pytorch · Feb 21, 2025
diff --git a/‎run_llama_train.sh
+4 b/‎run_llama_train.sh
+4
diff --git a/‎torchtitan/checkpoint.py
+99-27 b/‎torchtitan/checkpoint.py
+99-27
diff --git a/‎torchtitan/config_manager.py
+24 b/‎torchtitan/config_manager.py
+24
diff --git a/‎torchtitan/ft.py
+57 b/‎torchtitan/ft.py
+57
diff --git a/‎torchtitan/optimizer.py
+51-7 b/‎torchtitan/optimizer.py
+51-7
@@ -19,7 +19,11 @@ if [ $# -ne 0 ]; then
     overrides="$*"
 fi
 
+TORCHFT_MANAGER_PORT=${TORCHFT_MANAGER_PORT:-"29512"}
+
 PYTORCH_CUDA_ALLOC_CONF="expandable_segments:True" \
+TORCHFT_LIGHTHOUSE=http://localhost:29510 \
+TORCHFT_MANAGER_PORT=${TORCHFT_MANAGER_PORT} \
 torchrun --nproc_per_node=${NGPU} --rdzv_backend c10d --rdzv_endpoint="localhost:0" \
 --local-ranks-filter ${LOG_RANK} --role rank --tee 3 \
 train.py --job.config_file ${CONFIG_FILE} $overrides
@@ -33,6 +33,7 @@
 
 from torchtitan.config_manager import JobConfig, TORCH_DTYPE_MAP
 
+from torchtitan.ft import FTManager
 from torchtitan.logging import init_logger, logger
 from torchtitan.optimizer import LRSchedulersContainer, OptimizersContainer
 from torchtitan.utils import GarbageCollection
@@ -228,6 +229,7 @@ class CheckpointManager:
         states (Dict[str, Any]): The states that need to be saved, other than the
             previous 4 components.
         job_config (JobConfig): The job config used to configure the checkpointing.
+        ft_manager (Optional[FTManager]): The FTManager from TorchFT.
     """
 
     def __init__(
@@ -238,16 +240,40 @@ def __init__(
         lr_schedulers: LRSchedulersContainer,
         states: Dict[str, Any],
         job_config: JobConfig,
+        ft_manager: Optional[FTManager] = None,
     ) -> None:
         ckpt_config = job_config.checkpoint
         self.enable_checkpoint = ckpt_config.enable_checkpoint
+        self.ft_manager = ft_manager
+
+        if self.ft_manager:
+            optimizers.init_cache_state_dict()
+
+            def state_dict():
+                ret = {}
+                for k, v in self.states.items():
+                    if k in {
+                        CheckpointState.MODEL,
+                        CheckpointState.OPTIMIZER,
+                        CheckpointState.LR_SCHEDULER,
+                        CheckpointState.TRAIN_STATE,
+                    }:
+                        ret[k] = v.state_dict()
+                return ret
+
+            def load_state_dict(state_dict):
+                assert state_dict is not None
+                for k, v in state_dict.items():
+                    self.states[k].load_state_dict(v)
+
+            ft_manager.manager.set_state_dict_fns(load_state_dict, state_dict)
 
         async_mode = ckpt_config.async_mode.lower()
         self.enable_staging = (
             self.enable_checkpoint and async_mode == AsyncMode.ASYNC_WITH_PINNED_MEM
-        )
+        ) or self.ft_manager
 
-        if not self.enable_checkpoint:
+        if not self.enable_checkpoint and self.ft_manager is None:
             return
 
         self.states = states
@@ -259,6 +285,13 @@ def __init__(
                 LR_SCHEDULER: lr_schedulers,
             }
         )
+        self.ft_states = {CheckpointState.DATALOADER: dataloader}
+
+        self.staging = False
+        self.sending_to_checkpoint_mp = False
+        self.staging_id = None
+        self.cpu_offload_state_dict = None
+        self.staging_stream = torch.cuda.Stream() if self.enable_staging else None
 
         self.staging = False
         self.sending_to_checkpoint_mp = False
@@ -269,7 +302,7 @@ def __init__(
         self.folder = os.path.join(job_config.job.dump_folder, ckpt_config.folder)
         self.interval = ckpt_config.interval
         async_mode = ckpt_config.async_mode.lower()
-        if async_mode == AsyncMode.ASYNC:
+        if async_mode == AsyncMode.ASYNC or self.ft_manager:
             self.pg = dist.new_group(backend="gloo")
 
         self.keep_latest_k = ckpt_config.keep_latest_k
@@ -343,35 +376,42 @@ def save(self, curr_step: int, force: bool = False) -> None:
             None
         """
 
+        if self.ft_manager:
+            self._ft_save(curr_step)
+
         if not self._should_save(curr_step, force):
             return
 
         begin = time.monotonic()
-        logger.info("Saving the checkpoint (or staging if async is enabled).")
-        checkpoint_id = self._create_checkpoint_id(curr_step)
-        self._async_wait()
-        # This GC is called for async checkpoint as it is useless to do
-        # GC right after async_save -- the CPU memory is not able to be
-        # freed until _async_wait()
-        if force:
-            self._save_last_step(curr_step)
-        elif self.async_mode == AsyncMode.ASYNC_WITH_PINNED_MEM:
-            GarbageCollection.collect("GC collection invoked by checkpointer.")
-            self._async_with_pinned_memory(checkpoint_id)
-        elif self.async_mode == AsyncMode.ASYNC:
-            GarbageCollection.collect("GC collection invoked by checkpointer.")
-            self.async_future = dcp.async_save(
-                self.states, checkpoint_id=checkpoint_id, process_group=self.pg
-            )
-            GarbageCollection.collect("GC collection invoked by checkpointer.")
-        else:
-            save_with_gc(self.states, checkpoint_id=checkpoint_id)
-        self._purge_stale_checkpoints()
+        if not self.ft_manager or self.ft_manager.manager.participating_rank() == 0:
+            logger.info("Saving the checkpoint (or staging if async is enabled).")
+            checkpoint_id = self._create_checkpoint_id(curr_step)
+            self._async_wait()
+            # This GC is called for async checkpoint as it is useless to do
+            # GC right after async_save -- the CPU memory is not able to be
+            # freed until _async_wait()
+            if force:
+                self._save_last_step(curr_step)
+            elif self.async_mode == AsyncMode.ASYNC_WITH_PINNED_MEM:
+                GarbageCollection.collect("GC collection invoked by checkpointer.")
+                self._async_with_pinned_memory(checkpoint_id)
+            elif self.async_mode == AsyncMode.ASYNC:
+                GarbageCollection.collect("GC collection invoked by checkpointer.")
+                self.async_future = dcp.async_save(
+                    self.states, checkpoint_id=checkpoint_id, process_group=self.pg
+                )
+                GarbageCollection.collect("GC collection invoked by checkpointer.")
+            else:
+                save_with_gc(self.states, checkpoint_id=checkpoint_id)
+            self._purge_stale_checkpoints()
 
-        logger.info(
-            "Finished saving the checkpoint (or staging if async is enabled)"
-            f"in {time.monotonic() - begin:.2f} seconds."
-        )
+            logger.info(
+                "Finished saving the checkpoint (or staging if async is enabled)"
+                f"in {time.monotonic() - begin:.2f} seconds."
+            )
+        elif self.ft_manager:
+            logger.info("Waiting for replica 0 to save checkpoint.")
+            time.sleep(1)
 
     @torch.no_grad()
     def load(self, step: int = -1) -> bool:
@@ -388,6 +428,9 @@ def load(self, step: int = -1) -> bool:
             bool: Whether the checkpoint was loaded successfully.
         """
 
+        if self.ft_manager:
+            self._ft_load()
+
         if not self.enable_checkpoint or not os.path.isdir(self.folder):
             return False
 
@@ -471,10 +514,36 @@ def _find_load_step(self, folder: str = "") -> int:
             return -1
         return max(step_counts)
 
+    def _ft_folder(self) -> str:
+        return os.path.join(self.folder, f"ft-replicat-{self.ft_manager.replica_id}")
+
     def _create_checkpoint_id(self, step: int, folder: str = "") -> str:
         folder = folder if folder else self.folder
         return os.path.join(folder, f"step-{step}")
 
+    def _ft_save(self, step: int) -> None:
+        begin = time.monotonic()
+        self._async_wait()
+        checkpoint_id = self._create_checkpoint_id(step, folder=self._ft_folder())
+        self.async_future = dcp.async_save(
+            self.ft_states, checkpoint_id=checkpoint_id, process_group=self.pg
+        )
+        logger.info(f"Staging ft checkpoint took {time.monotonic() - begin} secs.")
+
+    def _ft_load(self) -> None:
+        step = self._find_load_step(folder=self._ft_folder())
+        if step == -1:
+            return
+
+        begin = time.monotonic()
+        logger.info(f"Loading the FT checkpoint at step {step}.")
+        checkpoint_id = self._create_checkpoint_id(step, folder=self._ft_folder())
+        dcp.load(self.ft_states, checkpoint_id=checkpoint_id)
+        GarbageCollection.collect("GC collection for checkpoint loading.")
+        logger.info(
+            f"Finished loading the ft checkpoint in {time.monotonic() - begin:.2f} seconds."
+        )
+
     def _states_to_load(self, step: int) -> Dict[str, Any]:
         """Determines which states to load for the given step.
 
