[moe] init mixtral impl

applications/ColossalMoE/colossal_moe/models/mixtral_checkpoint.py

+import logging
+import os
+from pathlib import Path
+
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+
+from colossalai.checkpoint_io import CheckpointIndexFile
+from colossalai.checkpoint_io.utils import is_safetensors_available, load_shard_state_dict, load_state_dict_into_model
+from colossalai.moe import MoECheckpintIO
+from colossalai.tensor.moe_tensor.api import get_dp_rank, get_ep_group, get_ep_rank, get_ep_size, is_moe_tensor
+
+
+class MixtralMoECheckpointIO(MoECheckpintIO):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    @torch.no_grad()
+    def pre_load_model(self, model: nn.Module, state_dict: dict) -> dict:
+        """
+        Preprocess state_dict before loading and slice the state_dict of MOE tensors.
+        """
+        model_param_dict = dict(model.named_parameters())
+        for name, param in list(state_dict.items()):
+            if ".gate.weight" in name:
+                new_name = "module." + name.replace(".gate.weight", ".gate_weight")
+                state_dict[new_name] = state_dict.pop(name)
+            elif ".experts." in name:
+                # if is moe tensor
+                # in our moe module, expert is cat as one tensor
+                # but mixtral's experts is not cat
+                # we will insert the loaded expert into the position of cat tensor
+
+                # get model param
+                str_idx = name.index(".experts.")
+                expert_idx = int(name.split(".")[-3])
+                if ".w1." in name:
+                    model_param_name = name.replace(name[str_idx:], ".experts.wi_gate")
+                elif ".w2." in name:
+                    model_param_name = name.replace(name[str_idx:], ".experts.wo")
+                elif ".w3." in name:
+                    model_param_name = name.replace(name[str_idx:], ".experts.wi_up")
+                model_param_name = "module." + model_param_name
+                # skip for pipeline
+                if model_param_name not in model_param_dict:
+                    continue
+                model_param = model_param_dict[model_param_name]
+                assert is_moe_tensor(model_param)
+                # get expert range
+                ep_rank = get_ep_rank(model_param)
+                ep_size = get_ep_size(model_param)
+                expert_num = 8 // ep_size
+                expert_range = list(range(ep_rank * expert_num, (ep_rank + 1) * expert_num))
+                # insert new param
+                if expert_idx in expert_range:
+                    new_param = model_param
+                    new_param[expert_idx - ep_rank * expert_num] = param.transpose(0, 1)
+                    state_dict[model_param_name] = new_param
+                state_dict.pop(name)
+            else:
+                new_name = "module." + name
+                state_dict[new_name] = state_dict.pop(name)
+
+        dist.barrier()
+        return state_dict
+
+    def load_sharded_model(self, model: nn.Module, checkpoint_index_file: Path, strict: bool = False):
+        """
+        Load sharded model with the given path to index file of checkpoint folder.
+
+        Args:
+            model (nn.Module): The model to be loaded.
+            checkpoint_index_file (str): Path to the index file of checkpointing folder.
+            strict (bool, optional): For name matching during loading state_dict. Defaults to False.
+                                     This argument should be manually set to False since params on same device might be stored in different files.
+        """
+
+        # Check whether the checkpoint uses safetensors.
+        use_safetensors = False
+        if "safetensors" in checkpoint_index_file.name:
+            use_safetensors = True
+
+        if use_safetensors and not is_safetensors_available():
+            raise ImportError("`safe_serialization` requires the `safetensors` library: `pip install safetensors`.")
+
+        # Read checkpoint index file.
+        ckpt_index_file = CheckpointIndexFile.from_file(checkpoint_index_file)
+        ckpt_root_path = ckpt_index_file.root_path
+        weight_map = ckpt_index_file.weight_map
+        strict = False
+
+        # Load params & buffers to model.
+        # Keep a record of loaded files so that file will not be repeatedly loaded.
+        loaded_file = set()
+
+        def _load(name: str):
+            if name not in weight_map:
+                raise ValueError(f"{name} is not stored in checkpoint, please check your checkpointing configuration!")
+            filename = weight_map[name]
+
+            # If this param/buffer has been loaded before, directly return.
+            if filename in loaded_file:
+                return
+
+            file_path = os.path.join(ckpt_root_path, filename)
+            state_dict = load_shard_state_dict(Path(file_path), use_safetensors)
+            state_dict = self.pre_load_model(model, state_dict)
+            missing_keys = []
+
+            load_state_dict_into_model(
+                model,
+                state_dict,
+                missing_keys=missing_keys,
+                strict=strict,
+                load_sub_module=True,
+            )
+            loaded_file.add(filename)
+
+        # Load parameters.
+        for name, _ in model.named_parameters():
+            name = name.replace("module.", "")
+            name = name.replace(".gate_weight", ".gate.weight")
+            if ".experts.wi_gate" in name:
+                for i in range(8):
+                    new_name = name.replace(".experts.wi_gate", f".experts.{i}.w1.weight")
+                    _load(new_name)
+            elif ".experts.wi_up" in name:
+                for i in range(8):
+                    new_name = name.replace(".experts.wi_up", f".experts.{i}.w3.weight")
+                    _load(new_name)
+            elif ".experts.wo" in name:
+                for i in range(8):
+                    new_name = name.replace(".experts.wo", f".experts.{i}.w2.weight")
+                    _load(new_name)
+            else:
+                _load(name)
+
+        if self.verbose:
+            logging.info(f"The model has been successfully loaded from sharded checkpoint: {ckpt_root_path}.")
+
+    @torch.no_grad()
+    def pre_save_model(self, model: nn.Module) -> dict:
+        torch.cuda.empty_cache()
+        state_dict = model.state_dict()
+        for name, param in list(model.named_parameters()):
+            if ".gate_weight" in name:
+                new_name = name.replace(".gate_weight", ".gate.weight")
+                state_dict[new_name] = state_dict.pop(name).cpu()
+            elif ".experts." in name:
+                ep_group = get_ep_group(param)
+                ep_rank = get_ep_rank(param)
+                ep_size = get_ep_size(param)
+                dp_rank = get_dp_rank(param)
+
+                if dp_rank == 0:
+                    param = param.data.cuda()
+                    all_param = [torch.zeros_like(param) for _ in range(ep_size)]
+                    # gather param from every ep rank
+                    dist.all_gather(all_param, param, group=ep_group)
+                    if ep_rank == 0:
+                        all_param = torch.cat(all_param, dim=0)
+                        assert all_param.shape[0] == 8
+                        for i in range(8):
+                            if ".wi_gate" in name:
+                                new_name = name.replace(".experts.wi_gate", f".experts.{i}.w1.weight")
+                            elif ".wi_up" in name:
+                                new_name = name.replace(".experts.wi_up", f".experts.{i}.w3.weight")
+                            elif ".wo" in name:
+                                new_name = name.replace(".experts.wo", f".experts.{i}.w2.weight")
+                            new_name = new_name.replace("module.", "")
+                            new_param = all_param[i].transpose(-1, -2)
+                            state_dict[new_name] = new_param.cpu()
+                        state_dict.pop(name)
+            else:
+                state_dict[name] = param.cpu()
+
+        for name, param in list(state_dict.items()):
+            new_name = name.replace("module.", "")
+            state_dict[new_name] = state_dict.pop(name)
+        
+        torch.cuda.empty_cache()
+        if self.pp_size > 1:
+            if self.dp_rank == 0:
+                # gather state_dict from every pp rank
+                # because ckpt is large, we split it into 10 parts
+                # and gather them one by one
+                new_state_dict = {}
+                state_dict_keys = list(state_dict.keys())
+                gap_key_num = min(30, len(state_dict_keys))
+                gap_keys = (len(state_dict_keys) + gap_key_num - 1) // gap_key_num
+                for i in range(gap_key_num):
+                    cur_keys = state_dict_keys[i * gap_keys : (i + 1) * gap_keys]
+                    cur_state_dict = {}
+                    for k in cur_keys:
+                        cur_state_dict[k] = state_dict[k]
+                    out = [None for _ in range(self.pp_size)]
+                    dist.all_gather_object(out, cur_state_dict, group=self.pp_group)
+                    if self.pp_rank == 0:
+                        for o in out:
+                            for k, v in o.items():
+                                new_state_dict[k] = v.cpu()
+                state_dict = new_state_dict
+        dist.barrier()
+        return state_dict

applications/ColossalMoE/colossal_moe/models/mixtral_layer.py

+import torch
+import torch.nn as nn
+from transformers.models.mixtral.modeling_mixtral import MixtralDecoderLayer, MixtralSparseMoeBlock
+
+from colossalai.lazy import LazyInitContext
+from colossalai.moe import SparseMLP
+
+
+class MixtralSparseMLP:
+    r"""
+    This is a wrapper around the apex fused layernorm implementation. It is meant to be used only with the from_native_module interface.
+    """
+
+    def __init__(self) -> None:
+        raise NotImplementedError(
+            "FusedLayerNorm is not implemented as a physical class. "
+            "It is meant to be used only with the from_native_module interface convert a native pytorch layer norm module to FusedLayerNorm module provided by apex."
+        )
+
+    @staticmethod
+    def from_native_module(module: MixtralSparseMoeBlock, enable_kernel: bool) -> nn.Module:
+        r"""
+        Convert a native pytorch layer norm module to FusedLayerNorm module provided by apex,
+        and optionally marking parameters for gradient aggregation.
+
+        Args:
+            module (nn.LayerNorm): The native PyTorch LayerNorm module to be converted.
+            sp_partial_derived (bool): Whether this module's gradients are partially derived in sequence parallelism.
+
+        Returns:
+            nn.Module: Union[FastLayerNorm, FusedLayerNorm].
+
+        Raises:
+            AssertionError: If the provided module is not an instance of nn.LayerNorm.
+        """
+        with torch.no_grad():
+            LazyInitContext.materialize(module)
+
+            # get the attributes of the module
+            moe_kwargs = dict(
+                num_experts=8,
+                hidden_size=module.hidden_dim,
+                intermediate_size=module.ffn_dim,
+                router_top_k=module.top_k,
+                router_norm=True,
+                router_loss=False,
+                # router_capacity_factor_train=
+                # router_capacity_factor_eval=
+                mlp_activation="silu",
+                mlp_gated=True,
+                # enable_load_balance=
+                # load_balance_tolerance=
+                # load_balance_beam_width=
+                # load_balance_group_swap_factor=
+                enable_kernel=enable_kernel,
+                # enable_comm_overlap=
+                # enable_hierarchical_comm=
+                return_gate_logits=True,
+            )
+            dtype = module.gate.weight.dtype
+            device = module.gate.weight.device
+            sparse_mlp = SparseMLP(**moe_kwargs).to(dtype).to(device)
+
+        return sparse_mlp
+
+
+def replace_moe_layer(model: nn.Module, enable_kernel: bool = False) -> nn.Module:
+    """
+    Reverse the replace layer operation
+
+    Args:
+        module (torch.nn.Module): The object of layer to shard
+    """
+    if isinstance(model, MixtralDecoderLayer):
+        model.block_sparse_moe = MixtralSparseMLP.from_native_module(
+            model.block_sparse_moe, enable_kernel=enable_kernel
+        )
+    else:
+        for _, child in model.named_children():
+            replace_moe_layer(child, enable_kernel)

