Refactor system architecture (#82)

cacheflow/model_executor/parallel_utils/README.md

@@ -0,0 +1 @@
+The files in this folder are ported from [Megatron-LM](https://github.com/NVIDIA/Megatron-LM/tree/main/megatron/core). We only keep the codes that are used in inference.

cacheflow/model_executor/parallel_utils/__init__.py

@@ -0,0 +1,12 @@
+import cacheflow.model_executor.parallel_utils.parallel_state
+import cacheflow.model_executor.parallel_utils.tensor_parallel
+import cacheflow.model_executor.parallel_utils.utils
+
+# Alias parallel_state as mpu, its legacy name
+mpu = parallel_state
+
+__all__ = [
+    "parallel_state",
+    "tensor_parallel",
+    "utils",
+]

cacheflow/model_executor/parallel_utils/parallel_state.py

@@ -0,0 +1,593 @@
+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+"""Model and data parallel groups."""
+
+import torch
+from typing import Optional
+
+from .utils import GlobalMemoryBuffer
+
+# Intra-layer model parallel group that the current rank belongs to.
+_TENSOR_MODEL_PARALLEL_GROUP = None
+# Inter-layer model parallel group that the current rank belongs to.
+_PIPELINE_MODEL_PARALLEL_GROUP = None
+# Model parallel group (both intra- and pipeline) that the current rank belongs to.
+_MODEL_PARALLEL_GROUP = None
+# Embedding group.
+_EMBEDDING_GROUP = None
+# Position embedding group.
+_POSITION_EMBEDDING_GROUP = None
+# Data parallel group that the current rank belongs to.
+_DATA_PARALLEL_GROUP = None
+
+_VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = None
+_VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
+_PIPELINE_MODEL_PARALLEL_SPLIT_RANK = None
+
+# These values enable us to change the mpu sizes on the fly.
+_MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE = None
+_MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
+_MPU_TENSOR_MODEL_PARALLEL_RANK = None
+_MPU_PIPELINE_MODEL_PARALLEL_RANK = None
+
+# A list of ranks that have a copy of the embedding.
+_EMBEDDING_GLOBAL_RANKS = None
+
+# A list of ranks that have a copy of the position embedding.
+_POSITION_EMBEDDING_GLOBAL_RANKS = None
+
+# A list of global ranks for each pipeline group to ease calculation of the source
+# rank when broadcasting from the first or last pipeline stage.
+_PIPELINE_GLOBAL_RANKS = None
+
+# A list of global ranks for each data parallel group to ease calculation of the source
+# rank when broadcasting weights from src to all other data parallel ranks
+_DATA_PARALLEL_GLOBAL_RANKS = None
+
+# Memory buffers to avoid dynamic memory allocation
+_GLOBAL_MEMORY_BUFFER = None
+
+_ALL_REDUCE_LAUNCHER: Optional['GraphAllReduce'] = None
+
+def initialize_model_parallel(
+    tensor_model_parallel_size: int = 1,
+    pipeline_model_parallel_size: int = 1,
+    virtual_pipeline_model_parallel_size: Optional[int] = None,
+    pipeline_model_parallel_split_rank: Optional[int] = None,
+) -> None:
+    """
+    Initialize model data parallel groups.
+
+    Arguments:
+        tensor_model_parallel_size: number of GPUs used for tensor model parallelism.
+        pipeline_model_parallel_size: number of GPUs used for pipeline model parallelism.
+        virtual_pipeline_model_parallel_size: number of virtual stages (interleaved
+                                              pipeline).
+        pipeline_model_parallel_split_rank: for models with both encoder and decoder,
+                                            rank in pipeline with split point.
+
+    Let's say we have a total of 16 GPUs denoted by g0 ... g15 and we
+    use 2 GPUs to parallelize the model tensor, and 4 GPUs to parallelize
+    the model pipeline. The present function will
+    create 8 tensor model-parallel groups, 4 pipeline model-parallel groups
+    and 8 data-parallel groups as:
+        8 data_parallel groups:
+            [g0, g2], [g1, g3], [g4, g6], [g5, g7], [g8, g10], [g9, g11], [g12, g14], [g13, g15]
+        8 tensor model-parallel groups:
+            [g0, g1], [g2, g3], [g4, g5], [g6, g7], [g8, g9], [g10, g11], [g12, g13], [g14, g15]
+        4 pipeline model-parallel groups:
+            [g0, g4, g8, g12], [g1, g5, g9, g13], [g2, g6, g10, g14], [g3, g7, g11, g15]
+    Note that for efficiency, the caller should make sure adjacent ranks
+    are on the same DGX box. For example if we are using 2 DGX-1 boxes
+    with a total of 16 GPUs, rank 0 to 7 belong to the first box and
+    ranks 8 to 15 belong to the second box.
+    """
+    # Get world size and rank. Ensure some consistencies.
+    assert torch.distributed.is_initialized()
+    world_size: int = torch.distributed.get_world_size()
+
+    if world_size % (tensor_model_parallel_size * pipeline_model_parallel_size) != 0:
+        raise RuntimeError(
+            f"world_size ({world_size}) is not divisible by tensor_model_parallel_size "
+            f"({tensor_model_parallel_size}) x pipeline_model_parallel_size ({pipeline_model_parallel_size})"
+        )
+
+    data_parallel_size: int = world_size // (tensor_model_parallel_size *
+                                             pipeline_model_parallel_size)
+
+    num_tensor_model_parallel_groups: int  = world_size // tensor_model_parallel_size
+    num_pipeline_model_parallel_groups: int = world_size // pipeline_model_parallel_size
+    num_data_parallel_groups: int = world_size // data_parallel_size
+
+    if virtual_pipeline_model_parallel_size is not None:
+        if not pipeline_model_parallel_size > 2:
+            raise RuntimeError("pipeline-model-parallel size should be greater than 2 with "
+                               "interleaved schedule")
+        global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
+        global _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+        _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = 0
+        _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = virtual_pipeline_model_parallel_size
+
+    if pipeline_model_parallel_split_rank is not None:
+        global _PIPELINE_MODEL_PARALLEL_SPLIT_RANK
+        _PIPELINE_MODEL_PARALLEL_SPLIT_RANK = pipeline_model_parallel_split_rank
+
+    rank = torch.distributed.get_rank()
+
+    # Build the data-parallel groups.
+    global _DATA_PARALLEL_GROUP
+    global _DATA_PARALLEL_GLOBAL_RANKS
+    assert _DATA_PARALLEL_GROUP is None, 'data parallel group is already initialized'
+    all_data_parallel_group_ranks = []
+    for i in range(pipeline_model_parallel_size):
+        start_rank = i * num_pipeline_model_parallel_groups
+        end_rank = (i + 1) * num_pipeline_model_parallel_groups
+        for j in range(tensor_model_parallel_size):
+            ranks = range(start_rank + j, end_rank, tensor_model_parallel_size)
+            all_data_parallel_group_ranks.append(list(ranks))
+            group = torch.distributed.new_group(ranks)
+            if rank in ranks:
+                _DATA_PARALLEL_GROUP = group
+                _DATA_PARALLEL_GLOBAL_RANKS = ranks
+
+    # Build the model-parallel groups.
+    global _MODEL_PARALLEL_GROUP
+    assert _MODEL_PARALLEL_GROUP is None, 'model parallel group is already initialized'
+    for i in range(data_parallel_size):
+        ranks = [data_parallel_group_ranks[i]
+                 for data_parallel_group_ranks in all_data_parallel_group_ranks]
+        group = torch.distributed.new_group(ranks)
+        if rank in ranks:
+            _MODEL_PARALLEL_GROUP = group
+
+    # Build the tensor model-parallel groups.
+    global _TENSOR_MODEL_PARALLEL_GROUP
+    assert _TENSOR_MODEL_PARALLEL_GROUP is None, \
+        'tensor model parallel group is already initialized'
+    for i in range(num_tensor_model_parallel_groups):
+        ranks = range(i * tensor_model_parallel_size,
+                      (i + 1) * tensor_model_parallel_size)
+        group = torch.distributed.new_group(ranks)
+        if rank in ranks:
+            _TENSOR_MODEL_PARALLEL_GROUP = group
+
+    # Build the pipeline model-parallel groups and embedding groups
+    # (first and last rank in each pipeline model-parallel group).
