feat: support for mcore optimizer (to enable MoE)

.github/workflows/cicd-main.yml

@@ -98,6 +98,8 @@ jobs:
           - sft-llama3-cp
           - rm-llama3
           - e2e-nemo2
+          - dpo-mixtral-ep
+          - dpo-mixtral-peft-tp-sp
     with:
       RUNNER: self-hosted-azure
       # Fairly aggresive timeout that all functional tests should try to adhere to

examples/nlp/gpt/conf/gpt_dpo.yaml

@@ -6,6 +6,7 @@ trainer:
   devices: 8
   accelerator: gpu
   precision: bf16
+  gradient_clip_val: 0.0  # No need to change. Megatron Core optimizer uses this value
 
   # dpo specific args
   dpo:
@@ -17,6 +18,7 @@ trainer:
 
     # how many GBS we loop over
     limit_val_batches: 1.0
+    # TODO: delete once Megatron Core optimizer becomes default
     gradient_clip_val: 1.0
 
   # do not change these

examples/nlp/gpt/conf/gpt_kto.yaml

@@ -6,6 +6,7 @@ trainer:
   devices: 8
   accelerator: gpu
   precision: bf16
+  gradient_clip_val: 0.0  # No need to change. Megatron Core optimizer uses this value
 
   # kto specific args
   kto:
@@ -17,6 +18,7 @@ trainer:
 
     # how many GBS we loop over
     limit_val_batches: 1.0
+    # TODO: delete once Megatron Core optimizer becomes default
     gradient_clip_val: 1.0
 
   # do not change these

examples/nlp/gpt/conf/gpt_ppo_actor.yaml

@@ -7,6 +7,7 @@ trainer:
   devices: 8
   accelerator: gpu
   precision: bf16
+  gradient_clip_val: 0.0  # No need to change. Megatron Core optimizer uses this value
 
   ppo:
     # How many steps we train warmup the critic for (without training the policy)
@@ -21,6 +22,7 @@ trainer:
     max_steps: -1  # max PPO steps (-1 to go through the whole train set)
     val_check_interval: 10
     save_interval: ${.val_check_interval}
+    # TODO: delete once Megatron Core optimizer becomes default
     gradient_clip_val: 1.0
 
     # PPO args to generate the data for training

examples/nlp/gpt/conf/gpt_ppo_critic.yaml

@@ -6,6 +6,7 @@ trainer:
   devices: 8
   accelerator: gpu
   precision: bf16
+  gradient_clip_val: 0.0  # No need to change. Megatron Core optimizer uses this value
 
   ppo:
     port: 5556
@@ -15,6 +16,7 @@ trainer:
 
     # used to set the learning rate scheduler
     max_steps: 10000
+    # TODO: delete once Megatron Core optimizer becomes default
     gradient_clip_val: 1.0
 
     # a PyTriton parameter to specify

examples/nlp/gpt/conf/gpt_rs_actor.yaml

@@ -7,12 +7,14 @@ trainer:
   devices: 8
   accelerator: gpu
   precision: bf16
+  gradient_clip_val: 0.0  # No need to change. Megatron Core optimizer uses this value
 
   rs:
     max_epochs: 1
     max_steps: -1  # max rs steps (-1 to go through the whole train set)
     val_check_interval: 10
     save_interval: ${.val_check_interval}
+    # TODO: delete once Megatron Core optimizer becomes default
     gradient_clip_val: 1.0
 
     # pick up from the model
@@ -178,4 +180,4 @@ model:
   # define fields from the base model's config that should be ignored when merging with this config.
   overwrite_base_config:
     data:
-      data_prefix: True
+      data_prefix: True

examples/nlp/gpt/conf/gpt_sft.yaml

@@ -5,6 +5,7 @@ trainer:
   devices: 1
   accelerator: gpu
   precision: bf16
+  gradient_clip_val: 0.0  # No need to change. Megatron Core optimizer uses this value
 
   sft:
     max_epochs: 1
@@ -15,6 +16,7 @@ trainer:
     limit_train_batches: 1.0
 
     limit_val_batches: 1.0
+    # TODO: delete once Megatron Core optimizer becomes default
     gradient_clip_val: 1.0
 
     # can be used to register any custom metrics that require token-by-token generation

examples/nlp/gpt/conf/gpt_spin.yaml

@@ -6,6 +6,7 @@ trainer:
   devices: 8
   accelerator: gpu
   precision: bf16-mixed
+  gradient_clip_val: 0.0  # No need to change. Megatron Core optimizer uses this value
 
   # spin specific args
   spin:
@@ -18,6 +19,7 @@ trainer:
 
     # how many GBS we loop over
     limit_val_batches: 1.0
+    # TODO: delete once Megatron Core optimizer becomes default
     gradient_clip_val: 1.0
 
   # do not change these

examples/nlp/gpt/conf/training_rm.yaml

@@ -6,6 +6,7 @@ trainer:
   devices: 8
   accelerator: gpu
   precision: bf16
+  gradient_clip_val: 0.0  # No need to change. Megatron Core optimizer uses this value
 
   # rm specific args
   rm:
@@ -20,6 +21,7 @@ trainer:
     # set to float for a percentage
     # of the validation dataset
     limit_val_batches: 1.0
+    # TODO: delete once Megatron Core optimizer becomes default
     gradient_clip_val: 1.0
 
   # do not change these

nemo_aligner/algorithms/critic_server_trainer.py

@@ -322,7 +322,7 @@ def run_training(self, tokens=None, returns=None, prev_values=None, mask=None):
             grad_norm = grad_norm.item() if torch.is_tensor(grad_norm) else grad_norm
             lr = self.optimizer.param_groups[0]["lr"]
 
-            self.optimizer.step()
+            self.optimizer.step(closure=None)
             self.scheduler.step()
 
             if grad_norm is not None:

nemo_aligner/algorithms/dpo.py

@@ -220,7 +220,7 @@ def train_single_step(self, global_batch):
         grad_norm = grad_norm.item() if torch.is_tensor(grad_norm) else grad_norm
         lr = self.optimizer.param_groups[0]["lr"]
 
-        self.optimizer.step()
+        self.optimizer.step(closure=None)
         self.scheduler.step()
 
         trainer_metrics = {}

nemo_aligner/algorithms/ppo.py

@@ -440,7 +440,7 @@ def run_training(self, dataloader_iter):
             grad_norm = grad_norm.item() if torch.is_tensor(grad_norm) else grad_norm
             lr = self.optimizer.param_groups[0]["lr"]
 
-            self.optimizer.step()
+            self.optimizer.step(closure=None)
             self.scheduler.step()
 
             if grad_norm is not None:

nemo_aligner/algorithms/rs.py

@@ -294,7 +294,7 @@ def run_training(self, dataloader_iter):
             grad_norm = grad_norm.item() if torch.is_tensor(grad_norm) else grad_norm
             lr = self.optimizer.param_groups[0]["lr"]
 
-            self.optimizer.step()
+            self.optimizer.step(closure=None)
             self.scheduler.step()
 
             if grad_norm is not None:

nemo_aligner/algorithms/spin.py

@@ -195,7 +195,7 @@ def train_single_step(self, global_batch):
         grad_norm = grad_norm.item() if torch.is_tensor(grad_norm) else grad_norm
         lr = self.optimizer.param_groups[0]["lr"]
 
-        self.optimizer.step()
+        self.optimizer.step(closure=None)
         self.scheduler.step()
 
         trainer_metrics = {}

nemo_aligner/algorithms/supervised.py

@@ -150,7 +150,7 @@ def train_single_step(self, batch):
         grad_norm = grad_norm.item() if torch.is_tensor(grad_norm) else grad_norm
         lr = self.optimizer.param_groups[0]["lr"]
 
-        self.optimizer.step()
+        self.optimizer.step(closure=None)
         self.scheduler.step()
 
         trainer_metrics = {}

nemo_aligner/utils/train_utils.py

@@ -101,31 +101,52 @@ def prepare_for_training_step(ptl_model, zero_grad=True):
                     param.data_ptr()
 
 
+# TODO: Delete this once API introduced in NeMo (https://github.com/NVIDIA/NeMo/pull/10803)
+# TODO: Update PR to move this logic into staticmethod in nemo/collections/nlp/models/language_modeling/megatron_gpt_model.py
 def grad_reductions(ptl_model):
     # when using sequence parallelism, the sequence parallel layernorm grads must be all-reduced
     if ptl_model.cfg.get("tensor_model_parallel_size", 1) > 1 and ptl_model.cfg.get("sequence_parallel", False):
-        ptl_model.allreduce_sequence_parallel_gradients()
-
-    if ptl_model.with_distributed_adam:
-        # synchronize asynchronous grad reductions
-        # note: not necessary, but reduces performance degradation
-        # from multiple simultaneous NCCL calls
-        ptl_model._optimizer._finish_bucket_grad_sync()
+        # Mcore DistOpt handles this, so we don't have to
+        if not ptl_model.use_mcore_dist_optim:
+            ptl_model.megatron_timer_start("allreduce_sequence_parallel_gradients", log_level=1)
+            ptl_model.allreduce_sequence_parallel_gradients()
+            ptl_model.megatron_timer_stop("allreduce_sequence_parallel_gradients")
+
+    ptl_model.megatron_timer_start("gradient_allreduce", log_level=1)
+    if ptl_model.use_fsdp:
+        # Reduce the gradients omitted from FSDP-sharding
+        ptl_model.allreduce_fsdp_sharding_omitted_gradients()
+    elif ptl_model.with_distributed_adam:
+        if not ptl_model.use_mcore_dist_optim:
+            # synchronize asynchronous grad reductions
+            # note: not necessary, but reduces performance degradation
+            # from multiple simultaneous NCCL calls
+            ptl_model._optimizer._finish_bucket_grad_sync()
+        # else: Mcore distributed optim calls finalize_model_grads to finish grad sync
     elif ptl_model.megatron_amp_O2:
         # when using pipeline parallelism grads must be all-reduced after the pipeline (not asynchronously)
-        if ptl_model.cfg.get("pipeline_model_parallel_size", 1) > 1 or ptl_model.cfg.get("sequence_parallel", False):
+        if (
+            ptl_model.cfg.get("pipeline_model_parallel_size", 1) > 1
+            or ptl_model.cfg.get("sequence_parallel", False)
+            or not ptl_model.cfg.get("async_grad_allreduce", True)
+        ):
             # main grads are stored in the MainParamsOptimizer wrapper
             ptl_model._optimizer.allreduce_main_grads()
     else:
         # async grad allreduce is not currently implemented for O1/autocasting mixed precision training
         # so we all-reduce gradients after the pipeline
         ptl_model.allreduce_gradients()  # @sangkug we think this is causing memory to blow up (hurts perf)
+    ptl_model.megatron_timer_stop("gradient_allreduce")
 
-    if ptl_model.cfg.get("pipeline_model_parallel_size", 1) > 1 and ptl_model.cfg.get(
-        "share_embeddings_and_output_weights", True
+    if (
+        not ptl_model.use_mcore_dist_optim
+        and ptl_model.cfg.get("pipeline_model_parallel_size", 1) > 1
+        and ptl_model.cfg.get("share_embeddings_and_output_weights", True)
     ):
+        ptl_model.megatron_timer_start("allreduce_first_last_embeddings", log_level=1)
         # when using pipeline parallelism the first and last stage must keep embeddings in sync
         ptl_model.allreduce_first_last_embeddings()
+        ptl_model.megatron_timer_stop("allreduce_first_last_embeddings")
 
 
 def prepare_for_validation_step(ptl_model):
@@ -155,14 +176,26 @@ def set_eval(ptl_model):
         ptl_model.eval()
 
 
+# TODO: adapt the version in /opt/NeMo/nemo/collections/nlp/models/language_modeling/megatron_base_model.py
 def clip_gradients(ptl_model, clip_val):
+    """PTL hook to configure gradients.
+    We use gradient clipping implementation from megatron-lm.
+    """
     if clip_val is None:
         return
 
     clip_val = float(clip_val)
     if clip_val <= 0:
         return
 
+    if ptl_model.with_megatron_fused_adam or ptl_model.use_mcore_dist_optim:
+        # Gradient clipping is done in optimizer step
+        return
+
+    if ptl_model.grad_clip_pl_default:
+        # use the default behavior
+        return super().configure_gradient_clipping(*args, **kwargs)
+
     if ptl_model.with_distributed_adam:
         grad_norm = clip_grad_norm_distributed_optimizer(ptl_model._optimizer, clip_val)
     else:
@@ -171,6 +204,5 @@ def clip_gradients(ptl_model, clip_val):
             parameters = ptl_model._optimizer.get_parameters_with_grad()
         else:
             parameters = ptl_model.get_parameters_with_grad()
-        grad_norm = clip_grad_norm_fp32(parameters=parameters, max_norm=clip_val)
-
+        grad_norm = clip_grad_norm_fp32(parameters=parameters, max_norm=clip_val, use_fsdp=ptl_model.use_fsdp,)
     return grad_norm

tests/functional/dpo.sh

@@ -1,14 +1,26 @@
 #!/bin/bash
 
+# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
 cd $SCRIPT_DIR
 set -eoux pipefail
 
 export NCCL_ALGO=Tree
-export NVTE_APPLY_QK_LAYER_SCALING=1
+export NVTE_APPLY_QK_LAYER_SCALING=${NVTE_APPLY_QK_LAYER_SCALING:-0}
 
-KL=${KL:-0.1}
-GBS=${GBS:-4}
 PRETRAINED_CHECKPOINT_NEMO_FILE=${PRETRAINED_CHECKPOINT_NEMO_FILE}
 
 
@@ -23,7 +35,6 @@ mkdir -p $RESULTS_DIR
 
 GPFS=$(git rev-parse --show-toplevel)
 
-# START HETEROGENEUS JOB 3
 CONF_DIR="${GPFS}/examples/nlp/gpt/conf/"
 CONF_NAME="gpt_dpo"
 
@@ -33,17 +44,16 @@ dpo() {
 export CUDA_VISIBLE_DEVICES=0,1
 export PYTHONPATH="${GPFS}:${PYTHONPATH:-}"
 export HYDRA_FULL_ERROR=1
-torchrun --nproc-per-node 2 ${GPFS}/examples/nlp/gpt/train_gpt_dpo.py \
+torchrun --nproc_per_node=2 ${GPFS}/examples/nlp/gpt/train_gpt_dpo.py \
     --config-path=${CONF_DIR} \
     --config-name=${CONF_NAME} \
     trainer.num_nodes=1 \
     trainer.devices=2 \
-    ++model.data.seq_length=128 \
     ++model.global_batch_size=${GBS} \
     ++model.micro_batch_size=1 \
     ++model.mcore_gpt=true \
     ++model.megatron_amp_O2=true \
-    ++model.dpo.ref_policy_kl_penalty=${KL} \
+    ++model.dpo.ref_policy_kl_penalty=0.1 \
     ++model.dpo.log_prob_forward_micro_batch_size=1 \
     ++model.dpo.average_log_probs=false \
     ++model.dpo.sft_loss_weight=0.1 \
@@ -52,6 +62,7 @@ torchrun --nproc-per-node 2 ${GPFS}/examples/nlp/gpt/train_gpt_dpo.py \
     "model.data.data_prefix={train: [${TRAIN_DATA_PATH}], validation: [${VALID_DATA_PATH}], test: [${VALID_DATA_PATH}]}" \
     exp_manager.create_checkpoint_callback=${CREATE_CHECKPOINT_CALLBACK:-False} \
     model.data.num_workers=2 \
+    ++model.data.seq_length=128 \
     ++model.tensor_model_parallel_size=1 \
     ++model.pipeline_model_parallel_size=1 \
     trainer.dpo.max_steps=${MAX_STEPS:-3} \
@@ -67,4 +78,5 @@ torchrun --nproc-per-node 2 ${GPFS}/examples/nlp/gpt/train_gpt_dpo.py \
 }
 
 log_file=$(mktemp /tmp/dpo-log-XXXXXX)
-dpo "$@" | tee $log_file
+dpo "$@" | tee $log_file
+echo "[Finished] $0"

tests/functional/test_cases/dpo-llama3

@@ -1,5 +1,6 @@
 #!/bin/bash
-# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
+
+# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -20,4 +21,5 @@ set -eoux pipefail
 
 PRETRAINED_CHECKPOINT_NEMO_FILE=${ALIGNER_CI_DIR}/checkpoints/tiny-llama3-results-nlayers2-hidden128-ffn448-nhead4-qgroup2-megatron_gpt.nemo \
 bash ../dpo.sh \
-    ++model.data.data_impl=jsonl \
+  ++model.optim.name=mcore_distributed_optim \
+  2>&1 | tee $(basename $0).log