[NeMo-UX] Add token drop callback and optimize mixtral configs (NVIDI…

…A#10361) * add token drop plugin Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com> * add checks Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com> * add expert parallel configs Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com> * Apply isort and black reformatting Signed-off-by: JimmyZhang12 <JimmyZhang12@users.noreply.github.com> * amend comment Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com> * Apply isort and black reformatting Signed-off-by: JimmyZhang12 <JimmyZhang12@users.noreply.github.com> * add comm overlap Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com> * fix rebase errors Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com> * Apply isort and black reformatting Signed-off-by: JimmyZhang12 <JimmyZhang12@users.noreply.github.com> * fix typo Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com> * add test configs Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com> * fix Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com> --------- Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com> Signed-off-by: JimmyZhang12 <JimmyZhang12@users.noreply.github.com> Co-authored-by: Jimmy Zhang <jiemingz@nvidia.com> Co-authored-by: JimmyZhang12 <JimmyZhang12@users.noreply.github.com> Co-authored-by: Pablo Garay <palenq@gmail.com> Co-authored-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com> Signed-off-by: George Armstrong <georgea@nvidia.com>
gwarmstrong · Sep 19, 2024 · 1035c70 · 1035c70
1 parent 34041b5
commit 1035c70
Show file tree

Hide file tree

Showing 8 changed files with 233 additions and 43 deletions.
diff --git a/nemo/collections/llm/gpt/model/mixtral.py b/nemo/collections/llm/gpt/model/mixtral.py
@@ -57,11 +57,10 @@ class MixtralConfig(GPTConfig):
     # MoE
     num_moe_experts: int = 8
     moe_aux_loss_coeff: float = 0.01
-    moe_expert_capacity_factor: float = 1.0
-    moe_pad_expert_input_to_capacity: bool = True
     moe_router_topk: int = 2
     moe_router_pre_softmax: bool = True
     moe_token_dispatcher_type: str = "alltoall"
+    moe_router_load_balancing_type: str = 'aux_loss'
 
     init_method_std: float = 0.02
     layernorm_epsilon: float = 1e-5

diff --git a/nemo/collections/llm/recipes/mixtral_8x22b.py b/nemo/collections/llm/recipes/mixtral_8x22b.py
@@ -29,6 +29,8 @@
 from nemo.collections.llm.peft.lora import LoRA
 from nemo.collections.llm.recipes.log.default import default_log, default_resume, tensorboard_logger
 from nemo.collections.llm.recipes.optim.adam import distributed_fused_adam_with_cosine_annealing
+from nemo.lightning.pytorch.callbacks.megatron_comm_overlap import MegatronCommOverlapCallback
+from nemo.lightning.pytorch.callbacks.moe_token_drop import MegatronTokenDropCallback
 from nemo.utils.exp_manager import TimingCallback
 
 NAME = "mixtral_8x22b"
@@ -54,14 +56,14 @@ def model() -> run.Config[pl.LightningModule]:
 
 
 def trainer(
-    tensor_parallelism: int = 8,
-    pipeline_parallelism: int = 8,
+    tensor_parallelism: int = 2,
+    pipeline_parallelism: int = 4,
     pipeline_parallelism_type: Optional[torch.dtype] = torch.bfloat16,
-    virtual_pipeline_parallelism: Optional[int] = 7,
-    context_parallelism: int = 1,
+    virtual_pipeline_parallelism: Optional[int] = 14,
+    context_parallelism: int = 2,
     sequence_parallelism: bool = True,
-    expert_parallelism: int = 1,
-    num_nodes: int = 8,
+    expert_parallelism: int = 8,
+    num_nodes: int = 16,
     num_gpus_per_node: int = 8,
     max_steps: int = 1168251,
     callbacks: Optional[list[run.Config[Callback]]] = None,
@@ -92,7 +94,7 @@ def trainer(
             $ nemo llm pretrain trainer=mixtral_8x22b ...
 
         Python API usage:
-            >>> trainer_config = trainer(num_nodes=8, num_gpus_per_node=8)
+            >>> trainer_config = trainer(num_nodes=16, num_gpus_per_node=8)
             >>> print(trainer_config)
 
     Note:
@@ -139,7 +141,7 @@ def trainer(
 
 @run.cli.factory(target=pretrain, name=NAME)
 def pretrain_recipe(
-    dir: Optional[str] = None, name: str = "default", num_nodes: int = 8, num_gpus_per_node: int = 8, fn=pretrain
+    dir: Optional[str] = None, name: str = "default", num_nodes: int = 16, num_gpus_per_node: int = 8, fn=pretrain
 ) -> run.Partial:
     """
     Create a pre-training recipe for Mixtral 8x22B model.
@@ -160,10 +162,10 @@ def pretrain_recipe(
     Examples:
         CLI usage:
             $ nemo llm pretrain --factory mixtral_8x22b
-            $ nemo llm pretrain --factory "mixtral_8x22b(num_nodes=2, name='my_mixtral_pretrain')"
+            $ nemo llm pretrain --factory "mixtral_8x22b(num_nodes=16, name='my_mixtral_pretrain')"
 
         Python API usage:
-            >>> recipe = pretrain_recipe(name="mixtral_pretrain", num_nodes=2)
+            >>> recipe = pretrain_recipe(name="mixtral_pretrain", num_nodes=16)
             >>> print(recipe)
     """
     return run.Partial(
@@ -179,6 +181,49 @@ def pretrain_recipe(
     )
 
 
+@run.cli.factory(target=pretrain, name=NAME + "_performance")
+def pretrain_recipe_performance(
+    dir: Optional[str] = None, name: str = "default", num_nodes: int = 8, num_gpus_per_node: int = 8, fn=pretrain
+) -> run.Partial:
+    """
+    Create a performance-optimized pre-training recipe for Mixtral 8x22B model.
+
+    This recipe enables performance optimizations that may not be suitable for all use cases.
+    It builds upon the standard pre-training recipe and adds additional performance enhancements.
+
+    Args:
+        dir (Optional[str]): Directory for saving logs and checkpoints.
+        name (str): Name of the pre-training run.
+        num_nodes (int): Number of compute nodes to use.
+        num_gpus_per_node (int): Number of GPUs per node.
+        fn (Callable): The pre-training function to use.
+
+    Returns:
+        run.Partial: Partial configuration for performance-optimized pre-training.
+
+    Examples:
+        CLI usage:
+            $ nemo llm pretrain --factory "mixtral_8x22b.pretrain_recipe_performance(num_nodes=8, name='perf_pretrain')"
+
+        Python API usage:
+            >>> recipe = pretrain_recipe_performance(name="mixtral_8x22b_perf", num_nodes=8)
+            >>> print(recipe)
+
+    Note:
+        Use this recipe with caution and only when you need maximum performance.
+        It may not be suitable for all hardware configurations or use cases.
+    """
+    recipe = pretrain_recipe(name=name, dir=dir, num_nodes=num_nodes, num_gpus_per_node=num_gpus_per_node, fn=fn)
+    recipe.trainer.callbacks.extend(
+        [
+            run.Config(MegatronTokenDropCallback),
+            run.Config(MegatronCommOverlapCallback),
+        ]
+    )
+
+    return recipe
+
+
 def hf_resume() -> run.Config[nl.AutoResume]:
     """
     Configure automatic resumption from a Hugging Face checkpoint for Mixtral 8x22B model.

diff --git a/nemo/collections/llm/recipes/mixtral_8x3b.py b/nemo/collections/llm/recipes/mixtral_8x3b.py
@@ -30,6 +30,8 @@
 from nemo.collections.llm.recipes.log.default import default_log, default_resume, tensorboard_logger
 from nemo.collections.llm.recipes.optim.adam import distributed_fused_adam_with_cosine_annealing
 from nemo.collections.llm.recipes.precision.mixed_precision import bf16_mixed
+from nemo.lightning.pytorch.callbacks.megatron_comm_overlap import MegatronCommOverlapCallback
+from nemo.lightning.pytorch.callbacks.moe_token_drop import MegatronTokenDropCallback
 from nemo.utils.exp_manager import TimingCallback
 
 NAME = "mixtral_8x3b"
@@ -55,14 +57,14 @@ def model() -> run.Config[pl.LightningModule]:
 
 
 def trainer(
-    tensor_parallelism: int = 4,
+    tensor_parallelism: int = 1,
     pipeline_parallelism: int = 1,
     pipeline_parallelism_type: Optional[torch.dtype] = None,
     virtual_pipeline_parallelism: Optional[int] = None,
     context_parallelism: int = 1,
-    sequence_parallelism: bool = True,
-    expert_parallelism: int = 1,
-    num_nodes: int = 1,
+    sequence_parallelism: bool = False,
+    expert_parallelism: int = 4,
+    num_nodes: int = 2,
     num_gpus_per_node: int = 8,
     max_steps: int = 1168251,
     callbacks: Optional[list[run.Config[Callback]]] = None,
@@ -93,7 +95,7 @@ def trainer(
             $ nemo llm pretrain trainer=mixtral_8x3b ...
 
         Python API usage:
-            >>> trainer_config = trainer(num_nodes=1, num_gpus_per_node=8)
+            >>> trainer_config = trainer(num_nodes=2, num_gpus_per_node=8)
             >>> print(trainer_config)
     """
     strategy = run.Config(
@@ -139,7 +141,7 @@ def trainer(
 
 @run.cli.factory(target=pretrain, name=NAME)
 def pretrain_recipe(
-    dir: Optional[str] = None, name: str = "default", num_nodes: int = 1, num_gpus_per_node: int = 8, fn=pretrain
+    dir: Optional[str] = None, name: str = "default", num_nodes: int = 2, num_gpus_per_node: int = 8, fn=pretrain
 ) -> run.Partial:
     """
     Create a pre-training recipe for Mixtral 8x3B model.
@@ -181,6 +183,50 @@ def pretrain_recipe(
     )
 
 
+@run.cli.factory(target=pretrain, name=NAME + "_performance")
+def pretrain_recipe_performance(
+    dir: Optional[str] = None, name: str = "default", num_nodes: int = 2, num_gpus_per_node: int = 8, fn=pretrain
+) -> run.Partial:
+    """
+    Create a performance-optimized pre-training recipe for Mixtral 8x3B model.
+
+    This recipe enables performance optimizations that may not be suitable for all use cases.
+    It builds upon the standard pre-training recipe and adds additional performance enhancements.
+
+    Args:
+        dir (Optional[str]): Directory for saving logs and checkpoints.
+        name (str): Name of the pre-training run.
+        num_nodes (int): Number of compute nodes to use.
+        num_gpus_per_node (int): Number of GPUs per node.
+        fn (Callable): The pre-training function to use.
+
+    Returns:
+        run.Partial: Partial configuration for performance-optimized pre-training.
+
+    Examples:
+        CLI usage:
+            $ nemo llm pretrain --factory "mixtral_8x3b.pretrain_recipe_performance(num_nodes=2, name='perf_pretrain')"
+
+        Python API usage:
+            >>> recipe = pretrain_recipe_performance(name="mixtral_8x3b", num_nodes=4)
+            >>> print(recipe)
+
+    Note:
+        Use this recipe with caution and only when you need maximum performance.
+        It may not be suitable for all hardware configurations or use cases.
+    """
+    recipe = pretrain_recipe(name=name, dir=dir, num_nodes=num_nodes, num_gpus_per_node=num_gpus_per_node, fn=fn)
+
+    recipe.trainer.callbacks.extend(
+        [
+            run.Config(MegatronTokenDropCallback),
+            run.Config(MegatronCommOverlapCallback),
+        ]
+    )
+
+    return recipe
+
+
 def hf_resume() -> run.Config[nl.AutoResume]:
     """
     Configure the Hugging Face model resuming for Mixtral 8x3B model.

