Upload train_autogrow.py with huggingface_hub

train_autogrow.py

@@ -0,0 +1,149 @@
+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 CONFIG ---
+MODEL_PATH = "./DiffReaper-Talk"
+REPO_ID = "darwinkernelpanic/DiffReaper-5"
+HF_TOKEN = "${HF_TOKEN}"
+OUTPUT_DIR = "./training_output"
+LOG_FILE = "training.log"
+BATCH_SIZE = 16 # Lower for 3090 VRAM
+LEARNING_RATE = 1e-4
+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")
+
+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)
+
+log("Initializing Autogrow Model...")
+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...")
+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)
+
+def run_test(step):
+    log(f"Running Cropmark Diagnostic [Step {step}]...")
+    model.eval()
+    with torch.no_grad():
+        prompt = "Hello! Who are you?"
+        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)
+            r_noise = 0.4 * r_noise + 0.6 * pred[:, MAX_PROMPT_LEN:, :]
+        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)
+        log(f"Prompt: '{prompt}' | [Cropmark]: '{tokenizer.decode(resp_ids[0], skip_special_tokens=True)}'")
+    model.train()
+
+log("Autonomous growth starting...")
+api = HfApi()
+start_time = time.time()
+step = 0
+while True: # Unlimited steps, controlled by your credit
+    for batch in dataloader:
+        optimizer.zero_grad()
+        input_ids = batch["input_ids"].to("cuda")
+        prompt_emb = model.token_embedding(input_ids[:, :MAX_PROMPT_LEN])
+        resp_emb = model.token_embedding(input_ids[:, MAX_PROMPT_LEN:])
+        
+        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:, :]
+        loss = 1 - F.cosine_similarity(pred_resp, resp_emb, dim=-1).mean()
+        loss.backward()
+        optimizer.step()
+        
+        if step % 100 == 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(step)
+        if step > 0 and step % SAVE_EVERY == 0:
+            ckpt_path = os.path.join(OUTPUT_DIR, f"cropmark_latest.pt")
+            torch.save(model.state_dict(), ckpt_path)
+            log("Syncing to HF...")
+            try: api.upload_file(path_or_fileobj=ckpt_path, path_in_repo="cropmark_latest.pt", repo_id=REPO_ID, token=HF_TOKEN)
+            except Exception as e: log(f"HF Sync Error: {e}")
+        step += 1