train_diffreaper_5l.py

import torch
import torch.nn as nn
import torch.nn.functional as F
from diffusers import DDPMScheduler
from transformers import AutoTokenizer
from datasets import load_dataset
import os
import time
import math
from huggingface_hub import HfApi

# --- FAILPROOF 5L CONFIG ---
MODEL_PATH = "./DiffReaper-Talk"
REPO_ID = "darwinkernelpanic/DiffReaper-5L"
HF_TOKEN = os.getenv("HF_TOKEN")
OUTPUT_DIR = "./training_output"
LOG_FILE = "training.log"
CHECKPOINT_LOG = "checkpoint_log.txt"
BATCH_SIZE = 16 # Back to higher batch for 5090 (32GB VRAM)
LEARNING_RATE = 1e-4
SAVE_EVERY = 2500
TEST_EVERY = 500

N_EMBD = 2048 # Back to full Large specs
N_HEAD = 32
N_LAYER = 24 # Full depth
MAX_PROMPT_LEN = 32
MAX_RESP_LEN = 128
TOTAL_LEN = MAX_PROMPT_LEN + MAX_RESP_LEN

os.makedirs(OUTPUT_DIR, exist_ok=True)

def log(msg):
    timestamp = time.strftime("%Y-%m-%d %H:%M:%S")
    formatted = f"[{timestamp}] {msg}"
    print(formatted)
    with open(LOG_FILE, "a") as f:
        f.write(formatted + "\n")

class TimeEmbedding(nn.Module):
    def __init__(self, n_embd):
        super().__init__()
        self.mlp = nn.Sequential(nn.Linear(n_embd, n_embd), nn.GELU(), nn.Linear(n_embd, n_embd))
    def forward(self, t):
        half_dim = N_EMBD // 2
        emb = math.log(10000) / (half_dim - 1)
        emb = torch.exp(torch.arange(half_dim, device=t.device) * -emb)
        emb = t[:, None] * emb[None, :]
        emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
        return self.mlp(emb)

class DiffReaperBlock(nn.Module):
    def __init__(self, n_embd, n_head):
        super().__init__()
        self.ln1 = nn.LayerNorm(n_embd)
        self.attn = nn.MultiheadAttention(n_embd, n_head, batch_first=True)
        self.ln2 = nn.LayerNorm(n_embd)
        self.mlp = nn.Sequential(nn.Linear(n_embd, 4 * n_embd), nn.GELU(), nn.Linear(4 * n_embd, n_embd))
        self.time_mlp = nn.Linear(n_embd, n_embd * 2)
    def forward(self, x, t_emb):
        time_params = self.time_mlp(t_emb).unsqueeze(1)
        scale, shift = time_params.chunk(2, dim=-1)
        x_norm = self.ln1(x) * (1 + scale) + shift
        attn_out, _ = self.attn(x_norm, x_norm, x_norm)
        x = x + attn_out
        x = x + self.mlp(self.ln2(x))
        return x

class DiffReaperModel(nn.Module):
    def __init__(self, vocab_size, n_embd, n_head, n_layer):
        super().__init__()
        self.token_embedding = nn.Embedding(vocab_size, n_embd)
        self.pos_embedding = nn.Parameter(torch.zeros(1, TOTAL_LEN, n_embd))
        self.time_embed = TimeEmbedding(n_embd)
        self.blocks = nn.ModuleList([DiffReaperBlock(n_embd, n_head) for _ in range(n_layer)])
        self.ln_f = nn.LayerNorm(n_embd)
    def forward(self, x_input, t):
        t_emb = self.time_embed(t)
        x = x_input + self.pos_embedding[:, :x_input.shape[1], :]
        for block in self.blocks: x = block(x, t_emb)
        return self.ln_f(x)

