merge, make cleaner, careful with gradient clipping when using grad scaler fp16 training

train.py

@@ -70,7 +70,7 @@ min_lr = 6e-5 # minimum learning rate, should be ~= learning_rate/10 per Chinchi
 backend = 'nccl' # 'nccl', 'gloo', etc.
 # system
 device = 'cuda' # examples: 'cpu', 'cuda', 'cuda:0', 'cuda:1' etc., or try 'mps' on macbooks
-dtype = 'bfloat16' # 'float32' or 'bfloat16'
+dtype = 'bfloat16' # 'float32', 'bfloat16', or 'float16', the latter will auto implement a GradScaler
 compile = True # use PyTorch 2.0 to compile the model to be faster
 # -----------------------------------------------------------------------------
 config_keys = [k for k,v in globals().items() if not k.startswith('_') and isinstance(v, (int, float, bool, str))]
@@ -98,8 +98,8 @@ torch.manual_seed(1337 + seed_offset)
 torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul
 torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn
 device_type = 'cuda' if 'cuda' in device else 'cpu' # for later use in torch.autocast
-# note: float16 would require us to change the code to use a GradScaler
+# note: float16 data type will automatically use a GradScaler
-ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16}[dtype]
+ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
 ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=ptdtype)
 
 # poor man's data loader, TODO evaluate need for actual DataLoader
@@ -173,6 +173,12 @@ if block_size < model.config.block_size:
     model.crop_block_size(block_size)
 model.to(device)
 
+# initialize a GradScaler if data type is float16
+scaler = None
+if dtype == 'float16':
+    print(f"Initializing Gradient Scaler to account for dtype: {dtype}")
+    scaler = torch.cuda.amp.GradScaler()
+
 # optimizer
 optimizer = model.configure_optimizers(weight_decay, learning_rate, (beta1, beta2))
 if init_from == 'resume':
@@ -188,6 +194,7 @@ if compile:
 if ddp:
     model = DDP(model, device_ids=[ddp_local_rank])
 
+# helps estimate an arbitrarily accurate loss over either split using many batches
 @torch.no_grad()
 def estimate_loss():
     out = {}
@@ -263,7 +270,9 @@ while True:
         break
 
     # forward backward update, with optional gradient accumulation to simulate larger batch size
+    # and using the GradScaler if data type is float16
     for micro_step in range(gradient_accumulation_steps):
+        # fetch a batch
         X, Y = get_batch('train')
         if ddp:
             # in DDP training we only need to sync gradients at the last micro step.
@@ -273,10 +282,19 @@ while True:
             model.require_backward_grad_sync = (micro_step == gradient_accumulation_steps - 1)
         with ctx:
             logits, loss = model(X, Y)
-        loss.backward()
+        # backward pass, with gradient scaling if training in fp16
-    if grad_clip != 0:
+        scaler.scale(loss).backward() if scaler else loss.backward()
+    # clip the gradient
+    if grad_clip != 0.0:
+        scaler.unscale_(optimizer) if scaler else None
         torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
-    optimizer.step()
+    # step the optimizer
+    if scaler:
+        scaler.step(optimizer)
+        scaler.update()
+    else:
+        optimizer.step()
+    # flush the gradients as soon as we can, no need for this memory anymore
     optimizer.zero_grad(set_to_none=True)
 
     # timing and logging