applications/ColossalMoE/colossal_moe/utils.py

+import json
+import os
+from typing import Any, Dict, Tuple, Union
+
+import torch
+from huggingface_hub import snapshot_download
+from torch.optim.lr_scheduler import _LRScheduler
+from torch.optim.optimizer import Optimizer
+
+from colossalai.booster import Booster
+from colossalai.cluster import DistCoordinator
+
+
+def move_to_cuda(batch, device):
+    return {k: v.to(device) for k, v in batch.items()}
+
+
+@torch.no_grad()
+def load_model(ckpt_path: str, model, booster: Booster, optimizer=None):
+    # pytorch ckpt
+    if os.path.exists(os.path.join(ckpt_path, "model.safetensors.index.json")):
+        ckpt_path = os.path.join(ckpt_path, "model.safetensors.index.json")
+    # saved ckpt
+    elif os.path.exists(os.path.join(ckpt_path, "pytorch_model.bin.index.json")):
+        ckpt_path = os.path.join(ckpt_path, "pytorch_model.bin.index.json")
+    # download
+    else:
+        ckpt_path = snapshot_download(ckpt_path)
+    booster.load_model(model, ckpt_path)
+    if optimizer is not None:
+        optimizer.sync_moe_master_param()
+        optimizer.update_master_params(model)
+
+
+def load_json(file_path: Union[str, os.PathLike]) -> Dict[str, Any]:
+    """
+    Load file in JSON format
+    """
+    with open(file=file_path, mode="r", encoding="utf-8") as fp:
+        return json.load(fp)
+
+
+def save_json(data: Dict[str, Any], file_path: Union[str, os.PathLike]) -> None:
+    """
+    Save as JSON format
+    """
+    with open(file=file_path, mode="w", encoding="utf-8") as fp:
+        json.dump(data, fp=fp, ensure_ascii=False, indent=4)
+
+
+def save_checkpoint(
+    save_dir: Union[str, os.PathLike],
+    booster: Booster,
+    model: torch.nn.Module,
+    optimizer: Optimizer,
+    lr_scheduler: _LRScheduler,
+    epoch: int,
+    step: int,
+    batch_size: int,
+    coordinator: DistCoordinator,
+) -> None:
+    """
+    Save model checkpoint, optimizer, LR scheduler and intermedidate running states.
+    """
+
+    save_dir = os.path.join(save_dir, f"epoch-{epoch}_step-{step}")
+    os.makedirs(os.path.join(save_dir, "modeling"), exist_ok=True)
+
+    booster.save_model(model, os.path.join(save_dir, "modeling"), shard=True)
+    booster.save_optimizer(optimizer, os.path.join(save_dir, "optimizer"), shard=True)
+    booster.save_lr_scheduler(lr_scheduler, os.path.join(save_dir, "lr_scheduler"))
+    running_states = {
+        "epoch": epoch,
+        "step": step,
+        "sample_start_index": step * batch_size,
+    }
+    if coordinator.is_master():
+        save_json(running_states, os.path.join(save_dir, "running_states.json"))
+
+
+def load_checkpoint(
+    load_dir: Union[str, os.PathLike],
+    booster: Booster,
+    model: torch.nn.Module,
+    optimizer: Optimizer,
+    lr_scheduler: _LRScheduler,
+) -> Tuple[int, int, int]:
+    """
+    Load model checkpoint, optimizer, LR scheduler and intermedidate running states.
+    """
+
+    # Update booster params states.
+    load_model(os.path.join(load_dir, "modeling"), model, booster, optimizer)
+    booster.load_optimizer(optimizer=optimizer, checkpoint=os.path.join(load_dir, "optimizer"))
+    booster.load_lr_scheduler(lr_scheduler=lr_scheduler, checkpoint=os.path.join(load_dir, "lr_scheduler"))
+
+    running_states = load_json(file_path=os.path.join(load_dir, "running_states.json"))
+    return (
+        running_states["epoch"],
+        running_states["step"],
+        running_states["sample_start_index"],
+    )

applications/ColossalMoE/infer.py

+import argparse
+
+import torch
+import torch.distributed as dist
+from colossal_moe.models.mixtral_checkpoint import MixtralMoECheckpointIO
+from colossal_moe.models.mixtral_layer import replace_moe_layer
+from colossal_moe.models.mixtral_policy import MixtralForCausalLMPolicy
+from colossal_moe.utils import load_model
+from transformers import AutoTokenizer
+from transformers.models.mixtral import MixtralConfig, MixtralForCausalLM
+
+import colossalai
+from colossalai.booster import Booster
+from colossalai.booster.plugin.moe_hybrid_parallel_plugin import MoeHybridParallelPlugin
+from colossalai.cluster import DistCoordinator
+from colossalai.moe import MOE_MANAGER
+from colossalai.moe.utils import skip_init
+from colossalai.utils import get_current_device
+
+
+def parse_args():
+    # basic settings
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model_name",
+        type=str,
+        default="mistralai/Mixtral-8x7B-v0.1",
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--plugin",
+        type=str,
+        default="hybrid",
+        choices=["ep"],
+        help="Parallel methos.",
+    )
+    parser.add_argument(
+        "--output_path",
+        type=str,
+        default="./outputs",
+        help="The path of your saved model after finetuning.",
+    )
+    parser.add_argument(
+        "--precision",
+        type=str,
+        default="bf16",
+        choices=["fp32", "bf16", "fp16"],
+        help="The mixed precision training.",
+    )
+    parser.add_argument("--seed", type=int, default=42, help="A seed for reproducible training.")
+
+    # kernel
+    parser.add_argument(
+        "--use_kernel",
+        action="store_true",
+        help="Use kernel optim. Need to install flash attention and triton to enable all kernel optimizations. Skip if not installed.",
+    )
+    parser.add_argument(
+        "--use_layernorm_kernel",
+        action="store_true",
+        help="Use layernorm kernel. Need to install apex. Raise error if not installed.",
+    )
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+
+    # Launch ColossalAI
+    colossalai.launch_from_torch(config={}, seed=args.seed)
+    coordinator = DistCoordinator()
+
+    # Set plugin
+    booster_kwargs = {}
+    hybrid_dict = {
+        "tp_size": 1,
+        "custom_policy": MixtralForCausalLMPolicy(),
+        "enable_fused_normalization": args.use_layernorm_kernel,
+        "enable_jit_fused": args.use_kernel,
+        "precision": args.precision,
+        "checkpoint_io": MixtralMoECheckpointIO,
+        "zero_stage": 1,
+    }
+    mgr_dict = {}
+    if args.plugin == "ep":
+        dp_size = dist.get_world_size()
+        plugin = MoeHybridParallelPlugin(
+            pp_size=1,
+            **hybrid_dict,
+        )
+        MOE_MANAGER.setup(
+            parallel="EP",
+            max_ep_size=dp_size,
+            **mgr_dict,
+        )
+    else:
+        raise ValueError(f"Invalid plugin {args.plugin}")
+    coordinator.print_on_master(f"Set plugin as {plugin.__class__.__name__}")
+
+    # Build mixtral model
+    config = MixtralConfig.from_pretrained(args.model_name)
+    config.num_local_experts = 1  # dont change this. it will not affect model
+    with skip_init():
+        model = MixtralForCausalLM(config)
+    model.num_experts = 8
+    model = model.to(torch.bfloat16) if args.precision == "bf16" else model.to(torch.float16)
+    model = model.to(get_current_device())
+    coordinator.print_on_master(f"Finish init model with config:\n{config}")
+
+    # Replace moe
+    with skip_init():
+        replace_moe_layer(model)
+    model.eval()
+    coordinator.print_on_master(f"Finish replace moe module")
+
+    # Prepare tokenizer and dataloader
+    tokenizer = AutoTokenizer.from_pretrained(args.model_name)
+
+    # Set booster
+    booster = Booster(plugin=plugin, **booster_kwargs)
+    model, _, _, _, _ = booster.boost(model=model)
+    coordinator.print_on_master(f"Finish init booster")
+
+    # load ckpt
+    load_model(args.model_name, model, booster)
+    coordinator.print_on_master(f"Finish load ckpt")
+
+    text = ["Hello my name is", "1+1=?"]
+    tokenizer.pad_token = tokenizer.unk_token
+    inputs = tokenizer(text, return_tensors="pt", padding=True).to(torch.cuda.current_device())
+    outputs = model.module.generate(**inputs, max_new_tokens=20)
+    outputs = tokenizer.batch_decode(outputs)[0]
+    print(outputs)
+
+
+if __name__ == "__main__":
+    main()

