Tensor Parallelism

llmfoundry/__init__.py

@@ -48,6 +48,7 @@
     models,
     optim,
     tokenizers,
+    tp,
     utils,
 )
 from llmfoundry._version import __version__
@@ -87,5 +88,6 @@
     'models',
     'optim',
     'tokenizers',
+    'tp',
     'utils',
 ]

llmfoundry/command_utils/train.py

@@ -5,6 +5,7 @@
 import os
 import time
 import warnings
+from copy import deepcopy
 from typing import Any, Optional, Union
 
 import torch
@@ -18,7 +19,11 @@
     TraceHandler,
     cyclic_schedule,
 )
-from composer.utils import dist, get_device, reproducibility
+from composer.utils import (
+    dist,
+    get_device,
+    reproducibility,
+)
 from omegaconf import DictConfig
 from omegaconf import OmegaConf as om
 
@@ -43,6 +48,7 @@
     build_save_planner,
     build_scheduler,
     build_tokenizer,
+    build_tp_strategies,
 )
 from llmfoundry.utils.config_utils import (
     TRAIN_CONFIG_KEYS,
@@ -329,16 +335,27 @@ def train(cfg: DictConfig) -> Trainer:
                 changing autoresume default to True...',
         )
 
-    # Warn if fsdp is enabled but user only has 1 GPU
-    if dist.get_world_size() == 1 and fsdp_config is not None:
+    # Optional tp config
+    tp_config: Optional[dict[str, Any]] = train_cfg.tp_config
+
+    # Warn if FSDP or TP is enabled but user only has 1 GPU
+    if dist.get_world_size(
+    ) == 1 and (fsdp_config is not None or tp_config is not None):
+        parallelism = ''
+        if fsdp_config is not None:
+            parallelism += 'FSDP'
+        if tp_config is not None:
+            parallelism += '+TP' if fsdp_config is not None else 'TP'
         warnings.warn(
-            'FSDP is not applicable for single-GPU training. Reverting to DDP.',
+            f'{parallelism} is not applicable for single-GPU training. Reverting to DDP.',
         )
         fsdp_config = None
+        tp_config = None
 
     # Initialize context
-    init_context = process_init_device(model_config, fsdp_config)
+    init_context = process_init_device(model_config, fsdp_config, tp_config)
     logged_cfg.update({'fsdp_config': fsdp_config}, merge=True)
+    logged_cfg.update({'tp_config': deepcopy(tp_config)}, merge=True)
 
     # Build tokenizer
     log.info('Building tokenizer...')
@@ -502,6 +519,15 @@ def train(cfg: DictConfig) -> Trainer:
 
     _log_num_params(model, logged_cfg)
 
+    # TP config
+    if tp_config is not None:
+        strategy = tp_config.pop('strategy', None)
+        assert isinstance(strategy, str), '`strategy` must be in `tp_config`.'
+        tp_config['layer_plan'] = build_tp_strategies(strategy, model)
+
+    # Parallelism config
+    parallelism_config = {'fsdp': fsdp_config, 'tp': tp_config}
+
     # Optimizer
     optimizer_name: str = train_cfg.optimizer.pop('name')
     optimizer_cfg = train_cfg.optimizer
@@ -546,7 +572,7 @@ def train(cfg: DictConfig) -> Trainer:
         precision=train_cfg.precision,
         algorithms=algorithms,
         device_train_microbatch_size=train_cfg.device_train_microbatch_size,
-        parallelism_config={'fsdp': fsdp_config},
+        parallelism_config=parallelism_config,
         save_folder=train_cfg.save_folder,
         save_filename=save_filename,
         save_latest_filename=save_latest_filename,

llmfoundry/registry.py

@@ -7,6 +7,7 @@
 from composer.models import ComposerModel
 from composer.optim import ComposerScheduler
 from torch.distributed.checkpoint import LoadPlanner, SavePlanner
+from torch.distributed.tensor.parallel.style import ParallelStyle
 from torch.optim import Optimizer
 from torch.utils.data import DataLoader as TorchDataloader
 from torch.utils.data import Dataset
@@ -389,6 +390,26 @@
     description=_save_planners_description,
 )
 
+_tp_strategies_description = (
+    """The tp_strategies registry is used to register strategies for tensor parallelism.
+
+    Args:
+        model (ComposerModel): The model.
+
+    Returns:
+        layer_plan (Dict[str, ParallelStyle]): The plan used to parallelize the model.
+        model (ComposerModel): The model.
+    """
+)
+
+tp_strategies = create_registry(
+    'llmfoundry',
+    'tp_strategies',
+    generic_type=Callable[[ComposerModel], dict[str, ParallelStyle]],
+    entry_points=True,
+    description=_tp_strategies_description,
+)
+
 __all__ = [
     'loggers',
     'callbacks',
@@ -416,4 +437,5 @@
     'config_transforms',
     'load_planners',
     'save_planners',
+    'tp_strategies',
 ]

llmfoundry/tp/__init__.py

@@ -0,0 +1,11 @@
+# Copyright 2024 MosaicML LLM Foundry authors
+# SPDX-License-Identifier: Apache-2.0
+
+from llmfoundry.registry import tp_strategies
+from llmfoundry.tp.ffn_tp_strategy import ffn_tp_strategy
+
+tp_strategies.register('ffn', func=ffn_tp_strategy)
+
+__all__ = [
+    'ffn_tp_strategy',
+]

