Merge a39c303912 into 63bb5831e2

scripts/chat_sft.py

@@ -216,18 +216,20 @@ for step in range(num_iterations):
         break
 
     # evaluate the gradient
+    steps = [next(train_loader) for _ in range(grad_accum_steps)]
     num_tokens = torch.tensor(0, device=device) # the number of "active" tokens of supervision seen
-    for micro_step in range(grad_accum_steps):
-        train_inputs, train_targets = next(train_loader)
-        with autocast_ctx:
-            loss = model(train_inputs, train_targets)
-        train_loss = loss.detach() # for logging
-        loss = loss / grad_accum_steps # each .backward() is a grad sum => normalize loss here
-        loss.backward() # accumulate the gradient
-        num_tokens += (train_targets >= 0).sum()
+    num_tokens += sum((targets >= 0).sum() for _, targets in steps)
+
     if ddp:
         dist.all_reduce(num_tokens, op=dist.ReduceOp.SUM) # sum over ranks
 
+    for micro_step, (train_inputs, train_targets) in enumerate(steps):
+        with autocast_ctx:
+            loss = model(train_inputs, train_targets, loss_reduction='sum')
+        loss = loss / num_tokens # normalize loss here
+        train_loss = loss.detach() # for logging
+        loss.backward() # accumulate the gradient
+
     # learning rate scheduler
     lrm = get_lr_multiplier(step)
     for opt in optimizers: