add reset learning rate functionality (NVIDIA#9372)

* add reset_lr functionality Signed-off-by: dimapihtar <dpihtar@gmail.com> * fix reset_lr logic Signed-off-by: dimapihtar <dpihtar@gmail.com> * Apply isort and black reformatting Signed-off-by: dimapihtar <dimapihtar@users.noreply.github.com> * move reset_lr from optim section Signed-off-by: dimapihtar <dpihtar@gmail.com> * Apply isort and black reformatting Signed-off-by: dimapihtar <dimapihtar@users.noreply.github.com> * add reset_lr value to config Signed-off-by: dimapihtar <dpihtar@gmail.com> * set reset_lr False by default Signed-off-by: dimapihtar <dpihtar@gmail.com> * remove extra line Signed-off-by: dimapihtar <dpihtar@gmail.com> * add reset_lr test Signed-off-by: dimapihtar <dpihtar@gmail.com> * add reset_lr test Signed-off-by: dimapihtar <dpihtar@gmail.com> * remove extra quote Signed-off-by: dimapihtar <dpihtar@gmail.com> * add ability to reset schedule's max_steps and decay_steps Signed-off-by: dimapihtar <dpihtar@gmail.com> * Apply isort and black reformatting Signed-off-by: dimapihtar <dimapihtar@users.noreply.github.com> * change scheduler's first step logic when using reset_lr Signed-off-by: dimapihtar <dpihtar@gmail.com> * revert config Signed-off-by: dimapihtar <dpihtar@gmail.com> * fix reset_lr logic Signed-off-by: dimapihtar <dpihtar@gmail.com> * Apply isort and black reformatting Signed-off-by: dimapihtar <dimapihtar@users.noreply.github.com> * revert config Signed-off-by: dimapihtar <dpihtar@gmail.com> * revert config Signed-off-by: dimapihtar <dpihtar@gmail.com> * update reset_lr comments Signed-off-by: dimapihtar <dpihtar@gmail.com> * add use cases for reset_lr feature Signed-off-by: dimapihtar <dpihtar@gmail.com> --------- Signed-off-by: dimapihtar <dpihtar@gmail.com> Signed-off-by: dimapihtar <dimapihtar@users.noreply.github.com> Co-authored-by: dimapihtar <dimapihtar@users.noreply.github.com> Signed-off-by: Alex Cui <alexcui1994@gmail.com>
BuyuanCui · Jul 12, 2024 · 11512aa · 11512aa
1 parent cd66e93
commit 11512aa
Show file tree

Hide file tree

Showing 5 changed files with 148 additions and 6 deletions.
diff --git a/.github/workflows/cicd-main.yml b/.github/workflows/cicd-main.yml
@@ -2630,6 +2630,89 @@ jobs:
     #    }
     #  }
 