+    global _PIPELINE_MODEL_PARALLEL_GROUP
+    global _PIPELINE_GLOBAL_RANKS
+    assert _PIPELINE_MODEL_PARALLEL_GROUP is None, \
+        'pipeline model parallel group is already initialized'
+    global _EMBEDDING_GROUP
+    global _EMBEDDING_GLOBAL_RANKS
+    assert _EMBEDDING_GROUP is None, 'embedding group is already initialized'
+    global _POSITION_EMBEDDING_GROUP
+    global _POSITION_EMBEDDING_GLOBAL_RANKS
+    assert _POSITION_EMBEDDING_GROUP is None, \
+        'position embedding group is already initialized'
+    for i in range(num_pipeline_model_parallel_groups):
+        ranks = range(i, world_size, num_pipeline_model_parallel_groups)
+        group = torch.distributed.new_group(ranks)
+        if rank in ranks:
+            _PIPELINE_MODEL_PARALLEL_GROUP = group
+            _PIPELINE_GLOBAL_RANKS = ranks
+        # Setup embedding group (to exchange gradients between
+        # first and last stages).
+        if len(ranks) > 1:
+            embedding_ranks = [ranks[0], ranks[-1]]
+            position_embedding_ranks = [ranks[0]]
+            if pipeline_model_parallel_split_rank is not None:
+                if ranks[pipeline_model_parallel_split_rank] not in embedding_ranks:
+                    embedding_ranks = [ranks[0],
+                                       ranks[pipeline_model_parallel_split_rank],
+                                       ranks[-1]]
+                if ranks[pipeline_model_parallel_split_rank] not in position_embedding_ranks:
+                    position_embedding_ranks = [ranks[0],
+                                       ranks[pipeline_model_parallel_split_rank]]
+        else:
+            embedding_ranks = ranks
+            position_embedding_ranks = ranks
+
+        group = torch.distributed.new_group(embedding_ranks)
+        if rank in embedding_ranks:
+            _EMBEDDING_GROUP = group
+        if rank in ranks:
+            _EMBEDDING_GLOBAL_RANKS = embedding_ranks
+
+        group = torch.distributed.new_group(position_embedding_ranks)
+        if rank in position_embedding_ranks:
+            _POSITION_EMBEDDING_GROUP = group
+        if rank in ranks:
+            _POSITION_EMBEDDING_GLOBAL_RANKS = position_embedding_ranks
+
+    # Initialize global memory buffer
+    # This isn't really "parallel state" but there isn't another good place to
+    # put this. If we end up with a more generic initialization of megatron-core
+    # we could stick it there
+    _set_global_memory_buffer()
+
+
+def initialize_all_reduce_launcher(
+    max_num_tokens: int,
+    hidden_size: int,
+    dtype: torch.dtype,
+    disable_graph: bool = False,
+) -> None:
+    global _ALL_REDUCE_LAUNCHER
+    _ALL_REDUCE_LAUNCHER = GraphAllReduce(
+        max_num_tokens=max_num_tokens,
+        hidden_size=hidden_size,
+        dtype=dtype,
+        disable_graph=disable_graph,
+    )
+
+def model_parallel_is_initialized():
+    """Check if model and data parallel groups are initialized."""
+    if _TENSOR_MODEL_PARALLEL_GROUP is None or \
+        _PIPELINE_MODEL_PARALLEL_GROUP is None or \
+        _DATA_PARALLEL_GROUP is None:
+        return False
+    return True
+
+
+def get_model_parallel_group():
+    """Get the model parallel group the caller rank belongs to."""
+    assert _MODEL_PARALLEL_GROUP is not None, \
+        'model parallel group is not initialized'
+    return _MODEL_PARALLEL_GROUP
+
+
+def get_tensor_model_parallel_group():
+    """Get the tensor model parallel group the caller rank belongs to."""
+    assert _TENSOR_MODEL_PARALLEL_GROUP is not None, \
+        'intra_layer_model parallel group is not initialized'
+    return _TENSOR_MODEL_PARALLEL_GROUP
+
+
+def get_pipeline_model_parallel_group():
+    """Get the pipeline model parallel group the caller rank belongs to."""
+    assert _PIPELINE_MODEL_PARALLEL_GROUP is not None, \
+        'pipeline_model parallel group is not initialized'
+    return _PIPELINE_MODEL_PARALLEL_GROUP
+
+
+def get_data_parallel_group():
+    """Get the data parallel group the caller rank belongs to."""
