Merge pull request #5278 from ver217/sync/npu

[sync] sync npu branch with main

colossalai/legacy/zero/gemini/__init__.py

+from .colo_init_context import ColoInitContext, post_process_colo_init_ctx
 from .ophooks import BaseOpHook, register_ophooks_recursively
 from .stateful_tensor import StatefulTensor
 from .stateful_tensor_mgr import StatefulTensorMgr
@@ -11,4 +12,6 @@ __all__ = [
     "AutoTensorPlacementPolicy",
     "register_ophooks_recursively",
     "BaseOpHook",
+    "ColoInitContext",
+    "post_process_colo_init_ctx",
 ]

colossalai/pipeline/p2p.py

@@ -4,23 +4,20 @@
 import io
 import pickle
 import re
-from typing import Any, List, Optional, Union
 from collections import namedtuple
+from typing import Any, Callable, List, Optional, Tuple, Union
 
 import torch
 import torch.distributed as dist
-from dataclasses import dataclass
-from enum import Enum
 from packaging.version import Version
 from torch.distributed import ProcessGroup
 from torch.distributed import distributed_c10d as c10d
+from torch.utils._pytree import tree_flatten, tree_unflatten
 
 from .stage_manager import PipelineStageManager
 
-_unpickler = pickle.Unpickler
 
-
-def _cuda_safe_tensor_to_object(tensor: torch.Tensor, tensor_size: torch.Size) -> object:
+def _cuda_safe_tensor_to_object(tensor: torch.Tensor, tensor_size: torch.Size) -> Any:
     """transform tensor to object with unpickle.
     Info of the device in bytes stream will be modified into current device before unpickling
 
@@ -42,27 +39,13 @@ def _cuda_safe_tensor_to_object(tensor: torch.Tensor, tensor_size: torch.Size) -
         buf = bytes(buf_array)
 
     io_bytes = io.BytesIO(buf)
-    byte_pickler = _unpickler(io_bytes)
+    byte_pickler = pickle.Unpickler(io_bytes)
     unpickle = byte_pickler.load()
 
     return unpickle
 
 
-def check_for_nccl_backend(group):
-    pg = group or c10d._get_default_group()
-    # Gate PG wrapper check on Gloo availability.
-    if c10d._GLOO_AVAILABLE:
-        # It is not expected for PG to be wrapped many times, but support it just
-        # in case
-        while isinstance(pg, c10d._ProcessGroupWrapper):
-            pg = pg.wrapped_pg
-
-    return (
-        c10d.is_nccl_available() and
-        pg.name() == c10d.Backend.NCCL
-    )
-
-
+# NOTE: FIXME: NPU DOES NOT support isend nor irecv, so broadcast is kept for future use
 def _broadcast_object_list(
     object_list: List[Any], src: int, group: ProcessGroup, device: Optional[Union[torch.device, str, int]] = None
 ):
@@ -70,20 +53,18 @@ def _broadcast_object_list(
     The only difference is that object will be move to correct device after unpickled.
     If local_rank = src, then object list will be sent to rank src. Otherwise, object list will
     be updated with data sent from rank src.
-
     Args:
         object_list (List[Any]): list of object to broadcast
         src (int): source rank to broadcast
         dst (int): dst rank to broadcast
         device (:class:`torch.device`): device to do broadcast. current device in default
-
     """
 
     if c10d._rank_not_in_group(group):
         c10d._warn_not_in_group("broadcast_object_list")
         return
 
-    is_nccl_backend = check_for_nccl_backend(group)
+    is_nccl_backend = _check_for_nccl_backend(group)
     current_device = None
 
     if device is not None:
@@ -131,7 +112,7 @@ def _broadcast_object_list(
 
     if my_rank != src:
         for i, obj_size in enumerate(object_sizes_tensor):
-            obj_view = object_tensor[offset: offset + obj_size]
+            obj_view = object_tensor[offset : offset + obj_size]
             obj_view = obj_view.type(torch.uint8)
             if obj_view.device != torch.device("cpu"):
                 obj_view = obj_view.cpu()
@@ -149,80 +130,107 @@ def _broadcast_object_list(
             object_list[i] = unpickle_object
 
 
-def check_device(group):
-    is_nccl_backend = check_for_nccl_backend(group)
-    current_device = None
+def _check_for_nccl_backend(group):
+    pg = group or c10d._get_default_group()
+    # Gate PG wrapper check on Gloo availability.
+    if c10d._GLOO_AVAILABLE:
+        # It is not expected for PG to be wrapped many times, but support it just in case
+        while isinstance(pg, c10d._ProcessGroupWrapper):
+            pg = pg.wrapped_pg
+
+    return c10d.is_nccl_available() and pg.name() == c10d.Backend.NCCL
 
+
+def _check_device(group):
+    is_nccl_backend = _check_for_nccl_backend(group)
     current_device = torch.device("cpu")
     if is_nccl_backend:
         current_device = torch.device("cuda", torch.cuda.current_device())
     return current_device, is_nccl_backend
 
 
-TensorMetadata = namedtuple('TensorMetadata', ['key', 'shape', 'dtype', 'requires_grad'])
-
+TensorMetadata = namedtuple("TensorMetadata", ["shape", "dtype", "requires_grad"])
+P2PMetadata = namedtuple("P2PMetadata", ["tree_spec", "tensor_metadata", "non_tensor_obj_idx", "non_tensor_objs"])
 
-class P2PDataType(Enum):
-    serialization = 0
-    tensor = 1
-    list = 2
-    dict = 3
 
+def create_send_metadata(
+    object: Any, strict: bool = True, return_tensor: bool = False
+) -> Union[P2PMetadata, Tuple[P2PMetadata, List[torch.Tensor]]]:
+    """
+    Args:
+        object (Any): object needed to be sent
+        strict (bool, optional): whether to check if the object is supported for fast send
+        return_tensor (bool, optional): whether to return tensor objects
+    """
+    objs, tree_spec = tree_flatten(object)
+    tensor_metadata, tensor_objs = [], []
+    non_tensor_obj_idx, non_tensor_objs = [], []
+    for idx, obj in enumerate(objs):
+        if isinstance(obj, torch.Tensor):
+            tensor_objs.append(obj)
+            tensor_metadata.append(TensorMetadata(obj.shape, obj.dtype, obj.requires_grad))
+        else:
+            non_tensor_obj_idx.append(idx)
+            non_tensor_objs.append(obj)
 
-@dataclass
-class P2PMetadata:
-    data_type: P2PDataType
-    content: Union[List[TensorMetadata], TensorMetadata, Any]
+    assert not strict or len(non_tensor_objs) == 0, "Only support tensor for fast send"
+    metadata = P2PMetadata(tree_spec, tensor_metadata, non_tensor_obj_idx, non_tensor_objs)
+    return metadata if not return_tensor else (metadata, tensor_objs)
 
 
-def filling_ops_queue(obj, comm_op, comm_rank, ops_queue, group):
+def _filling_ops_queue(
+    obj: Union[torch.Tensor, List[torch.Tensor]],
+    comm_op: Callable,
+    comm_rank: int,
+    ops_queue: List,
+    group: ProcessGroup,
+):
     if isinstance(obj, torch.Tensor):
         obj = obj.contiguous()
         op_to_add = dist.P2POp(comm_op, obj, comm_rank, group)
         ops_queue.append(op_to_add)
     else:
         for tensor_to_comm in obj:
-            tensor_to_comm = tensor_to_comm.contiguous()
-            op_to_add = dist.P2POp(comm_op, tensor_to_comm, comm_rank, group)
-            ops_queue.append(op_to_add)
-
-
-def create_recv_buffer(p2p_metadata: P2PMetadata, current_device):
-    if p2p_metadata.data_type == P2PDataType.tensor:
-        metadata = p2p_metadata.content
-        tensor_recv = torch.empty(metadata.shape, requires_grad=metadata.requires_grad, device=current_device, dtype=metadata.dtype)
-        return tensor_recv
-    elif p2p_metadata.data_type in (P2PDataType.list, P2PDataType.dict):
-        buffer_recv = []
-        for metadata in p2p_metadata.content:
-            tensor_recv = torch.empty(metadata.shape, requires_grad=metadata.requires_grad, device=current_device, dtype=metadata.dtype)
-            buffer_recv.append(tensor_recv)
-        return buffer_recv
-    else:
-        raise ValueError(f"Unknown data_type: {p2p_metadata.data_type}")
+            assert isinstance(tensor_to_comm, torch.Tensor)
+            _filling_ops_queue(tensor_to_comm, comm_op, comm_rank, ops_queue, group)
+
 
+def _create_recv_buffer(tensor_metadata: List[TensorMetadata], current_device) -> List[torch.Tensor]:
+    buffer_recv = []
+    for metadata in tensor_metadata:
+        tensor_recv = torch.empty(
+            metadata.shape, requires_grad=metadata.requires_grad, device=current_device, dtype=metadata.dtype
+        )
+        buffer_recv.append(tensor_recv)
+    return buffer_recv
 
-def _batch_send_recv_tensor(send_tensor_list, recv_tensor_metadata, send_dst, recv_src, send_group, recv_group, current_device):
+
+def _batch_send_recv_tensor(
+    send_tensor_list: Optional[List[torch.Tensor]],
+    recv_tensor_metadata: Optional[List[TensorMetadata]],
+    send_dst: Optional[int],
+    recv_src: Optional[int],
+    send_group: Optional[ProcessGroup],
+    recv_group: Optional[ProcessGroup],
+    current_device: Any,
+) -> Optional[Union[torch.Tensor, List[torch.Tensor]]]:
     buffer_recv = None
     if recv_tensor_metadata is not None:
-        buffer_recv = create_recv_buffer(recv_tensor_metadata, current_device)
+        buffer_recv = _create_recv_buffer(recv_tensor_metadata, current_device)
 
     ops = []
-
-    if send_dst is not None:
-        filling_ops_queue(send_tensor_list, dist.isend, send_dst, ops, send_group)
-
-    if recv_src is not None:
-        assert buffer_recv is not None
-        filling_ops_queue(buffer_recv, dist.irecv, recv_src, ops, recv_group)
+    if send_dst is not None and send_tensor_list is not None:
+        assert send_group is not None
+        _filling_ops_queue(send_tensor_list, dist.isend, send_dst, ops, send_group)
+    if recv_src is not None and buffer_recv is not None:
+        assert recv_group is not None
+        _filling_ops_queue(buffer_recv, dist.irecv, recv_src, ops, recv_group)
 
     if len(ops) > 0:
         reqs = dist.batch_isend_irecv(ops)
         for req in reqs:
             req.wait()
 
-    torch.cuda.synchronize()
-
     # Remove synchronization according to Pytorch's documentation
     # However, the Megatron-LM does synchronization here
     # https://github.com/microsoft/Megatron-DeepSpeed/blob/ef13d099c2a1609225a4ce4c1a1753cc76dd90a1/megatron/p2p_communication.py#L111-L112
@@ -233,12 +241,16 @@ def _batch_send_recv_tensor(send_tensor_list, recv_tensor_metadata, send_dst, re
 
 
 def _send_recv_serialization_object(
-        object: Any,
-        send_dst: Optional[int], recv_src: Optional[int],
-        send_group: Optional[ProcessGroup], recv_group: Optional[ProcessGroup],
-        current_device,
-        is_nccl_backend):
+    object: Optional[P2PMetadata],
+    send_dst: Optional[int],
+    recv_src: Optional[int],
+    send_group: Optional[ProcessGroup],
+    recv_group: Optional[ProcessGroup],
+    current_device: Any,
+    is_nccl_backend: bool,
+) -> Optional[P2PMetadata]:
     ops = []
+
     send_object_tensor = None
     if object is not None and send_dst is not None:
         if Version(torch.__version__) >= Version("1.13.0"):
@@ -250,44 +262,40 @@ def _send_recv_serialization_object(
             send_object_size_tensor = send_object_size_tensor.to(current_device)
             send_object_tensor = send_object_tensor.to(current_device)
 
-        filling_ops_queue(send_object_size_tensor, dist.isend, send_dst, ops, send_group)
+        _filling_ops_queue(send_object_size_tensor, dist.isend, send_dst, ops, send_group)
 
     recv_object_size_tensor = None
     if recv_src is not None:
         recv_object_size_tensor = torch.empty(1, dtype=torch.long)
         if is_nccl_backend:
             recv_object_size_tensor = recv_object_size_tensor.to(current_device)
-        filling_ops_queue(recv_object_size_tensor, dist.irecv, recv_src, ops, recv_group)
+        _filling_ops_queue(recv_object_size_tensor, dist.irecv, recv_src, ops, recv_group)
 
     if len(ops) > 0:
         reqs = dist.batch_isend_irecv(ops)
         for req in reqs:
             req.wait()
 
-    torch.cuda.synchronize()
-
     # See the comment in `_batch_send_recv_tensor`
     # torch.cuda.synchronize()
 
     ops = []
 
     if send_dst is not None and send_object_tensor is not None:
-        filling_ops_queue(send_object_tensor, dist.isend, send_dst, ops, send_group)
+        _filling_ops_queue(send_object_tensor, dist.isend, send_dst, ops, send_group)
 
     recv_object_tensor = None
     if recv_src is not None and recv_object_size_tensor is not None:
         recv_object_tensor = torch.empty(recv_object_size_tensor.item(), dtype=torch.uint8)
         if is_nccl_backend:
             recv_object_tensor = recv_object_tensor.to(current_device)
-        filling_ops_queue(recv_object_tensor, dist.irecv, recv_src, ops, recv_group)
+        _filling_ops_queue(recv_object_tensor, dist.irecv, recv_src, ops, recv_group)
 
     if len(ops) > 0:
         reqs = dist.batch_isend_irecv(ops)
         for req in reqs:
             req.wait()
 
-    torch.cuda.synchronize()
-
     # See the comment in `_batch_send_recv_tensor`
     # torch.cuda.synchronize()
 
@@ -296,112 +304,119 @@ def _send_recv_serialization_object(
         if recv_object_tensor.device != torch.device("cpu"):
             recv_object_tensor = recv_object_tensor.cpu()
 
-        unpickle_object = _cuda_safe_tensor_to_object(
-            recv_object_tensor, recv_object_size_tensor.item())
+        unpickle_object = _cuda_safe_tensor_to_object(recv_object_tensor, recv_object_size_tensor.item())
 
-        if (
-            isinstance(unpickle_object, torch.Tensor)
-            and unpickle_object.device.index != torch.cuda.current_device()
-        ):
+        if isinstance(unpickle_object, torch.Tensor) and unpickle_object.device.index != torch.cuda.current_device():
             unpickle_object = unpickle_object.cuda()
 
         return unpickle_object
 
 
-def _check_if_fast_send_available(object):
-    if type(object) is torch.Tensor:
-        return True
-    elif type(object) is list:
-        is_list_of_tensor = all([type(v) is torch.Tensor for v in object])
-        return is_list_of_tensor
-    elif type(object) is dict:
-        is_dict_of_tensor = all([type(k) is str and type(
-            v) is torch.Tensor for k, v in object.items()])
-
-        return is_dict_of_tensor
-    return False
-
-
 def _communicate(
-    object,
+    object: Any,
     send_dst: Optional[int],
     recv_src: Optional[int],
     send_group: Optional[ProcessGroup] = None,
     recv_group: Optional[ProcessGroup] = None,
+    send_metadata: bool = True,
+    metadata_recv: Optional[P2PMetadata] = None,
+    send_prior_fallback: Optional[bool] = None,
 ) -> Any:
-    if c10d._rank_not_in_group(send_group) or c10d._rank_not_in_group(recv_group):
-        c10d._warn_not_in_group("_communicate")
-        return
+    """
+    Send and receive object from send_dst and recv_src respectively
 
-    current_send_device, is_send_nccl_backend = check_device(send_group)
-    current_recv_device, is_recv_nccl_backend = check_device(recv_group)
+    Args:
+        object (Any): object needed to be sent
+        send_dst (int): rank of the destination
+        recv_src (int): rank of the source
+        send_group (ProcessGroup, optional): process group of sender
+        recv_group (ProcessGroup, optional): process group of receiver
+        send_metadata (bool, optional): whether to send metadata
+        metadata_recv (P2PMetadata, optional): metadata of the object to be received
+    """
+    assert send_dst is not None or recv_src is not None, "send_dst and recv_src cannot be both None"
+    assert send_dst is None or send_group is not None, "send_group must be specified when send_dst is not None"
+    assert recv_src is None or recv_group is not None, "recv_group must be specified when recv_src is not None"
+    assert (
+        metadata_recv is None or len(metadata_recv.non_tensor_obj_idx) == 0
+    ), "metadata_recv should not contain non-tensor objects"
+
+    metadata_send, tensor_objs = None, None
+    if object is not None:
+        # NOTE: if object contains non-tensor objects, we have to send metadata
+        metadata_send, tensor_objs = create_send_metadata(object, strict=False, return_tensor=True)
+        send_metadata = send_metadata or len(metadata_send.non_tensor_obj_idx) > 0
+
+    # NOTE: send & recv should be atomic operations. However, if we need to send metadata or receive metadata,
+    #   we are not able to do that (1. send & recv metadata 2. send & recv). So we need to split the send & recv into two parts in this case.
+    if (send_dst is not None and recv_src is not None) and (send_metadata or metadata_recv is None):
+        assert send_prior_fallback is not None, "Priority must be set if fallback happens"
+        if send_prior_fallback:
+            _communicate(object, send_dst=send_dst, recv_src=None, send_group=send_group, send_metadata=send_metadata)
+            return _communicate(
+                None, send_dst=None, recv_src=recv_src, recv_group=recv_group, metadata_recv=metadata_recv
+            )
+        else:
+            recv_data = _communicate(
+                None, send_dst=None, recv_src=recv_src, recv_group=recv_group, metadata_recv=metadata_recv
+            )
+            _communicate(object, send_dst=send_dst, recv_src=None, send_group=send_group, send_metadata=send_metadata)
+            return recv_data
+
+    # NOTE: only the following 5 cases are valid:
+    #   1. send() [needs extra metadata] and no recv()
+    #   2. recv() [needs extra metadata] and no send()
+    #   3. neither send() nor recv() need extra metadata
+    assert not (send_dst is not None and send_metadata) or recv_src is None
+    assert not (recv_src is not None and metadata_recv is None) or send_dst is None
+    assert not (send_dst is not None and recv_src is not None) or (not send_metadata and metadata_recv is not None)
+    assert not c10d._rank_not_in_group(send_group) and not c10d._rank_not_in_group(recv_group)
+
+    current_send_device, is_send_nccl_backend = _check_device(send_group)
+    current_recv_device, is_recv_nccl_backend = _check_device(recv_group)
 
     is_nccl_backend = is_send_nccl_backend and is_recv_nccl_backend
 
     assert current_send_device == current_recv_device
     current_device = current_send_device
 
-    assert (send_dst is not None) or (recv_src is not None)
+    if (send_dst is not None and send_metadata) or (recv_src is not None and metadata_recv is None):
+        # Send and receive metadata
+        _metadata_recv = _send_recv_serialization_object(
+            object=metadata_send,
+            send_dst=send_dst if send_metadata else None,
+            recv_src=recv_src if metadata_recv is None else None,
+            send_group=send_group if send_metadata else None,
+            recv_group=recv_group if metadata_recv is None else None,
+            current_device=current_device,
+            is_nccl_backend=is_nccl_backend,
+        )
+        assert metadata_recv is None or _metadata_recv is None
+        metadata_recv = _metadata_recv if metadata_recv is None else metadata_recv
 
-    can_fast_send = False
-    send_metadata = None
-    if send_dst is not None:
-        can_fast_send = _check_if_fast_send_available(object) and is_nccl_backend
-        if not can_fast_send:
-            send_metadata = P2PMetadata(P2PDataType.serialization, object)
-        else:
-            if type(object) is torch.Tensor:
-                data_type = P2PDataType.tensor
-                content = TensorMetadata(None, object.shape, object.dtype, object.requires_grad)
-            elif type(object) is list:
-                data_type = P2PDataType.list
-                content = []
-                for v in object:
-                    content.append(TensorMetadata(None, v.shape, v.dtype, v.requires_grad))
-            elif type(object) is dict:
-                data_type = P2PDataType.dict
-                content = []
-                for k, v in object.items():
-                    content.append(TensorMetadata(k, v.shape, v.dtype, v.requires_grad))
-            else:
-                raise ValueError('Cannot send object of type {}'.format(type(object)))
-            send_metadata = P2PMetadata(data_type, content)
-
-    recv_metadata = _send_recv_serialization_object(send_metadata, send_dst, recv_src, send_group, recv_group, current_device, is_nccl_backend)
-    if recv_metadata is not None:
-        assert type(recv_metadata) is P2PMetadata
-        if recv_metadata.data_type == P2PDataType.serialization:
-            return recv_metadata.content
-    if not can_fast_send and send_dst is not None:
-        return
+    # Send and receive data
+    recv_tensor_metadata = None if metadata_recv is None else metadata_recv.tensor_metadata
+    recv_tensor_objs = _batch_send_recv_tensor(
+        tensor_objs, recv_tensor_metadata, send_dst, recv_src, send_group, recv_group, current_device
+    )
 
-    send_tensor_list = None
-    if type(object) is torch.Tensor:
-        send_tensor_list = object
-    elif type(object) is list:
-        send_tensor_list = object
-    elif type(object) is dict:
-        send_tensor_list = list(object.values())
-
-    recv_buffer = _batch_send_recv_tensor(send_tensor_list, recv_metadata, send_dst, recv_src, send_group, recv_group, current_device)
-
-    if recv_metadata is not None:
-        assert recv_buffer is not None
-        if recv_metadata.data_type in [P2PDataType.tensor, P2PDataType.list]:
-            return recv_buffer
-        elif recv_metadata.data_type == P2PDataType.dict:
-            return {
-                k: v
-                for k, v in zip(
-                    [m.key for m in recv_metadata.content],
-                    recv_buffer,
-                )
-            }
-        else:
-            raise ValueError('Unknown data type {}'.format(recv_metadata.data_type))
+    if metadata_recv is not None:
+        assert isinstance(metadata_recv, P2PMetadata)
+        tree_spec = metadata_recv.tree_spec
+        non_tensor_obj_idx = metadata_recv.non_tensor_obj_idx
+        non_tensor_objs = metadata_recv.non_tensor_objs
 
+        if recv_tensor_objs is None:
+            recv_tensor_objs = []
 
-def _send_object(object: Any, src: int, dst: int, group: ProcessGroup) -> None:
+        for idx in non_tensor_obj_idx:
+            recv_tensor_objs.insert(idx, non_tensor_objs.pop(0))
+        recv_object = tree_unflatten(recv_tensor_objs, tree_spec)
+
+        return recv_object
+
+
+def _send_object(object: Any, src: int, dst: int, group: ProcessGroup, **kwargs) -> None:
     """send anything to dst rank
 
     Args:
@@ -411,10 +426,10 @@ def _send_object(object: Any, src: int, dst: int, group: ProcessGroup) -> None:
     Returns:
         None
     """
-    _communicate(object, send_dst=dst, recv_src=None, send_group=group)
+    _communicate(object, send_dst=dst, recv_src=None, send_group=group, **kwargs)
 
 
-def _recv_object(src: int, dst: int, group: ProcessGroup) -> Any:
+def _recv_object(src: int, dst: int, group: ProcessGroup, **kwargs) -> Any:
     """recv anything from src
 
     Args:
@@ -423,7 +438,7 @@ def _recv_object(src: int, dst: int, group: ProcessGroup) -> Any:
     Returns:
         Any: Object received from src.
     """
-    return _communicate(None, send_dst=None, recv_src=src, recv_group=group)
+    return _communicate(None, send_dst=None, recv_src=src, recv_group=group, **kwargs)
 
 
 def _p2p_comm(
@@ -436,7 +451,7 @@ def _p2p_comm(
     """
     Send and recv tensor using P2P communication, used when pipeline size is 2 to solve the race communication.
 
-    Agrs:
+    Args:
         tensor_send_next (torch.Tensor): tensor to be sent to next stage
         recv_prev (bool): whether to receive tensor from previous stage
         peer (int): rank of the peer
@@ -467,7 +482,6 @@ def _p2p_comm(
             group=group,
         )
         ops.append(recv_prev_op)
-
     if len(ops) > 0:
         reqs = dist.batch_isend_irecv(ops)
         for req in reqs:
@@ -490,7 +504,6 @@ def _p2p_comm(
             group=group,
         )
         ops.append(send_next_op)
-
     if tensor_recv_prev is not None:
         recv_prev_op = dist.P2POp(
             dist.irecv,
@@ -510,7 +523,7 @@ class PipelineP2PCommunication:
     def __init__(self, stage_manager: PipelineStageManager) -> None:
         self.stage_manager = stage_manager
 
-    def recv_forward(self, prev_rank: int = None) -> Any:
+    def recv_forward(self, prev_rank: Optional[int] = None, metadata_recv: Optional[P2PMetadata] = None) -> Any:
         """Copy the forward output from the previous stage in pipeline as the input tensor of this stage.
 
         Args:
@@ -522,11 +535,16 @@ class PipelineP2PCommunication:
         if prev_rank is None:
             prev_rank = self.stage_manager.get_prev_rank()
         cur_rank = self.stage_manager.get_rank()
-        input_tensor = _recv_object(prev_rank, cur_rank, self.stage_manager.get_p2p_process_group(prev_rank, cur_rank))
+        input_tensor = _recv_object(
+            prev_rank,
+            cur_rank,
+            self.stage_manager.get_p2p_process_group(prev_rank, cur_rank),
+            metadata_recv=metadata_recv,
+        )
 
         return input_tensor
 
-    def recv_backward(self, next_rank: int = None) -> Any:
+    def recv_backward(self, next_rank: Optional[int] = None, metadata_recv: Optional[P2PMetadata] = None) -> Any:
         """Copy the gradient tensor from the next stage in pipeline as the input gradient of this stage.
 
         Args:
@@ -539,12 +557,15 @@ class PipelineP2PCommunication:
             next_rank = self.stage_manager.get_next_rank()
         cur_rank = self.stage_manager.get_rank()
         output_tensor_grad = _recv_object(
-            next_rank, cur_rank, self.stage_manager.get_p2p_process_group(next_rank, cur_rank)
+            next_rank,
+            cur_rank,
+            self.stage_manager.get_p2p_process_group(next_rank, cur_rank),
+            metadata_recv=metadata_recv,
         )
 
         return output_tensor_grad
 
-    def send_forward(self, output_object: Any, next_rank: int = None) -> None:
+    def send_forward(self, output_object: Any, next_rank: Optional[int] = None, send_metadata: bool = True) -> None:
         """Sends the input tensor to the next stage in pipeline.
 
         Args:
@@ -554,9 +575,15 @@ class PipelineP2PCommunication:
         if next_rank is None:
             next_rank = self.stage_manager.get_next_rank()
         cur_rank = self.stage_manager.get_rank()
-        _send_object(output_object, cur_rank, next_rank, self.stage_manager.get_p2p_process_group(cur_rank, next_rank))
+        _send_object(
+            output_object,
+            cur_rank,
+            next_rank,
+            self.stage_manager.get_p2p_process_group(cur_rank, next_rank),
+            send_metadata=send_metadata,
+        )
 
-    def send_backward(self, input_object: Any, prev_rank: int = None) -> None:
+    def send_backward(self, input_object: Any, prev_rank: Optional[int] = None, send_metadata: bool = True) -> None:
         """Sends the gradient tensor to the previous stage in pipeline.
 
         Args:
@@ -566,9 +593,22 @@ class PipelineP2PCommunication:
         if prev_rank is None:
             prev_rank = self.stage_manager.get_prev_rank()
         cur_rank = self.stage_manager.get_rank()
-        _send_object(input_object, cur_rank, prev_rank, self.stage_manager.get_p2p_process_group(cur_rank, prev_rank))
+        _send_object(
+            input_object,
+            cur_rank,
+            prev_rank,
+            self.stage_manager.get_p2p_process_group(cur_rank, prev_rank),
+            send_metadata=send_metadata,
+        )
 
-    def send_forward_recv_backward(self, input_object: Any, next_rank: int = None) -> Any:
+    def send_forward_recv_backward(
+        self,
+        input_object: Any,
+        next_rank: Optional[int] = None,
+        send_metadata: bool = True,
+        metadata_recv: Optional[P2PMetadata] = None,
+        send_prior_fallback: Optional[bool] = None,
+    ) -> Any:
         """Sends the gradient tensor to and copy the gradient tensor from the next stage in pipeline
 
         Args:
@@ -581,11 +621,24 @@ class PipelineP2PCommunication:
         cur_rank = self.stage_manager.get_rank()
         group = self.stage_manager.get_p2p_process_group(cur_rank, next_rank)
         return _communicate(
-            input_object, next_rank, next_rank,
-            send_group=group, recv_group=group,
+            input_object,
+            next_rank,
+            next_rank,
+            send_group=group,
+            recv_group=group,
+            send_metadata=send_metadata,
+            metadata_recv=metadata_recv,
+            send_prior_fallback=send_prior_fallback,
         )
 
-    def send_backward_recv_forward(self, input_object: Any, prev_rank: int = None) -> Any:
+    def send_backward_recv_forward(
+        self,
+        input_object: Any,
+        prev_rank: Optional[int] = None,
+        send_metadata: bool = True,
+        metadata_recv: Optional[P2PMetadata] = None,
+        send_prior_fallback: Optional[bool] = None,
+    ) -> Any:
         """Sends the gradient tensor to and copy the gradient tensor from the previous stage in pipeline
 
         Args:
@@ -597,37 +650,23 @@ class PipelineP2PCommunication:
 
         cur_rank = self.stage_manager.get_rank()
         group = self.stage_manager.get_p2p_process_group(prev_rank, cur_rank)
-        return _communicate(
-            input_object, prev_rank, prev_rank,
-            send_group=group, recv_group=group,
-        )
-
-    def send_forward_recv_forward(self, input_object: Any, prev_rank: int = None, next_rank: int = None) -> Any:
-        """Sends the gradient tensor to the previous stage and copy the input tensor from the previous stage in pipeline.
-
-        Args:
-            input_object (Any): Object to be sent.
-            prev_rank (int, optional): The rank of the sender of the tensor
-            next_rank (int, optional): The rank of the recipient of the tensor
-        """
-        if prev_rank is None:
-            prev_rank = self.stage_manager.get_prev_rank()
-        if next_rank is None:
-            next_rank = self.stage_manager.get_next_rank()
-
-        cur_rank = self.stage_manager.get_rank()
-        recv_group = self.stage_manager.get_p2p_process_group(prev_rank, cur_rank)
-        send_group = self.stage_manager.get_p2p_process_group(cur_rank, next_rank)
         return _communicate(
             input_object,
-            send_dst=next_rank,
-            recv_src=prev_rank,
-            send_group=send_group,
-            recv_group=recv_group,
+            prev_rank,
+            prev_rank,
+            send_group=group,
+            recv_group=group,
+            send_metadata=send_metadata,
+            metadata_recv=metadata_recv,
+            send_prior_fallback=send_prior_fallback,
         )
 
     def p2p_communicate(
-        self, output_object: Any, recv_pre: bool, peer: int = None, comm_dtype: torch.dtype = torch.float16
+        self,
+        output_object: Any,
+        recv_pre: bool,
+        next_rank: Optional[int] = None,
+        comm_dtype: torch.dtype = torch.float16,
     ) -> None:
         """
         Sends the input tensor to the next stage in pipeline, using `P2Pop` in torch.
@@ -636,10 +675,14 @@ class PipelineP2PCommunication:
             output_object (Any): Object to be sent.
             next_rank (int, optional): The rank of the recipient of the tensor.
         """
-        if peer is None:
-            peer = self.stage_manager.get_next_rank()
+        if next_rank is None:
+            next_rank = self.stage_manager.get_next_rank()
         cur_rank = self.stage_manager.get_rank()
         recv_tensor = _p2p_comm(
-            output_object, recv_pre, peer, self.stage_manager.get_p2p_process_group(cur_rank, peer), comm_dtype
+            output_object,
+            recv_pre,
+            next_rank,
+            self.stage_manager.get_p2p_process_group(cur_rank, next_rank),
+            comm_dtype,
         )
         return recv_tensor

colossalai/pipeline/schedule/interleaved_pp.py

 from functools import partial
-from typing import Any, Callable, Iterable, List, Optional, Union
+from typing import Any, Callable, Dict, Iterable, List, Optional, Union
 
 import torch
 import torch.cuda
-from torch.nn import Module
+from torch.nn import Module, ModuleList
 from torch.utils._pytree import tree_map
 
 from colossalai.accelerator import get_accelerator
 from colossalai.interface import OptimizerWrapper
-from colossalai.pipeline.p2p import PipelineP2PCommunication
+from colossalai.pipeline.p2p import PipelineP2PCommunication, create_send_metadata
 from colossalai.pipeline.stage_manager import PipelineStageManager
 
 from ._utils import detach, get_batch_size, get_micro_batch, merge_batch, model_forward, retain_grad, to_device
@@ -16,18 +16,35 @@ from .base import PipelineSchedule
 
 
 class InterleavedSchedule(PipelineSchedule):
-    def __init__(self, num_microbatches: int, num_model_chunks: int, stage_manager: PipelineStageManager) -> None:
-        self.num_model_chunks = num_model_chunks
-        assert (
-            num_microbatches % self.num_model_chunks == 0
-        ), "Number of microbatches should be an integer multiple of number of model chunks"
+    def __init__(
+        self,
+        stage_manager: PipelineStageManager,
+        num_model_chunks: int,
+        num_microbatch: Optional[int] = None,
+        microbatch_size: Optional[int] = None,
+        enable_metadata_cache: bool = True,
+    ) -> None:
         super().__init__(stage_manager)
+        assert (
+            num_microbatch is not None or microbatch_size is not None
+        ), "Either num_microbatch or microbatch_size should be provided"
+
         self.comm = PipelineP2PCommunication(stage_manager)
-        self.num_microbatches = num_microbatches
-        self.batch: Optional[Any] = None
-        self.batch_size: Optional[int] = None
-        self.microbatch_offset: Optional[int] = None
-        self.microbatch_size: Optional[int] = None
+        self.num_microbatch = num_microbatch
+        self.microbatch_size = microbatch_size
+        self.num_model_chunks = num_model_chunks
+
+        self.batch: Any
+        self.batch_size: int
+        self.last_batch_size: Optional[int] = None
+        self.microbatch_offset: List[int]
+
+        # P2PMeta cache
+        self.enable_metadata_cache = enable_metadata_cache
+        self.send_tensor_metadata = True
+        self.send_grad_metadata = True
+        self.tensor_metadata_recv = None
+        self.grad_metadata_recv = None
 
     def load_batch(self, data_iter: Iterable, device: Optional[torch.device] = None) -> None:
         """Load a batch from data iterator.
@@ -39,11 +56,37 @@ class InterleavedSchedule(PipelineSchedule):
         batch = next(data_iter)
         if device is not None:
             batch = tree_map(partial(to_device, device=device), batch)
+
+        self.microbatch_offset = [0 for _ in range(self.num_model_chunks)]
         self.batch = batch
         self.batch_size = get_batch_size(batch)
-        self.microbatch_offset = [0 for _ in range(self.num_model_chunks)]
-        assert self.batch_size % self.num_microbatches == 0, "Batch size should divided by the number of microbatches"
-        self.microbatch_size = self.batch_size // self.num_microbatches
+
+        if self.microbatch_size is None:
+            assert self.batch_size % self.num_microbatch == 0, "Batch size should divided by the number of microbatch"
+            self.microbatch_size = self.batch_size // self.num_microbatch
+        if self.num_microbatch is None:
+            assert self.batch_size % self.microbatch_size == 0, "Batch size should divided by the microbatch size"
+            self.num_microbatch = self.batch_size // self.microbatch_size
+
+        if not self.forward_only:
+            assert self.last_batch_size is None or self.last_batch_size == self.batch_size
+            assert self.batch_size == self.microbatch_size * self.num_microbatch
+
+            assert (
+                self.num_microbatch % self.stage_manager.num_stages == 0
+            ), "Number of microbatch should be an integer multiple of number of pipeline parallel devices"
+
+        if self.forward_only:
+            self.num_microbatch = (self.batch_size - 1) // self.microbatch_size + 1
+            # NOTE: disable metadata cache when batch size changes (not valid anymore)
+            if self.batch_size != self.last_batch_size:
+                self.enable_metadata_cache = False
+                self.send_tensor_metadata = True
+                self.send_grad_metadata = True
+                self.tensor_metadata_recv = None
+                self.grad_metadata_recv = None
+
+        self.last_batch_size = self.batch_size
 
     def load_micro_batch(self, model_chunk_id: int) -> Any:
         """Load a micro batch from the current batch.
@@ -54,11 +97,12 @@ class InterleavedSchedule(PipelineSchedule):
         Returns:
             Any: Micro batch.
         """
+        assert self.microbatch_offset[model_chunk_id] <= self.batch_size, "Microbatches exhausted"
         micro_batch = get_micro_batch(self.batch, self.microbatch_offset[model_chunk_id], self.microbatch_size)
         self.microbatch_offset[model_chunk_id] += self.microbatch_size
         return tree_map(partial(to_device, device=get_accelerator().get_current_device()), micro_batch)
 
-    def get_model_chunk_id(self, microbatch_id: int, forward: bool) -> int:
+    def get_model_chunk_id(self, microbatch_id: int, is_forward: bool) -> int:
         """Helper method to get the model chunk ID given the iteration number.
 
         Args:
@@ -68,38 +112,13 @@ class InterleavedSchedule(PipelineSchedule):
         Returns:
             int: The model chunk idx of the input microbatch_id
         """
-        microbatch_id_in_group = (microbatch_id) % (self.stage_manager.num_stages * self.num_model_chunks)
+        assert microbatch_id < self.num_microbatch * self.num_model_chunks
+        microbatch_id_in_group = microbatch_id % (self.stage_manager.num_stages * self.num_model_chunks)
         model_chunk_id = microbatch_id_in_group // self.stage_manager.num_stages
-        if not forward:
+        if not is_forward:
             model_chunk_id = self.num_model_chunks - model_chunk_id - 1
         return model_chunk_id
 
-    def is_first_stage(self, model_chunk_id: int) -> bool:
-        """Is the current virtual stage the first stage
-
-        Args:
-            model_chunk_id (int): The current model chunk idx.
-
-        Returns:
-            bool: Whether the current virtual stage is the first stage.
-        """
-        if self.stage_manager.is_first_stage() and model_chunk_id == 0:
-            return True
-        return False
-
-    def is_last_stage(self, model_chunk_id: int) -> bool:
-        """Is the current virtual stage the last stage
-
-        Args:
-            model_chunk_id (int): The current model chunk idx.
-
-        Returns:
-            bool: Whether the current virtual stage is the last stage.
-        """
-        if self.stage_manager.is_last_stage() and model_chunk_id == self.num_model_chunks - 1:
-            return True
-        return False
-
     def recv_forward(self, model_chunk_id: int, prev_rank: int = None) -> Any:
         """Copy the forward output from the previous stage in pipeline as the input tensor of this stage.
            For interleaved 1F1B.
@@ -111,12 +130,13 @@ class InterleavedSchedule(PipelineSchedule):
         Returns:
             Any: The input tensor or input tensor list.
         """
-        if self.is_first_stage(model_chunk_id):
-            input_tensor = None
-        else:
-            input_tensor = self.comm.recv_forward(prev_rank)
+        with self.stage_manager.switch_model_chunk_id(model_chunk_id):
+            if not self.stage_manager.is_first_stage():
+                input_tensor = self.comm.recv_forward(prev_rank, metadata_recv=self.tensor_metadata_recv)
+                if self.enable_metadata_cache and self.tensor_metadata_recv is None:
+                    self.tensor_metadata_recv = create_send_metadata(input_tensor)
 
-        return input_tensor
+                return input_tensor
 
     def recv_backward(self, model_chunk_id: int, next_rank: int = None) -> Any:
         """Copy the gradient tensor from the next stage in pipeline as the input gradient of this stage.
@@ -129,14 +149,15 @@ class InterleavedSchedule(PipelineSchedule):
         Returns:
             Any: The input gradient tensor or gradient tensor list.
         """
-        if self.is_last_stage(model_chunk_id):
-            output_tensor_grad = None
-        else:
-            output_tensor_grad = self.comm.recv_backward(next_rank)
+        with self.stage_manager.switch_model_chunk_id(model_chunk_id):
+            if not self.stage_manager.is_last_stage():
+                output_tensor_grad = self.comm.recv_backward(next_rank, metadata_recv=self.grad_metadata_recv)
+                if self.enable_metadata_cache and self.grad_metadata_recv is None:
+                    self.grad_metadata_recv = create_send_metadata(output_tensor_grad)
 
-        return output_tensor_grad
+                return output_tensor_grad
 
-    def send_forward(self, model_chunk_id, output_object: Any, next_rank: int = None) -> None:
+    def send_forward(self, model_chunk_id: int, output_tensor: Any, next_rank: int = None) -> None:
         """Sends the input tensor to the next stage in pipeline.
            For interleaved 1F1B.
 
@@ -145,10 +166,12 @@ class InterleavedSchedule(PipelineSchedule):
             output_object (Any): Object to be sent.
             next_rank (int, optional): The rank of the recipient of the tensor.
         """
-        if not self.is_last_stage(model_chunk_id):
-            self.comm.send_forward(output_object, next_rank)
+        with self.stage_manager.switch_model_chunk_id(model_chunk_id):
+            if not self.stage_manager.is_last_stage():
+                self.comm.send_forward(output_tensor, next_rank, send_metadata=self.send_tensor_metadata)
+                self.send_tensor_metadata = not self.enable_metadata_cache
 
-    def send_backward(self, model_chunk_id, input_object: Any, prev_rank: int = None) -> None:
+    def send_backward(self, model_chunk_id: int, input_tensor_grad: Any, prev_rank: int = None) -> None:
         """Sends the gradient tensor to the previous stage in pipeline.
            For interleaved 1F1B.
 
@@ -157,12 +180,102 @@ class InterleavedSchedule(PipelineSchedule):
             input_object (Any): Object to be sent.
             prev_rank (int, optional): The rank of the recipient of the tensor
         """
-        if not self.is_first_stage(model_chunk_id):
-            self.comm.send_backward(input_object, prev_rank)
+        with self.stage_manager.switch_model_chunk_id(model_chunk_id):
+            if not self.stage_manager.is_first_stage():
+                self.comm.send_backward(input_tensor_grad, prev_rank, send_metadata=self.send_grad_metadata)
+                self.send_grad_metadata = not self.enable_metadata_cache
+
+    def send_forward_recv_backward(
+        self,
+        model_chunk_id_send: int,
+        model_chunk_id_recv: int,
+        output_tensor: Any,
+        next_rank: Optional[int] = None,
+        send_prior_fallback: Optional[bool] = None,
+    ) -> Any:
+        with self.stage_manager.switch_model_chunk_id(model_chunk_id_send):
+            send_data = not self.stage_manager.is_last_stage()
+        with self.stage_manager.switch_model_chunk_id(model_chunk_id_recv):
+            recv_data = not self.stage_manager.is_last_stage()
+
+        if send_data and recv_data:
+            if not self.send_forward_recv_backward and self.grad_metadata_recv is not None:
+                send_prior_fallback = None  # must not fallback
+            output_tensor_grad = self.comm.send_forward_recv_backward(
+                output_tensor,
+                next_rank,
+                send_metadata=self.send_tensor_metadata,
+                metadata_recv=self.grad_metadata_recv,
+                send_prior_fallback=send_prior_fallback,
+            )
+            self.send_tensor_metadata = not self.enable_metadata_cache
+            if self.enable_metadata_cache and self.grad_metadata_recv is None:
+                self.grad_metadata_recv = create_send_metadata(output_tensor_grad)
+            return output_tensor_grad
+
+        # send only or recv only
+        self.send_forward(model_chunk_id_send, output_tensor)
+        return self.recv_backward(model_chunk_id_recv)
+
+    def send_backward_recv_forward(
+        self,
+        model_chunk_id_send: int,
+        model_chunk_id_recv: int,
+        input_tensor_grad: Any,
+        prev_rank: Optional[int] = None,
+        send_prior_fallback: Optional[bool] = None,
+    ) -> Any:
+        with self.stage_manager.switch_model_chunk_id(model_chunk_id_send):
+            send_data = not self.stage_manager.is_first_stage()
+        with self.stage_manager.switch_model_chunk_id(model_chunk_id_recv):
+            recv_data = not self.stage_manager.is_first_stage()
+
+        if send_data and recv_data:
+            if not self.send_backward_recv_backward and self.tensor_metadata_recv is not None:
+                send_prior_fallback = None  # must not fallback
+            input_tensor = self.comm.send_backward_recv_forward(
+                input_tensor_grad,
+                prev_rank,
+                send_metadata=self.send_grad_metadata,
+                metadata_recv=self.tensor_metadata_recv,
+                send_prior_fallback=send_prior_fallback,
+            )
+            self.send_grad_metadata = not self.enable_metadata_cache
+            if self.enable_metadata_cache and self.tensor_metadata_recv is None:
+                self.tensor_metadata_recv = create_send_metadata(input_tensor)
+            return input_tensor
+
+        # send only or recv only
+        self.send_backward(model_chunk_id_send, input_tensor_grad)
+        return self.recv_forward(model_chunk_id_recv)
+
+    def send_forward_recv_forward(
+        self, model_chunk_id_send: int, model_chunk_id_recv: int, output_tensor: Any, send_prior: bool
+    ):
+        if send_prior:
+            self.send_forward(model_chunk_id_send, output_tensor)
+            input_tensor = self.recv_forward(model_chunk_id_recv)
+        else:
+            input_tensor = self.recv_forward(model_chunk_id_recv)
+            self.send_forward(model_chunk_id_send, output_tensor)
+
+        return input_tensor
+
+    def send_backward_recv_backward(
+        self, model_chunk_id_send: int, model_chunk_id_recv: int, input_tensor_grad: Any, send_prior: bool
+    ):
+        if send_prior:
+            self.send_backward(model_chunk_id_send, input_tensor_grad)
+            output_tensor_grad = self.recv_backward(model_chunk_id_recv)
+        else:
+            output_tensor_grad = self.recv_backward(model_chunk_id_recv)
+            self.send_backward(model_chunk_id_send, input_tensor_grad)
+
+        return output_tensor_grad
 
     def forward_step(
         self,
-        model_chunk: Module,
+        model_chunk: Union[ModuleList, Module],
         model_chunk_id: int,
         input_obj: Optional[dict],
         criterion: Callable,
@@ -171,7 +284,7 @@ class InterleavedSchedule(PipelineSchedule):
     ) -> Union[torch.Tensor, dict]:
         """Forward one step of the pipeline
         Args:
-            model (Module): Model Chunk to be run
+            model (ModuleList or Module): Model Chunk to be run
             input_obj (Optional[dict]): The output from the previous stage. If it is the first stage, the `input_obj` is None.
             criterion (Callable): Criterion to calculate loss.
             accum_loss (Optional[torch.Tensor], optional): Accumulated loss. Defaults to None.
@@ -184,17 +297,25 @@ class InterleavedSchedule(PipelineSchedule):
 
         # for the first stage, input_obj is None
         # for the non-first stage, input_obj is the output of the previous stage and it's must be a dict
-        output_obj = model_forward(model_chunk[model_chunk_id], micro_batch, input_obj)
-
-        if self.is_last_stage(model_chunk_id):
-            loss = criterion(output_obj, micro_batch) / self.num_microbatches
-            if accum_loss is not None:
-                accum_loss.add_(loss.detach())
-            if outputs is not None:
-                outputs.append(tree_map(detach, output_obj))
-            return loss
-        else:
-            return output_obj
+
+        with self.stage_manager.switch_model_chunk_id(model_chunk_id):
+            if isinstance(model_chunk, ModuleList):
+                output_obj = model_forward(model_chunk[model_chunk_id], micro_batch, input_obj)
+            else:
+                # NOTE: in shardformer, each device still has the entire model, so we need to use relevant stage layers
+                internal_inputs = {} if input_obj is None else input_obj
+                internal_inputs["stage_index"] = self.stage_manager.stage_indices[model_chunk_id]
+                output_obj = model_forward(model_chunk, micro_batch, internal_inputs)
+
+            if self.stage_manager.is_last_stage():
+                loss = criterion(output_obj, micro_batch) / self.num_microbatch
+                if accum_loss is not None:
+                    accum_loss.add_(loss.detach())
+                if outputs is not None:
+                    outputs.append(tree_map(detach, output_obj))
+                return loss
+            else:
+                return output_obj
 
     def backward_step(
         self,
@@ -241,140 +362,211 @@ class InterleavedSchedule(PipelineSchedule):
                     input_obj_grad[k] = v.grad
         return input_obj_grad
 
-    def forward_backward_step(
+    def run_forward_only(
         self,
-        model_chunk: Module,
+        model_chunk: Union[ModuleList, Module],
         data_iter: Iterable,
         criterion: Callable[..., Any],
-        optimizer: Optional[OptimizerWrapper] = None,
         return_loss: bool = False,
         return_outputs: bool = False,
-    ) -> dict:
-        """Runs interleaved 1F1B schedule, with communication between pipeline stages.
+    ) -> Dict:
+        assert self.forward_only
 
-        Args:
-            model_chunk (List[Module]): Model Chunk to be trained.
-            data_iter (Iterable): Data iterator.
-            criterion (Callable[[Any, Any], Tensor]): Criterion to be used. It should take two arguments: model outputs and inputs, and returns loss tensor.
-            optimizer (OptimizerWrapper, optional): Optimizer to be used. Can be None when only forward is executed. Defaults to None.
-            return_loss (bool, optional): Whether to return loss. Defaults to False. Whether to return loss.
-            return_outputs (bool, optional): Whether to return model outputs. Defaults to False. Whether to return model outputs.
+        self.load_batch(data_iter)
 
-        Returns:
-            dict: A dict with keys: 'loss' and 'outputs'.
-        """
-        forward_only = not torch.is_grad_enabled()
-        if optimizer is None:
-            assert forward_only, "Optimizer should be passed when doing backward."
+        outputs = [] if return_outputs and self.stage_manager.is_last_stage(ignore_chunk=True) else None
 
-        self.load_batch(data_iter)
-        num_model_chunks = len(model_chunk)
+        accum_loss = None
+        if return_loss and self.stage_manager.is_last_stage(ignore_chunk=True):
+            accum_loss = torch.scalar_tensor(0, device=get_current_device())
 
-        # num_warmup_microbatches is the step when not all the processes are working
-        num_microbatches = self.num_microbatches * num_model_chunks
-        if forward_only:
-            num_warmup_microbatches = num_microbatches
-        else:
-            num_warmup_microbatches = (self.stage_manager.num_stages - self.stage_manager.stage - 1) * 2
-            num_warmup_microbatches += (num_model_chunks - 1) * self.stage_manager.num_stages
-            num_warmup_microbatches = min(num_warmup_microbatches, num_microbatches)
+        model_chunk_id = self.get_model_chunk_id(0, is_forward=True)
+        input_obj = self.recv_forward(model_chunk_id)
 
-        num_microbatches_remaining = num_microbatches - num_warmup_microbatches
+        for i in range(self.num_microbatch * self.num_model_chunks):
+            last_iteration = i == self.num_microbatch * self.num_model_chunks - 1
+            model_chunk_id = self.get_model_chunk_id(i, is_forward=True)
+            output_obj = self.forward_step(model_chunk, model_chunk_id, input_obj, criterion, accum_loss, outputs)
 
-        # Input, output tensors only need to be saved when doing backward passes
-        input_objs = None
-        output_objs = None
+            if not last_iteration:
+                input_obj = self.send_forward_recv_forward(
+                    model_chunk_id_send=model_chunk_id,
+                    model_chunk_id_recv=self.get_model_chunk_id(i + 1, is_forward=True),
+                    output_tensor=output_obj,
+                    send_prior=self.stage_manager.stage % 2 == 0,
+                )
+            else:
+                self.send_forward(model_chunk_id, output_obj)
 
-        if not forward_only:
-            input_objs = [[] for _ in range(num_model_chunks)]
-            output_objs = [[] for _ in range(num_model_chunks)]
+        if outputs is not None:
+            outputs = merge_batch(outputs)
+        return {"loss": accum_loss, "outputs": outputs}
 
-        outputs = [] if return_outputs and self.stage_manager.is_last_stage() else None
+    def run_forward_backward(
+        self,
+        model_chunk: Union[ModuleList, Module],
+        data_iter: Iterable,
+        criterion: Callable[..., Any],
+        optimizer: Optional[OptimizerWrapper] = None,
+        return_loss: bool = False,
+        return_outputs: bool = False,
+    ) -> Dict:
+        """
+        Runs interleaved schedule, with communication between pipeline stages.
+        """
+        assert not self.forward_only
 
-        if return_loss and self.stage_manager.is_last_stage():
-            accum_loss = torch.zeros(1, device=get_accelerator().get_current_device())
-        else:
-            accum_loss = None
+        self.load_batch(data_iter)
 
-        # for ranks except the first one, get into recv state
-        # print(self.stage_manager.stage,num_microbatches, num_warmup_microbatches, num_microbatches_remaining)
-        input_obj = self.recv_forward(0)
-        input_objs[0].append(input_obj)
-        # Run warmup forward passes.
-        for i in range(num_warmup_microbatches):
-            model_chunk_id = self.get_model_chunk_id(i, forward=True)
+        num_microbatch = self.num_microbatch * self.num_model_chunks
+        num_warmup_microbatch = (self.stage_manager.num_stages - self.stage_manager.stage - 1) * 2
+        num_warmup_microbatch += (self.num_model_chunks - 1) * self.stage_manager.num_stages
+        num_warmup_microbatch = min(num_warmup_microbatch, num_microbatch)
+        num_microbatch_remaining = num_microbatch - num_warmup_microbatch
 
-            # recv first on first rank to avoid sending or recving at the same time
-            if self.stage_manager.is_first_stage():
-                input_obj = self.recv_forward(model_chunk_id)
-                output_obj = self.forward_step(model_chunk, model_chunk_id, input_obj, criterion, accum_loss, outputs)
-                self.send_forward(model_chunk_id, output_obj)
-                if not forward_only:
-                    input_objs[model_chunk_id].append(input_obj)
-                    output_objs[model_chunk_id].append(output_obj)
-            else:
-                output_obj = self.forward_step(model_chunk, model_chunk_id, input_obj, criterion, accum_loss, outputs)
-                if not forward_only:
-                    output_objs[model_chunk_id].append(output_obj)
-                self.send_forward(model_chunk_id, output_obj)
-                if num_microbatches_remaining == 0 and i + 1 == num_warmup_microbatches:
-                    break
-                else:
-                    model_chunk_id = self.get_model_chunk_id(i + 1, forward=True)
+        # Input, output tensors only need to be saved when doing backward passes
+        input_objs = [[] for _ in range(self.num_model_chunks)]
+        output_objs = [[] for _ in range(self.num_model_chunks)]
 
-                    input_obj = self.recv_forward(model_chunk_id)
-                    if not forward_only:
-                        input_objs[model_chunk_id].append(input_obj)
+        outputs = [] if return_outputs and self.stage_manager.is_last_stage(ignore_chunk=True) else None
 
-        # Run 1F1B in steady state.
-        for i in range(num_microbatches_remaining):
-            model_chunk_id = self.get_model_chunk_id(i + num_warmup_microbatches, forward=True)
-            last_iteration = i == (num_microbatches_remaining - 1)
+        accum_loss = None
+        if return_loss and self.stage_manager.is_last_stage(ignore_chunk=True):
+            accum_loss = torch.scalar_tensor(0, device=get_current_device())
 
+        model_chunk_id = self.get_model_chunk_id(0, is_forward=True)
+        input_obj = self.recv_forward(model_chunk_id)
+        # Run warmup forward passes.
+        for i in range(num_warmup_microbatch):
+            last_iteration = i == num_warmup_microbatch - 1
+            model_chunk_id = self.get_model_chunk_id(i, is_forward=True)
             output_obj = self.forward_step(model_chunk, model_chunk_id, input_obj, criterion, accum_loss, outputs)
-            if forward_only:
-                self.send_forward(model_chunk_id, output_obj)
+            input_objs[model_chunk_id].append(input_obj)
+            output_objs[model_chunk_id].append(output_obj)
 
-                if not last_iteration:
-                    input_obj = self.recv_forward(model_chunk_id)
-
-            else:
+            if last_iteration and num_microbatch_remaining == 0:
                 self.send_forward(model_chunk_id, output_obj)
-                # Add input_obj and output_obj to end of list.
-                input_objs[model_chunk_id].append(input_obj)
-                output_objs[model_chunk_id].append(output_obj)
-
-                model_chunk_id = self.get_model_chunk_id(i, forward=False)
-                output_obj_grad = self.recv_backward(model_chunk_id)
+            else:
+                input_obj = self.send_forward_recv_forward(
+                    model_chunk_id_send=model_chunk_id,
+                    model_chunk_id_recv=self.get_model_chunk_id(i + 1, is_forward=True),
+                    output_tensor=output_obj,
+                    send_prior=self.stage_manager.stage % 2 == 0,
+                )
 
-                # Pop output_obj and output_obj from the start of the list for
-                # the backward pass.
-                input_obj = input_objs[model_chunk_id].pop(0)
-                output_obj = output_objs[model_chunk_id].pop(0)
+        if num_microbatch_remaining > 0:
+            model_chunk_id = self.get_model_chunk_id(0, is_forward=False)
+            output_obj_grad = self.recv_backward(model_chunk_id)
 
-                # backward
-                input_obj_grad = self.backward_step(optimizer, input_obj, output_obj, output_obj_grad)
+        # Run 1F1B in steady state.
+        for i in range(num_microbatch_remaining):
+            last_iteration = i == num_microbatch_remaining - 1
 
+            model_chunk_id = self.get_model_chunk_id(i + num_warmup_microbatch, is_forward=True)
+            output_obj = self.forward_step(model_chunk, model_chunk_id, input_obj, criterion, accum_loss, outputs)
+            # Add input_obj and output_obj to end of list.
+            input_objs[model_chunk_id].append(input_obj)
+            output_objs[model_chunk_id].append(output_obj)
+
+            model_chunk_id = self.get_model_chunk_id(i, is_forward=False)
+            # Pop output_obj and output_obj from the start of the list for the backward pass.
+            _input_obj = input_objs[model_chunk_id].pop(0)
+            _output_obj = output_objs[model_chunk_id].pop(0)
+            input_obj_grad = self.backward_step(optimizer, _input_obj, _output_obj, output_obj_grad)
+
+            # NOTE: perform 2x communication for forward and backward
+            def send_forward_recv_backward():
+                if last_iteration and num_microbatch == num_microbatch_remaining:
+                    model_chunk_id = self.get_model_chunk_id(i + num_warmup_microbatch, is_forward=True)
+                    self.send_forward(model_chunk_id, output_obj)
+                else:
+                    output_obj_grad = self.send_forward_recv_backward(
+                        model_chunk_id_send=self.get_model_chunk_id(i + num_warmup_microbatch, is_forward=True),
+                        model_chunk_id_recv=self.get_model_chunk_id(i + 1, is_forward=False),
+                        output_tensor=output_obj,
+                        send_prior_fallback=self.stage_manager.stage % 2 == 0,
+                    )
+                    return output_obj_grad
+
+            def send_backward_recv_forward():
                 if last_iteration:
-                    input_obj = None
+                    model_chunk_id = self.get_model_chunk_id(i, is_forward=False)
+                    self.send_backward(model_chunk_id, input_obj_grad)
                 else:
-                    model_chunk_id = self.get_model_chunk_id(i + num_warmup_microbatches + 1, forward=True)
-                    input_obj = self.recv_forward(model_chunk_id)
-                model_chunk_id = self.get_model_chunk_id(i, forward=False)
-                self.send_backward(model_chunk_id, input_obj_grad)
+                    input_obj = self.send_backward_recv_forward(
+                        model_chunk_id_send=self.get_model_chunk_id(i, is_forward=False),
+                        model_chunk_id_recv=self.get_model_chunk_id(i + num_warmup_microbatch + 1, is_forward=True),
+                        input_tensor_grad=input_obj_grad,
+                        send_prior_fallback=self.stage_manager.stage % 2 == 0 and i > 0,
+                    )
+                    return input_obj
+
+            if self.stage_manager.stage % 2 == 0:
+                output_obj_grad = send_forward_recv_backward()
+                input_obj = send_backward_recv_forward()
+            else:
+                input_obj = send_backward_recv_forward()
+                output_obj_grad = send_forward_recv_backward()
 
