fix: adds missing support for mcore dist opt and adds test for moe

Signed-off-by: Terry Kong <terryk@nvidia.com>

moe test is all2all

Signed-off-by: Terry Kong <terryk@nvidia.com>

other params

Signed-off-by: Terry Kong <terryk@nvidia.com>

fix peft mixtral

Signed-off-by: Terry Kong <terryk@nvidia.com>

dockerfile bump to be on dev

Signed-off-by: Terry Kong <terryk@nvidia.com>

just take dockerfile on dev

Signed-off-by: Terry Kong <terryk@nvidia.com>

.github/workflows/cicd-main.yml

@@ -95,6 +95,8 @@ jobs:
           - kd-llama3
           - sft-llama3
           - rm-llama3
+          - 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

Dockerfile

@@ -13,8 +13,8 @@ ARG MAX_JOBS=8
 # Git refs for dependencies
 ARG TE_TAG=7d576ed25266a17a7b651f2c12e8498f67e0baea
 ARG PYTRITON_VERSION=0.5.10
-ARG NEMO_TAG=19668e5320a2e2af0199b6d5e0b841993be3a634  # On: main
-ARG MLM_TAG=25059d3bbf68be0751800f3644731df12a88f3f3   # On: main
+ARG NEMO_TAG=06eae2895c0fea09f8dd7c34feff0163e55c419a  # On: main
+ARG MLM_TAG=844119f5c856a3037ec7c7f6d6ef7b3518ceee6b   # On: main
 ARG ALIGNER_COMMIT=main
 ARG TRTLLM_VERSION=v0.13.0
 ARG PROTOBUF_VERSION=4.24.4
@@ -123,19 +123,19 @@ RUN cd /opt/NeMo-Aligner && \
 
 RUN cd TensorRT-LLM && patch -p1 < ../NeMo-Aligner/setup/trtllm.patch
 
-# TODO(terryk): This layer should be deleted ASAP after NeMo is bumped to include all of these PRs
+# NOTE: Comment this layer out if it is not needed
+# NOTE: This section exists to allow cherry-picking PRs in cases where
+#  we do not wish to simply update to the top-of-tree. Sometimes PRs
+#  cannot be cherry-picked cleanly if rebased a few times to top-of-tree
+#  so this logic also requires you to select a SHA (can be dangling) from
+#  the PR.
 RUN <<"EOF" bash -exu
 cd NeMo
 # Ensures we don't cherry-pick "future" origin/main commits
 git fetch -a
-# 0c92fe17df4642ffc33d5d8c0c83fda729e3910c: [fix] Ensures disabling exp_manager with exp_manager=null does not error NeMo#10651
-# 60e677423667c029dd05875da72bf0719774f844: [feat] Update get_model_parallel_src_rank to support tp-pp-dp ordering NeMo#10652
-# 0deaf6716cb4f20766c995ce25d129795f1ae200: fix[export]: update API for disabling device reassignment in TRTLLM for Aligner NeMo#10863
-# (superceded by 10863) 148543d6e9c66ff1f8562e84484448202249811d: feat: Migrate GPTSession refit path in Nemo export to ModelRunner for Aligner NeMo#10654
+# d27dd28b4186f6ecd9f46f1c5679a5eef9bad14e: fix: export weight name mapping if model is nemo model#11497
 for pr_and_commit in \
-  "10651 0c92fe17df4642ffc33d5d8c0c83fda729e3910c" \
-  "10652 60e677423667c029dd05875da72bf0719774f844" \
-  "10863 0deaf6716cb4f20766c995ce25d129795f1ae200" \
+  "11497 d27dd28b4186f6ecd9f46f1c5679a5eef9bad14e" \
 ; do
   pr=$(cut -f1 -d' ' <<<"$pr_and_commit")
   head_pr_commit=$(cut -f2 -d' ' <<<"$pr_and_commit")

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
@@ -177,4 +179,4 @@ 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