applications/ColossalMoE/infer.sh

+NUM_GPU=2
+MODEL="mistralai/Mixtral-8x7B-v0.1"
+
+# ep
+torchrun --standalone --nproc_per_node $NUM_GPU infer.py \
+    --model_name $MODEL \
+    --plugin "ep" \

applications/ColossalMoE/requirements.txt

+colossalai >= 0.3.3
+torch >= 1.8.1
+transformers == 4.36.0
+sentencepiece
+datasets

applications/ColossalMoE/setup.py

+from setuptools import find_packages, setup
+
+
+def fetch_requirements(path):
+    with open(path, "r") as fd:
+        return [r.strip() for r in fd.readlines()]
+
+
+def fetch_readme():
+    with open("README.md", encoding="utf-8") as f:
+        return f.read()
+
+
+def fetch_version():
+    with open("version.txt", "r") as f:
+        return f.read().strip()
+
+
+setup(
+    name="colossal_moe",
+    version=fetch_version(),
+    packages=find_packages(
+        exclude=(
+            "tests",
+            "benchmarks",
+            "*.egg-info",
+        )
+    ),
+    description="Colossal-AI MoE",
+    long_description=fetch_readme(),
+    long_description_content_type="text/markdown",
+    license="Apache Software License 2.0",
+    url="https://github.com/hpcaitech",
+    install_requires=fetch_requirements("requirements.txt"),
+    python_requires=">=3.6",
+    classifiers=[
+        "Programming Language :: Python :: 3",
+        "License :: OSI Approved :: Apache Software License",
+        "Environment :: GPU :: NVIDIA CUDA",
+        "Topic :: Scientific/Engineering :: Artificial Intelligence",
+        "Topic :: System :: Distributed Computing",
+    ],
+)

applications/ColossalMoE/tests/test_moe_checkpoint.py

+import os
+import shutil
+
+import pytest
+import torch
+import torch.distributed as dist
+from colossal_moe.models.mixtral_checkpoint import MixtralMoECheckpointIO
+from colossal_moe.models.mixtral_layer import replace_moe_layer
+from colossal_moe.models.mixtral_policy import MixtralForCausalLMPolicy
+from transformers.models.mixtral import MixtralConfig, MixtralForCausalLM
+
+import colossalai
+from colossalai.booster import Booster
+from colossalai.booster.plugin.moe_hybrid_parallel_plugin import MoeHybridParallelPlugin
+from colossalai.moe.manager import MOE_MANAGER
+from colossalai.testing import DummyDataloader, check_state_dict_equal, rerun_if_address_is_in_use, spawn
+from colossalai.utils import get_current_device
+
+
+def data_gen_fn(batch_size: int = 2, max_length: int = 4, vocab_size: int = 20):
+    input_ids = torch.randint(0, vocab_size, (batch_size, max_length), device=get_current_device())
+    attention_mask = torch.ones_like(input_ids)
+    return {
+        "input_ids": input_ids,
+        "attention_mask": attention_mask,
+        "labels": input_ids,
+    }
+
+
+def run_fwd_bwd(
+    model, data, label, criterion, optimizer, enable_autocast=False, pipeline=False, booster=None, plugin=None
+):
+    model.train()
+    if pipeline:
+        train_dataloader_iter = DummyDataloader(data_gen_fn, length=1)
+        is_pp_last_stage = booster.plugin.stage_manager.is_last_stage()
+        y = booster.execute_pipeline(
+            train_dataloader_iter,
+            model,
+            lambda x, y: x.loss,
+            optimizer,
+            return_loss=True,
+            return_outputs=True,
+        )
+        # Backward and optimize
+        if is_pp_last_stage:
+            loss = y["loss"]
+    else:
+        if criterion:
+            y = model(data).logits
+            loss = criterion(y)
+        else:
+            loss = model(data, label)
+        loss = loss.float()
+
+        if optimizer is not None:
+            optimizer.backward(loss)
+        else:
+            loss.backward()
+    return y
+
+
+def get_config():
+    config = MixtralConfig(
+        vocab_size=300,
+        hidden_size=32,
+        intermediate_size=16,
+        num_hidden_layers=2,
+        dropout_rate=0.0,
+    )
+    return config
+
+
+def get_model(parallel):
+    config = get_config()
+    model = MixtralForCausalLM(config).to(torch.bfloat16)
+    replace_moe_layer(model)
+    optim = torch.optim.Adam(model.parameters())
+    args = dict(
+        precision="bf16",
+        tp_size=1,
+        zero_stage=1,
+        custom_policy=MixtralForCausalLMPolicy(),
+        checkpoint_io=MixtralMoECheckpointIO,
+    )
+    if parallel == "ep":
+        plugin = MoeHybridParallelPlugin(
+            pp_size=1,
+            **args,
+        )
+    elif parallel == "hybrid":
+        plugin = MoeHybridParallelPlugin(
+            pp_size=2,
+            microbatch_size=1,
+            **args,
+        )
+    booster = Booster(plugin=plugin)
+    model, optim, _, _, _ = booster.boost(model=model, optimizer=optim)
+    return model, booster, optim
+
+
+def _test_moe_checkpoint(parallel):
+    if dist.get_rank() == 0:
+        if os.path.exists("./tmp_ckpt1"):
+            shutil.rmtree("./tmp_ckpt1")
+        if os.path.exists("./tmp_ckpt2"):
+            shutil.rmtree("./tmp_ckpt2")
+    dist.barrier()
+
+    if parallel == None:
+        MOE_MANAGER.setup(
+            parallel=None,
+        )
+    elif parallel == "ep":
+        MOE_MANAGER.setup(
+            parallel="EP",
+        )
+    elif parallel == "hybrid":
+        MOE_MANAGER.setup(
+            parallel="EP",
+            mode="fixed",
+            fixed_dp_size=1,
+            fixed_ep_size=2,
+            fixed_pp_size=2,
+        )
+    model1, booster1, optim1 = get_model(parallel)
+    model2, booster2, optim2 = get_model(parallel)
+    # param ckpt
+    # check not equal
+    try:
+        check_state_dict_equal(model1.state_dict(), model2.state_dict(), False)
+        raise AssertionError("state_dict should not be equal")
+    except:
+        pass
+    # shard
+    booster1.save_model(model1, "./tmp_ckpt1", shard=True, size_per_shard=1)
+    booster2.load_model(model2, "./tmp_ckpt1")
+    # check
+    check_state_dict_equal(model1.state_dict(), model2.state_dict(), False)
+
+    # optim ckpt
+    criterion = lambda x: x.mean()
+    data = torch.randint(0, 4, (2, 4)).cuda()
+    label = torch.randint(0, 4, (2,)).cuda()
+    if parallel == "hybrid":
+        kwargs = {"pipeline": True, "booster": booster1, "plugin": booster1.plugin}
+    else:
+        kwargs = {}
+    run_fwd_bwd(model1, data, label, criterion, optim1, **kwargs)
+    optim1.step()
+    optim1.zero_grad()
+    # shard
+    booster1.save_optimizer(optim1, "./tmp_ckpt2", shard=True, size_per_shard=1)
+    dist.barrier()
+    booster2.load_optimizer(optim2, "./tmp_ckpt2")
+    # check
+    check_state_dict_equal(optim1.optim.state_dict(), optim2.optim.state_dict(), False)
+
+    if dist.get_rank() == 0:
+        shutil.rmtree("./tmp_ckpt1")
+        shutil.rmtree("./tmp_ckpt2")
+
+
+def _run_dist(rank, world_size, port, parallel):
+    colossalai.launch(
+        config=dict(),
+        rank=rank,
+        world_size=world_size,
+        host="localhost",
+        port=port,
+        backend="nccl",
+    )
+    _test_moe_checkpoint(parallel)
+
+
+@pytest.mark.dist
+@pytest.mark.parametrize("world_size", [4])
+@pytest.mark.parametrize("parallel", ["ep", "hybrid"])
+@rerun_if_address_is_in_use()
+def test_moe_checkpoint(world_size, parallel):
+    spawn(_run_dist, world_size, parallel=parallel)
+
+
+if __name__ == "__main__":
+    test_moe_checkpoint(world_size=4, parallel="hybrid")