diff --git a/nemo/collections/llm/recipes/mixtral_8x7b.py b/nemo/collections/llm/recipes/mixtral_8x7b.py
@@ -29,6 +29,8 @@
 from nemo.collections.llm.peft.lora import LoRA
 from nemo.collections.llm.recipes.log.default import default_log, default_resume, tensorboard_logger
 from nemo.collections.llm.recipes.optim.adam import distributed_fused_adam_with_cosine_annealing
+from nemo.lightning.pytorch.callbacks.megatron_comm_overlap import MegatronCommOverlapCallback
+from nemo.lightning.pytorch.callbacks.moe_token_drop import MegatronTokenDropCallback
 from nemo.utils.exp_manager import TimingCallback
 
 NAME = "mixtral_8x7b"
@@ -54,14 +56,14 @@ def model() -> run.Config[pl.LightningModule]:
 
 
 def trainer(
-    tensor_parallelism: int = 8,
-    pipeline_parallelism: int = 2,
+    tensor_parallelism: int = 1,
+    pipeline_parallelism: int = 4,
     pipeline_parallelism_type: Optional[torch.dtype] = torch.bfloat16,
-    virtual_pipeline_parallelism: Optional[int] = None,
+    virtual_pipeline_parallelism: Optional[int] = 8,
     context_parallelism: int = 1,
-    sequence_parallelism: bool = True,
-    expert_parallelism: int = 1,
-    num_nodes: int = 2,
+    sequence_parallelism: bool = False,
+    expert_parallelism: int = 8,
+    num_nodes: int = 8,
     num_gpus_per_node: int = 8,
     max_steps: int = 1168251,
     callbacks: Optional[list[run.Config[Callback]]] = None,
@@ -138,7 +140,7 @@ def trainer(
 
 @run.cli.factory(target=pretrain, name=NAME)
 def pretrain_recipe(
-    dir: Optional[str] = None, name: str = "default", num_nodes: int = 2, num_gpus_per_node: int = 8, fn=pretrain
+    dir: Optional[str] = None, name: str = "default", num_nodes: int = 8, num_gpus_per_node: int = 8, fn=pretrain
 ) -> run.Partial:
     """
     Create a pre-training recipe for Mixtral 8x7B model.
@@ -159,10 +161,10 @@ def pretrain_recipe(
     Examples:
         CLI usage:
             $ nemo llm pretrain --factory mixtral_8x7b
-            $ nemo llm pretrain --factory "mixtral_8x7b(num_nodes=2, name='my_mixtral_pretrain')"
+            $ nemo llm pretrain --factory "mixtral_8x7b(num_nodes=8, name='my_mixtral_pretrain')"
 
         Python API usage:
-            >>> recipe = pretrain_recipe(name="mixtral_8x7b_pretrain", num_nodes=2)
+            >>> recipe = pretrain_recipe(name="mixtral_8x7b_pretrain", num_nodes=8)
             >>> print(recipe)
     """
     return run.Partial(
@@ -178,6 +180,49 @@ def pretrain_recipe(
     )
 
 
+@run.cli.factory(target=pretrain, name=NAME + "_performance")
+def pretrain_recipe_performance(
+    dir: Optional[str] = None, name: str = "default", num_nodes: int = 8, num_gpus_per_node: int = 8, fn=pretrain
+) -> run.Partial:
+    """
+    Create a performance-optimized pre-training recipe for Mixtral 8x7B model.
+
+    This recipe enables performance optimizations that may not be suitable for all use cases.
+    It builds upon the standard pre-training recipe and adds additional performance enhancements.
+
+    Args:
+        dir (Optional[str]): Directory for saving logs and checkpoints.
+        name (str): Name of the pre-training run.
+        num_nodes (int): Number of compute nodes to use.
+        num_gpus_per_node (int): Number of GPUs per node.
+        fn (Callable): The pre-training function to use.
+
+    Returns:
+        run.Partial: Partial configuration for performance-optimized pre-training.
+
+    Examples:
+        CLI usage:
+            $ nemo llm pretrain --factory "mixtral_8x3b.pretrain_recipe_performance(num_nodes=8, name='perf_pretrain')"
+
+        Python API usage:
+            >>> recipe = pretrain_recipe_performance(name="mixtral_8x3b_perf", num_nodes=8)
+            >>> print(recipe)
+
+    Note:
+        Use this recipe with caution and only when you need maximum performance.
+        It may not be suitable for all hardware configurations or use cases.
+    """
+    recipe = pretrain_recipe(name=name, dir=dir, num_nodes=num_nodes, num_gpus_per_node=num_gpus_per_node, fn=fn)
+    recipe.trainer.callbacks.extend(
+        [
+            run.Config(MegatronTokenDropCallback),
+            run.Config(MegatronCommOverlapCallback),
+        ]
+    )
+
+    return recipe
+
+
 def hf_resume() -> run.Config[nl.AutoResume]:
     """
     Configure automatic resumption from a Hugging Face checkpoint for Mixtral 8x7B model.

diff --git a/nemo/lightning/pytorch/callbacks/moe_token_drop.py b/nemo/lightning/pytorch/callbacks/moe_token_drop.py
@@ -0,0 +1,55 @@
+import pytorch_lightning as pl
+from megatron.core import ModelParallelConfig
+from pytorch_lightning.callbacks.callback import Callback
+
+from nemo.lightning.pytorch.strategies.megatron_strategy import MegatronStrategy
+
+
+class MegatronTokenDropCallback(Callback):
+    """
+    A PyTorch Lightning callback to enable token drop for MOEs. Token drop improves performance by better
+    balancing work across experts, but may affect convergence.
+
+    Args:
+        moe_expert_capacity_factor (float): The capacity factor for all experts
+        moe_pad_expert_input_to_capacity (bool): Pad the input for each expert to the expert capacity lengt
+
+    Example:
+        >>> callback = MegatronCommOverlapCallback()
+        >>> trainer = Trainer(callbacks=[callback])
+    """
+
+    def __init__(
+        self,
+        moe_expert_capacity_factor: float = 1.0,
+        moe_pad_expert_input_to_capacity: bool = True,
+    ):
+
+        if moe_expert_capacity_factor < 0:
+            moe_expert_capacity_factor = None
+        self.moe_expert_capacity_factor = moe_expert_capacity_factor
+        self.moe_pad_expert_input_to_capacity = moe_pad_expert_input_to_capacity
+
+    def _set_cfgs(self, cfg):
+        cfg.moe_expert_capacity_factor = self.moe_expert_capacity_factor
+        cfg.moe_pad_expert_input_to_capacity = self.moe_pad_expert_input_to_capacity
+
+    def setup(self, trainer: pl.Trainer, pl_module: pl.LightningModule, stage: str) -> None:
+        assert isinstance(trainer.strategy, MegatronStrategy), "MegatronTokenDrop requires MegatronStrategy"
+        if hasattr(trainer.model, "config") and isinstance(trainer.model.config, ModelParallelConfig):
+            assert trainer.model.config.moe_token_dispatcher_type in [
+                "alltoall",
+                "alltoall_seq",
+            ], 'moe_expert_capacity_factor only works with alltoall token dispatcher'
+            assert trainer.model.config.moe_router_load_balancing_type in [
+                "aux_loss",
+                "none",
+            ], 'moe_expert_capacity_factor only works with aux_loss or none load balancing'
+
+            if self.moe_pad_expert_input_to_capacity:
+                if self.moe_expert_capacity_factor is None:
+                    raise ValueError('moe_expert_capacity_factor must be set to use moe_pad_expert_input_to_capacity')
+
+            self._set_cfgs(trainer.model.config)
+            if hasattr(trainer.model, '__io__'):
+                self._set_cfgs(trainer.model.__io__.config)