+        if num_microbatch_remaining == 0:
+            model_chunk_id = self.get_model_chunk_id(0, is_forward=False)
+            output_obj_grad = self.recv_backward(model_chunk_id)
         # Run cooldown backward passes.
-        if not forward_only:
-            for i in range(num_microbatches_remaining, num_microbatches):
-                model_chunk_id = self.get_model_chunk_id(i, forward=False)
-                # print(f"{self.stage_manager.stage}/{model_chunk_id}: {len(input_objs[model_chunk_id])} {len(output_objs[model_chunk_id])} {i}")
-                input_obj = input_objs[model_chunk_id].pop(0)
-                output_obj = output_objs[model_chunk_id].pop(0)
-
-                output_obj_grad = self.recv_backward(model_chunk_id)
-                input_obj_grad = self.backward_step(optimizer, input_obj, output_obj, output_obj_grad)
+        for i in range(num_microbatch_remaining, num_microbatch):
+            last_iteration = i == num_microbatch - 1
+            model_chunk_id = self.get_model_chunk_id(i, is_forward=False)
+            _input_obj = input_objs[model_chunk_id].pop(0)
+            _output_obj = output_objs[model_chunk_id].pop(0)
+            # output_obj_grad = self.recv_backward(model_chunk_id)
+            input_obj_grad = self.backward_step(optimizer, _input_obj, _output_obj, output_obj_grad)
+
+            if not last_iteration:
+                output_obj_grad = self.send_backward_recv_backward(
+                    model_chunk_id_send=self.get_model_chunk_id(i, is_forward=False),
+                    model_chunk_id_recv=self.get_model_chunk_id(i + 1, is_forward=False),
+                    input_tensor_grad=input_obj_grad,
+                    send_prior=self.stage_manager.stage % 2 == 0 and i > num_microbatch_remaining,
+                )
+            else:
+                model_chunk_id = self.get_model_chunk_id(i, is_forward=False)
                 self.send_backward(model_chunk_id, input_obj_grad)
 
+        assert all(len(v) == 0 for v in input_objs) and all(len(v) == 0 for v in output_objs)
+
         if outputs is not None:
             outputs = merge_batch(outputs)
         return {"loss": accum_loss, "outputs": outputs}
+
+    def forward_backward_step(
+        self,
+        model_chunk: Union[ModuleList, Module],
+        data_iter: Iterable,
+        criterion: Callable[..., Any],
+        optimizer: Optional[OptimizerWrapper] = None,
+        return_loss: bool = False,
+        return_outputs: bool = False,
+    ) -> dict:
+        """
+        Args:
+            model_chunk (ModuleList or Module): Model Chunk to be trained. Original interleaved uses a module list whereas shardformer uses entire model + layer specification
+            data_iter (Iterable): Data iterator.
+            criterion (Callable[[Any, Any], Tensor]): Criterion to be used. It should take two arguments: model outputs and inputs, and returns loss tensor.
+            optimizer (OptimizerWrapper, optional): Optimizer to be used. Can be None when only forward is executed. Defaults to None.
+            return_loss (bool, optional): Whether to return loss. Defaults to False. Whether to return loss.
+            return_outputs (bool, optional): Whether to return model outputs. Defaults to False. Whether to return model outputs.
+
+        Returns:
+            dict: A dict with keys: 'loss' and 'outputs'.
+        """
+        self.forward_only = not torch.is_grad_enabled()
+        if optimizer is None:
+            assert self.forward_only, "Optimizer should be passed when doing backward."
+
+        if self.forward_only:
+            result = self.run_forward_only(model_chunk, data_iter, criterion, return_loss, return_outputs)
+        else:
+            result = self.run_forward_backward(
+                model_chunk, data_iter, criterion, optimizer, return_loss, return_outputs
+            )
+
+        return result

colossalai/pipeline/schedule/one_f_one_b.py

 from functools import partial
-from typing import Any, Callable, Iterable, List, Optional, Union
+from typing import Any, Callable, Dict, Iterable, List, Optional, Union
 
 import torch
 import torch.cuda
@@ -8,7 +8,7 @@ from torch.utils._pytree import tree_map
 
 from colossalai.accelerator import get_accelerator
 from colossalai.interface import ModelWrapper, OptimizerWrapper
-from colossalai.pipeline.p2p import PipelineP2PCommunication
+from colossalai.pipeline.p2p import PipelineP2PCommunication, create_send_metadata
 from colossalai.pipeline.stage_manager import PipelineStageManager
 
 from ._utils import (
@@ -30,6 +30,7 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
         stage_manager: PipelineStageManager,
         num_microbatches: Optional[int] = None,
         microbatch_size: Optional[int] = None,
+        enable_metadata_cache: bool = True,
     ) -> None:
         """1F1B pipeline schedule.
 
@@ -42,13 +43,21 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
         assert (
             num_microbatches is not None or microbatch_size is not None
         ), "Either num_microbatches or microbatch_size should be provided"
+
         self.comm = PipelineP2PCommunication(stage_manager)
         self.num_microbatches = num_microbatches
         self.microbatch_size = microbatch_size
         self.batch: Optional[Any] = None
         self.batch_size: Optional[int] = None
+        self.last_batch_size: Optional[int] = None
         self.microbatch_offset: Optional[int] = None
-        self._use_microbatch_size = num_microbatches is None
+
+        # P2PMeta cache
+        self.enable_metadata_cache = enable_metadata_cache
+        self.send_tensor_metadata = True
+        self.send_grad_metadata = True
+        self.tensor_metadata_recv = None
+        self.grad_metadata_recv = None
 
     def load_batch(self, data_iter: Iterable, device: Optional[torch.device] = None) -> None:
         """Load a batch from data iterator.
@@ -60,24 +69,45 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
         batch = next(data_iter)
         if device is not None:
             batch = tree_map(partial(to_device, device=device), batch)
+
+        self.microbatch_offset = 0
         self.batch = batch
         self.batch_size = get_batch_size(batch)
-        self.microbatch_offset = 0
-        if not self._use_microbatch_size:
-            assert (
-                self.batch_size % self.num_microbatches == 0
-            ), "Batch size should divided by the number of microbatches"
+
+        if self.microbatch_size is None:
+            assert self.batch_size % self.num_microbatches == 0, "Batch size should divided by # microbatches"
             self.microbatch_size = self.batch_size // self.num_microbatches
-        else:
+        if self.num_microbatches is None:
             assert self.batch_size % self.microbatch_size == 0, "Batch size should divided by the microbatch size"
             self.num_microbatches = self.batch_size // self.microbatch_size
 
+        if not self.forward_only:
+            assert self.last_batch_size is None or self.last_batch_size == self.batch_size
+            assert self.batch_size == self.microbatch_size * self.num_microbatches
+
+            assert (
+                self.num_microbatches >= self.stage_manager.num_stages
+            ), "Number of microbatch should be larger than number of stages"
+
+        if self.forward_only:
+            self.num_microbatches = (self.batch_size - 1) // self.microbatch_size + 1
+            # NOTE: disable metadata cache when batch size changes (not valid anymore)
+            if self.batch_size != self.last_batch_size:
+                self.enable_metadata_cache = False
+                self.send_tensor_metadata = True
+                self.send_grad_metadata = True
+                self.tensor_metadata_recv = None
+                self.grad_metadata_recv = None
+
+        self.last_batch_size = self.batch_size
+
     def load_micro_batch(self) -> Any:
         """Load a micro batch from the current batch.
 
         Returns:
             Any: Micro batch.
         """
+        assert self.microbatch_offset <= self.batch_size, "Microbatches exhausted"
         micro_batch = get_micro_batch(self.batch, self.microbatch_offset, self.microbatch_size)
         self.microbatch_offset += self.microbatch_size
         return tree_map(partial(to_device, device=get_accelerator().get_current_device()), micro_batch)
@@ -92,12 +122,12 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
         Returns:
             Any: The input tensor or input tensor list.
         """
-        if self.stage_manager.is_first_stage():
-            input_tensor = None
-        else:
-            input_tensor = self.comm.recv_forward(prev_rank)
+        if not self.stage_manager.is_first_stage():
+            input_tensor = self.comm.recv_forward(prev_rank, metadata_recv=self.tensor_metadata_recv)
+            if self.enable_metadata_cache and self.tensor_metadata_recv is None:
+                self.tensor_metadata_recv = create_send_metadata(input_tensor)
 
-        return input_tensor
+            return input_tensor
 
     def recv_backward(self, next_rank: int = None) -> Any:
         """Copy the gradient tensor from the next stage in pipeline as the input gradient of this stage.
@@ -109,14 +139,14 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
         Returns:
             Any: The input gradient tensor or gradient tensor list.
         """
-        if self.stage_manager.is_last_stage():
-            output_tensor_grad = None
-        else:
-            output_tensor_grad = self.comm.recv_backward(next_rank)
+        if not self.stage_manager.is_last_stage():
+            output_tensor_grad = self.comm.recv_backward(next_rank, metadata_recv=self.grad_metadata_recv)
+            if self.enable_metadata_cache and self.grad_metadata_recv is None:
+                self.grad_metadata_recv = create_send_metadata(output_tensor_grad)
 
-        return output_tensor_grad
+            return output_tensor_grad
 
-    def send_forward(self, output_object: Any, next_rank: int = None) -> None:
+    def send_forward(self, output_tensor: Any, next_rank: int = None) -> None:
         """Sends the input tensor to the next stage in pipeline.
            For 1F1B.
 
@@ -125,20 +155,10 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
             next_rank (int, optional): The rank of the recipient of the tensor.
         """
         if not self.stage_manager.is_last_stage():
-            self.comm.send_forward(output_object, next_rank)
-
-    def send_forward_recv_backward(self, output_object: Any, next_rank: int = None) -> Any:
-        """Sends the input tensor to the next stage and copy the gradient tensor from the next stage in pipeline.
-           For 1F1B.
-
-        Args:
-            output_object (Any): Object to be sent.
-            next_rank (int, optional): The rank of the recipient of the tensor.
-        """
-        if not self.stage_manager.is_last_stage():
-            return self.comm.send_forward_recv_backward(output_object, next_rank)
+            self.comm.send_forward(output_tensor, next_rank, send_metadata=self.send_tensor_metadata)
+            self.send_tensor_metadata = not self.enable_metadata_cache
 
-    def send_backward(self, input_object: Any, prev_rank: int = None) -> None:
+    def send_backward(self, input_tensor_grad: Any, prev_rank: int = None) -> None:
         """Sends the gradient tensor to the previous stage in pipeline.
            For 1F1B.
 
@@ -147,34 +167,60 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
             prev_rank (int, optional): The rank of the recipient of the tensor
         """
         if not self.stage_manager.is_first_stage():
-            self.comm.send_backward(input_object, prev_rank)
+            self.comm.send_backward(input_tensor_grad, prev_rank, send_metadata=self.send_grad_metadata)
+            self.send_grad_metadata = not self.enable_metadata_cache
 
-    def send_backward_recv_forward(self, output_object: Any, prev_rank: int = None) -> Any:
-        """Sends the gradient tensor to the previous stage and copy the input tensor from the previous stage in pipeline.
+    def send_forward_recv_backward(
+        self, output_tensor: Any, next_rank: int = None, send_prior_fallback: Optional[bool] = None
+    ) -> Any:
+        """Sends the input tensor to the next stage and copy the gradient tensor from the next stage in pipeline.
            For 1F1B.
 
         Args:
             output_object (Any): Object to be sent.
-            prev_rank (int, optional): The rank of the recipient of the tensor.
+            next_rank (int, optional): The rank of the recipient of the tensor.
         """
-        if not self.stage_manager.is_first_stage():
-            return self.comm.send_backward_recv_forward(output_object, prev_rank)
-
-    def send_forward_recv_forward(self, input_object: Any, prev_rank: int = None, next_rank: int = None) -> Any:
-        """Sends the input tensor to the next stage and copy the input tensor from the previous stage in pipeline.
+        if not self.stage_manager.is_last_stage():
+            if not self.send_tensor_metadata and self.grad_metadata_recv is not None:
+                send_prior_fallback = None  # must not fallback
+            output_tensor_grad = self.comm.send_forward_recv_backward(
+                output_tensor,
+                next_rank,
+                send_metadata=self.send_tensor_metadata,
+                metadata_recv=self.grad_metadata_recv,
+                send_prior_fallback=send_prior_fallback,
+            )
+            self.send_tensor_metadata = not self.enable_metadata_cache
+            if self.enable_metadata_cache and self.grad_metadata_recv is None:
+                self.grad_metadata_recv = create_send_metadata(output_tensor_grad)
+
+            return output_tensor_grad
+
+    def send_backward_recv_forward(
+        self, input_tensor_grad: Any, prev_rank: int = None, send_prior_fallback: Optional[bool] = None
+    ) -> Any:
+        """Sends the gradient tensor to the previous stage and copy the input tensor from the previous stage in pipeline.
            For 1F1B.
 
         Args:
-            input_object (Any): Object to be sent.
-            prev_rank (int, optional): The previous rank of the recipient of the tensor.
-            next_rank (int, optional): The next rank of the recipient of the tensor.
+            output_object (Any): Object to be sent.
+            prev_rank (int, optional): The rank of the recipient of the tensor.
         """
-        if self.stage_manager.is_first_stage():
-            return self.comm.send_forward(input_object, next_rank)
-        elif self.stage_manager.is_last_stage():
-            return self.comm.recv_forward(prev_rank)
-        else:
-            return self.comm.send_forward_recv_forward(input_object, prev_rank, next_rank)
+        if not self.stage_manager.is_first_stage():
+            if not self.send_grad_metadata and self.tensor_metadata_recv is not None:
+                send_prior_fallback = None  # must not fallback
+            input_tensor = self.comm.send_backward_recv_forward(
+                input_tensor_grad,
+                prev_rank,
+                send_metadata=self.send_grad_metadata,
+                metadata_recv=self.tensor_metadata_recv,
+                send_prior_fallback=send_prior_fallback,
+            )
+            self.send_grad_metadata = not self.enable_metadata_cache
+            if self.enable_metadata_cache and self.tensor_metadata_recv is None:
+                self.tensor_metadata_recv = create_send_metadata(input_tensor)
+
+            return input_tensor
 
     def forward_step(
         self,
@@ -254,31 +300,50 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
                     input_obj_grad[k] = v.grad
         return input_obj_grad
 
-    def forward_backward_step(
+    def run_forward_only(
         self,
         model: Module,
         data_iter: Iterable,
         criterion: Callable[..., Any],
-        optimizer: Optional[OptimizerWrapper] = None,
         return_loss: bool = False,
         return_outputs: bool = False,
-    ) -> dict:
-        """Runs non-interleaved 1F1B schedule, with communication between pipeline stages.
+    ) -> Dict:
+        """
+        Runs forward only schedule, with communication between pipeline stages.
+        """
+        assert self.forward_only
 
-        Args:
-            model (Module): Model to be trained.
-            data_iter (Iterable): Data iterator.
-            criterion (Callable[[Any, Any], Tensor]): Criterion to be used. It should take two arguments: model outputs and inputs, and returns loss tensor.
-            optimizer (OptimizerWrapper, optional): Optimizer to be used. Can be None when only forward is executed. Defaults to None.
-            return_loss (bool, optional): Whether to return loss. Defaults to False. Whether to return loss.
-            return_outputs (bool, optional): Whether to return model outputs. Defaults to False. Whether to return model outputs.
+        self.load_batch(data_iter)
 
-        Returns:
-            dict: A dict with keys: 'loss' and 'outputs'.
+        accum_loss = None
+        if return_loss and self.stage_manager.is_last_stage():
+            accum_loss = torch.scalar_tensor(0, device=get_accelerator().get_current_device())
+        outputs = [] if return_outputs and self.stage_manager.is_last_stage() else None
+
+        for _ in range(self.num_microbatches):
+            input_obj = self.recv_forward()
+            output_obj = self.forward_step(model, input_obj, criterion, accum_loss, outputs)
+            self.send_forward(output_obj)
+
+        if outputs is not None:
+            if isinstance(model, ModelWrapper):
+                model = model.unwrap()
+            outputs = merge_batch(outputs, getattr(model, "batch_size_dim", 0))
+        return {"loss": accum_loss, "outputs": outputs}
+
+    def run_forward_backward(
+        self,
+        model: Module,
+        data_iter: Iterable,
+        criterion: Callable[..., Any],
+        optimizer: Optional[OptimizerWrapper] = None,
+        return_loss: bool = False,
+        return_outputs: bool = False,
+    ) -> Dict:
         """
-        forward_only = not torch.is_grad_enabled()
-        if optimizer is None:
-            assert forward_only, "Optimizer should be passed when doing backward."
+        Runs non-interleaved 1F1B schedule, with communication between pipeline stages.
+        """
+        assert not self.forward_only
 
         self.load_batch(data_iter)
 
@@ -288,30 +353,20 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
         num_microbatches_remaining = self.num_microbatches - num_warmup_microbatches
 
         # Input, output tensors only need to be saved when doing backward passes
-        input_objs = None
-        output_objs = None
-
-        if not forward_only:
-            input_objs = []
-            output_objs = []
+        input_objs, output_objs = [], []
 
-        outputs = [] if return_outputs and self.stage_manager.is_last_stage() else None
+        accum_loss = None
         if return_loss and self.stage_manager.is_last_stage():
-            accum_loss = torch.zeros(1, device=get_accelerator().get_current_device())
-        else:
-            accum_loss = None
+            accum_loss = torch.scalar_tensor(0, device=get_current_device())
+        outputs = [] if return_outputs and self.stage_manager.is_last_stage() else None
 
         # Run warmup forward passes.
         for i in range(num_warmup_microbatches):
             input_obj = self.recv_forward()
-
             output_obj = self.forward_step(model, input_obj, criterion, accum_loss, outputs)
-
             self.send_forward(output_obj)
-
-            if not forward_only:
-                input_objs.append(input_obj)
-                output_objs.append(output_obj)
+            input_objs.append(input_obj)
+            output_objs.append(output_obj)
 