+  L2_Megatron_GPT_with_ResetLR_Pretraining_and_Resume_Training_TP2:
+    needs: [cicd-test-container-setup]
+    runs-on: self-hosted-azure
+    timeout-minutes: 10
+    container:
+      image: nemoci.azurecr.io/nemo_container_${{ github.run_id }}
+      options: 
+        # --user 0:128
+        --device=/dev/nvidia0
+        --gpus all
+        --shm-size=8g
+        --env TRANSFORMERS_OFFLINE=0 
+        --env HYDRA_FULL_ERROR=1
+        --volume /mnt/datadrive/TestData:/home/TestData
+    steps:
+        - name: Checkout repository
+          uses: actions/checkout@v4
+        - run: |
+           python examples/nlp/language_modeling/megatron_gpt_pretraining.py \
+           trainer.devices=2 \
+           trainer.accelerator=gpu \
+           trainer.log_every_n_steps=1 \
+           trainer.val_check_interval=3 \
+           trainer.limit_val_batches=2 \
+           trainer.accumulate_grad_batches=1 \
+           trainer.max_steps=3 \
+           trainer.precision=bf16 \
+           trainer.gradient_clip_val=1.0 \
+           exp_manager.exp_dir=examples/nlp/language_modeling/gpt_pretrain_results \
+           model.tensor_model_parallel_size=2 \
+           model.megatron_amp_O2=True \
+           model.optim.name=distributed_fused_adam \
+           model.optim.lr=2e-4 \
+           model.optim.sched.warmup_steps=2 \
+           model.optim.sched.constant_steps=2 \
+           model.optim.sched.min_lr=8e-5 \
+           model.max_position_embeddings=128 \
+           model.encoder_seq_length=128 \
+           model.data.seq_length=128 \
+           model.tokenizer.vocab_file=/home/TestData/nlp/megatron_gpt/data/gpt/vocab.json \
+           model.tokenizer.merge_file=/home/TestData/nlp/megatron_gpt/data/gpt/merges.txt \
+           model.num_layers=8 \
+           model.hidden_size=256 \
+           model.num_attention_heads=8 \
+           model.data.data_prefix=[.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document,.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document] \
+           model.data.index_mapping_dir=examples/nlp/language_modeling/gpt_index_mappings
+        
+           python examples/nlp/language_modeling/megatron_gpt_pretraining.py \
+           trainer.devices=2 \
+           trainer.accelerator=gpu \
+           trainer.log_every_n_steps=1 \
+           trainer.val_check_interval=3 \
+           trainer.limit_val_batches=2 \
+           trainer.accumulate_grad_batches=1 \
+           trainer.max_steps=6 \
+           trainer.precision=bf16 \
+           trainer.gradient_clip_val=1.0 \
+           exp_manager.exp_dir=examples/nlp/language_modeling/gpt_pretrain_results \
+           exp_manager.resume_if_exists=True \
+           model.reset_lr=True \
+           model.tensor_model_parallel_size=2 \
+           model.megatron_amp_O2=True \
+           model.optim.name=distributed_fused_adam \
+           model.optim.lr=2e-4 \
+           model.optim.sched.warmup_steps=2 \
+           model.optim.sched.constant_steps=2 \
+           model.optim.sched.min_lr=8e-5 \
+           model.max_position_embeddings=128 \
+           model.encoder_seq_length=128 \
+           model.data.seq_length=128 \
+           model.tokenizer.vocab_file=/home/TestData/nlp/megatron_gpt/data/gpt/vocab.json \
+           model.tokenizer.merge_file=/home/TestData/nlp/megatron_gpt/data/gpt/merges.txt \
+           model.num_layers=8 \
+           model.hidden_size=256 \
+           model.num_attention_heads=8 \
+           model.data.data_prefix=[.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document,.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document] \
+           model.data.index_mapping_dir=examples/nlp/language_modeling/gpt_index_mappings
+
+           rm -rf examples/nlp/language_modeling/gpt_pretrain_results
+           rm -rf examples/nlp/language_modeling/gpt_index_mappings
+        - uses: "NVIDIA/NeMo/.github/actions/cancel-workflow@main"
+          if: "failure()"
+
   L2_Megatron_GPT_with_ALiBi_Pretraining_and_Resume_Training_TP2:
     needs: [cicd-test-container-setup]
     runs-on: self-hosted-azure
@@ -4296,6 +4379,7 @@ jobs:
       - L2_BioMegatron_Bert_NER_Task
       - L2_Megatron_GPT_Pretraining_and_Resume_Training_TP2
       - L2_Megatron_GPT_with_Rope_Pretraining_and_Resume_Training_TP2
+      - L2_Megatron_GPT_with_ResetLR_Pretraining_and_Resume_Training_TP2
       - L2_Megatron_GPT_with_ALiBi_Pretraining_and_Resume_Training_TP2
       - L2_Megatron_GPT_with_KERPLE_Pretraining_and_Resume_Training_TP2
       - L2_Megatron_GPT_Pretraining_and_Resume_Training_PP2

diff --git a/examples/nlp/language_modeling/conf/megatron_gpt_config.yaml b/examples/nlp/language_modeling/conf/megatron_gpt_config.yaml
@@ -115,6 +115,14 @@ model:
   seq_len_interpolation_factor: null # RoPE Interpolation factor for sequence length. This is used to build long-context models with RoPE ex: https://arxiv.org/abs/2306.15595.
   num_query_groups: null # Number of query groups for group query attention. If None, normal attention is used.
 
+  ## Reset learning rate schedule.
+  # 1. reset_lr=True, reset_lr_steps=False. When pre-training an existing checkpoint "from scratch" on a different dataset.
+  # 2. reset_lr=True, reset_lr_steps=True. When continuing training from an existing checkpoint with the same configuration.
+  #    Learning rate's max_steps and decay_steps will be recalculated as follows: max_steps -= completed_steps, decay_steps -= completed_steps where completed_steps is the number of steps already completed at the checkpoint.
+  #    This will help to reach the min_lr value by the end of training without changing trainer.max_steps.
+  reset_lr: False # Set to True to reset learning rate to initial learning rate. Only supported with distributed optmizer and megatron_amp_O2.
+  reset_lr_steps: False # Set to True to adjust learning rate's max_steps and decay_steps by subtracting number of steps already completed at the checkpoint.
+
   tokenizer:
     library: 'megatron'
     type: 'GPT2BPETokenizer'

diff --git a/nemo/collections/nlp/models/language_modeling/megatron_base_model.py b/nemo/collections/nlp/models/language_modeling/megatron_base_model.py
@@ -846,7 +846,9 @@ def configure_optimizers(self):
             if hasattr(self._cfg.optim, 'sched'):
                 sched_config = self._cfg.optim.sched
                 self._scheduler = prepare_lr_scheduler(
-                    optimizer=self._optimizer, scheduler_config=sched_config, train_dataloader=self._train_dl
+                    optimizer=self._optimizer,
+                    scheduler_config=sched_config,
+                    train_dataloader=self._train_dl,
                 )
 
         if getattr(self._cfg.optim, 'sched', None) is not None and self._scheduler is None:

diff --git a/nemo/collections/nlp/models/language_modeling/megatron_gpt_model.py b/nemo/collections/nlp/models/language_modeling/megatron_gpt_model.py
@@ -397,6 +397,15 @@ def __init__(self, cfg: DictConfig, trainer: Trainer):
 
         self.inference_params = None
 
+        # Reset learning rate params
+        self.if_init_step = True
+        self.reset_lr = self.cfg.get('reset_lr', False)
+        self.reset_lr_steps = self.cfg.get('reset_lr_steps', False)
+        if self.reset_lr and (not self.with_distributed_adam or not self.megatron_amp_O2):
+            raise ValueError(
+                'Learning rate reset feature is only supported with the distributed optmizer and megatron_amp_O2 for now.'
+            )
+
         # default to false since this doesn't work with sequence parallelism currently
         self.use_loss_mask = self.cfg.get('use_loss_mask', False)
 
