Fix gradient accumulation for variable length sequences

scripts/chat_sft.py

@@ -34,7 +34,7 @@ from tasks.smoltalk import SmolTalk
 # SFT Hyperparameters
 run = "dummy" # wandb run name default ("dummy" is special - we won't log to wandb)
 # input model options
-source = "mid" # base|mid , which checkpoint to load the model from (base model or midtrained model)
+source = "base" # base|mid , which checkpoint to load the model from (base model or midtrained model)
 model_tag = None # model tag to load the model from (base model or midtrained model)
 step = None # step to load the model from (base model or midtrained model)
 # compute/precision
@@ -208,18 +208,24 @@ for step in range(num_iterations):
         break
 
     # evaluate the gradient
+    total_loss_sum = torch.tensor(0.0, device=device)  # sum of losses
     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_iter)
         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 = model(train_inputs, train_targets, loss_reduction='sum')
+        total_loss_sum += loss.detach() # for logging
         loss.backward() # accumulate the gradient
         num_tokens += (train_targets >= 0).sum()
     if ddp:
+        dist.all_reduce(total_loss_sum, op=dist.ReduceOp.SUM)
         dist.all_reduce(num_tokens, op=dist.ReduceOp.SUM) # sum over ranks
 
+    # Scale gradients by total number of tokens
+    for param in model.parameters():
+        if param.grad is not None:
+            param.grad.div_(num_tokens.item())
+
     # learning rate scheduler
     lrm = get_lr_multiplier(step)
     for opt in optimizers:
@@ -232,8 +238,8 @@ for step in range(num_iterations):
     model.zero_grad(set_to_none=True)
 
     # logging
-    train_loss_item = train_loss.item()
     num_tokens_item = num_tokens.item()
+    train_loss_item = total_loss_sum.item() / num_tokens_item
     print0(f"Step {step:05d}/{num_iterations:05d} | Training loss: {train_loss_item:.6f}| lrm: {lrm:.6f}| num_tokens: {num_tokens_item:,}")
     wandb_run.log({
         "step": step,