         # Before running 1F1B, need to receive first forward tensor.
         # If all microbatches are run in warmup / cooldown phase, then no need to
@@ -324,44 +379,72 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
             last_iteration = i == (num_microbatches_remaining - 1)
 
             output_obj = self.forward_step(model, input_obj, criterion, accum_loss, outputs)
-            if forward_only:
-                self.send_forward(output_obj)
-
-                if not last_iteration:
-                    input_obj = self.recv_forward()
-            else:
-                # TODO adjust here
-                self.send_forward(output_obj)
-                output_obj_grad = self.recv_backward()
-
-                # Add input_obj and output_obj to end of list.
-                input_objs.append(input_obj)
-                output_objs.append(output_obj)
-
-                # Pop output_obj and output_obj from the start of the list for
-                # the backward pass.
-                input_obj = input_objs.pop(0)
-                output_obj = output_objs.pop(0)
-                input_obj_grad = self.backward_step(optimizer, input_obj, output_obj, output_obj_grad)
-
-                if last_iteration:
-                    input_obj = None
-                else:
-                    input_obj = self.recv_forward()
+            output_obj_grad = self.send_forward_recv_backward(
+                output_obj, send_prior_fallback=self.stage_manager.stage % 2 == 0
+            )
+            # Add input_obj and output_obj to end of list.
+            input_objs.append(input_obj)
+            output_objs.append(output_obj)
+
+            # Pop output_obj and output_obj from the start of the list for
+            # the backward pass.
+            input_obj = input_objs.pop(0)
+            output_obj = output_objs.pop(0)
+            input_obj_grad = self.backward_step(optimizer, input_obj, output_obj, output_obj_grad)
+
+            if last_iteration:
                 self.send_backward(input_obj_grad)
+            else:
+                input_obj = self.send_backward_recv_forward(
+                    input_obj_grad, send_prior_fallback=self.stage_manager.stage % 2 == 0
+                )
 
         # Run cooldown backward passes.
-        if not forward_only:
-            for i in range(num_warmup_microbatches):
-                input_obj = input_objs.pop(0)
-                output_obj = output_objs.pop(0)
+        for i in range(num_warmup_microbatches):
+            input_obj = input_objs.pop(0)
+            output_obj = output_objs.pop(0)
 
-                output_obj_grad = self.recv_backward()
-                input_obj_grad = self.backward_step(optimizer, input_obj, output_obj, output_obj_grad)
-                self.send_backward(input_obj_grad)
+            output_obj_grad = self.recv_backward()
+            input_obj_grad = self.backward_step(optimizer, input_obj, output_obj, output_obj_grad)
+            self.send_backward(input_obj_grad)
+
+        assert all(len(v) == 0 for v in input_objs) and all(len(v) == 0 for v in output_objs)
 
         if outputs is not None:
             if isinstance(model, ModelWrapper):
                 model = model.unwrap()
             outputs = merge_batch(outputs, getattr(model, "batch_size_dim", 0))
         return {"loss": accum_loss, "outputs": outputs}
+
+    def forward_backward_step(
+        self,
+        model: Module,
+        data_iter: Iterable,
+        criterion: Callable[..., Any],
+        optimizer: Optional[OptimizerWrapper] = None,
+        return_loss: bool = False,
+        return_outputs: bool = False,
+    ) -> dict:
+        """
+        Args:
+            model (Module): Model to be trained.
+            data_iter (Iterable): Data iterator.
+            criterion (Callable[[Any, Any], Tensor]): Criterion to be used. It should take two arguments: model outputs and inputs, and returns loss tensor.
+            optimizer (OptimizerWrapper, optional): Optimizer to be used. Can be None when only forward is executed. Defaults to None.
+            return_loss (bool, optional): Whether to return loss. Defaults to False. Whether to return loss.
+            return_outputs (bool, optional): Whether to return model outputs. Defaults to False. Whether to return model outputs.
+
+        Returns:
+            dict: Dictionary containing loss and outputs.
+        """
+
+        self.forward_only = not torch.is_grad_enabled()
+        if optimizer is None:
+            assert self.forward_only, "Optimizer should be passed when doing backward."
+
+        if self.forward_only:
+            result = self.run_forward_only(model, data_iter, criterion, return_loss, return_outputs)
+        else:
+            result = self.run_forward_backward(model, data_iter, criterion, optimizer, return_loss, return_outputs)
+
+        return result

colossalai/pipeline/stage_manager.py

+import contextlib
 from typing import Dict, List, Optional, Tuple
 
 import torch.distributed as dist
@@ -19,7 +20,15 @@ class PipelineStageManager:
         stage (int): The current stage.
     """
 
-    def __init__(self, pg_mesh: ProcessGroupMesh, pipeline_axis: int, is_virtual: bool = False) -> None:
+    def __init__(
+        self,
+        pg_mesh: ProcessGroupMesh,
+        pipeline_axis: int,
+        enable_interleave: bool = False,
+        num_model_chunks: int = 1,
+    ) -> None:
+        assert enable_interleave or num_model_chunks == 1, "num_model_chunks must be 1 when enable_interleave is False"
+
         self.pg_mesh = pg_mesh
         self.pipeline_axis = pipeline_axis
         self.prev_rank: Optional[Tuple[int, ...]] = None
@@ -43,29 +52,56 @@ class PipelineStageManager:
                 ranks_in_group = self.pg_mesh.get_ranks_in_group(group)
                 self.p2p_groups[tuple(ranks_in_group)] = group
 
-        if is_virtual:
+        self.is_interleave = enable_interleave
+        if enable_interleave:
+            # use circle p2p communication
             # add the process group of the first rank and the last rank
-            # only used in interleaved pipeline for now
             group = self.pg_mesh.get_group_along_axis(self.pipeline_axis, [stages[0], stages[-1]])
             if self.stage in [stages[0], stages[-1]]:
                 ranks_in_group = self.pg_mesh.get_ranks_in_group(group)
                 self.p2p_groups[tuple(ranks_in_group)] = group
 
-    def is_first_stage(self) -> bool:
+            # for interleaved pipeline parallel, each device is responsible for multiple chunk of layers
+            self.num_model_chunks: int = num_model_chunks
+
+            # for shardformer, hold stage indices of model
+            self.stage_indices: List[Tuple[int, int]]
+            # for shardformer, hold model chunk id
+            self.model_chunk_id: Optional[int] = None
+
+    def is_first_stage(self, ignore_chunk: bool = False) -> bool:
         """Is the current stage the first stage.
 
+        NOTE:
+            1. if using interleaved pipeline parallel, the first stage is the first chunk of the first device.
+            2. invoke is_first_stage() with ignore_chunk=True is equivalent to invoke is_first_device()
+
         Returns:
             bool: Whether the current stage is the first stage.
         """
-        return self.stage == 0
-
-    def is_last_stage(self) -> bool:
+        assert isinstance(ignore_chunk, bool)
+        assert not self.is_interleave or (ignore_chunk or self.model_chunk_id is not None)
+        if not self.is_interleave or ignore_chunk:
+            return self.stage == 0
+        else:
+            return self.stage == 0 and self.model_chunk_id == 0
+
+    def is_last_stage(self, ignore_chunk: bool = False) -> bool:
         """Is the current stage the last stage.
 
+        NOTE:
+            1. if using interleaved pipeline parallel, the last stage is the last chunk of the last device.
+            2. invoke is_last_stage() with ignore_chunk=True is equivalent to invoke is_last_device()
+
         Returns:
             bool: Whether the current stage is the last stage.
         """
-        return self.stage == self.num_stages - 1
+        assert isinstance(ignore_chunk, bool)
+        assert not self.is_interleave or (ignore_chunk or self.model_chunk_id is not None)
+        if not self.is_interleave or ignore_chunk:
+            return self.stage == self.num_stages - 1
+        else:
+            return self.stage == self.num_stages - 1 and self.model_chunk_id == self.num_model_chunks - 1
 
     @property
     def num_stages(self) -> int:
@@ -133,3 +169,10 @@ class PipelineStageManager:
             ProcessGroup: Process group of the given stages.
         """
         return self.pg_mesh.get_group_along_axis(self.pipeline_axis, stages)
+
+    @contextlib.contextmanager
+    def switch_model_chunk_id(self, model_chunk_id: int):
+        old_model_chunk_id = self.model_chunk_id
+        self.model_chunk_id = model_chunk_id
+        yield
+        self.model_chunk_id = old_model_chunk_id

colossalai/shardformer/README.md

@@ -79,9 +79,9 @@ Following are the description `ShardConfig`'s arguments:
 
 - `enable_sequence_overlap`: Whether to turn on sequence overlap, which overlap the computation and communication in sequence parallelism. It can only be used when `enable_sequence_parallelism` is True. Defaults to False.
 
--  `enable_all_optimization`: Whether to turn on all optimization tools including `fused normalizaion`, `flash attention`, `JIT fused operators`, `sequence parallelism` and `sequence overlap`. Defaults to False.
+-  `enable_all_optimization`: Whether to turn on all optimization tools including `fused normalization`, `flash attention`, `JIT fused operators`, `sequence parallelism` and `sequence overlap`. Defaults to False.
 
-- `extra_kwargs`: A dict to store extra kwargs for ShardFomer.
+- `extra_kwargs`: A dict to store extra kwargs for ShardFormer.
 
 ### Write your own policy
 
@@ -116,17 +116,18 @@ We will follow this roadmap to develop Shardformer:
 
 | model |   tensor parallel    |  pipeline parallel   |   lazy initialization |  xformer   |  flash attn2 | jit fused operator | fused layernorm |  sequence parallel |  overlap |
 | :------: | :-----: | :-----: | :--------: | :---------: | :------: | :-----: | :-----: | :--------: | :---------: |
-| bert |   [x]   |  [x]   |   [x] |  [x]   |  [x] | [x] | [x] |  [x] |  [x] |
-| t5 |   [x]   |  [x]   |   [x] |  [x]   |  [x] | [x] | [x] |  [ ] |  [ ] |
-| llama V1/V2 |   [x]   |  [x]   |   [x] |  [x]   |  [x] | [x] | [x] |  [ ] |  [ ] |
-| gpt2 |   [x]   |  [x]   |   [x] |  [x]   |  [x] | [x] | [x] |  [x] |  [x] |
-| opt |   [x]   |  [x]   |   [x] |  [x]   |  [x] | [x] | [x] |  [ ] |  [ ] |
-| bloom |   [x]   |  [x]   |   [x] |  [x]   |  [x] | [x] | [x] |  [x] |  [x] |
-| chatglm2 |   [x]   |  [x]   |   [x] |  [x]   |  [x] | [x] | [x] |  [x] |  [x] |
-| vit |   [x]   |  [x]   |   [ ] |  [x]   |  [x] | [x] | [x] |  [ ] |  [ ] |
-| whisper |   [x]   |  [x]   |   [x] |  [x]   |  [x] | [ ] | [x] |  [ ] |  [ ] |
-| sam |   [x]   |  [ ]   |   [ ] |  [x]   |  [x] | [x] | [x] |  [ ] |  [ ] |
-| blip2 |   [x]   |  [ ]   |   [ ] |  [x]   |  [x] | [x] | [x] |  [ ] |  [ ] |
+| bert |   [√]   |  [√]   |   [√] |  [√]   |  [√] | [√] | [√] |  [√] |  [√] |
+| t5 |   [√]   |  [√]   |   [√] |  [√]   |  [√] | [√] | [√] |  [ ] |  [ ] |
+| llama V1/V2 |   [√]   |  [√]   |   [√] |  [√]   |  [√] | [√] | [√] |  [ ] |  [ ] |
+| gpt2 |   [√]   |  [√]   |   [√] |  [√]   |  [√] | [√] | [√] |  [√] |  [√] |
+| opt |   [√]   |  [√]   |   [√] |  [√]   |  [√] | [√] | [√] |  [ ] |  [ ] |
+| bloom |   [√]   |  [√]   |   [√] |  [√]   |  [√] | [√] | [√] |  [√] |  [√] |
+| chatglm2 |   [√]   |  [√]   |   [√] |  [√]   |  [√] | [√] | [√] |  [√] |  [√] |
+| vit |   [√]   |  [√]   |   [ ] |  [√]   |  [√] | [√] | [√] |  [ ] |  [ ] |
+| whisper |   [√]   |  [√]   |   [√] |  [√]   |  [√] | [ ] | [√] |  [ ] |  [ ] |
+| sam |   [√]   |  [ ]   |   [ ] |  [√]   |  [√] | [√] | [√] |  [ ] |  [ ] |
+| blip2 |   [√]   |  [ ]   |   [ ] |  [√]   |  [√] | [√] | [√] |  [ ] |  [ ] |
+| falcon |   [√]   |  [√]   |   [√] |  [√]   |  [√] | [ ] | [√] |  [ ] |  [ ] |
 | roberta |   [ ]   |  [ ]   |   [ ] |  [ ]   |  [ ] | [ ] | [ ] |  [ ] |  [ ] |
 | albert |   [ ]   |  [ ]   |   [ ] |  [ ]   |  [ ] | [ ] | [ ] |  [ ] |  [ ] |
 | ernie |   [ ]   |  [ ]   |   [ ] |  [ ]   |  [ ] | [ ] | [ ] |  [ ] |  [ ] |
@@ -136,6 +137,7 @@ We will follow this roadmap to develop Shardformer:
 | swin |   [ ]   |  [ ]   |   [ ] |  [ ]   |  [ ] | [ ] | [ ] |  [ ] |  [ ] |
 | swin V2 |   [ ]   |  [ ]   |   [ ] |  [ ]   |  [ ] | [ ] | [ ] |  [ ] |  [ ] |
 | qwen |   [ ]   |  [ ]   |   [ ] |  [ ]   |  [ ] | [ ] | [ ] |  [ ] |  [ ] |
+| mistral |   [√]   |  [ ]   |   [ ] |  [√]   |  [√] | [√] | [√] |  [ ] |  [ ] |
 
 
 ## 💡 API Design

colossalai/shardformer/_utils.py

@@ -32,7 +32,7 @@ def set_obj_list_element(obj, attr: str, value):
     r"""
     Set the element to value of a list object
 
-    It used like set_obj_list_element(obj, 'lyaers[0]', new_layer), it will set obj.layers[0] to value
+    It used like set_obj_list_element(obj, 'layers[0]', new_layer), it will set obj.layers[0] to value
 
     Args:
         obj (object): The object to set

colossalai/shardformer/layer/_operation.py

@@ -7,6 +7,12 @@ try:
 except:
     fused_mix_prec_layer_norm_cuda = None
 
+try:
+    import fused_weight_gradient_mlp_cuda
+    _grad_accum_fusion_available = True
+except ImportError:
+    _grad_accum_fusion_available = False
+
 
 class FusedLayerNormAffineFunction1D(torch.autograd.Function):
     r"""Layernorm
@@ -141,7 +147,19 @@ class LinearWithAsyncCommunication(torch.autograd.Function):
             # all-reduce scheduled first and have GPU resources allocated
             _ = torch.empty(1, device=grad_output.device) + 1
 
-        grad_weight = grad_output.t().matmul(total_input)
+        if _grad_accum_fusion_available and weight.grad is not None:
+            grad = weight.grad
+            if grad.dtype == torch.float32:
+                fused_weight_gradient_mlp_cuda.wgrad_gemm_accum_fp32(total_input, grad_output, grad)
+                grad_weight = None
+            elif grad.dtype == torch.float16:
+                fused_weight_gradient_mlp_cuda.wgrad_gemm_accum_fp16(total_input, grad_output, grad)
+                grad_weight = None
+            else:
+                grad_weight = grad_output.t().matmul(total_input)
+        else:
+            grad_weight = grad_output.t().matmul(total_input)
+
         grad_bias = grad_output.sum(dim=0) if use_bias else None
 
         if ctx.async_grad_allreduce:
@@ -214,7 +232,19 @@ class _LinearWithGatherForwardReduceScatterBackward(torch.autograd.Function):
                 # reduce-scatter scheduled first and have GPU resources allocated
                 _ = torch.empty(1, device=grad_output.device) + 1
 
-            grad_weight = grad_output.t().matmul(total_input)
+            if _grad_accum_fusion_available and weight.grad is not None:
+                grad = weight.grad
+                if grad.dtype == torch.float32:
+                    fused_weight_gradient_mlp_cuda.wgrad_gemm_accum_fp32(total_input, grad_output, grad)
+                    grad_weight = None
+                elif grad.dtype == torch.float16:
+                    fused_weight_gradient_mlp_cuda.wgrad_gemm_accum_fp16(total_input, grad_output, grad)
+                    grad_weight = None
+                else:
+                    grad_weight = grad_output.t().matmul(total_input)
+            else:
+                grad_weight = grad_output.t().matmul(total_input)
+
             grad_bias = grad_output.sum(dim=0) if use_bias else None
 
             if ctx.async_grad_reduce_scatter:
@@ -249,7 +279,20 @@ class _LinearWithGatherForwardReduceScatterBackward(torch.autograd.Function):
             # calculate gradient
             if len(input_parallel.shape) > 2:
                 input_parallel = input_parallel.view(-1, input_parallel.shape[-1])
-            grad_weight = grad_output.t().matmul(input_parallel)
+
+            if _grad_accum_fusion_available and weight.grad is not None:
+                grad = weight.grad
+                if grad.dtype == torch.float32:
+                    fused_weight_gradient_mlp_cuda.wgrad_gemm_accum_fp32(input_parallel, grad_output, grad)
+                    grad_weight = None
+                elif grad.dtype == torch.float16:
+                    fused_weight_gradient_mlp_cuda.wgrad_gemm_accum_fp16(input_parallel, grad_output, grad)
+                    grad_weight = None
+                else:
+                    grad_weight = grad_output.t().matmul(input_parallel)
+            else:
+                grad_weight = grad_output.t().matmul(input_parallel)
+            # grad_weight = grad_output.t().matmul(input_parallel)
             # wait until reduce-scatter finished
             reducescatter_handle.wait()
 
@@ -388,7 +431,7 @@ class _MatmulWithGatherForwardReduceScatterBackward(torch.autograd.Function):
             input_parallel = torch.cat(tensor_list, dim=dim).contiguous()
             # calculate gradient
             if len(input_parallel.shape) > 2:
-                input_parallel = input_parallel.view(-1, input_parallel.shape[-1])
+                input_parallel = input_parallel.view(-1, input_parallel.shape[-1])   
             grad_weight = input_parallel.t().matmul(grad_output)
             # wait until reduce-scatter finished
             reducescatter_handle.wait()
@@ -473,16 +516,17 @@ class _GatherForwardSplitBackward(torch.autograd.Function):
     @staticmethod
     def backward(ctx, grad_output):
         return _split(grad_output, ctx.dim, ctx.process_group), None, None
-    
+
 
 class HookParameter(torch.autograd.Function):
     """In order to be hooked into Gemini's '__torch_function__', adding a view operation to weight and bias. Used in FusedLayerNorm"""
+
     @staticmethod
     def forward(ctx, input, weight, bias):
         ctx.save_for_backward(weight, bias)
         output = input
         return output
-    
+
     @staticmethod
     def backward(ctx, grad_output):
         weight, bias = ctx.saved_tensors
@@ -491,13 +535,12 @@ class HookParameter(torch.autograd.Function):
         if bias is not None:
             bias = bias.view(bias.shape)
         return grad_output, None, None
-    
+
 
 def hook_paramter_in_backward(input, weight=None, bias=None):
     return HookParameter.apply(input, weight, bias)
 
 
-
 def _reduce(input_, process_group):
     # skip if only one rank involved
     if dist.get_world_size(process_group) == 1:
@@ -522,7 +565,7 @@ def _split(input_, dim=-1, process_group=None):
 
     tensor_list = torch.split(input_, dim_size // world_size, dim=dim)
     rank = dist.get_rank(process_group)
-    output = tensor_list[rank].contiguous()
+    output = tensor_list[rank].clone().contiguous()
 
     return output
 

colossalai/shardformer/layer/linear.py

@@ -408,7 +408,7 @@ class Linear1D_Row(ParallelModule):
                     handle.wait()
                 output = torch.cat(output_parallel_list, dim=-1)
         else:
-            output_parallel = F.linear(input_, self.weight)
+            output_parallel = linear_with_async_comm(input_, self.weight, None, None, False)
             if self.seq_parallel:
                 output = linear_reducescatter_forward_gather_backward(
                     output_parallel, self.process_group, self.seq_parallel_dim

colossalai/shardformer/layer/loss.py

@@ -78,10 +78,13 @@ class DistCrossEntropy(Function):
         # calculate the loss
         # loss = log(sum(exp(x[i]))) - x[class]
         loss = torch.where(target == ignore_index, 0.0, torch.log(sum_exp_logits) - pred_logits)
-        loss = torch.sum(loss).div_(torch.sum(loss != 0.0))
+        num_non_zero = torch.sum(loss != 0.0)
+        ctx.inv_num_non_zero = 1.0 / num_non_zero
+        loss = torch.sum(loss).div_(num_non_zero)
 