@@ -763,6 +772,20 @@ def training_step(self, dataloader_iter):
         if self.initialize_ub:
             self.initialize_ub_func()
 
+        # Reset learning rate
+        if self.if_init_step and self.reset_lr:
+            num_groups = len(self._optimizer.param_groups)
+            for group in range(num_groups):
+                self._optimizer.param_groups[group]['lr'] = (
+                    0.0 if self.cfg.optim.sched.warmup_steps > 0 else self.cfg.optim.lr
+                )
+            self._optimizer.param_groups[0]['reset_lr'] = {
+                'num_steps': self.trainer.global_step,
+                'reset_lr_steps': True if self.reset_lr_steps else False,
+                'if_init_step': self.if_init_step,
+            }
+            self.if_init_step = False
+
         if self.rampup_batch_size:
             num_microbatch_calculator = apex.transformer.pipeline_parallel.utils._GLOBAL_NUM_MICROBATCHES_CALCULATOR
             current_global_batch_size = num_microbatch_calculator.current_global_batch_size

diff --git a/nemo/core/optim/lr_scheduler.py b/nemo/core/optim/lr_scheduler.py
@@ -97,7 +97,14 @@ class SquareRootConstantPolicy(_LRScheduler):
     """
 
     def __init__(
-        self, optimizer, *, constant_steps=None, constant_ratio=None, max_steps=None, min_lr=0.0, last_epoch=-1
+        self,
+        optimizer,
+        *,
+        constant_steps=None,
+        constant_ratio=None,
+        max_steps=None,
+        min_lr=0.0,
+        last_epoch=-1,
     ):
         assert not (
             constant_steps is not None and constant_ratio is not None
@@ -114,7 +121,7 @@ def __init__(
         else:
             self.constant_steps = 0
 
-        self.constant_lr = 1 / (constant_steps ** 0.5)
+        self.constant_lr = 1 / (constant_steps**0.5)
         self.min_lr = min_lr
         super().__init__(optimizer, last_epoch)
 
@@ -280,6 +287,16 @@ def get_lr(self):
 
         step = self.last_epoch
 
+        # Reset learning rate
+        if 'reset_lr' in self.optimizer.param_groups[0].keys():
+            reset_lr = self.optimizer.param_groups[0]['reset_lr']
+            num_steps = reset_lr['num_steps']
+            step -= num_steps
+            if reset_lr['if_init_step'] and reset_lr['reset_lr_steps']:
+                self.decay_steps -= num_steps
+                self.max_steps -= num_steps
+                self.optimizer.param_groups[0]['reset_lr']['if_init_step'] = False
+
         # Warmup steps
         if self.warmup_steps > 0 and step <= self.warmup_steps:
             return self._get_warmup_lr(step)
@@ -364,7 +381,7 @@ def _poly_decay(initial_lr, step, decay_steps, power, min_lr, cycle):
 
 def _noam_hold_annealing(initial_lr, step, warmup_steps, hold_steps, decay_rate, min_lr):
     # hold_steps = total number of steps to hold the LR, not the warmup + hold steps.
-    T_warmup_decay = max(1, warmup_steps ** decay_rate)
+    T_warmup_decay = max(1, warmup_steps**decay_rate)
     T_hold_decay = max(1, (step - hold_steps) ** decay_rate)
     lr = (initial_lr * T_warmup_decay) / T_hold_decay
     lr = max(lr, min_lr)
@@ -453,7 +470,15 @@ def _get_linear_warmup_with_cosine_annealing_lr(self, step):
 
 class NoamAnnealing(_LRScheduler):
     def __init__(
-        self, optimizer, *, d_model, warmup_steps=None, warmup_ratio=None, max_steps=None, min_lr=0.0, last_epoch=-1
+        self,
+        optimizer,
+        *,
+        d_model,
+        warmup_steps=None,
+        warmup_ratio=None,
+        max_steps=None,
+        min_lr=0.0,
+        last_epoch=-1,
     ):
         self._normalize = d_model ** (-0.5)
         assert not (
@@ -593,7 +618,7 @@ def __init__(self, optimizer, *, max_steps, last_epoch=-1, min_lr=0.0, **kwargs)
         super().__init__(optimizer=optimizer, max_steps=max_steps, **kwargs, last_epoch=last_epoch, min_lr=min_lr)
 
     def _get_lr(self, step):
-        return [1 / (step ** 0.5) for _ in self.base_lrs]
+        return [1 / (step**0.5) for _ in self.base_lrs]
 
 
 class PolynomialDecayAnnealing(WarmupPolicy):