def run_test(model, tokenizer, step):
    log(f"Running Cropmark Diagnostic [Step {step}]...")
    model.eval()
    with torch.no_grad():
        prompt = "Hello! Who are you today?"
        p_tokens = tokenizer(prompt, return_tensors="pt").input_ids.to("cuda")[:, :MAX_PROMPT_LEN]
        p_padded = torch.full((1, MAX_PROMPT_LEN), tokenizer.pad_token_id, device="cuda")
        p_padded[:, :p_tokens.shape[1]] = p_tokens
        p_emb = model.token_embedding(p_padded)
        r_noise = torch.randn(1, MAX_RESP_LEN, N_EMBD).to("cuda")
        for i in range(10):
            t = torch.tensor([1000 - (i*100) - 1], device="cuda").long()
            pred = model(torch.cat([p_emb, r_noise], dim=1), t)[:, MAX_PROMPT_LEN:, :]
            r_noise = 0.4 * r_noise + 0.6 * pred
        norm_weights = F.normalize(model.token_embedding.weight, dim=-1)
        norm_r = F.normalize(r_noise, dim=-1)
        logits = torch.matmul(norm_r, norm_weights.T)
        resp_ids = torch.argmax(logits, dim=-1)
        result = tokenizer.decode(resp_ids[0], skip_special_tokens=True)
        log(f"Prompt: '{prompt}' | [Cropmark]: '{result}'")
        with open(CHECKPOINT_LOG, "a") as f:
            f.write(f"Step {step} - Prompt: '{prompt}' | [Cropmark]: '{result}'\n")
    model.train()

if __name__ == "__main__":
    log("Initializing DiffReaper-5L (Autopilot)...")
    tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
    if tokenizer.pad_token is None: tokenizer.pad_token = tokenizer.eos_token
    model = DiffReaperModel(tokenizer.vocab_size, N_EMBD, N_HEAD, N_LAYER).to("cuda")
    
    # Try to load existing checkpoint from HF or local if possible (manual step)
    
    noise_scheduler = DDPMScheduler(num_train_timesteps=1000, beta_schedule="squaredcos_cap_v2")
    optimizer = torch.optim.AdamW(model.parameters(), lr=LEARNING_RATE)

    log("Loading OpenAssistant...")
    dataset = load_dataset("OpenAssistant/oasst1", split="train")
    def tokenize_function(examples):
        return tokenizer(examples["text"], padding="max_length", truncation=True, max_length=TOTAL_LEN)
    tokenized_dataset = dataset.map(tokenize_function, batched=True, remove_columns=dataset.column_names)
    tokenized_dataset.set_format("torch")
    dataloader = torch.utils.data.DataLoader(tokenized_dataset, batch_size=BATCH_SIZE, shuffle=True)

    log("DiffReaper-5L training starting...")
    api = HfApi()
    start_time = time.time()
    step = 0
    while True:
        for batch in dataloader:
            optimizer.zero_grad()
            input_ids = batch["input_ids"].to("cuda")
            prompt_ids = input_ids[:, :MAX_PROMPT_LEN]
            resp_ids = input_ids[:, MAX_PROMPT_LEN:]
            prompt_emb = model.token_embedding(prompt_ids)
            resp_emb = model.token_embedding(resp_ids)
            
            noise = torch.randn_like(resp_emb)
            t = torch.randint(0, 1000, (input_ids.shape[0],), device="cuda").long()
            noisy_resp = noise_scheduler.add_noise(resp_emb, noise, t)
            
            pred_resp = model(torch.cat([prompt_emb, noisy_resp], dim=1), t)[:, MAX_PROMPT_LEN:, :]
            mask = (resp_ids != tokenizer.pad_token_id).float().unsqueeze(-1)
            loss = 1 - (F.cosine_similarity(pred_resp, resp_emb, dim=-1) * mask.squeeze(-1)).sum() / (mask.sum() + 1e-8)
            
            loss.backward()
            optimizer.step()
            
            if step % 50 == 0:
                elapsed = time.time() - start_time
                log(f"Step {step} - Loss: {loss.item():.6f} - Speed: {(step+1)/elapsed:.2f} s/s")
            if step > 0 and step % TEST_EVERY == 0: run_test(model, tokenizer, step)
            if step > 0 and step % SAVE_EVERY == 0:
                ckpt_path = os.path.join(OUTPUT_DIR, f"diffreaper5l_{step}.pt")
                torch.save(model.state_dict(), ckpt_path)
                log("Syncing to HF...")
                try:
                    api.upload_file(path_or_fileobj=ckpt_path, path_in_repo=f"diffreaper5l_{step}.pt", repo_id=REPO_ID, token=HF_TOKEN)
                    api.upload_file(path_or_fileobj=CHECKPOINT_LOG, path_in_repo="checkpoint_log.txt", repo_id=REPO_ID, token=HF_TOKEN)
                    api.upload_file(path_or_fileobj="train_diffreaper_5l.py", path_in_repo="train_diffreaper_5l.py", repo_id=REPO_ID, token=HF_TOKEN)
                except Exception as e: log(f"HF Sync Error: {e}")
            step += 1