         # calculate the softmax
         exp_logits.div_(sum_exp_logits.unsqueeze(dim=-1))
+        exp_logits[target == ignore_index] = 0.0
         ctx.save_for_backward(exp_logits, mask, masked_target_1d)
 
         return loss
@@ -89,6 +92,7 @@ class DistCrossEntropy(Function):
     @staticmethod
     def backward(ctx, grad_output):
         # retrieve the saved tensors
+        grad_output = grad_output * ctx.inv_num_non_zero
         exp_logits, mask, masked_target_1d = ctx.saved_tensors
 
         # use exp logits as the input grad
@@ -100,7 +104,7 @@ class DistCrossEntropy(Function):
         grad_logits_2d[torch.arange(0, grad_logits_2d.shape[0]), masked_target_1d] -= update
 
         grad_logits.mul_(grad_output.unsqueeze(dim=-1))
-        return grad_logits, None, None
+        return grad_logits, None, None, None
 
 
 def cross_entropy_1d(

colossalai/shardformer/layer/normalization.py

@@ -275,8 +275,8 @@ class FusedRMSNorm(BaseLayerNorm):
             )
 
         LazyInitContext.materialize(module)
-        # to check if it is huggingface LlamaRMSNorm
-        if module.__class__.__name__ == "LlamaRMSNorm":
+        # to check if it is huggingface LlamaRMSNorm or MistralRMSNorm
+        if module.__class__.__name__ in ["LlamaRMSNorm", "MistralRMSNorm"]:
             normalized_shape = module.weight.shape[0]
             eps = module.variance_epsilon
             elementwise_affine = True

colossalai/shardformer/modeling/falcon.py

+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.distributed as dist
+from torch.distributed import ProcessGroup
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.models.falcon.modeling_falcon import (
+    FalconForCausalLM,
+    FalconForQuestionAnswering,
+    FalconForSequenceClassification,
+    FalconForTokenClassification,
+    FalconModel,
+    build_alibi_tensor,
+)
+from transformers.utils import logging
+
+from colossalai.pipeline.stage_manager import PipelineStageManager
+from colossalai.shardformer.shard import ShardConfig
+
+
+def build_falcon_alibi_tensor_fn(process_group: ProcessGroup) -> torch.Tensor:
+    def build_falcon_alibi_tensor(
+        self, attention_mask: torch.Tensor, num_heads: int, dtype: torch.dtype
+    ) -> torch.Tensor:
+        """
+        Link to paper: https://arxiv.org/abs/2108.12409 Alibi tensor is not causal as the original paper mentions, it
+        relies on a translation invariance of softmax for quick implementation: with l being a tensor, and a fixed value
+        `softmax(l+a) = softmax(l)`. Based on
+        https://github.com/ofirpress/attention_with_linear_biases/blob/a35aaca144e0eb6b789dfcb46784c4b8e31b7983/fairseq/models/transformer.py#L742
+        TODO @thomasw21 this doesn't work as nicely due to the masking strategy, and so masking varies slightly.
+
+        Args:
+        Returns tensor shaped (batch_size * num_heads, 1, max_seq_len)
+            attention_mask (`torch.Tensor`):
+                Token-wise attention mask, this should be of shape (batch_size, max_seq_len).
+            num_heads (`int`, *required*):
+                number of heads
+            dtype (`torch.dtype`, *optional*, default=`torch.bfloat16`):
+                dtype of the output tensor
+        """
+        import math
+
+        if dist.is_initialized():
+            world_size = dist.get_world_size(process_group)
+            num_heads = num_heads * world_size
+
+        batch_size, seq_length = attention_mask.shape
+        closest_power_of_2 = 2 ** math.floor(math.log2(num_heads))
+        base = torch.tensor(
+            2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32
+        )
+        powers = torch.arange(1, 1 + closest_power_of_2, device=attention_mask.device, dtype=torch.int32)
+        slopes = torch.pow(base, powers)
+
+        if closest_power_of_2 != num_heads:
+            extra_base = torch.tensor(
+                2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))),
+                device=attention_mask.device,
+                dtype=torch.float32,
+            )
+            num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2)
+            extra_powers = torch.arange(
+                1, 1 + 2 * num_remaining_heads, 2, device=attention_mask.device, dtype=torch.int32
+            )
+            slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)
+
+        # Note: alibi will added to the attention bias that will be applied to the query, key product of attention
+        # => therefore alibi will have to be of shape (batch_size, num_heads, query_length, key_length)
+        # => here we set (batch_size=1, num_heads=num_heads, query_length=1, key_length=max_length)
+        # => the query_length dimension will then be broadcasted correctly
+        # This is more or less identical to T5's relative position bias:
+        # https://github.com/huggingface/transformers/blob/f681437203baa7671de3174b0fa583c349d9d5e1/src/transformers/models/t5/modeling_t5.py#L527
+        arange_tensor = ((attention_mask.cumsum(dim=-1) - 1) * attention_mask)[:, None, :]
+        alibi = slopes[..., None] * arange_tensor
+        if dist.is_initialized():
+            num_heads_per_rank = int(num_heads / dist.get_world_size(process_group))
+            offset = dist.get_rank(process_group) * num_heads_per_rank
+            alibi = alibi.view(batch_size, num_heads, 1, seq_length)
+            alibi = alibi[:, offset : num_heads_per_rank + offset, :, :]
+            return alibi.reshape(batch_size * num_heads_per_rank, 1, seq_length).to(dtype)
+        else:
+            return alibi.reshape(batch_size * num_heads, 1, seq_length).to(dtype)
+
+    return build_falcon_alibi_tensor
+
+
+def get_tp_falcon_decoder_layer_forward():
+    from transformers.models.falcon.modeling_falcon import FalconDecoderLayer, dropout_add
+
+    def forward(
+        self: FalconDecoderLayer,
+        hidden_states: torch.Tensor,
+        alibi: Optional[torch.Tensor],
+        attention_mask: torch.Tensor,
+        layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        use_cache: bool = False,
+        output_attentions: bool = False,
+    ):
+        residual = hidden_states
+
+        if self.config.new_decoder_architecture:
+            attention_layernorm_out = self.ln_attn(hidden_states)
+            mlp_layernorm_out = self.ln_mlp(hidden_states)
+        else:
+            attention_layernorm_out = self.input_layernorm(hidden_states)
+
+        # Self attention.
+        attn_outputs = self.self_attention(
+            attention_layernorm_out,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            alibi=alibi,
+            head_mask=head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+        )
+
+        attention_output = attn_outputs[0]
+
+        if not self.config.new_decoder_architecture:
+            if self.config.parallel_attn:
+                mlp_layernorm_out = attention_layernorm_out
+            else:
+                residual = dropout_add(
+                    attention_output, residual, self.config.attention_dropout, training=self.training
+                )
+                mlp_layernorm_out = self.post_attention_layernorm(residual)
+
+        outputs = attn_outputs[1:]
+
+        # MLP.
+        mlp_output = self.mlp(mlp_layernorm_out)
+
+        if self.config.new_decoder_architecture or self.config.parallel_attn:
+            mlp_output = mlp_output + attention_output
+
+        output = dropout_add(mlp_output, residual, self.config.hidden_dropout, training=self.training)
+
+        if use_cache:
+            outputs = (output,) + outputs
+        else:
+            outputs = (output,) + outputs[1:]
+
+        return outputs  # hidden_states, present, attentions
+
+    return forward
+
+
+def get_falcon_flash_attention_forward():
+    try:
+        from xformers.ops import memory_efficient_attention as me_attention
+    except:
+        raise ImportError("Error: xformers module is not installed. Please install it to use flash attention.")
+    from transformers.models.falcon.modeling_falcon import FalconAttention
+
+    def forward(
+        self: FalconAttention,
+        hidden_states: torch.Tensor,
+        alibi: Optional[torch.Tensor],
+        attention_mask: torch.Tensor,
+        layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        use_cache: bool = False,
+        output_attentions: bool = False,
+    ):
+        fused_qkv = self.query_key_value(hidden_states)  # [batch_size, seq_length, 3 x hidden_size]
+        num_kv_heads = self.num_heads if self.new_decoder_architecture else self.num_kv_heads
+        # 3 x [batch_size, seq_length, num_heads, head_dim]
+        (query_layer, key_layer, value_layer) = self._split_heads(fused_qkv)
+
+        batch_size, query_length, _, _ = query_layer.shape
+
+        query_layer = query_layer.transpose(1, 2).reshape(batch_size * self.num_heads, query_length, self.head_dim)
+        key_layer = key_layer.transpose(1, 2).reshape(
+            batch_size * num_kv_heads,
+            query_length,
+            self.head_dim,
+        )
+        value_layer = value_layer.transpose(1, 2).reshape(batch_size * num_kv_heads, query_length, self.head_dim)
+
+        past_kv_length = 0 if layer_past is None else layer_past[0].shape[1]
+        query_layer, key_layer = self.maybe_rotary(query_layer, key_layer, past_kv_length)
+
+        if layer_past is not None:
+            past_key, past_value = layer_past
+            # concatenate along seq_length dimension:
+            #  - key: [batch_size * self.num_heads, kv_length, head_dim]
+            #  - value: [batch_size * self.num_heads, kv_length, head_dim]
+            key_layer = torch.cat((past_key, key_layer), dim=1)
+            value_layer = torch.cat((past_value, value_layer), dim=1)
+
+        _, kv_length, _ = key_layer.shape
+        if use_cache:
+            present = (key_layer, value_layer)
+        else:
+            present = None
+
+        attention_mask_float = (attention_mask * 1.0).masked_fill(attention_mask, float("-1e9")).to(query_layer.dtype)
+
+        query_layer_ = query_layer.reshape(batch_size, self.num_heads, -1, self.head_dim).transpose(1, 2).contiguous()
+        key_layer_ = key_layer.reshape(batch_size, num_kv_heads, -1, self.head_dim).transpose(1, 2).contiguous()
+        value_layer_ = value_layer.reshape(batch_size, num_kv_heads, -1, self.head_dim).transpose(1, 2).contiguous()
+
+        if alibi is not None:
+            attention_mask_float = (
+                attention_mask_float + alibi.view(batch_size, self.num_heads, 1, kv_length) * self.beta
+            )
+
+        batch_size, src_len = query_layer_.size()[0], query_layer_.size()[1]
+        tgt_len = key_layer_.size()[1]
+        attention_mask_float = attention_mask_float.expand(batch_size, self.num_heads, src_len, tgt_len).contiguous()
+        context_layer = me_attention(
+            query_layer_,
+            key_layer_,
+            value_layer_,
+            attn_bias=attention_mask_float,
+            scale=self.inv_norm_factor,
+            p=self.attention_dropout.p,
+        )
+        batch_size, seq_length, _, _ = context_layer.shape
+        context_layer = context_layer.reshape(batch_size, seq_length, -1)
+
+        output_tensor = self.dense(context_layer)
+
+        return output_tensor, present
+
+    return forward
+
+
+class FalconPipelineForwards:
+    """
+    This class serves as a micro library for falcon pipeline forwards.
+    """
+
+    @staticmethod
+    def falcon_model_forward(
+        self: FalconModel,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        stage_manager: Optional[PipelineStageManager] = None,
+        hidden_states: Optional[torch.FloatTensor] = None,
+        stage_index: Optional[List[int]] = None,
+        shard_config: ShardConfig = None,
+    ) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPastAndCrossAttentions]:
+        logger = logging.get_logger(__name__)
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        if use_cache:
+            logger.warning_once("use_cache=True is not supported for pipeline models at the moment.")
+            use_cache = False
+
+        if past_key_values is not None:
+            logger.warning_once("past_key_values is not supported for pipeline models at the moment.")
+            past_key_values = None
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if past_key_values is None:
+            past_key_values = tuple([None] * len(self.h))
+        else:
+            past_key_values = self._convert_to_rw_cache(past_key_values)
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape batch_size x num_heads x N x N
+        # head_mask has shape n_layer x batch x num_heads x N x N
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        # case: First stage of training
+        if stage_manager.is_first_stage():
+            if input_ids is not None and inputs_embeds is not None:
+                raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+            elif input_ids is not None:
+                batch_size, seq_length = input_ids.shape
+            elif inputs_embeds is not None:
+                batch_size, seq_length, _ = inputs_embeds.shape
+            else:
+                raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+            if inputs_embeds is None:
+                inputs_embeds = self.word_embeddings(input_ids)
+
+            hidden_states = inputs_embeds
+
+        else:
+            input_shape = hidden_states.shape[:-1]
+            batch_size, seq_length = input_shape
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        # Compute alibi tensor: check build_alibi_tensor documentation
+        past_key_values_length = 0
+        if past_key_values[0] is not None:
+            past_key_values_length = past_key_values[0][0].shape[1]  # 1 because RW-cache, not standard format
+        if attention_mask is None:
+            attention_mask = torch.ones((batch_size, seq_length + past_key_values_length), device=hidden_states.device)
+        else:
+            attention_mask = attention_mask.to(hidden_states.device)
+
+        if self.use_alibi:
+            alibi = build_alibi_tensor(attention_mask, self.num_heads, dtype=hidden_states.dtype)
+        else:
+            alibi = None
+
+        causal_mask = self._prepare_attn_mask(
+            attention_mask,
+            input_shape=(batch_size, seq_length),
+            past_key_values_length=past_key_values_length,
+        )
+
+        start_idx, end_idx = stage_index[0], stage_index[1]
+        for i, (block, layer_past) in enumerate(
+            zip(self.h[start_idx:end_idx], past_key_values[start_idx:end_idx]), start=start_idx
+        ):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, use_cache=use_cache, output_attentions=output_attentions)
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    alibi,
+                    causal_mask,
+                    head_mask[i],
+                )
+            else:
+                outputs = block(
+                    hidden_states,
+                    layer_past=layer_past,
+                    attention_mask=causal_mask,
+                    head_mask=head_mask[i],
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    alibi=alibi,
+                )
+
+            hidden_states = outputs[0]
+            if use_cache is True:
+                presents = presents + (outputs[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+
+        if stage_manager.is_last_stage():
+            # Add last hidden state
+            hidden_states = self.ln_f(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if presents is not None:
+            presents = self._convert_cache_to_standard_format(presents, batch_size)
+
+        if stage_manager.is_last_stage():
+            if not return_dict:
+                return tuple(
+                    v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None
+                )
+            return BaseModelOutputWithPastAndCrossAttentions(
+                last_hidden_state=hidden_states,
+                past_key_values=presents,
+                hidden_states=all_hidden_states,
+                attentions=all_self_attentions,
+            )
+        else:
+            # always return dict for imediate stage
+            return {"hidden_states": hidden_states}
+
+    @staticmethod
+    def falcon_for_causal_lm_forward(
+        self: FalconForCausalLM,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        stage_manager: Optional[PipelineStageManager] = None,
+        hidden_states: Optional[torch.FloatTensor] = None,
+        stage_index: Optional[List[int]] = None,
+        shard_config: ShardConfig = None,
+    ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        """
+        logger = logging.get_logger(__name__)
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if output_attentions:
+            logger.warning_once("output_attentions=True is not supported for pipeline models at the moment.")
+            output_attentions = False
+        if output_hidden_states:
+            logger.warning_once("output_hidden_states=True is not supported for pipeline models at the moment.")
+            output_hidden_states = False
+
+        transformer_outputs = FalconPipelineForwards.falcon_model_forward(
+            self.transformer,
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            stage_manager=stage_manager,
+            hidden_states=hidden_states,
+            stage_index=stage_index,
+            shard_config=shard_config,
+        )
+
+        past_key_values = None
+        if stage_manager.is_last_stage():
+            hidden_states = transformer_outputs[0]
+            lm_logits = self.lm_head(hidden_states)
+
+            loss = None
+            if labels is not None:
+                # Shift so that tokens < n predict n
+                shift_logits = lm_logits[..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous()
+                batch_size, seq_length, vocab_size = shift_logits.shape
+                # Flatten the tokens
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(
+                    shift_logits.view(batch_size * seq_length, vocab_size), shift_labels.view(batch_size * seq_length)
+                )
+
+            if not return_dict:
+                output = (lm_logits,) + transformer_outputs[1:]
+                return ((loss,) + output) if loss is not None else output
+
+            return CausalLMOutputWithCrossAttentions(
+                loss=loss,
+                logits=lm_logits,
+                past_key_values=transformer_outputs.past_key_values,
+                hidden_states=transformer_outputs.hidden_states,
+                attentions=transformer_outputs.attentions,
+            )
+
+        else:
+            hidden_states = transformer_outputs.get("hidden_states")
+            return {"hidden_states": hidden_states}
+
+    @staticmethod
+    def falcon_for_sequence_classification_forward(
+        self: FalconForSequenceClassification,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        stage_manager: Optional[PipelineStageManager] = None,
+        hidden_states: Optional[torch.FloatTensor] = None,
+        stage_index: Optional[List[int]] = None,
+        shard_config: ShardConfig = None,
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        logger = logging.get_logger(__name__)
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if output_attentions:
+            logger.warning_once("output_attentions=True is not supported for pipeline models at the moment.")
+            output_attentions = False
+        if output_hidden_states:
+            logger.warning_once("output_hidden_states=True is not supported for pipeline models at the moment.")
+            output_hidden_states = False
+
+        transformer_outputs = FalconPipelineForwards.falcon_model_forward(
+            self.transformer,
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            stage_manager=stage_manager,
+            hidden_states=hidden_states,
+            stage_index=stage_index,
+            shard_config=shard_config,
+        )
+
+        past_key_values = None
+        if stage_manager.is_last_stage():
+            batch_size = hidden_states.shape[0]
+            hidden_states = transformer_outputs[0]
+            logits = self.score(hidden_states)
+
+            if self.config.pad_token_id is None and batch_size != 1:
+                raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+            if self.config.pad_token_id is None:
+                sequence_lengths = -1
+            else:
+                if input_ids is not None:
+                    sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(dim=-1) - 1).to(logits.device)
+                else:
+                    sequence_lengths = -1
+                    logger.warning(
+                        f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                        "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+                    )
+
+            pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+            loss = None
+            if labels is not None:
+                if self.config.problem_type is None:
+                    if self.num_labels == 1:
+                        self.config.problem_type = "regression"
+                    elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                        self.config.problem_type = "single_label_classification"
+                    else:
+                        self.config.problem_type = "multi_label_classification"
+
+                if self.config.problem_type == "regression":
+                    loss_fct = MSELoss()
+                    if self.num_labels == 1:
+                        loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                    else:
+                        loss = loss_fct(pooled_logits, labels)
+                elif self.config.problem_type == "single_label_classification":
+                    loss_fct = CrossEntropyLoss()
+                    loss = loss_fct(pooled_logits, labels)
+                elif self.config.problem_type == "multi_label_classification":
+                    loss_fct = BCEWithLogitsLoss()
+                    loss = loss_fct(pooled_logits, labels)
+            if not return_dict:
+                output = (pooled_logits,) + transformer_outputs[1:]
+                return ((loss,) + output) if loss is not None else output
+
+            return SequenceClassifierOutputWithPast(
+                loss=loss,
+                logits=pooled_logits,
+                past_key_values=transformer_outputs.past_key_values,
+                hidden_states=transformer_outputs.hidden_states,
+                attentions=transformer_outputs.attentions,
+            )
+        else:
+            hidden_states = transformer_outputs.get("hidden_states")
+            return {"hidden_states": hidden_states}
+
+    @staticmethod
+    def falcon_for_token_classification_forward(
+        self: FalconForTokenClassification,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        stage_manager: Optional[PipelineStageManager] = None,
+        hidden_states: Optional[torch.FloatTensor] = None,
+        stage_index: Optional[List[int]] = None,
+        shard_config: ShardConfig = None,
+    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        logger = logging.get_logger(__name__)
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if output_attentions:
+            logger.warning_once("output_attentions=True is not supported for pipeline models at the moment.")
+            output_attentions = False
+        if output_hidden_states:
+            logger.warning_once("output_hidden_states=True is not supported for pipeline models at the moment.")
+            output_hidden_states = False
+
+        transformer_outputs = FalconPipelineForwards.falcon_model_forward(
+            self.transformer,
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            stage_manager=stage_manager,
+            hidden_states=hidden_states,
+            stage_index=stage_index,
+            shard_config=shard_config,
+        )
+
+        past_key_values = None
+
+        if stage_manager.is_last_stage():
+            hidden_states = transformer_outputs[0]
+            hidden_states = self.dropout(hidden_states)
+            logits = self.classifier(hidden_states)
+
+            loss = None
+            if labels is not None:
+                batch_size, seq_length = labels.shape
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(
+                    logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length)
+                )
+
+            if not return_dict:
+                output = (logits,) + transformer_outputs[2:]
+                return ((loss,) + output) if loss is not None else output
+
+            return TokenClassifierOutput(
+                loss=loss,
+                logits=logits,
+                hidden_states=transformer_outputs.hidden_states,
+                attentions=transformer_outputs.attentions,
+            )
+
+        else:
+            hidden_states = transformer_outputs.get("hidden_states")
+            return {"hidden_states": hidden_states}
+
+    @staticmethod
+    def falcon_for_question_answering_forward(
+        self: FalconForQuestionAnswering,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        stage_manager: Optional[PipelineStageManager] = None,
+        hidden_states: Optional[torch.FloatTensor] = None,
+        stage_index: Optional[List[int]] = None,
+        shard_config: ShardConfig = None,
+    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        """
+
+        logger = logging.get_logger(__name__)
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if output_attentions:
+            logger.warning_once("output_attentions=True is not supported for pipeline models at the moment.")
+            output_attentions = False
+        if output_hidden_states:
+            logger.warning_once("output_hidden_states=True is not supported for pipeline models at the moment.")
+            output_hidden_states = False
+
+        outputs = FalconPipelineForwards.falcon_model_forward(
+            self.transformer,
+            input_ids,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            stage_manager=stage_manager,
+            hidden_states=hidden_states,
+            stage_index=stage_index,
+            shard_config=shard_config,
+        )
+
+        if stage_manager.is_last_stage():
+            sequence_output = outputs[0]
+            logits = self.qa_outputs(sequence_output)
+            start_logits, end_logits = logits.split(1, dim=-1)
+            start_logits = start_logits.squeeze(-1).contiguous()
+            end_logits = end_logits.squeeze(-1).contiguous()
+
+            total_loss = None
+            if start_positions is not None and end_positions is not None:
+                # If we are on multi-GPU, split add a dimension
+                if len(start_positions.size()) > 1:
+                    start_positions = start_positions.squeeze(-1)
+                if len(end_positions.size()) > 1:
+                    end_positions = end_positions.squeeze(-1)
+                # sometimes the start/end positions are outside our model inputs, we ignore these terms
+                ignored_index = start_logits.size(1)
+                start_positions = start_positions.clamp(0, ignored_index)
+                end_positions = end_positions.clamp(0, ignored_index)
+
+                loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+                start_loss = loss_fct(start_logits, start_positions)
+                end_loss = loss_fct(end_logits, end_positions)
+                total_loss = (start_loss + end_loss) / 2
+
+            if not return_dict:
+                output = (start_logits, end_logits) + outputs[2:]
+                return ((total_loss,) + output) if total_loss is not None else output
+
+            return QuestionAnsweringModelOutput(
+                loss=total_loss,
+                start_logits=start_logits,
+                end_logits=end_logits,
+                hidden_states=outputs.hidden_states,
+                attentions=outputs.attentions,
+            )
+        else:
+            hidden_states = outputs.get("hidden_states")
+            return {"hidden_states": hidden_states}

colossalai/shardformer/modeling/gptj.py

+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutputWithPast,
+)
+from transformers.models.gptj.modeling_gptj import (
+    GPTJForCausalLM,
+    GPTJForQuestionAnswering,
+    GPTJForSequenceClassification,
+    GPTJModel,
+    apply_rotary_pos_emb,
+    get_embed_positions,
+)
+from transformers.utils import is_torch_fx_proxy, logging
+
+from colossalai.pipeline.stage_manager import PipelineStageManager
+from colossalai.shardformer.layer._operation import gather_forward_split_backward, split_forward_gather_backward
+from colossalai.shardformer.shard import ShardConfig
+
+
+class GPTJPipelineForwards:
+    """
+    This class serves as a micro library for forward function substitution of GPTJ models
+    under pipeline setting.
+    """
+
+    @staticmethod
+    def gptj_model_forward(
+        self: GPTJModel,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        stage_manager: Optional[PipelineStageManager] = None,
+        hidden_states: Optional[torch.FloatTensor] = None,
+        stage_index: Optional[List[int]] = None,
+        shard_config: ShardConfig = None,
+    ) -> Union[Dict, Tuple, BaseModelOutputWithPast]:
+        # This function is modified on the basis of transformers.models.gptj.modeling_gptj.GPTJModel.forward.
+        # Please refer to original code of transformers for more details.
+        # GPTJ has no cross attention in comparison to GPT2
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        logger = logging.get_logger(__name__)
+
+        # Preprocess passed in arguments
+        # TODO(baizhou): left the recording kv-value tensors as () or None type, this feature may be added in the future.
+        if past_key_values:
+            logger.warning_once("Non-empty past_key_values is not supported for pipeline models at the moment.")
+            past_key_values = None
+        if output_attentions:
+            logger.warning_once("output_attentions=True is not supported for pipeline models at the moment.")
+            output_attentions = False
+        if output_hidden_states:
+            logger.warning_once("output_hidden_states=True is not supported for pipeline models at the moment.")
+            output_hidden_states = False
+        if use_cache:
+            logger.warning_once("use_cache=True is not supported for pipeline models at the moment.")
+            use_cache = False
+
+        if stage_manager.is_first_stage():
+            if input_ids is not None and inputs_embeds is not None:
+                raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+            elif input_ids is not None:
+                batch_size, seq_length = input_ids.shape
+                input_shape = input_ids.size()
+                input_ids = input_ids.view(-1, seq_length)
+
+            elif inputs_embeds is not None:
+                input_shape = inputs_embeds.size()[:-1]
+                batch_size = inputs_embeds.shape[0]
+            else:
+                raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids.view(-1, seq_length)
+        else:
+            if hidden_states is None:
+                raise ValueError("hidden_states shouldn't be None for stages other than the first stage.")
+            input_shape = hidden_states.size()[:-1]
+            batch_size, seq_length = input_shape[0], input_shape[1]
+            device = hidden_states.device
+
+        # Attention mask.
+        if attention_mask is not None:
+            if batch_size <= 0:
+                raise ValueError("batch_size has to be defined and > 0")
+            attention_mask = attention_mask.view(batch_size, -1)
+            # We create a 3D attention mask from a 2D tensor mask.
+            # Sizes are [batch_size, 1, 1, to_seq_length]
+            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+            # this attention mask is more simple than the triangular masking of causal attention
+            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+            attention_mask = attention_mask[:, None, None, :]
+
+            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+            # masked positions, this operation will create a tensor which is 0.0 for
+            # positions we want to attend and the dtype's smallest value for masked positions.
+            # Since we are adding it to the raw scores before the softmax, this is
+            # effectively the same as removing these entirely.
+            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+            attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x num_attention_heads x N x N
+        # head_mask has shape n_layer x batch x num_attention_heads x N x N
+        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
+
+        # position id to be asssigned not just for the first stage for attn input
+        if position_ids is not None:
+            position_ids = position_ids.view(-1, seq_length)
+        else:
+            position_ids = torch.arange(0, seq_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
+        if stage_manager.is_first_stage():
+            if inputs_embeds is None:
+                inputs_embeds = self.wte(input_ids)
+            hidden_states = inputs_embeds
+            if token_type_ids is not None:
+                token_type_embeds = self.wte(token_type_ids)
+                hidden_states = hidden_states + token_type_embeds
+            hidden_states = self.drop(hidden_states)
+
+        output_shape = input_shape + (hidden_states.size(-1),)
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        # split the input tensor along sequence dimension
+        # [batch_size, seq_len, hidden_size] -> [batch_size, seq_len/TP_size, hidden_size]
+        if shard_config.enable_sequence_parallelism:
+            hidden_states = split_forward_gather_backward(
+                hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group
+            )
+
+        # Going through held blocks.
+        start_idx, end_idx = stage_index[0], stage_index[1]
+        for i in range(start_idx, end_idx):
+            block = self.h[i]
+            torch.cuda.set_device(hidden_states.device)
+
+            # Ensure that attention_mask is always on the same device as hidden_states
+            if attention_mask is not None:
+                attention_mask = attention_mask.to(hidden_states.device)
+            if isinstance(head_mask, torch.Tensor):
+                head_mask = head_mask.to(hidden_states.device)
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, use_cache, output_attentions)
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    None,
+                    attention_mask,
+                    position_ids,
+                    head_mask[i],
+                )
+            else:
+                outputs = block(
+                    hidden_states=hidden_states,
+                    layer_past=None,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    head_mask=head_mask[i],
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = outputs[0]
+            if use_cache is True:
+                presents = presents + (outputs[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+
+        # When sequence parallelism done, gather the output tensor in forward and split it in backward
+        if shard_config.enable_sequence_parallelism:
+            hidden_states = gather_forward_split_backward(
+                hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group
+            )
+
+        if stage_manager.is_last_stage():
+            hidden_states = self.ln_f(hidden_states)
+
+        hidden_states = hidden_states.view(output_shape)
+        # Add last hidden state
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if stage_manager.is_last_stage():
+            if not return_dict:
+                return tuple(
+                    v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None
+                )
+
+            return BaseModelOutputWithPast(
+                last_hidden_state=hidden_states,
+                past_key_values=presents,
+                hidden_states=all_hidden_states,
+                attentions=all_self_attentions,
+            )
+        else:
+            # always return dict for intermediate stage
+            return {"hidden_states": hidden_states}
+
+    @staticmethod
+    def gptj_causallm_model_forward(
+        self: GPTJForCausalLM,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        stage_manager: Optional[PipelineStageManager] = None,
+        hidden_states: Optional[torch.FloatTensor] = None,
+        stage_index: Optional[List[int]] = None,
+        shard_config: ShardConfig = None,
+    ) -> Union[Dict, Tuple, CausalLMOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+
+        # This function is modified on the basis of transformers.models.gptj.modeling_gptj.GPTJForCausalLM.forward.
+        # Please refer to original code of transformers for more details.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = GPTJPipelineForwards.gptj_model_forward(
+            self.transformer,
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            stage_manager=stage_manager,
+            hidden_states=hidden_states,
+            stage_index=stage_index,
+            shard_config=shard_config,
+        )
+
+        # If not at the last stage, return hidden_states as in GPTJModel
+        if not stage_manager.is_last_stage():
+            return {"hidden_states": transformer_outputs["hidden_states"]}
+
+        hidden_states = transformer_outputs[0]
+        lm_logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(lm_logits.device)
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+            loss = loss.to(hidden_states.dtype)
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @staticmethod
+    def gptj_for_sequence_classification_forward(
+        self: GPTJForSequenceClassification,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        stage_manager: Optional[PipelineStageManager] = None,
+        hidden_states: Optional[torch.FloatTensor] = None,
+        stage_index: Optional[List[int]] = None,
+        shard_config: ShardConfig = None,
+    ) -> Union[Dict, Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+
+        # This function is modified on the basis of transformers.models.gptj.modeling_gptj.GPTJForSequenceClassification.forward.
+        # Please refer to original code of transformers for more details.
+        """
+        logger = logging.get_logger(__name__)
+
+        if input_ids is not None:
+            batch_size, _ = input_ids.shape[:2]
+        else:
+            batch_size, _ = hidden_states.shape[:2]
+        assert (
+            self.config.pad_token_id is not None or batch_size == 1
+        ), "Cannot handle batch sizes > 1 if no padding token is defined."
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = GPTJPipelineForwards.gptj_model_forward(
+            self.transformer,
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            stage_manager=stage_manager,
+            hidden_states=hidden_states,
+            stage_index=stage_index,
+            shard_config=shard_config,
+        )
+
+        # If not at the last stage, return hidden_states as in GPTJModel
+        if not stage_manager.is_last_stage():
+            return {"hidden_states": transformer_outputs["hidden_states"]}
+
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
+            else:
+                sequence_lengths = -1
+                logger.warning_once(
+                    f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                    "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+                )
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(pooled_logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @staticmethod
+    def gptj_for_question_answering_forward(
+        self: GPTJForQuestionAnswering,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        stage_manager: Optional[PipelineStageManager] = None,
+        hidden_states: Optional[torch.FloatTensor] = None,
+        stage_index: Optional[List[int]] = None,
+        shard_config: ShardConfig = None,
+    ) -> Union[Dict, Tuple, QuestionAnsweringModelOutput]:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+
+        # This function is modified on the basis of transformers.models.gptj.modeling_gptj.GPTJForQuestionAnswering.forward.
+        # Please refer to original code of transformers for more details.
+
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = GPTJPipelineForwards.gptj_model_forward(
+            self.transformer,
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            stage_manager=stage_manager,
+            hidden_states=hidden_states,
+            stage_index=stage_index,
+            shard_config=shard_config,
+        )
+
+        # If not at the last stage, return hidden_states as in GPTJModel
+        if not stage_manager.is_last_stage():
+            return {"hidden_states": outputs["hidden_states"]}
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1).to(start_logits.device)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1).to(end_logits.device)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+def get_gptj_flash_attention_forward():
+    from transformers.models.gptj.modeling_gptj import GPTJAttention
+
+    from colossalai.kernel.cuda_native import AttnMaskType, ColoAttention
+
+    def split_heads(tensor, num_attention_heads, attn_head_size, rotary):
+        """
+        Splits hidden dim into attn_head_size and num_attention_heads
+        """
+        new_shape = tensor.size()[:-1] + (num_attention_heads, attn_head_size)
+        tensor = tensor.view(new_shape)
+        if rotary or len(tensor.shape) in [4, 5]:
+            return tensor
+        else:
+            raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}")
+
+    def forward(
+        self: GPTJAttention,
+        hidden_states: torch.FloatTensor,
+        layer_past: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+    ) -> Union[
+        Tuple[torch.Tensor, Tuple[torch.Tensor]],
+        Optional[Tuple[torch.Tensor, Tuple[torch.Tensor], Tuple[torch.Tensor, ...]]],
+    ]:
+        query = self.q_proj(hidden_states)
+        key = self.k_proj(hidden_states)
+        value = self.v_proj(hidden_states)
+
+        query = split_heads(query, self.num_attention_heads, self.head_dim, True)
+        key = split_heads(key, self.num_attention_heads, self.head_dim, True)
+        value = split_heads(value, self.num_attention_heads, self.head_dim, False)
+
+        if is_torch_fx_proxy(position_ids) or torch.jit.is_tracing():
+            # The logic to conditionally copy to GPU could not be traced, so we do this
+            # every time in the torch.fx case
+            embed_positions = get_embed_positions(self.embed_positions, position_ids)
+        else:
+            embed_positions = self._get_embed_positions(position_ids)
+
+        repeated_position_ids = position_ids.unsqueeze(-1).repeat(1, 1, embed_positions.shape[-1])
+        sincos = torch.gather(embed_positions, 1, repeated_position_ids)
+        sin, cos = torch.split(sincos, sincos.shape[-1] // 2, dim=-1)
+
+        if self.rotary_dim is not None:
+            k_rot = key[:, :, :, : self.rotary_dim]
+            k_pass = key[:, :, :, self.rotary_dim :]
+
+            q_rot = query[:, :, :, : self.rotary_dim]
+            q_pass = query[:, :, :, self.rotary_dim :]
+
+            k_rot = apply_rotary_pos_emb(k_rot, sin, cos)
+            q_rot = apply_rotary_pos_emb(q_rot, sin, cos)
+
+            key = torch.cat([k_rot, k_pass], dim=-1)
+            query = torch.cat([q_rot, q_pass], dim=-1)
+        else:
+            key = apply_rotary_pos_emb(key, sin, cos)
+            query = apply_rotary_pos_emb(query, sin, cos)
+
+        # key = key.permute(0, 2, 1, 3)
+        # query = query.permute(0, 2, 1, 3)
+        key = key.to(dtype=value.dtype)  # fp16 compatability
+        query = query.to(dtype=value.dtype)
+
+        if layer_past is not None:
+            past_key = layer_past[0]
+            past_value = layer_past[1]
+            key = torch.cat((past_key, key), dim=1)
+            value = torch.cat((past_value, value), dim=1)
+
+        if use_cache is True:
+            present = (key, value)
+        else:
+            present = None
+
+        # use AttnMaskType and ColoAttention
+        attn_mask_type = AttnMaskType.causal
+        flash_attention_mask = None
+        if attention_mask != None:
+            if attn_mask_type == AttnMaskType.causal:
+                attn_mask_type == AttnMaskType.paddedcausal
+            else:
+                attn_mask_type = AttnMaskType.padding
+            flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool)).contiguous()
+
+        # use coloattention
+        scale = value.size(-1) ** -0.5
+
+        attention = ColoAttention(
+            embed_dim=self.embed_dim, num_heads=self.num_attention_heads, dropout=self.attn_dropout.p, scale=scale
+        )
+
+        attn_output = attention(query, key, value, attn_mask=flash_attention_mask, attn_mask_type=attn_mask_type)
+
+        attn_output = self.out_proj(attn_output)
+        attn_output = self.resid_dropout(attn_output)
+        outputs = (attn_output, present, None)
+
+        return outputs  # a, present, (attentions)
+
+    return forward
+
+
+def gptj_sequence_parallel_forward_fn(shard_config: ShardConfig):
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+            batch_size = input_ids.shape[0]
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size = inputs_embeds.shape[0]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.view(-1, input_shape[-1])
+
+        if position_ids is not None:
+            position_ids = position_ids.view(-1, input_shape[-1]).long()
+
+        if past_key_values is None:
+            past_length = 0
+            past_key_values = tuple([None] * len(self.h))
+        else:
+            past_length = past_key_values[0][0].size(-2)
+
+        if position_ids is None:
+            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
+
+        # Attention mask.
+        if attention_mask is not None:
+            if batch_size <= 0:
+                raise ValueError("batch_size has to be defined and > 0")
+            attention_mask = attention_mask.view(batch_size, -1)
+            # We create a 3D attention mask from a 2D tensor mask.
+            # Sizes are [batch_size, 1, 1, to_seq_length]
+            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+            # this attention mask is more simple than the triangular masking of causal attention
+            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+            attention_mask = attention_mask[:, None, None, :]
+
+            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+            # masked positions, this operation will create a tensor which is 0.0 for
+            # positions we want to attend and the dtype's smallest value for masked positions.
+            # Since we are adding it to the raw scores before the softmax, this is
+            # effectively the same as removing these entirely.
+            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+            attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x num_attention_heads x N x N
+        # head_mask has shape n_layer x batch x num_attention_heads x N x N
+        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+
+        hidden_states = inputs_embeds
+
+        if token_type_ids is not None:
+            token_type_embeds = self.wte(token_type_ids)
+            hidden_states = hidden_states + token_type_embeds
+
+        hidden_states = self.drop(hidden_states)
+
+        output_shape = input_shape + (hidden_states.size(-1),)
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        # split the input tensor along sequence dimension
+        # [batch_size, seq_len, hidden_size] -> [batch_size, seq_len/TP_size, hidden_size]
+        hidden_states = split_forward_gather_backward(
+            hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group
+        )
+
+        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+            # Model parallel
+            if self.model_parallel:
+                torch.cuda.set_device(hidden_states.device)
+                # Ensure layer_past is on same device as hidden_states (might not be correct)
+                if layer_past is not None:
+                    layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
+                # Ensure that attention_mask is always on the same device as hidden_states
+                if attention_mask is not None:
+                    attention_mask = attention_mask.to(hidden_states.device)
+                if isinstance(head_mask, torch.Tensor):
+                    head_mask = head_mask.to(hidden_states.device)
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, use_cache, output_attentions)
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    None,
+                    attention_mask,
+                    position_ids,
+                    head_mask[i],
+                )
+            else:
+                outputs = block(
+                    hidden_states=hidden_states,
+                    layer_past=layer_past,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    head_mask=head_mask[i],
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = outputs[0]
+            if use_cache is True:
+                presents = presents + (outputs[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+
+            # Model Parallel: If it's the last layer for that device, put things on the next device
+            if self.model_parallel:
+                for k, v in self.device_map.items():
+                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
+                        hidden_states = hidden_states.to("cuda:" + str(k + 1))
+
+        # When sequence parallelism done, gather the output tensor in forward and split it in backward
+        hidden_states = gather_forward_split_backward(
+            hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group
+        )
+
+        hidden_states = self.ln_f(hidden_states)
+
+        hidden_states = hidden_states.view(output_shape)
+        # Add last hidden state
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+    return forward

colossalai/shardformer/modeling/llama.py

@@ -2,6 +2,7 @@ import warnings
 from typing import List, Optional, Tuple
 
 import torch
+import torch.nn.functional as F
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 from transformers.modeling_outputs import (
     BaseModelOutputWithPast,
@@ -12,6 +13,9 @@ from transformers.models.llama.modeling_llama import LlamaForCausalLM, LlamaForS
 from transformers.utils import logging
 
 from colossalai.pipeline.stage_manager import PipelineStageManager
+from colossalai.shardformer.shard import ShardConfig
+
+from ..layer import cross_entropy_1d
 
 try:
     from transformers.models.llama.modeling_llama import _prepare_4d_causal_attention_mask
@@ -42,6 +46,7 @@ class LlamaPipelineForwards:
         stage_manager: Optional[PipelineStageManager] = None,
         hidden_states: Optional[torch.FloatTensor] = None,
         stage_index: Optional[List[int]] = None,
+        shard_config: ShardConfig = None,
     ):
         logger = logging.get_logger(__name__)
 
@@ -200,6 +205,7 @@ class LlamaPipelineForwards:
         stage_manager: Optional[PipelineStageManager] = None,
         hidden_states: Optional[torch.FloatTensor] = None,
         stage_index: Optional[List[int]] = None,
+        shard_config: ShardConfig = None,
     ):
         r"""
         Args:
@@ -269,11 +275,18 @@ class LlamaPipelineForwards:
                 shift_labels = labels[..., 1:].contiguous()
                 # Flatten the tokens
                 loss_fct = CrossEntropyLoss()
-                shift_logits = shift_logits.view(-1, self.config.vocab_size)
                 shift_labels = shift_labels.view(-1)
                 # Enable model parallelism
                 shift_labels = shift_labels.to(shift_logits.device)
-                loss = loss_fct(shift_logits, shift_labels)
+                if shard_config.enable_tensor_parallelism:
+                    new_vocab_size = logits.shape[-1]
+                    shift_logits = shift_logits.view(-1, new_vocab_size)
+                    loss = cross_entropy_1d(
+                        shift_logits, shift_labels, process_group=shard_config.tensor_parallel_process_group
+                    )
+                else:
+                    shift_logits = shift_logits.view(-1, self.config.vocab_size)
+                    loss = loss_fct(shift_logits, shift_labels)
 
             if not return_dict:
                 output = (logits,) + outputs[1:]
@@ -306,6 +319,7 @@ class LlamaPipelineForwards:
         stage_manager: Optional[PipelineStageManager] = None,
         hidden_states: Optional[torch.FloatTensor] = None,
         stage_index: Optional[List[int]] = None,
+        shard_config: ShardConfig = None,
     ):
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
@@ -403,7 +417,7 @@ class LlamaPipelineForwards:
             return {"hidden_states": hidden_states}
 
 
-def get_llama_flash_attention_forward():
+def get_llama_flash_attention_forward(shard_config: ShardConfig):
     from transformers.models.llama.modeling_llama import LlamaAttention, apply_rotary_pos_emb
 
     from colossalai.kernel import AttnMaskType, ColoAttention
@@ -459,14 +473,13 @@ def get_llama_flash_attention_forward():
 
         flash_attention_mask = None
         attn_mask_type = AttnMaskType.causal
-        if attention_mask != None:
+        if not getattr(shard_config, "causal_lm", False) and attention_mask != None:
             if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
                 raise ValueError(
                     f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                 )
             flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool)).contiguous()
-            if not torch.all(flash_attention_mask):
-                attn_mask_type = AttnMaskType.paddedcausal
+            attn_mask_type = AttnMaskType.paddedcausal
 
         attention = ColoAttention(embed_dim=self.hidden_size, num_heads=self.num_heads)
         attn_output = attention(
@@ -483,3 +496,108 @@ def get_llama_flash_attention_forward():
         return attn_output, None, past_key_value
 
     return forward
+
+
+def get_lm_forward_with_dist_cross_entropy(shard_config: ShardConfig):
+    from transformers import LlamaForCausalLM
+
+    def forward(
+        self: LlamaForCausalLM,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, LlamaForCausalLM
+
+        >>> model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        if self.config.pretraining_tp > 1:
+            lm_head_slices = self.lm_head.weight.split(self.vocab_size // self.config.pretraining_tp, dim=0)
+            logits = [F.linear(hidden_states, lm_head_slices[i]) for i in range(self.config.pretraining_tp)]
+            logits = torch.cat(logits, dim=-1)
+        else:
+            logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            if shard_config.enable_tensor_parallelism:
+                new_vocab_size = logits.shape[-1]
+                shift_logits = shift_logits.view(-1, new_vocab_size)
+                loss = cross_entropy_1d(
+                    shift_logits, shift_labels, process_group=shard_config.tensor_parallel_process_group
+                )
+            else:
+                shift_logits = shift_logits.view(-1, self.config.vocab_size)
+                loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    return forward

colossalai/shardformer/modeling/mistral.py

+from typing import Optional, Tuple
+
+import torch
+
+
+def get_mistral_flash_attention_forward():
+    from transformers.models.mistral.modeling_mistral import MistralAttention, apply_rotary_pos_emb, repeat_kv
+
+    from colossalai.kernel.cuda_native import AttnMaskType, ColoAttention
+
+    def forward(
+        self: MistralAttention,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+        assert q_len % 4 == 0, "Flash Attention Error: The sequence length should be a multiple of 4."
+
+        query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = (
+            self.k_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        )
+        value_states = (
+            self.v_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        )
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        if past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        me_input_shape = (bsz, q_len, self.num_heads, self.head_dim)
+        query_states = query_states.transpose(1, 2).contiguous().view(*me_input_shape)
+        key_states = key_states.transpose(1, 2).contiguous().view(*me_input_shape)
+        value_states = value_states.transpose(1, 2).contiguous().view(*me_input_shape)
+
+        flash_attention_mask = None
+        attn_mask_type = AttnMaskType.causal
+        if attention_mask != None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool)).contiguous()
+            attn_mask_type = AttnMaskType.paddedcausal
+
+        attention = ColoAttention(embed_dim=self.hidden_size, num_heads=self.num_heads)
+        attn_output = attention(
+            query_states, key_states, value_states, attn_mask=flash_attention_mask, attn_mask_type=attn_mask_type
+        )
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+    return forward

colossalai/shardformer/policies/auto_policy.py

@@ -85,6 +85,17 @@ _POLICY_LIST = {
     "transformers.models.gpt2.modeling_gpt2.GPT2ForSequenceClassification": PolicyLocation(
         file_name="gpt2", class_name="GPT2ForSequenceClassificationPolicy"
     ),
+    # GPTJ
+    "transformers.models.gptj.modeling_gptj.GPTJModel": PolicyLocation(file_name="gptj", class_name="GPTJModelPolicy"),
+    "transformers.models.gptj.modeling_gptj.GPTJForCausalLM": PolicyLocation(
+        file_name="gptj", class_name="GPTJForCausalLMPolicy"
+    ),
+    "transformers.models.gptj.modeling_gptj.GPTJForQuestionAnswering": PolicyLocation(
+        file_name="gptj", class_name="GPTJForQuestionAnsweringPolicy"
+    ),
+    "transformers.models.gptj.modeling_gptj.GPTJForSequenceClassification": PolicyLocation(
+        file_name="gptj", class_name="GPTJForSequenceClassificationPolicy"
+    ),
     # ViT
     "transformers.models.vit.modeling_vit.ViTModel": PolicyLocation(file_name="vit", class_name="ViTModelPolicy"),
     "transformers.models.vit.modeling_vit.ViTForImageClassification": PolicyLocation(
@@ -146,6 +157,31 @@ _POLICY_LIST = {
     "colossalai.shardformer.modeling.chatglm2_6b.modeling_chatglm.ChatGLMForConditionalGeneration": PolicyLocation(
         file_name="chatglm2", class_name="ChatGLMForConditionalGenerationPolicy"
     ),
+    # Falcon
+    "transformers.models.falcon.modeling_falcon.FalconModel": PolicyLocation(
+        file_name="falcon", class_name="FalconModelPolicy"
+    ),
+    "transformers.models.falcon.modeling_falcon.FalconForCausalLM": PolicyLocation(
+        file_name="falcon", class_name="FalconForCausalLMPolicy"
+    ),
+    "transformers.models.falcon.modeling_falcon.FalconForSequenceClassification": PolicyLocation(
+        file_name="falcon", class_name="FalconForSequenceClassificationPolicy"
+    ),
+    "transformers.models.falcon.modeling_falcon.FalconForTokenClassification": PolicyLocation(
+        file_name="falcon", class_name="FalconForTokenClassificationPolicy"
+    ),
+    "transformers.models.falcon.modeling_falcon.FalconForQuestionAnswering": PolicyLocation(
+        file_name="falcon", class_name="FalconForQuestionAnsweringPolicy"
+    ),
+    "transformers.models.mistral.modeling_mistral.MistralModel": PolicyLocation(
+        file_name="mistral", class_name="MistralModelPolicy"
+    ),
+    "transformers.models.mistral.modeling_mistral.MistralForCausalLM": PolicyLocation(
+        file_name="mistral", class_name="MistralForCausalLMPolicy"
+    ),
+    "transformers.models.mistral.modeling_mistral.MistralForSequenceClassification": PolicyLocation(
+        file_name="mistral", class_name="MistralForSequenceClassificationPolicy"
+    ),
 }
 
 

colossalai/shardformer/policies/base_policy.py

@@ -2,7 +2,7 @@
 
 from abc import ABC, abstractmethod
 from dataclasses import dataclass
-from typing import Any, Callable, Dict, List, Optional, Union
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 
 import numpy as np
 import torch.nn as nn
@@ -214,13 +214,32 @@ class Policy(ABC):
         return layers_per_stage
 
     @staticmethod
-    def get_stage_index(layers_per_stage: List[int], stage: int) -> List[int]:
+    def get_stage_index(
+        layers_per_stage: List[int],
+        stage: int,
+        num_model_chunks: int = 1,
+        num_stages: int = 0,
+    ) -> Union[Tuple[int, int], List[Tuple[int, int]]]:
         """
-        get the start index and end index of layers for each stage.
+        Get the start index and end index of layers for each stage.
+
+        Args:
+            layers_per_stage (List[int]): number of layers for each stage
+            stage (int): the stage index
+            num_stages (int): number of stages
+            num_model_chunks (int): number of model chunks
+
+        Returns:
+            - Tuple[int, int]: the start index and end index of this stage
+            - List[Tuple[int, int]]: the start index and end index of this stage for each model chunk
+
         """
         num_layers_per_stage_accumulated = np.insert(np.cumsum(layers_per_stage), 0, 0)
 
-        start_idx = num_layers_per_stage_accumulated[stage]
-        end_idx = num_layers_per_stage_accumulated[stage + 1]
+        stage_indices = []
+        for model_chunk in range(num_model_chunks):
+            start_idx = num_layers_per_stage_accumulated[stage + model_chunk * num_stages]
+            end_idx = num_layers_per_stage_accumulated[stage + model_chunk * num_stages + 1]
+            stage_indices.append([start_idx, end_idx])
 
-        return [start_idx, end_idx]
+        return stage_indices[0] if num_model_chunks == 1 else stage_indices

colossalai/shardformer/policies/bert.py

@@ -21,7 +21,7 @@ __all__ = [
     "BertPolicy",
     "BertModelPolicy",
     "BertForPreTrainingPolicy",
-    "BertLMdHeadModelPolicy",
+    "BertLMHeadModelPolicy",
     "BertForMaskedLMPolicy",
     "BertForNextSentencePredictionPolicy",
     "BertForSequenceClassificationPolicy",
@@ -249,15 +249,34 @@ class BertPolicy(Policy):
         return self.model
 
     def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, policy: Dict) -> None:
-        """If under pipeline parallel setting, replacing the original forward method of huggingface
-        to customized forward method, and add this changing to policy."""
-        if self.pipeline_stage_manager:
-            stage_manager = self.pipeline_stage_manager
-            if self.model.__class__.__name__ == "BertModel":
-                module = self.model
-            else:
-                module = self.model.bert
+        """
+        If under pipeline parallel setting, replacing the original forward method of huggingface
+        to customized forward method, and add this changing to policy.
+        """
+        if self.pipeline_stage_manager is None:
+            return
 
+        stage_manager = self.pipeline_stage_manager
+        if self.model.__class__.__name__ == "BertModel":
+            module = self.model
+        else:
+            module = self.model.bert
+
+        if stage_manager.is_interleave:
+            layers_per_stage = self.distribute_layers(
+                len(module.encoder.layer), stage_manager.num_stages * stage_manager.num_model_chunks
+            )
+            stage_manager.stage_indices = Policy.get_stage_index(
+                layers_per_stage,
+                stage_manager.stage,
+                num_model_chunks=stage_manager.num_model_chunks,
+                num_stages=stage_manager.num_stages,
+            )
+            method_replacement = {
+                "forward": partial(new_forward, stage_manager=stage_manager, shard_config=self.shard_config)
+            }
+
+        else:
             layers_per_stage = Policy.distribute_layers(len(module.encoder.layer), stage_manager.num_stages)
             stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage)
             method_replacement = {
@@ -265,11 +284,8 @@ class BertPolicy(Policy):
                     new_forward, stage_manager=stage_manager, stage_index=stage_index, shard_config=self.shard_config
                 )
             }
-            self.append_or_create_method_replacement(
-                description=method_replacement, policy=policy, target_key=model_cls
-            )
 
-        return
+        self.append_or_create_method_replacement(description=method_replacement, policy=policy, target_key=model_cls)
 
     def get_held_layers(self) -> List[Module]:
         """Get pipeline layers for current stage."""
@@ -282,13 +298,32 @@ class BertPolicy(Policy):
         stage_manager = self.pipeline_stage_manager
 
         held_layers = []
-        layers_per_stage = self.distribute_layers(len(module.encoder.layer), stage_manager.num_stages)
-        if stage_manager.is_first_stage():
-            held_layers.append(module.embeddings)
-        start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage)
-        held_layers.extend(module.encoder.layer[start_idx:end_idx])
-        if stage_manager.is_last_stage():
-            held_layers.append(module.pooler)
+        if stage_manager.is_interleave:
+            assert stage_manager.num_model_chunks is not None
+            layers_per_stage = self.distribute_layers(
+                len(module.encoder.layer), stage_manager.num_stages * stage_manager.num_model_chunks
+            )
+            stage_indices = Policy.get_stage_index(
+                layers_per_stage,
+                stage_manager.stage,
+                num_model_chunks=stage_manager.num_model_chunks,
+                num_stages=stage_manager.num_stages,
+            )
+            if stage_manager.is_first_stage(ignore_chunk=True):
+                held_layers.append(module.embeddings)
+            for start_idx, end_idx in stage_indices:
+                held_layers.extend(module.encoder.layer[start_idx:end_idx])
+            if stage_manager.is_last_stage(ignore_chunk=True):
+                held_layers.append(module.pooler)
+
+        else:
+            layers_per_stage = self.distribute_layers(len(module.encoder.layer), stage_manager.num_stages)
+            if stage_manager.is_first_stage():
+                held_layers.append(module.embeddings)
+            start_idx, end_idx = Policy.get_stage_index(layers_per_stage, stage_manager.stage)
+            held_layers.extend(module.encoder.layer[start_idx:end_idx])
+            if stage_manager.is_last_stage():
+                held_layers.append(module.pooler)
 
         return held_layers
 
@@ -335,7 +370,7 @@ class BertForPreTrainingPolicy(BertPolicy):
         """Get pipeline layers for current stage"""
         held_layers = super().get_held_layers()
         stage_manager = self.pipeline_stage_manager
-        if stage_manager.is_last_stage():
+        if stage_manager.is_last_stage(ignore_chunk=True):
             held_layers.append(self.model.cls)
 
         return held_layers
@@ -374,7 +409,7 @@ class BertLMHeadModelPolicy(BertPolicy):
         """
         held_layers = super().get_held_layers()
         stage_manager = self.pipeline_stage_manager
-        if stage_manager.is_last_stage():
+        if stage_manager.is_last_stage(ignore_chunk=True):
             held_layers.append(self.model.cls)
         return held_layers
 
@@ -412,7 +447,7 @@ class BertForMaskedLMPolicy(BertPolicy):
         """
         held_layers = super().get_held_layers()
         stage_manager = self.pipeline_stage_manager
-        if stage_manager.is_last_stage():
+        if stage_manager.is_last_stage(ignore_chunk=True):
             held_layers.append(self.model.cls)
         return held_layers
 
@@ -464,7 +499,7 @@ class BertForSequenceClassificationPolicy(BertPolicy):
         """
         held_layers = super().get_held_layers()
         stage_manager = self.pipeline_stage_manager
-        if stage_manager.is_last_stage():
+        if stage_manager.is_last_stage(ignore_chunk=True):
             held_layers.append(self.model.dropout)
             held_layers.append(self.model.classifier)
         return held_layers
@@ -508,7 +543,7 @@ class BertForTokenClassificationPolicy(BertPolicy):
         """
         held_layers = super().get_held_layers()
         stage_manager = self.pipeline_stage_manager
-        if stage_manager.is_last_stage():
+        if stage_manager.is_last_stage(ignore_chunk=True):
             held_layers.append(self.model.dropout)
             held_layers.append(self.model.classifier)
         return held_layers
@@ -539,7 +574,7 @@ class BertForNextSentencePredictionPolicy(BertPolicy):
         """
         held_layers = super().get_held_layers()
         stage_manager = self.pipeline_stage_manager
-        if stage_manager.is_last_stage():
+        if stage_manager.is_last_stage(ignore_chunk=True):
             held_layers.append(self.model.cls)
         return held_layers
 
@@ -582,7 +617,7 @@ class BertForMultipleChoicePolicy(BertPolicy):
         """
         held_layers = super().get_held_layers()
         stage_manager = self.pipeline_stage_manager
-        if stage_manager.is_last_stage():
+        if stage_manager.is_last_stage(ignore_chunk=True):
             held_layers.append(self.model.dropout)
             held_layers.append(self.model.classifier)
         return held_layers
@@ -612,7 +647,7 @@ class BertForQuestionAnsweringPolicy(BertPolicy):
         """
         held_layers = super().get_held_layers()
         stage_manager = self.pipeline_stage_manager
-        if stage_manager.is_last_stage():
+        if stage_manager.is_last_stage(ignore_chunk=True):
             held_layers.append(self.model.qa_outputs)
         return held_layers
 

colossalai/shardformer/policies/bloom.py

@@ -21,6 +21,15 @@ from .base_policy import ModulePolicyDescription, Policy, SubModuleReplacementDe
 
 
 class BloomPolicy(Policy):
+    def __init__(self) -> None:
+        super().__init__()
+        import transformers
+        from packaging.version import Version
+
+        assert Version(transformers.__version__) <= Version(
+            "4.33.0"
+        ), "The Bloom model should run on a transformers version not greater than 4.33.0."
+
     def config_sanity_check(self):
         pass