llmfoundry/tp/ffn_tp_strategy.py

@@ -0,0 +1,56 @@
+# Copyright 2024 MosaicML LLM Foundry authors
+# SPDX-License-Identifier: Apache-2.0
+
+from composer.models import ComposerModel
+from torch.distributed._tensor import Replicate, Shard
+from torch.distributed.tensor.parallel import (
+    ColwiseParallel,
+    PrepareModuleInput,
+    RowwiseParallel,
+)
+from torch.distributed.tensor.parallel.style import ParallelStyle
+
+
+def ffn_tp_strategy(model: ComposerModel) -> dict[str, ParallelStyle]:
+    TP_LAYERS = {'ffn', 'ffn.up_proj', 'ffn.down_proj'}
+
+    # Validate that all TP_LAYERS are in model
+    tp_layers_in_model = {
+        layer for layer in TP_LAYERS for name, _ in model.named_modules()
+        if layer in name
+    }
+    if tp_layers_in_model != TP_LAYERS:
+        raise RuntimeError(
+            f'The FFN tensor parallelism strategy requires `model` to have layers {TP_LAYERS}. But `model` is missing layers {TP_LAYERS - tp_layers_in_model}.',
+        )
+
+    # Generate layer plan
+    layer_plan: dict[str, ParallelStyle] = {}
+    for name, _ in model.named_modules():
+        # Before the ffn layer starts, distribute the input data for proper TP use
+        # Inputs are currently sharded across the batch dimension (dim 0) as is done in standard DDP
+        # Inputs will be replicated across hidden dimension (dim 1) via allgather
+        if name.split('.')[-1] == 'ffn':
+            layer_plan[name] = PrepareModuleInput(
+                input_layouts=Shard(0),
+                desired_input_layouts=Replicate(),
+                use_local_output=True,
+            )
+        # Shard the ffn.up_proj weight matrix across its columns
+        # Inputs are already replicated across each TP group
+        # Outputs will be sharded along the hidden dimension (dim 1) via allgather
+        elif name.split('.')[-2:] == ['ffn', 'up_proj']:
+            layer_plan[name] = ColwiseParallel(
+                input_layouts=Replicate(),
+                output_layouts=Shard(-1),
+            )
+        # Shard the ffn.down_proj weight matrix across its rows
+        # Inputs are sharded along the hidden dimension (dim 1)
+        # Outputs will be sharded along batch dimension (dim 0) via allreduce
+        elif name.split('.')[-2:] == ['ffn', 'down_proj']:
+            layer_plan[name] = RowwiseParallel(
+                input_layouts=Shard(-1),
+                output_layouts=Shard(0),
+            )
+
+    return layer_plan

llmfoundry/utils/builders.py

@@ -25,6 +25,7 @@
 from omegaconf import DictConfig
 from omegaconf import OmegaConf as om
 from torch.distributed.checkpoint import LoadPlanner, SavePlanner
+from torch.distributed.tensor.parallel.style import ParallelStyle
 from torch.optim.optimizer import Optimizer
 from torchmetrics import Metric
 from transformers import AutoTokenizer, PreTrainedTokenizerBase
@@ -37,6 +38,7 @@
 )
 from llmfoundry.utils.config_utils import to_dict_container, to_list_container
 from llmfoundry.utils.registry_utils import construct_from_registry
+from llmfoundry.utils.warnings import experimental_function
 
 log = logging.getLogger(__name__)
 
@@ -52,6 +54,7 @@
     'build_tokenizer',
     'build_composer_model',
     'build_metric',
+    'build_tp_strategies',
 ]
 
 
@@ -701,3 +704,16 @@ def _validate_cfg(icl_cfg: dict[str, Any]):
                 )
 
     return evaluators, logger_keys
+
+
+@experimental_function('tp_strategies')
+def build_tp_strategies(
+    name: str,
+    model: ComposerModel,
+) -> dict[str, ParallelStyle]:
+    return construct_from_registry(
+        name=name,
+        registry=registry.tp_strategies,
+        partial_function=False,
+        kwargs={'model': model},
+    )

llmfoundry/utils/config_utils.py

@@ -120,6 +120,7 @@ class TrainConfig:
     # Distributed training parameters
     dist_timeout: Union[int, float] = 600.0
     fsdp_config: Optional[dict[str, Any]] = None
+    tp_config: Optional[dict[str, Any]] = None
 
     # Evaluation parameters
     eval_interval: Union[int, str] = 1
@@ -501,7 +502,11 @@ def update_batch_size_info(cfg: dict[str, Any]) -> dict[str, Any]:
     return cfg
 
 
-def process_init_device(model_cfg: dict[str, Any], fsdp_config: Optional[dict]):
+def process_init_device(
+    model_cfg: dict[str, Any],
+    fsdp_config: Optional[dict] = None,
+    tp_config: Optional[dict] = None,
+):
     # Restrict model init_device to 'meta' and 'cpu',
     # using 'cuda' vs. 'cuda:id' is tricky and can lead to common user errors
     # when multiple GPUs are available.
@@ -533,6 +538,13 @@ def process_init_device(model_cfg: dict[str, Any], fsdp_config: Optional[dict]):
             # Set defaults for mixed initialization
             fsdp_config.setdefault('load_monolith_rank0_only', True)
 
+    # Check we are not using tensor parallelism with MoEs
+    if tp_config is not None and 'ffn_config' in model_cfg and model_cfg[
+        'ffn_config'].get('ffn_type', None) in ffns_with_megablocks:
+        raise ValueError(
+            'Tensor Parallelism is not currently supported for MoE models.',
+        )
+
     # Set ffn_config.device_mesh using fsdp_config
     if fsdp_config is not None and 'ffn_config' in model_cfg and model_cfg[
         'ffn_config'].get('ffn_type', None) in ffns_with_megablocks: