Integrate TorchFT

**Summary**
This is a WIP TorchFT integration PR.

**Current Issues**

This doesn't work at this moment as there are hanged groups when a new
group joins.

**Issue 1:**
~Group 0 and group 1 will hang during the first `should_commit` after
group 1 applying the pending state_dict from group 0.~

Fixed with: pytorch/torchft#83

**Issue 2:**
~Group 0 and group 1 will pass the `should_commit` but group 0 needs
healing which is wrong and the healing process will cause another hang.~

Fixed with: pytorch/torchft#83

**Issue 3:**
~The byproduct of issue 1 and issue 2: group 1 will continue to print
out~
```
[rank0]:devgpu051:76838:80357 [0] misc/socket.cc:50 NCCL WARN
socketProgress: Connection closed by remote peer
devgpu051.cln3.svc.fbinfra.net<33618>
```

Fixed with pytorch/torchft#91 and several other
fixes.

**Issue 4:**
When there are 3 groups, everyone requests the state dict every step.
***How to reproduce?***
Using the `Reproduce steps` to run 2 groups, then add another group by
modifying the command.

Seems to be fixed, will need more tests.

**Issue 5:**
Hang will happen if using functional collective.
***How to reproduce?***
Pull the latest version of this PR and comment out line 41 and uncomment
line 42 in `torchtitan/utils.py`

**Reproduce steps:**

1. Patch TorchFT with pytorch/torchft#82
2. Execute lighthouse
3. Execute the following command in one terminal:
```
TORCHFT_MANAGER_PORT=29520 REPLICA_GROUP_ID=0 CUDA_VISIBLE_DEVICES=0,1
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=0
```
4. Wait 10 seconds, execute following command in another terminal:
```
TORCHFT_MANAGER_PORT=29522 REPLICA_GROUP_ID=1 CUDA_VISIBLE_DEVICES=2,3
NGPU=2 ./run_llama_train.sh --training.data_parallel_shard_degree=2
--experimental.enable_torchft --experimental.ft_replica_group_id=1
```

ghstack-source-id: 40d49640772abac268fa97147d889aa652559116
Pull Request resolved: #834

Author: fegin

run_llama_train.sh

@@ -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

torchtitan/checkpoint.py

@@ -19,6 +19,7 @@
 import torch.distributed as dist
 import torch.distributed.checkpoint as dcp
 import torch.nn as nn
+from torch.distributed._state_dict_utils import _copy_state_dict, _create_cpu_state_dict
 from torch.distributed.checkpoint.state_dict import (
     get_model_state_dict,
     set_model_state_dict,
@@ -144,50 +145,29 @@ def __init__(
         lr_schedulers: LRSchedulersContainer,
         states: Dict[str, Any],
         job_config: JobConfig,
+        ft_manager: Optional[Any] = None,
     ) -> None:
         ckpt_config = job_config.checkpoint
         self.enable_checkpoint = ckpt_config.enable_checkpoint
-        self.keep_latest_k = ckpt_config.keep_latest_k
+        self.ft_manager = ft_manager
+        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
-        """
-        Note: Pipeline Parallelism and Virtual Stages
-
-        1. even for simple PP schedules, there is a separate optimizer each PP rank.
-        rank0's optimizer would have a param_group[0] which refers to layers.0 in the original model.
-        rank1's would _also_ have a param_group[0], since it's index based, but referring to layers.1.
-        When saving, these collide and one of them is lost.  Then when reloading, only one stage can
-        restore its optimizer states, others will error.
-
-            The solution to this problem is optimizer flattening: it landed in #127071 and is enabled in TorchTitan
-            by passing the 'flatten_optimizer_state_dict' kwarg to DCP functions called in the OptimizerContainer.
-
-        2. With complex PP schedules, we have multiple model chunks per pp rank. This compounds challenge (1) by also
-        requiring us to reason about multiple 'optim' objects locally.
-
-            We solve this in the Model and Optimizer wrapper classes by flattening the state dicts from each object
-            into one state dict before saving/loading. We rely on the individual state_dicts to not collide,
-            which is gauranteed for the model by correct pipeline splitting and for the optimizer by the flattening
-            support described in (1).
-
-        3. LR schedulers also index model states like optimizers and would need to be flattened properly to support
-        resharding.  Unfortunately, the implementations of different lr_schedulers do not follow a clear pattern like
-        optimizers do, so it's hard to write a generic 'flattener' utility.
-
-            TODO: This is currently unsolved and needs a fix.
-        """
-        self.states = states
 
-        self.states.update(
-            {
-                "model": ModelWrapper(model_parts),
-                "optimizer": optimizers,
-                "dataloader": dataloader,
-                "lr_scheduler": lr_schedulers,
-            }
+        self._initialize_states(
+            states, dataloader, model_parts, optimizers, lr_schedulers
         )
 
