Jst/support multi epoch pp qlora finetuning

large_language_models/alpaca-qlora/finetune_pp.py

@@ -120,7 +120,7 @@ def tokenize(prompt):
         batch_size=args.micro_batch_size,
         sampler=RandomSampler(train_dataset, generator=generator),
         collate_fn=data_collator,
-        drop_last=False,
+        drop_last=True,
         num_workers=0,
         pin_memory=True,
         worker_init_fn=seed_worker,
@@ -148,7 +148,7 @@ def tokenize(prompt):
     lr_scheduler = get_linear_schedule_with_warmup(
         optimizer=optimizer,
         num_warmup_steps=100,
-        num_training_steps=len(epoch_iterator) // GRADIENT_ACCUMULATION_STEPS,
+        num_training_steps=EPOCHS * len(epoch_iterator) // GRADIENT_ACCUMULATION_STEPS,
     )
 
     scaler = torch.cuda.amp.GradScaler()
@@ -158,47 +158,48 @@ def tokenize(prompt):
     model.eval()
     model.train()
 
-    step = 0
-    accumulated_loss = 0
-    for inputs in tqdm(epoch_iterator):
-        step += 1
-        with torch.cuda.amp.autocast(cache_enabled=True, dtype=torch.float16):
-            labels = inputs.pop("labels")
-            outputs = model(**inputs)
-            outputs["logits"] = outputs["logits"].float()
-            labels = labels.to(outputs["logits"].device)
-            shift_logits = outputs["logits"][..., :-1, :].contiguous()
-            shift_labels = labels[..., 1:].contiguous()
-            # Flatten the tokens
-            shift_logits = shift_logits.view(-1, shift_logits.shape[-1])
-            shift_labels = shift_labels.view(-1)
-            # Enable model parallelism
-            shift_labels = shift_labels.to(shift_logits.device)
-            loss = loss_fct(shift_logits, shift_labels)
-
-        loss /= GRADIENT_ACCUMULATION_STEPS
-        accumulated_loss += loss.item()
-        scaler.scale(loss).backward()
-
-        if step % GRADIENT_ACCUMULATION_STEPS == 0:
-            scaler.unscale_(optimizer)
-            torch.nn.utils.clip_grad_norm_(
-                filter(lambda p: p.requires_grad, model.parameters()), max_norm=1.0
-            )
-            scaler.step(optimizer)
-            scaler.update()
-            lr_scheduler.step()
-            model.zero_grad()
-            tqdm.write(
-                "{"
-                + "'loss': {0:1.4f}, 'learning_rate': {1:2.6f}, 'epoch': {2:3.2f}".format(
-                    accumulated_loss,
-                    optimizer.param_groups[0]["lr"],
-                    step / len(epoch_iterator),
+    for epoch in range(EPOCHS):
+        step = 0
+        accumulated_loss = 0
+        for inputs in tqdm(epoch_iterator):
+            step += 1
+            with torch.cuda.amp.autocast(cache_enabled=True, dtype=torch.float16):
+                labels = inputs.pop("labels")
+                outputs = model(**inputs)
+                outputs["logits"] = outputs["logits"].float()
+                labels = labels.to(outputs["logits"].device)
+                shift_logits = outputs["logits"][..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous()
+                # Flatten the tokens
+                shift_logits = shift_logits.view(-1, shift_logits.shape[-1])
+                shift_labels = shift_labels.view(-1)
+                # Enable model parallelism
+                shift_labels = shift_labels.to(shift_logits.device)
+                loss = loss_fct(shift_logits, shift_labels)
+
+            loss /= GRADIENT_ACCUMULATION_STEPS
+            accumulated_loss += loss.item()
+            scaler.scale(loss).backward()
+
+            if step % GRADIENT_ACCUMULATION_STEPS == 0:
+                scaler.unscale_(optimizer)
+                torch.nn.utils.clip_grad_norm_(
+                    filter(lambda p: p.requires_grad, model.parameters()), max_norm=1.0
                 )
-                + "}"
-            )
-            accumulated_loss = 0
+                scaler.step(optimizer)
+                scaler.update()
+                lr_scheduler.step()
+                model.zero_grad()
+                tqdm.write(
+                    "{"
+                    + "'loss': {0:1.4f}, 'learning_rate': {1:2.6f}, 'epoch': {2:3.2f}".format(
+                        accumulated_loss,
+                        optimizer.param_groups[0]["lr"],
+                        step / len(epoch_iterator) + epoch,
+                    )
+                    + "}"
+                )
+                accumulated_loss = 0
 
     print("Peak memory usage for GPUs: ", end="")
     for i in range(len(model.model.devices)):