@@ -495,6 +564,8 @@ def _states_to_load(self, step: int) -> Dict[str, Any]:
         for exclude_key in self.exclude_from_loading:
             if exclude_key not in states:
                 raise ValueError(f"{exclude_key} not found in state_dict.")
+        if self.ft_manager:
+            states_to_load.pop(CheckpointState.DATALOADER)
         return states_to_load
 
     def _save_last_step(self, curr_step: int) -> None:
@@ -579,6 +650,7 @@ def _cpu_staging(self, checkpoint_id: Optional[str]) -> None:
     def _purge_stale_checkpoints(self):
         if (
             self.keep_latest_k > 0
+            and self.ft_manager.manager.participating_rank() == 0
             and dist.get_rank() == 0
             and os.path.isdir(self.folder)
         ):
 
@@ -658,6 +658,30 @@ def __init__(self):
             action="store_true",
         )
 
+        self.parser.add_argument(
+            "--experimental.enable_torchft",
+            action="store_true",
+            help="Enable TorchFT integration.",
+        )
+
+        self.parser.add_argument(
+            "--experimental.ft_replica_id",
+            type=int,
+            default=0,
+            help="The TorchFT replica ID of this run.",
+        )
+
+        self.parser.add_argument(
+            "--experimental.ft_group_size",
+            type=int,
+            default=1,
+            help="""
+                The number of TorchFT replicate groups. This number will be used for
+                dataloader to split the dataset across the replicate groups and FSDP
+                dimension.
+            """,
+        )
+
     def to_dict(self):
         return self.args_dict
 
 
@@ -0,0 +1,57 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import importlib
+from dataclasses import dataclass
+from typing import Optional
+
+from torchtitan.config_manager import JobConfig
+
+if importlib.util.find_spec("torchft") is not None:
+    import torchft as ft
+
+    has_torchft = True
+else:
+    has_torchft = False
+
+
+@dataclass
+class FTManager:
+    manager: ft.Manager
+    replica_id: int
+    group_size: int
+
+
+def init_ft_manager(job: JobConfig) -> Optional[FTManager]:
+    """Initialize the FT manager if TorchFT is enabled.
+
+    Args:
+        job (JobConfig): The job configuration.
+
+    Returns:
+        Optional[FTManager]: The FT manager if TorchFT is enabled, otherwise None.
+    """
+    if not job.experimental.enable_torchft:
+        return None
+
+    if not has_torchft:
+        raise ImportError("torchft is not installed. Please install it.")
+
+    pg = ft.ProcessGroupBabyNCCL()
+    manager = ft.Manager(
+        pg=pg,
+        min_replica_size=1,
+        load_state_dict=None,
+        state_dict=None,
+        use_async_quorum=True,
+        replica_id=f"torchtitan_ft_{job.experimental.ft_replica_id}",
+    )
+
+    return FTManager(
+        manager,
+        job.experimental.ft_replica_id,
+        job.experimental.ft_group_size,
+    )
@@ -6,7 +6,7 @@
 
 import copy
 import functools
-from typing import Any, Callable, Dict, Iterable, List
+from typing import Any, Callable, Dict, Iterable, List, Optional
 
 import torch
 import torch.nn as nn
@@ -177,8 +177,49 @@ def zero_grad(self) -> None:
         pass
 
 
+class FTOptimizersContainer(OptimizersContainer):
+    def __init__(
+        self,
+        model_parts: List[nn.Module],
+        optimizer_kwargs: Dict[str, Any],
+        name: str,
+        ft_manager: Any,
+    ) -> None:
+        import torchft as ft
+
+        super().__init__(model_parts, optimizer_kwargs, name)
+
+        # Force to initialize the optimizer state so that `optim.step()`
+        # won't be called by state_dict() and load_state_dict().
+        _ = {
+            k: v
+            for sd in map(get_optimizer_state_dict, model_parts, self.optimizers)
+            for k, v in sd.items()
+        }
+        self.optimizers = [
+            ft.Optimizer(ft_manager.manager, optim) for optim in self.optimizers
+        ]
+        self.cache_state_dict: Dict[str, Any] = {}
+
+    def init_cache_state_dict(self) -> None:
+        self.cache_state_dict = super().state_dict()
+
+    def state_dict(self) -> Dict[str, Any]:
+        return self.cache_state_dict
+
+    def load_state_dict(self, state_dict: Dict[str, Any]) -> None:
+        # We have to invalidate the `cache_state_dict` because optimizer uses
+        # assign instead of copy when doing `load_state_dict()`. Without
+        # invalidating the `cache_state_dict`, there will be memory leakage.
+        self.cache_state_dict = {}
+        super().load_state_dict(state_dict)
+        self.init_cache_state_dict()
+
+
 def build_optimizers(
-    model_parts: List[nn.Module], job_config: JobConfig
+    model_parts: List[nn.Module],
+    job_config: JobConfig,
+    ft_manager: Optional[Any] = None,
 ) -> OptimizersContainer:
     """Create a OptimizersContainer for the given model parts and job config.
 
@@ -213,11 +254,14 @@ def build_optimizers(
         "foreach": not fused,
     }
 
-    return (
-        OptimizersContainer(model_parts, optimizer_kwargs, name)
-        if not optim_in_bwd
-        else OptimizersInBackwardContainer(model_parts, optimizer_kwargs, name)
-    )
+    if optim_in_bwd and ft_manager:
+        raise ValueError("TorchFT is not supported with optimizers in backward.")
+    elif optim_in_bwd:
+        return OptimizersInBackwardContainer(model_parts, optimizer_kwargs, name)
+    elif ft_manager:
+        return FTOptimizersContainer(model_parts, optimizer_kwargs, name, ft_manager)
+    else:
+        return OptimizersContainer(model_parts, optimizer_kwargs, name)
 
 
 class LRSchedulersContainer(Stateful):