+        async_mode = ckpt_config.async_mode.lower()
+        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.folder = os.path.join(job_config.job.dump_folder, ckpt_config.folder)
         self.interval_type = (
             IntervalType.SECONDS
@@ -202,6 +182,7 @@ def __init__(
         if async_mode == AsyncMode.ASYNC or self.interval_type == IntervalType.SECONDS:
             self.pg = dist.new_group(backend="gloo")
 
+        self.keep_latest_k = ckpt_config.keep_latest_k
         self.model_weights_only = ckpt_config.model_weights_only
         self.export_dtype = TORCH_DTYPE_MAP[ckpt_config.export_dtype]
 
@@ -225,10 +206,6 @@ def __init__(
                 daemon=True,
             )
             self.mp.start()
-            self.cpu_offload_state_dict = None
-            self.staging = False
-            self.staging_id = None
-            self.staging_stream = torch.cuda.Stream()
         else:
             raise ValueError(f"Unkown checkpoint async_mode {ckpt_config.async_mode}")
 
@@ -242,8 +219,61 @@ def __del__(self):
             self.mp.join()
 
     def reset(self) -> None:
+        # We need to stage the local state if another replicate joins during the
+        # first step.
+        if self.ft_manager:
+            self.cpu_staging(None)
         self.begin_time = time.monotonic()
 
+    def _initialize_states(
+        self,
+        states: Dict[str, Any],
+        dataloader: DataLoader,
+        model_parts: List[nn.Module],
+        optimizers: OptimizersContainer,
+        lr_schedulers: LRSchedulersContainer,
+    ) -> None:
+        """
+        Note: Pipeline Parallelism and Virtual Stages
+
+        1. Even for simple PP schedules, there is a separate optimizer each PP rank.
+        rank0's optimizer would have a param_group[0] which refers to layers.0 in the
+        original model. rank1's would _also_ have a param_group[0], since it's index based,
+        but referring to layers.1.
+        When saving, these collide and one of them is lost.  Then when reloading, only one
+        stage can restore its optimizer states, others will error.
+
+            The solution to this problem is optimizer flattening: it landed in #127071
+            and is enabled in TorchTitan by passing the 'flatten_optimizer_state_dict'
+            kwarg to DCP functions called in the OptimizerContainer.
+
+        2. With complex PP schedules, we have multiple model chunks per pp rank. This
+        compounds challenge (1) by also requiring us to reason about multiple 'optim'
+        objects locally.
+
+            We solve this in the Model and Optimizer wrapper classes by flattening the
+            state dicts from each object into one state dict before saving/loading.
+            We rely on the individual state_dicts to not collide, which is gauranteed for
+            the model by correct pipeline splitting and for the optimizer by the flattening
+            support described in (1).
+
+        3. LR schedulers also index model states like optimizers and would need to be
+        flattened properly to support resharding. Unfortunately, the implementations of
+        different lr_schedulers do not follow a clear pattern like optimizers do, so it's
+        hard to write a generic 'flattener' utility.
+
+            TODO: This is currently unsolved and needs a fix.
+        """
+        self.states = states
+        self.states.update(
+            {
+                "model": ModelWrapper(model_parts),
+                "optimizer": optimizers,
+                "dataloader": dataloader,
+                "lr_scheduler": lr_schedulers,
+            }
+        )
+
     def _create_checkpoint_id(self, step: int) -> str:
         return os.path.join(self.folder, f"step-{step}")
 
@@ -326,31 +356,8 @@ def _async_wait(self) -> None:
                 self.async_future.result()
 
     def _async_with_pinned_memory(self, checkpoint_id: str) -> None:
-        try:
-            from torch.distributed._state_dict_utils import (
-                _copy_state_dict,
-                _create_cpu_state_dict,
-            )
-        except ImportError as e:
-            raise ImportError(
-                "Please install the latest PyTorch nightly to use async checkpointing with pinned memory."
-            ) from e
-        state_dict = dcp.state_dict_saver._stateful_to_state_dict(self.states)
-        if self.cpu_offload_state_dict is None:
-            logger.debug(f"Preparing the CPU memory, {time.monotonic()=}.:.2f")
-            self.cpu_offload_state_dict = _create_cpu_state_dict(
-                state_dict, pin_memory=True, share_memory=True
-            )
-
-        logger.debug(f"Staging the state_dict, {time.monotonic()=}.:.2f")
-        with torch.cuda.stream(self.staging_stream):
-            self.cpu_offload_state_dict = _copy_state_dict(
-                state_dict,
-                self.cpu_offload_state_dict,
-                non_blocking=True,
-            )
-            self.staging = True
-            self.staging_id = checkpoint_id
+        self.cpu_staging(checkpoint_id)
+        self.sending_to_checkpoint_mp = True
 
     def save(self, curr_step: int, force: bool = False) -> None:
         """
@@ -360,6 +367,8 @@ def save(self, curr_step: int, force: bool = False) -> None:
         for initial seed checkpoint.
         """
         if not self._should_save(curr_step, force):
+            if self.ft_manager:
+                self.cpu_staging(None)
             return
 
         begin = time.monotonic()
@@ -383,26 +392,51 @@ def save(self, curr_step: int, force: bool = False) -> None:
             f"in {time.monotonic() - begin:.2f} seconds."
         )
 
+    def cpu_staging(self, checkpoint_id: Optional[str]) -> None:
+        """Offload state_dict to CPU memory"""
+        state_dict = dcp.state_dict_saver._stateful_to_state_dict(self.states)
+        if self.cpu_offload_state_dict is None:
+            logger.debug(f"Preparing the CPU memory, {time.monotonic()=}.:.2f")
+            self.cpu_offload_state_dict = _create_cpu_state_dict(
+                state_dict, pin_memory=True, share_memory=True
+            )
+
+        logger.debug(f"Staging the state_dict, {time.monotonic()=}.:.2f")
+        with torch.cuda.stream(self.staging_stream):
+            self.cpu_offload_state_dict = _copy_state_dict(
+                state_dict,
+                self.cpu_offload_state_dict,
+                non_blocking=True,
+            )
+            self.staging = True
+            self.staging_id = checkpoint_id
+
+    def wait_for_staging(self) -> None:
+        if not self.staging_stream.query():
+            self.staging_stream.synchronize()
+        self.staging = False
+
+    def staging_results(self) -> Dict[str, Any]:
+        self.maybe_wait_for_staging()
+        return self.cpu_offload_state_dict
+
     def maybe_wait_for_staging(self) -> None:
-        if (
-            self.enable_checkpoint
-            and self.async_mode == AsyncMode.ASYNC_WITH_PINNED_MEM
-            and self.staging
-        ):
-            if not self.staging_stream.query():
-                self.staging_stream.synchronize()
-
-            def sync_func():
-                self.mp_queue_send.put_nowait(
-                    (self.cpu_offload_state_dict, self.staging_id)
-                )
-
-            # This may be a faster way to do zero-overhead checkpointing staging
-            # checkpointing but we need more thorough investigation before
-            # swithing to this method.
-            # self.my_thread = threading.Thread(target=func).start()
-            sync_func()
-            self.staging = False
+        if self.enable_staging and self.staging:
+            self.wait_for_staging()
+
+            if self.sending_to_checkpoint_mp:
+                # Copy the sync staging result to another process.
+                def sync_func():
+                    self.mp_queue_send.put_nowait(
+                        (self.cpu_offload_state_dict, self.staging_id)
+                    )
+
+                # This may be a faster way to do zero-overhead checkpointing staging
+                # checkpointing but we need more thorough investigation before
+                # swithing to this method.
+                # self.my_thread = threading.Thread(target=func).start()
+                sync_func()
+                self.sending_to_checkpoint_mp = False
 
     def load(self, step: int = -1) -> bool:
         if not self.enable_checkpoint:

torchtitan/config_manager.py

@@ -620,6 +620,19 @@ 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_group_id",
+            type=int,
+            default=-1,
+            help="The FT replicate group of this run.",
+        )
+
     def to_dict(self):
         return self.args_dict
 

torchtitan/ft.py

@@ -0,0 +1,58 @@
+import importlib
+from typing import Any, Callable, Optional
+
+from torch.distributed._state_dict_utils import _copy_state_dict, _create_cpu_state_dict
+
+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
+
+
+def init_ft_manager(job: JobConfig) -> Optional["ft.Manager"]:
+    """
+    Initialize the FT manager for the given job.
+    """
+    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_group_id}",
+    )
+
+    return manager
+
+
+def set_ft_state_dict_fns(manager: Optional["ft.Manager"], ckpt_manager) -> None:
+    """
+    Set the state dict for the given manager.
+    """
+    if manager is None:
+        return
+
+    def state_dict():
+        ret = {}
+        for k, v in ckpt_manager.staging_results().items():
+            if k in {"model", "optimizer", "lr_schedulers"}:
+                ret[k] = v
+        return ret
+
+    def load_state_dict(state_dict):
+        assert state_dict is not None
+        for k, v in state_dict.items():
+            ckpt_manager.states[k].load_state_dict(v)
+
+    manager.set_state_dict_fns(load_state_dict, state_dict)