applications/ColossalMoE/tests/test_moe_layer.py

+import copy
+
+import torch
+from colossal_moe.models.mixtral_layer import MixtralSparseMLP
+from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock
+
+
+class Config:
+    def __init__(self, hidden_size, intermediate_size, num_local_experts, num_experts_per_tok, hidden_act):
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_local_experts = num_local_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.hidden_act = hidden_act
+
+
+def test_moe_layer():
+    config = Config(hidden_size=4, intermediate_size=8, num_local_experts=32, num_experts_per_tok=2, hidden_act="silu")
+    mistral_moe = MixtralSparseMoeBlock(config).cuda()
+    colossal_moe = MixtralSparseMLP.from_native_module(copy.deepcopy(mistral_moe)).cuda()
+
+    data = torch.randn(2, 8, 4).cuda()
+    mistral_output = mistral_moe(data)[0]
+    colossal_output = colossal_moe(data)[0]
+    assert torch.allclose(
+        mistral_output, colossal_output
+    ), f"mistral_output: {mistral_output}\ncolossal_output: {colossal_output}"
+
+
+if __name__ == "__main__":
+    test_moe_layer()

applications/ColossalMoE/train.py

+import argparse
+
+import torch
+import torch.distributed as dist
+from colossal_moe.models.mixtral_checkpoint import MixtralMoECheckpointIO
+from colossal_moe.models.mixtral_layer import replace_moe_layer
+from colossal_moe.models.mixtral_policy import MixtralForCausalLMPolicy
+from colossal_moe.utils import load_checkpoint, load_model, move_to_cuda, save_checkpoint
+from torch.utils.data import Dataset
+from tqdm import tqdm
+from transformers import AutoTokenizer
+from transformers.models.mixtral import MixtralConfig, MixtralForCausalLM
+
+import colossalai
+from colossalai.booster import Booster
+from colossalai.booster.plugin.moe_hybrid_parallel_plugin import MoeHybridParallelPlugin
+from colossalai.cluster import DistCoordinator
+from colossalai.moe import MOE_MANAGER, apply_load_balance
+from colossalai.moe.layers import apply_load_balance
+from colossalai.moe.manager import MOE_MANAGER
+from colossalai.nn.lr_scheduler import CosineAnnealingWarmupLR
+from colossalai.nn.optimizer import HybridAdam
+from colossalai.utils import get_current_device
+
+
+@torch.no_grad()
+def get_global_loss(loss, booster):
+    global_loss = loss.clone().detach()
+    dist.all_reduce(tensor=global_loss, op=dist.ReduceOp.SUM, group=booster.plugin.dp_group)
+    global_loss.div_(booster.plugin.dp_size)
+    return global_loss
+
+
+class RandomDataset(Dataset):
+    def __init__(self, num_samples: int = 1000, max_length: int = 2048, vocab_size: int = 100, tokenizer=None):
+        self.num_samples = num_samples
+        self.max_length = max_length
+        self.input_ids = torch.randint(0, vocab_size, (num_samples, max_length), device=get_current_device())
+        self.attention_mask = torch.ones_like(self.input_ids)
+
+    def __len__(self):
+        return self.num_samples
+
+    def __getitem__(self, idx):
+        return {
+            "input_ids": self.input_ids[idx],
+            "attention_mask": self.attention_mask[idx],
+            "labels": self.input_ids[idx],
+        }
+
+
+def parse_args():
+    # basic settings
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model_name",
+        type=str,
+        default="mistralai/Mixtral-8x7B-v0.1",
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument("--load_checkpoint", type=str, default=None, help="Load checkpoint")
+    parser.add_argument(
+        "--plugin",
+        type=str,
+        default="hybrid",
+        choices=["hybrid"],
+        help="Parallel methods.",
+    )
+    parser.add_argument(
+        "--output_path",
+        type=str,
+        default="./outputs",
+        help="The path of your saved model after finetuning.",
+    )
+    parser.add_argument("--num_epoch", type=int, default=1, help="Number of epochs.")
+    parser.add_argument(
+        "--batch_size",
+        type=int,
+        default=1,
+        help="Batch size (per dp group) for the training dataloader.",
+    )
+    parser.add_argument(
+        "--save_interval",
+        type=int,
+        default=1000,
+        help=" The interval (steps) of saving checkpoints.",
+    )
+    parser.add_argument(
+        "--precision",
+        type=str,
+        default="bf16",
+        choices=["fp32", "bf16", "fp16"],
+        help="The mixed precision training.",
+    )
+    parser.add_argument("--max_length", type=int, default=2048, help="Max sequence length.")
+    parser.add_argument("--seed", type=int, default=42, help="A seed for reproducible training.")
+
+    # optim
+    parser.add_argument("--lr", type=float, default=1e-5, help="Learning rate.")
+    parser.add_argument("--weight_decay", type=float, default=0.0, help="Weight decay to use.")
+
+    # lr scheduler
+    parser.add_argument("--num_epochs", type=int, default=1, help="Number of training epochs")
+    parser.add_argument("--warmup_steps", type=int, default=None, help="Warmup steps")
+
+    # zero stage for all plugins
+    parser.add_argument("--zero_stage", type=int, default=2, help="zero stage.")
+    # hybrid plugin
+    parser.add_argument("--pp_size", type=int, default=2, help="pp size for hybrid plugin")
+    parser.add_argument("--dp_size", type=int, default=1, help="dp size for hybrid plugin")
+    parser.add_argument("--ep_size", type=int, default=2, help="ep size for hybrid plugin")
+    parser.add_argument("--microbatch_size", type=int, default=1, help="Microbatch size in pipeline for hybrid plugin")
+
+    # kernel
+    parser.add_argument(
+        "--use_kernel",
+        action="store_true",
+        help="Use kernel optim. Need to install flash attention and triton to enable all kernel optimizations. Skip if not installed.",
+    )
+    parser.add_argument(
+        "--use_layernorm_kernel",
+        action="store_true",
+        help="Use layernorm kernel. Need to install apex. Raise error if not installed.",
+    )
+
+    # load balance
+    parser.add_argument(
+        "--load_balance", action="store_true", help="Expert load balance. Defaults to False. Recommend to enable."
+    )
+    parser.add_argument("--load_balance_interval", type=int, default=1000, help="Expert load balance interval.")
+    # communicate overlap
+    parser.add_argument(
+        "--comm_overlap",
+        action="store_true",
+        help="Use communication overlap for MoE. Recommended to enable for muiti-node training.",
+    )
+    # hierarchical all-to-all
+    parser.add_argument(
+        "--hierarchical_alltoall",
+        action="store_true",
+        help="Use hierarchical all-to-all for MoE. Recommended to enable for muiti-node training.",
+    )
+
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+
+    # Launch ColossalAI
+    colossalai.launch_from_torch(config={}, seed=args.seed)
+    coordinator = DistCoordinator()
+
+    # Set plugin
+    booster_kwargs = {}
+    hybrid_dict = {
+        "tp_size": 1,
+        "custom_policy": MixtralForCausalLMPolicy(),
+        "enable_fused_normalization": args.use_layernorm_kernel,
+        "enable_jit_fused": args.use_kernel,
+        "precision": args.precision,
+        "zero_stage": args.zero_stage,
+        "checkpoint_io": MixtralMoECheckpointIO,
+    }
+    mgr_dict = {}
+    if args.plugin == "hybrid":
+        plugin = MoeHybridParallelPlugin(
+            pp_size=args.pp_size,
+            microbatch_size=args.microbatch_size,
+            **hybrid_dict,
+        )
+        MOE_MANAGER.setup(
+            parallel="EP",
+            mode="fixed",
+            fixed_dp_size=args.dp_size,
+            fixed_ep_size=args.ep_size,
+            fixed_pp_size=args.pp_size,
+            **mgr_dict,
+        )
+    else:
+        raise ValueError(f"Invalid plugin {args.plugin}")
+    coordinator.print_on_master(f"Set plugin as {plugin.__class__.__name__}")
+
+    # Build Mixtral model
+    config = MixtralConfig.from_pretrained(args.model_name)
+    config.use_cache = False
+    config.num_local_experts = 1
+    model = MixtralForCausalLM(config)
+    model.num_experts = 8
+    model = model.to(torch.bfloat16) if args.precision == "bf16" else model.to(torch.float16)
+    model = model.to(get_current_device())
+    replace_moe_layer(model, enable_kernel=args.use_kernel)
+    coordinator.print_on_master(f"Finish init model with config:\n{config}")
+
+    # Enable gradient checkpointing
+    model.gradient_checkpointing_enable()
+
+    # Prepare tokenizer and dataloader
+    tokenizer = AutoTokenizer.from_pretrained(args.model_name)
+    dataset = RandomDataset(num_samples=100, tokenizer=tokenizer)
+    collate_fn = None
+    dataloader = plugin.prepare_dataloader(
+        dataset, batch_size=args.batch_size, shuffle=True, drop_last=True, collate_fn=collate_fn
+    )
+
+    # Set optimizer
+    optimizer = HybridAdam(
+        model_params=model.parameters(),
+        lr=args.lr,
+        betas=(0.9, 0.95),
+        weight_decay=args.weight_decay,
+        adamw_mode=True,
+    )
+
+    # Set lr scheduler
+    lr_scheduler = CosineAnnealingWarmupLR(
+        optimizer=optimizer,
+        total_steps=args.num_epochs * len(dataloader),
+        warmup_steps=args.warmup_steps
+        if args.warmup_steps is not None
+        else int(args.num_epochs * len(dataloader) * 0.025),
+        eta_min=0.1 * args.lr,
+    )
+
+    # Set booster
+    booster = Booster(plugin=plugin, **booster_kwargs)
+    model, optimizer, _, dataloader, lr_scheduler = booster.boost(
+        model=model,
+        optimizer=optimizer,
+        lr_scheduler=lr_scheduler,
+        dataloader=dataloader,
+    )
+    use_pipeline = isinstance(booster.plugin, MoeHybridParallelPlugin) and booster.plugin.pp_size > 1
+    is_pp_last_stage = use_pipeline and booster.plugin.stage_manager.is_last_stage()
+    coordinator.print_on_master(f"Finish init booster")
+
+    # Load ckpt
+    if args.load_checkpoint is None:
+        load_model(args.model_name, model, booster, optimizer)
+        coordinator.print_on_master(f"Finish load checkpoint")
+    else:
+        load_checkpoint(args.load_checkpoint, booster, model, optimizer, lr_scheduler)
+        coordinator.print_on_master(f"Finish load optimizer")
+
+    # Start finetuning
+    coordinator.print_on_master(f"Start finetuning")
+    for epoch in range(args.num_epoch):
+        model.train()
+        train_dataloader_iter = iter(dataloader)
+        total_len = len(train_dataloader_iter)
+        with tqdm(
+            range(total_len),
+            desc=f"Epoch [{epoch + 1}/{args.num_epoch}]",
+            disable=not coordinator.is_master() if use_pipeline == False else not is_pp_last_stage,
+        ) as pbar:
+            for step in pbar:
+                if use_pipeline:
+                    # Forward pass
+                    outputs = booster.execute_pipeline(
+                        train_dataloader_iter,
+                        model,
+                        lambda x, y: x.loss,
+                        optimizer,
+                        return_loss=True,
+                        return_outputs=True,
+                    )
+                    # Backward and optimize
+                    if is_pp_last_stage:
+                        loss = outputs["loss"]
+                        global_loss = get_global_loss(loss, booster)
+                        if coordinator._local_rank == "0":
+                            pbar.set_postfix({"Loss": global_loss.item()})
+                else:
+                    # Forward pass
+                    data = next(train_dataloader_iter)
+                    data = move_to_cuda(data, torch.cuda.current_device())
+                    outputs = model(**data)
+                    loss = outputs["loss"]
+                    # Backward
+                    booster.backward(loss, optimizer)
+                    pbar.set_postfix({"loss": loss.item()})
+
+                optimizer.step()
+                lr_scheduler.step()
+                optimizer.zero_grad()
+
+                # Apply load balance
+                if (
+                    args.load_balance
+                    and args.load_balance_interval > 0
+                    and (step + 1) % args.load_balance_interval == 0
+                ):
+                    coordinator.print_on_master(f"Apply load balance")
+                    apply_load_balance(model, optimizer)
+                # save ckeckpoint
+                if (step + 1) % args.save_interval == 0:
+                    coordinator.print_on_master(f"Saving model checkpoint to {args.output_path}")
+                    save_checkpoint(
+                        args.output_path,
+                        booster,
+                        model,
+                        optimizer,
+                        lr_scheduler,
+                        epoch,
+                        step,
+                        args.batch_size,
+                        coordinator,
+                    )
+
+        # save checkpoint at the end of each epochs
+        booster.save_model(model, args.output_path, shard=True, size_per_shard=5120)
+        coordinator.print_on_master(f"Saving model checkpoint to {args.output_path}")
+
+    # Finish training
+    coordinator.print_on_master(f"Finish training")
+
+
+if __name__ == "__main__":
+    main()

applications/ColossalMoE/train.sh

+NUM_GPU=8
+MODEL="mistralai/Mixtral-8x7B-v0.1"
+SEQ_LENGTH=2048
+BATCH_SIZE=1
+LR=0.00001
+
+# hybrid
+# torchrun --standalone --nproc_per_node $NUM_GPU \
+colossalai run --nproc_per_node $NUM_GPU --hostfile "hostfile" \
+    train.py \
+    --num_epoch 1 \
+    --model_name $MODEL \
+    --plugin "hybrid" \
+    --batch_size $BATCH_SIZE \
+    --lr $LR \
+    --zero_stage 1 \
+    --pp_size 2 \
+    --dp_size 1 \
+    --ep_size 8 \

applications/ColossalMoE/version.txt

+1.0.0

colossalai/booster/plugin/moe_hybrid_parallel_plugin.py

@@ -181,6 +181,7 @@ class MoeHybridParallelPlugin(HybridParallelPlugin):
         overlap_communication: bool = True,
         use_ep_inside: bool = True,
         custom_policy: Policy = None,
+        checkpoint_io: Optional[MoECheckpintIO] = None,
     ) -> None:
         assert (
             dist.get_world_size() % (tp_size * pp_size) == 0
@@ -200,6 +201,7 @@ class MoeHybridParallelPlugin(HybridParallelPlugin):
         self.enable_flash_attention = enable_flash_attention
         self.enable_jit_fused = enable_jit_fused
         self.enable_sequence_parallelism = enable_sequence_parallelism
+        self.checkpoint_io = checkpoint_io
         # we change pg mesh to (pp, dp, tp) for better moe performance
         self.pg_mesh = ProcessGroupMesh(self.pp_size, self.dp_size, self.tp_size)
 
@@ -323,7 +325,10 @@ class MoeHybridParallelPlugin(HybridParallelPlugin):
         )
 
     def get_checkpoint_io(self) -> MoECheckpintIO:
+        if self.checkpoint_io is None:
             self.checkpoint_io = MoECheckpintIO(self.dp_group, self.pp_group, self.tp_group, self.zero_stage)
+        else:
+            self.checkpoint_io = self.checkpoint_io(self.dp_group, self.pp_group, self.tp_group, self.zero_stage)
         return self.checkpoint_io
 
     def configure(

colossalai/moe/__init__.py

 from .checkpoint import MoECheckpintIO
 from .experts import MLPExperts
-from .layers import SparseMLP
+from .layers import SparseMLP, apply_load_balance
+from .manager import MOE_MANAGER
 from .routers import MoeRouter, Top1Router, Top2Router, TopKRouter
 from .utils import NormalNoiseGenerator, UniformNoiseGenerator
 
@@ -14,4 +15,6 @@ __all__ = [
     "UniformNoiseGenerator",
     "SparseMLP",
     "MoECheckpintIO",
+    "MOE_MANAGER",
+    "apply_load_balance",
 ]

colossalai/moe/checkpoint.py

@@ -224,6 +224,7 @@ class MoECheckpintIO(HybridParallelCheckpointIO):
             size_per_shard (int, optional): Size per shard in MB. Defaults to 1024.
             use_safetensors (bool, optional): Whether to use safe tensors. Defaults to False.
         """
+        torch.cuda.empty_cache()
         if os.path.isfile(checkpoint):
             logging.error(f"Provided path ({checkpoint}) should be a directory, not a file")
             return
@@ -265,6 +266,7 @@ class MoECheckpintIO(HybridParallelCheckpointIO):
                         f"index located at {save_index_file}."
                     )
         dist.barrier()
+        torch.cuda.empty_cache()
 
     # ========================================================
     # Abstract methods for optimizer loading/saving implementation
@@ -332,10 +334,12 @@ class MoECheckpintIO(HybridParallelCheckpointIO):
         assert isinstance(optimizer, OptimizerWrapper), "Please boost the optimizer before loading!"
 
         def _get_param_id_from_optimizer_param(
-            param: torch.Tensor, master_to_working_map: Optional[Dict[int, torch.Tensor]] = None
+            param: torch.Tensor, master_to_working_map: Optional[Dict[int, torch.Tensor]] = None, optimizer=None
         ):
             if master_to_working_map is not None and id(param) in master_to_working_map:
                 working_param = master_to_working_map[id(param)]
+            elif hasattr(optimizer, "moe_master_to_working_map") and id(param) in optimizer.moe_master_to_working_map:
+                working_param = optimizer.moe_master_to_working_map[id(param)]
             else:
                 working_param = param
             return optimizer.param_info["param2id"][id(working_param)]
@@ -347,7 +351,7 @@ class MoECheckpintIO(HybridParallelCheckpointIO):
         master_to_working_map = optimizer.get_master_to_working_map()
         for pg in optimizer.optim.param_groups:
             for param in pg["params"]:
-                param_id = _get_param_id_from_optimizer_param(param, master_to_working_map)
+                param_id = _get_param_id_from_optimizer_param(param, master_to_working_map, optimizer)
                 id_map[param_id] = param
 
         # Read checkpoint index file.
@@ -371,14 +375,10 @@ class MoECheckpintIO(HybridParallelCheckpointIO):
             new_pg = copy.deepcopy(saved_pg)
             new_pg["params"] = old_pg["params"]  # The parameters in the same group shouln't change.
             updated_groups.append(new_pg)
-        # ep extra group
-        if MOE_MANAGER.parallel == "EP":
+        # ep param group
+        if len(optimizer.optim.param_groups) > len(saved_groups):
             new_pg = copy.deepcopy(saved_pg)
-            new_pg["params"] = optimizer.optim.param_groups[-1][
-                "params"
-            ]  # Only keep the parameters kept by current pipeline stage.
-            for param in new_pg["params"]:
-                param.data = param.data.to(torch.float32)
+            new_pg["params"] = optimizer.optim.param_groups[-1]["params"]
             updated_groups.append(new_pg)
         optimizer.optim.__dict__.update({"param_groups": updated_groups})
 
@@ -389,7 +389,7 @@ class MoECheckpintIO(HybridParallelCheckpointIO):
             for param in pg["params"]:
                 if param is None:
                     continue
-                param_id = _get_param_id_from_optimizer_param(param, master_to_working_map)
+                param_id = _get_param_id_from_optimizer_param(param, master_to_working_map, optimizer)
                 if param_id not in weight_map:
                     continue
                 filename = weight_map[param_id]
@@ -400,26 +400,33 @@ class MoECheckpintIO(HybridParallelCheckpointIO):
 
                 file_path = os.path.join(ckpt_root_path, filename)
                 state_dict = load_shard_state_dict(Path(file_path), use_safetensors=False)
-                load_states_into_optimizer(optimizer.optim, state_dict, id_map, strict=True)
-                loaded_file.add(filename)
 
                 # Then shard the loaded optimizer states if using tp/zero.
-        for param, state in optimizer.optim.state.items():
-            device = param.device
+                for pid, state in list(state_dict.items()):
+                    if pid in id_map:
+                        param = id_map[pid]
                         if master_to_working_map is not None and id(param) in master_to_working_map:
                             working_param = master_to_working_map[id(param)]
+                        elif (
+                            hasattr(optimizer, "moe_master_to_working_map")
+                            and id(param) in optimizer.moe_master_to_working_map
+                        ):
+                            working_param = optimizer.moe_master_to_working_map[id(param)]
                         else:
                             working_param = param
                         original_shape = optimizer.param_info["param2shape"][id(working_param)]
                         sharded_state = self.pre_load_optim(
                             state,
-                param,
+                            working_param,
                             current_shape=working_param.shape,
                             original_shape=original_shape,
-                device=device,
+                            device="cpu",
                             inplace=True,
                         )
-            optimizer.optim.state[param] = sharded_state
+                        state_dict[pid] = sharded_state
+
+                load_states_into_optimizer(optimizer.optim, state_dict, id_map, strict=True)
+                loaded_file.add(filename)
 
         sharded_optimizer_loading_epilogue(optimizer.optim)
         if self.verbose and self.coordinator.is_master():
@@ -576,6 +583,8 @@ class MoECheckpintIO(HybridParallelCheckpointIO):
 
             if master_to_working_map is not None and id(param) in master_to_working_map:
                 working_param = master_to_working_map[id(param)]
+            elif hasattr(optimizer, "moe_master_to_working_map") and id(param) in optimizer.moe_master_to_working_map:
+                working_param = optimizer.moe_master_to_working_map[id(param)]
             else:
                 working_param = param
 
@@ -618,6 +627,7 @@ class MoECheckpintIO(HybridParallelCheckpointIO):
             prefix (str): Perfix of file to save
             size_per_shard (int): Max file size of each file shard that store state tensors
         """
+        torch.cuda.empty_cache()
         assert isinstance(optimizer, OptimizerWrapper), "Please boost the optimizer before saving!"
         if os.path.isfile(checkpoint):
             logging.error(f"Provided path ({checkpoint}) should be a directory, not a file")
@@ -723,6 +733,7 @@ class MoECheckpintIO(HybridParallelCheckpointIO):
                         f"You can find where each parameters has been saved in the "
                         f"index located at {final_index_file_path}."
                     )
+        torch.cuda.empty_cache()
 
     def save_unsharded_optimizer(self, optimizer: OptimizerWrapper, checkpoint: str, gather_dtensor: bool):
         """