[Enhancement] Support ZeRO-3 when using BAdam

.gitignore

@@ -163,3 +163,5 @@ cython_debug/
 user.config
 saves/
 cache/
+wandb
+ds_badam_exp

examples/extras/badam/llama3_badam_sft.yaml

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+use_badam: true
+badam_switch_mode: ascending
+badam_switch_interval: 50
+badam_verbose: 2
+
+### dataset
+dataset: identity,alpaca_en_demo
+template: llama3
+cutoff_len: 1024
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: saves/llama3-8b/full/sft
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-6
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 1
+eval_strategy: steps
+eval_steps: 500

examples/extras/badam/train_single_gpu.sh

@@ -0,0 +1,37 @@
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=0
+
+cd ../../..
+
+llamafactory-cli train \
+    --stage sft \
+    --do_train True \
+    --model_name_or_path meta-llama/Llama-2-13b-hf \
+    --preprocessing_num_workers 16 \
+    --finetuning_type full \
+    --template default \
+    --flash_attn auto \
+    --dataset_dir data \
+    --dataset alpaca_en_demo \
+    --cutoff_len 1024 \
+    --learning_rate 1e-6 \
+    --num_train_epochs 3.0 \
+    --max_samples 100000 \
+    --per_device_train_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --max_grad_norm 1.0 \
+    --logging_steps 5 \
+    --save_steps 100 \
+    --warmup_steps 0 \
+    --optim adamw_torch \
+    --packing False \
+    --report_to none \
+    --use_badam True \
+    --output_dir saves/LLaMA2-13B/full/BAdam \
+    --plot_loss True \
+    --ddp_timeout 180000000 \
+    --include_num_input_tokens_seen True \
+    --badam_mode layer \
+    --badam_switch_mode ascending \
+    --badam_switch_interval 50

examples/extras/badam/train_zero3.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+
+cd ../../..
+
+llamafactory-cli train \
+    --stage sft \
+    --do_train True \
+    --model_name_or_path meta-llama/Llama-2-13b-hf \
+    --preprocessing_num_workers 16 \
+    --finetuning_type full \
+    --template default \
+    --flash_attn auto \
+    --dataset_dir data \
+    --dataset alpaca_en_demo \
+    --cutoff_len 1024 \
+    --learning_rate 1e-6 \
+    --num_train_epochs 3.0 \
+    --max_samples 100000 \
+    --per_device_train_batch_size 8 \
+    --gradient_accumulation_steps 2 \
+    --lr_scheduler_type cosine \
+    --max_grad_norm 1.0 \
+    --logging_steps 5 \
+    --save_steps 100 \
+    --warmup_steps 0 \
+    --optim adamw_torch \
+    --packing False \
+    --report_to none \
+    --use_badam True \
+    --output_dir saves/LLaMA2-13B/full/BAdam \
+    --fp16 True \
+    --plot_loss True \
+    --ddp_timeout 180000000 \
+    --include_num_input_tokens_seen True \
+    --badam_mode layer \
+    --badam_switch_mode ascending \
+    --badam_switch_interval 50 \
+    --deepspeed cache/ds_z3_config.json 

src/llamafactory/hparams/parser.py

@@ -211,13 +211,15 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
 
     if (
         finetuning_args.use_badam
-        and finetuning_args.badam_mode == "layer"
-        and training_args.parallel_mode == ParallelMode.DISTRIBUTED
+        and training_args.parallel_mode.value == "distributed"
     ):
-        raise ValueError("Layer-wise BAdam does not yet support distributed training, use ratio-wise BAdam.")
+        if finetuning_args.badam_mode == "ratio":
+            raise ValueError("Ratio-wise BAdam does not yet support distributed training, use layer-wise BAdam: --badam_mode layer")
+        if finetuning_args.badam_mode == "layer" and (not is_deepspeed_zero3_enabled()):
+            raise ValueError(f"Layer-wise BAdam only supports DeepSpeed ZeRO 3 stage.")
 
-    if (finetuning_args.use_galore or finetuning_args.use_badam) and training_args.deepspeed is not None:
-        raise ValueError("GaLore and BAdam are incompatible with DeepSpeed yet.")
+    if (finetuning_args.use_galore) and training_args.deepspeed is not None:
+        raise ValueError("GaLore are incompatible with DeepSpeed yet.")
 
     if model_args.infer_backend == "vllm":
         raise ValueError("vLLM backend is only available for API, CLI and Web.")

