update

applications/ColossalMoE/train_moe.py

@@ -1,6 +1,7 @@
 import argparse
 import os
 import json
+from typing import Dict, Union, List
 
 import torch
 from colossal_moe.models.mixtral_checkpoint import MixtralMoECheckpointIO
@@ -8,7 +9,7 @@ from colossal_moe.models.mixtral_layer import replace_moe_layer
 from colossal_moe.models.mixtral_policy import MixtralForCausalLMPolicy
 from colossal_moe.utils import load_checkpoint, load_model, move_to_cuda, save_checkpoint
 from tqdm import tqdm
-from transformers import AutoTokenizer
+from transformers import AutoTokenizer, PreTrainedModel
 from transformers.models.mixtral import MixtralConfig, MixtralForCausalLM
 import torch.distributed as dist
 
@@ -29,6 +30,35 @@ from colossal_moe.dataset.loader import (
     DataCollatorForSupervisedDataset,   
 )
 
+
+def get_optimizer_grouped_parameters(model: Union[torch.nn.Module, PreTrainedModel],
+                                     weight_decay: float = 0.0,
+                                     disable_decay_names: List[str] = None
+                                     ) -> List[Dict[str, Union[List[torch.nn.Parameter], float]]]:
+    if disable_decay_names is None:
+        disable_decay_names = ["bias", "LayerNorm.weight"]
+    optimizer_grouped_parameters = [
+        {
+            "params": [
+                p for n, p in model.named_parameters()
+                if p.requires_grad is True and not any(
+                    name.lower() in n.lower() for name in disable_decay_names
+                )
+            ],
+            "weight_decay": weight_decay
+        },
+        {
+            "params": [
+                p for n, p in model.named_parameters()
+                if p.requires_grad is True and any(
+                    name.lower() in n.lower() for name in disable_decay_names
+                )
+            ],
+            "weight_decay": 0.0
+        }
+    ]
+    return optimizer_grouped_parameters
+
 @torch.no_grad()
 def get_global_loss(loss, booster):
     global_loss = loss.clone().detach()
@@ -84,7 +114,7 @@ def parse_args():
 
     # optim
     parser.add_argument("--lr", type=float, default=1e-5, help="Learning rate.")
-    parser.add_argument("--weight_decay", type=float, default=0.1, help="Weight decay to use.")
+    parser.add_argument("--weight_decay", type=float, default=0.01, help="Weight decay to use.")
 
     # zero stage for all plugins
     parser.add_argument("--zero_stage", type=int, default=2, help="zero stage.")
@@ -194,7 +224,8 @@ def main():
 
     # Prepare tokenizer and dataloader
     tokenizer = AutoTokenizer.from_pretrained(args.model_name)
-    tokenizer.pad_token_id = tokenizer.eos_token_id
+    tokenizer.padding_side = "right"
+    tokenizer.pad_token = tokenizer.unk_token
     
     
     data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer, max_length=args.max_length)
@@ -205,10 +236,11 @@ def main():
 
     # Set optimizer
     optimizer = HybridAdam(
-        model_params=model.parameters(),
+        model_params=get_optimizer_grouped_parameters(
+            model=model, weight_decay=args.weight_decay
+        ),
         lr=args.lr,
         betas=(0.9, 0.95),
-        weight_decay=args.weight_decay,
         adamw_mode=True,
     )