train_diffreaper.py

import torch
import torch.nn as nn
from diffusers import DDPMScheduler
from transformers import AutoTokenizer
from datasets import load_dataset
import os
import time
import math

# --- DiffReaper Configuration ---
MODEL_PATH = "./DiffReaper-Talk"
OUTPUT_DIR = "./training_output"
LOG_FILE = "training.log"
BATCH_SIZE = 32
LEARNING_RATE = 1e-4
NUM_STEPS = 50000 
SAVE_EVERY = 2500
TEST_EVERY = 500

N_EMBD = 1024
N_HEAD = 16
N_LAYER = 12
MAX_PROMPT_LEN = 32
MAX_RESP_LEN = 32
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")

# --- Time Embedding (Sinusoidal) ---
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)

# --- Conditioned Diffusion Block ---
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)
        x_norm = x_norm * (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)

log("Initializing DiffReaper Parallel Form (Conditioned Diffusion)...")
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")
noise_scheduler = DDPMScheduler(num_train_timesteps=1000, beta_schedule="squaredcos_cap_v2")
optimizer = torch.optim.AdamW(model.parameters(), lr=LEARNING_RATE)

log("Loading Dataset (Prompt/Response focus)...")
dataset = load_dataset("OpenAssistant/oasst1", split="train")

def tokenize_function(examples):
    tokens = tokenizer(examples["text"], padding="max_length", truncation=True, max_length=TOTAL_LEN)
    return tokens

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)
data_iter = iter(dataloader)

def get_batch():
    global data_iter
    try: batch = next(data_iter)
    except StopIteration:
        data_iter = iter(dataloader)
        batch = next(data_iter)
    return batch["input_ids"].to("cuda")

def run_test():
    log("Running Cropmark Diagnostic...")
    model.eval()
    with torch.no_grad():
        prompt = "How are you?"
        p_tokens = tokenizer(prompt, return_tensors="pt").input_ids.to("cuda")
        p_tokens = p_tokens[:, :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()
            combined = torch.cat([p_emb, r_noise], dim=1)
            pred = model(combined, t)
            r_0_pred = pred[:, MAX_PROMPT_LEN:, :]
            r_noise = 0.4 * r_noise + 0.6 * r_0_pred
            
        logits = torch.matmul(r_noise, model.token_embedding.weight.T)
        resp_ids = torch.argmax(logits, dim=-1)
        log(f"Prompt: '{prompt}' | [Cropmark]: '{tokenizer.decode(resp_ids[0], skip_special_tokens=False)}'")

    model.train()

log("Starting Conditioned Cropmark Training...")
start_time = time.time()

for step in range(NUM_STEPS):
    optimizer.zero_grad()
    
    input_ids = get_batch()
    prompt_ids = input_ids[:, :MAX_PROMPT_LEN]
    resp_ids = input_ids[:, MAX_PROMPT_LEN:]
    
    with torch.no_grad():
        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)
    
    combined_input = torch.cat([prompt_emb, noisy_resp], dim=1)
    
    predicted_clean_combined = model(combined_input, t)
    predicted_resp = predicted_clean_combined[:, MAX_PROMPT_LEN:, :]
    
    mask = (resp_ids != tokenizer.pad_token_id).unsqueeze(-1).expand_as(resp_emb).float()
    loss = torch.nn.functional.mse_loss(predicted_resp * mask, resp_emb * mask, reduction='sum') / (mask.sum() + 1e-8)
    
    loss.backward()
    optimizer.step()
    
    if step % 50 == 0:
        elapsed = time.time() - start_time
        log(f"Step {step}/{NUM_STEPS} - Loss: {loss.item():.6f} - Speed: {(step+1)/elapsed:.2f} s/s")
        
    if step > 0 and step % TEST_EVERY == 0:
        run_test()
        
log("Training complete.")