src/llamafactory/train/dpo/trainer.py

@@ -96,9 +96,9 @@ def __init__(
             self.save_model(os.path.join(self.args.output_dir, "pissa_init"))
 
         if finetuning_args.use_badam:
-            from badam import clip_grad_norm_for_sparse_tensor
-
-            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_for_sparse_tensor, self.accelerator)
+            from badam import clip_grad_norm_old_version, BAdamCallback
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.callback_handler.add_callback(BAdamCallback)
 
     def create_optimizer(self) -> "torch.optim.Optimizer":
         if self.optimizer is None:

src/llamafactory/train/kto/trainer.py

@@ -91,9 +91,9 @@ def __init__(
                 self.ref_model.eval()
 
         if finetuning_args.use_badam:
-            from badam import clip_grad_norm_for_sparse_tensor
-
-            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_for_sparse_tensor, self.accelerator)
+            from badam import clip_grad_norm_old_version, BAdamCallback
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.callback_handler.add_callback(BAdamCallback)
 
     def create_optimizer(self) -> "torch.optim.Optimizer":
         if self.optimizer is None:

src/llamafactory/train/ppo/trainer.py

@@ -166,9 +166,9 @@ def __init__(
                 self.reward_model = self.accelerator.prepare_model(self.reward_model, evaluation_mode=True)
 
         if finetuning_args.use_badam:
-            from badam import clip_grad_norm_for_sparse_tensor
-
-            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_for_sparse_tensor, self.accelerator)
+            from badam import clip_grad_norm_old_version, BAdamCallback
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.callback_handler.add_callback(BAdamCallback)
 
     def ppo_train(self, resume_from_checkpoint: Optional[str] = None) -> None:
         r"""

src/llamafactory/train/pt/trainer.py

@@ -48,9 +48,9 @@ def __init__(
             self.save_model(os.path.join(self.args.output_dir, "pissa_init"))
 
         if finetuning_args.use_badam:
-            from badam import clip_grad_norm_for_sparse_tensor
-
-            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_for_sparse_tensor, self.accelerator)
+            from badam import clip_grad_norm_old_version, BAdamCallback
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.callback_handler.add_callback(BAdamCallback)
 
     def create_optimizer(self) -> "torch.optim.Optimizer":
         if self.optimizer is None:

src/llamafactory/train/rm/trainer.py

@@ -72,9 +72,9 @@ def __init__(
         self.processor = processor
         self.can_return_loss = True  # override property to return eval_loss
         if finetuning_args.use_badam:
-            from badam import clip_grad_norm_for_sparse_tensor
-
-            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_for_sparse_tensor, self.accelerator)
+            from badam import clip_grad_norm_old_version, BAdamCallback
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.callback_handler.add_callback(BAdamCallback)
 
     def create_optimizer(self) -> "torch.optim.Optimizer":
         if self.optimizer is None:

src/llamafactory/train/sft/trainer.py

@@ -56,9 +56,9 @@ def __init__(
             self.save_model(os.path.join(self.args.output_dir, "pissa_init"))
 
         if finetuning_args.use_badam:
-            from badam import clip_grad_norm_for_sparse_tensor
-
-            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_for_sparse_tensor, self.accelerator)
+            from badam import clip_grad_norm_old_version, BAdamCallback
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.callback_handler.add_callback(BAdamCallback)
 
     def create_optimizer(self) -> "torch.optim.Optimizer":
         if self.optimizer is None:

src/llamafactory/train/trainer_utils.py

@@ -371,6 +371,9 @@ def _create_badam_optimizer(
         dict(params=decay_params, weight_decay=training_args.weight_decay),
     ]
 
+    from transformers.integrations import is_deepspeed_zero3_enabled
+    ds_zero3_enabled = is_deepspeed_zero3_enabled()
+
     if finetuning_args.badam_mode == "layer":
         from badam import BlockOptimizer
 
@@ -383,6 +386,7 @@ def _create_badam_optimizer(
             start_block=finetuning_args.badam_start_block,
             switch_mode=finetuning_args.badam_switch_mode,
             verbose=finetuning_args.badam_verbose,
+            ds_zero3_enabled=ds_zero3_enabled
         )
         logger.info(
             f"Using BAdam optimizer with layer-wise update, switch mode is {finetuning_args.badam_switch_mode}, "
@@ -393,6 +397,7 @@ def _create_badam_optimizer(
     elif finetuning_args.badam_mode == "ratio":
         from badam import BlockOptimizerRatio
 
+        assert not ds_zero3_enabled, "BAdam with ratio-based update does not support Deepspeed ZeRO-3 yet, use layer-wise update instead: --badam_mode layer."
         assert finetuning_args.badam_update_ratio > 1e-6
         optimizer = BlockOptimizerRatio(
             param_groups=param_groups,
@@ -404,7 +409,7 @@ def _create_badam_optimizer(
             **optim_kwargs,
         )
         logger.info(
-            f"Using BAdam optimizer with ratio-wise update, update ratio is {finetuning_args.badam_update_ratio}, "
+            f"Using BAdam optimizer with ratio-based update, update ratio is {finetuning_args.badam_update_ratio}, "
             f"mask mode is {finetuning_args.badam_mask_mode}"
         )