Quick training script

data.py

@@ -4,18 +4,19 @@ import queue
 
 def startDataWorker(dataset,encoding,batch_size,block_size):
     data_q = queue.Queue(maxsize=100)
-    t = threading.Thread(target=data_worker, args=(data_q, dataset, encoding, batch_size, block_size), daemon=True)
+    t = threading.Thread(target=dataWorker, args=(data_q, dataset, encoding, batch_size, block_size), daemon=True)
     t.start()
     while (1):
         try:
             bx, by = data_q.get(timeout=30)
         except queue.Empty:
+            print("queue empty ...")
             continue
         yield (bx,by)
 
 def dataWorker(q, dataset, encoding, batch_size, block_size):
     batch_x, batch_y = [], []
-    while(1):
+    while True:
         for text in dataset["text"]:
             tokens = encoding.encode(text)
             for i in range(0, len(tokens)-block_size-1,block_size):
@@ -33,7 +34,7 @@ def dataWorker(q, dataset, encoding, batch_size, block_size):
                 batch_x.append(x)
                 batch_y.append(y)
 
-            if len(batch_x) == batch_size:
+                if len(batch_x) == batch_size:
-                q.put((np.array(batch_x, dtype=np.int32),
+                    q.put((np.array(batch_x, dtype=np.int32),
-                       np.array(batch_y, dtype=np.int32)))
+                           np.array(batch_y, dtype=np.int32)))
-                batch_x, batch_y = [], []
+                    batch_x, batch_y = [], []

model.py

@@ -58,10 +58,10 @@ class Block:
         return self
 
 class Transformer():
-    def __init__(self,vocab_size,embed_size,n_heads,n_blocks,max_len):
+    def __init__(self,vocab_size,embed_size,n_heads,n_blocks,block_size):
         self.tok_embed = nn.Embedding(vocab_size,embed_size)
         self.pos_embed = nn.Embedding(block_size,embed_size)
-        self.pos_idx = Tensor.arange(max_len, requires_grad=False)
+        self.pos_idx = Tensor.arange(block_size, requires_grad=False)
 
         self.blocks = [Block(embed_size,n_heads) for _ in range(n_blocks)]
         self.norm = nn.RMSNorm(embed_size)

optm.py

@@ -1,16 +1,18 @@
-from tinygrad import Tensor
+from tinygrad import Tensor,nn
 import math
 
 class CosineLR:
-    def __init__(self,optm,totalSteps,minlr):
+    def __init__(self,optm,totalSteps,maxlr,minlr):
         self.optm = optm
-        self.maxlr = optm.lr
+        self.maxlr = maxlr
         self.minlr = minlr
         self.totalSteps = totalSteps
         self.steps = 0
 
     def step(self):
-        self.optm.lr = self.minlr + 0.5 * (self.maxlr - self.minlr) * (1 + math.cos((step / self.totalSteps) * math.pi))
+        lr = self.minlr + 0.5 * (self.maxlr - self.minlr) * (1 + math.cos((self.steps / self.totalSteps) * math.pi))
+        for o in self.optm:
+            o.lr = lr
         self.optm.step()
         self.steps += 1
 
@@ -18,11 +20,11 @@ class CosineLR:
         self.optm.zero_grad()
 
 
-def llmOptimizer(params,steps,minlr):
+def llmOptimizer(params,steps,maxlr,minlr):
     muon_params = [p for p in params if len(p.shape) >= 2]
     adamw_params = [p for p in params if len(p.shape) < 2]
 
-    o1 = nn.optim.Muon(muon_params, lr=hypr["starting_lr"])
+    o1 = nn.optim.Muon(muon_params, lr=maxlr)
-    o2 = nn.optim.AdamW(adamw_params, lr=hypr["starting_lr"])
+    o2 = nn.optim.AdamW(adamw_params, lr=maxlr)
-    optimizer = nn.optim.OptimizerGroup([o1,o2])
+    optimizer = nn.optim.OptimizerGroup(o1,o2)
-    return CosineLR(optimizer,steps,minlr)
+    return CosineLR(optimizer,steps,maxlr,minlr)

train.py

@@ -0,0 +1,77 @@
+from concurrent.futures import ThreadPoolExecutor
+from tinygrad import Tensor,TinyJit,Device,nn
+from tinygrad.nn.state import get_state_dict
+from model import Transformer
+from transformers import AutoTokenizer
+from datasets import load_dataset
+from tqdm import tqdm
+import optm
+import data
+import log
+
+hypr = {
+    "embed_size":  256,
+    "n_heads":     4,
+    "n_blocks":    4,
+    "block_size":  256,
+    "batch_size":  16,
+    "starting_lr": 3e-4,
+    "minimum_lr":  3e-5,
+    "warmup":      1_000,
+    "steps":       5_000,
+    "encoding":    "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
+    "dataset":     "HuggingFaceTB/smollm-corpus",
+    "subset":      "cosmopedia-v2",
+}
+
+print(Device.DEFAULT)
+
+#for loging
+loger = ThreadPoolExecutor(max_workers=2)
+
+dataset = load_dataset(hypr["dataset"],
+                       hypr["subset"],
+                       split="train",
+                       streaming=True)
+encoding = AutoTokenizer.from_pretrained(hypr["encoding"])
+hypr["vocab_size"] = encoding.vocab_size
+model = Transformer(hypr["vocab_size"],hypr["embed_size"],hypr["n_heads"],hypr["n_blocks"],hypr["block_size"])
+batch = data.startDataWorker(dataset,encoding,hypr["batch_size"],hypr["block_size"])
+
+params = nn.state.get_parameters(model)
+optimizer = optm.llmOptimizer(params,hypr["steps"],hypr["starting_lr"],hypr["minimum_lr"])
+
+@TinyJit
+def step(x,y):
+    optimizer.zero_grad()
+
+    logits = model(x)
+    B,T,C = logits.shape
+    logits = logits.view(B*T,C)
+    y = y.view(B*T)
+    loss = logits.cross_entropy(y)
+
+    loss.backward()
+    optimizer.step()
+    return loss
+
+Tensor.training=True
+bar = tqdm(range(hypr["steps"]))
+
+for steps in bar:
+    nx, ny = next(batch)
+    x = Tensor(nx, device=Device.DEFAULT).realize()
+    y = Tensor(ny, device=Device.DEFAULT).realize()
+    loss = step(x, y)
+    if steps % 10 == 0:
+        l = loss.numpy()
+        loger.submit(log.logLoss, steps, l)
+        bar.set_postfix(loss= f"{l:.4f}")
+    if steps % 500 == 0:
+        loss.realize()
+        m = get_state_dict(model)
+        log.logModel(steps,m)
+        #TODO non sycronus safetensor loging
+        #loger.submit(log.logModel,steps,m)
+
+loger.shutdown(wait=True)