "data/logs/loss.log" @@ -15,6 +14,7 @@ def log_loss(value: float): def train_on_message(text: str): expand_model_if_needed() + model.train() context_texts = get_recent_context(3) augmented_text = " ".join(context_texts + [text]) @@ -23,16 +23,23 @@ def train_on_message(text: str): if len(tokens) < 2: return - tokens = tokens[:128] - words = tokenizer.detokenize(tokens).split() - update_brainmap(words) + # ✋ Clamp to model's known vocab + max_token_id = model.head.out_features - 1 + tokens = [t for t in tokens if t <= max_token_id] + + if len(tokens) < 2: + return # after filtering, too short to train + + tokens = tokens[:128] # safety clamp input_tensor = torch.tensor(tokens[:-1], dtype=torch.long, device=DEVICE).unsqueeze(0) target_tensor = torch.tensor(tokens[1:], dtype=torch.long, device=DEVICE).unsqueeze(0) + opt = get_optimizer() + output = model(input_tensor) + loss = loss_fn(output.view(-1, output.size(-1)), target_tensor.view(-1)) - opt = get_optimizer() opt.zero_grad() loss.backward() opt.step() diff --git a/reader/reader.py b/reader/reader.py index be7dc2c..d8b0166 100644 --- a/reader/reader.py +++ b/reader/reader.py @@ -1,7 +1,9 @@ import os import asyncio from model.trainer import train_on_message +from model.scheduler import set_next_action from reader.filter import is_valid_line +import json BOOK_DIR = "data/books" PROGRESS_FILE = "data/memory/book_progress.json" @@ -45,4 +47,5 @@ async def read_books_forever(): if is_valid_line(line): train_on_message(line) + set_next_action(READ_DELAY, "Reading") await asyncio.sleep(READ_DELAY)