+    assert _DATA_PARALLEL_GROUP is not None, \
+        'data parallel group is not initialized'
+    return _DATA_PARALLEL_GROUP
+
+
+def get_embedding_group():
+    """Get the embedding group the caller rank belongs to."""
+    assert _EMBEDDING_GROUP is not None, \
+        'embedding group is not initialized'
+    return _EMBEDDING_GROUP
+
+
+def get_position_embedding_group():
+    """Get the position embedding group the caller rank belongs to."""
+    assert _POSITION_EMBEDDING_GROUP is not None, \
+        'position embedding group is not initialized'
+    return _POSITION_EMBEDDING_GROUP
+
+
+def set_tensor_model_parallel_world_size(world_size):
+    """Set the tensor model parallel size"""
+    global _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE
+    _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE = world_size
+
+
+def set_pipeline_model_parallel_world_size(world_size):
+    """Set the pipeline model parallel size"""
+    global _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+    _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = world_size
+
+
+def get_tensor_model_parallel_world_size():
+    """Return world size for the tensor model parallel group."""
+    global _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE
+    if _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE is not None:
+        return _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE
+    return torch.distributed.get_world_size(group=get_tensor_model_parallel_group())
+
+
+def get_pipeline_model_parallel_world_size():
+    """Return world size for the pipeline model parallel group."""
+    global _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+    if _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE is not None:
+        return _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+    return torch.distributed.get_world_size(group=get_pipeline_model_parallel_group())
+
+
+def set_tensor_model_parallel_rank(rank):
+    """Set tensor model parallel rank."""
+    global _MPU_TENSOR_MODEL_PARALLEL_RANK
+    _MPU_TENSOR_MODEL_PARALLEL_RANK = rank
+
+
+def set_pipeline_model_parallel_rank(rank):
+    """Set pipeline model parallel rank."""
+    global _MPU_PIPELINE_MODEL_PARALLEL_RANK
+    _MPU_PIPELINE_MODEL_PARALLEL_RANK = rank
+
+
+def set_pipeline_model_parallel_split_rank(rank):
+    """Set pipeline model parallel split rank."""
+    global _MPU_PIPELINE_MODEL_PARALLEL_SPLIT_RANK
+    _MPU_PIPELINE_MODEL_PARALLEL_SPLIT_RANK = rank
+
+
+def get_tensor_model_parallel_rank():
+    """Return my rank for the tensor model parallel group."""
+    global _MPU_TENSOR_MODEL_PARALLEL_RANK
+    if _MPU_TENSOR_MODEL_PARALLEL_RANK is not None:
+        return _MPU_TENSOR_MODEL_PARALLEL_RANK
+    return torch.distributed.get_rank(group=get_tensor_model_parallel_group())
+
+
+def get_pipeline_model_parallel_rank():
+    """Return my rank for the pipeline model parallel group."""
+    global _MPU_PIPELINE_MODEL_PARALLEL_RANK
+    if _MPU_PIPELINE_MODEL_PARALLEL_RANK is not None:
+        return _MPU_PIPELINE_MODEL_PARALLEL_RANK
+    return torch.distributed.get_rank(group=get_pipeline_model_parallel_group())
+
+
+
+def is_pipeline_first_stage(ignore_virtual=False):
+    """Return True if in the first pipeline model-parallel stage, False otherwise."""
+    if not ignore_virtual:
+        if get_virtual_pipeline_model_parallel_world_size() is not None and \
+            get_virtual_pipeline_model_parallel_rank() != 0:
+            return False
+    return get_pipeline_model_parallel_rank() == 0
+
+
+def is_pipeline_last_stage(ignore_virtual=False):
+    """Return True if in the last pipeline model-parallel stage, False otherwise."""
+    if not ignore_virtual:
+        virtual_pipeline_model_parallel_world_size = \
+            get_virtual_pipeline_model_parallel_world_size()
+        if virtual_pipeline_model_parallel_world_size is not None and \
+            get_virtual_pipeline_model_parallel_rank() != (
+                virtual_pipeline_model_parallel_world_size - 1):
+            return False
+    return get_pipeline_model_parallel_rank() == (
+        get_pipeline_model_parallel_world_size() - 1)
+
+
+def is_rank_in_embedding_group(ignore_virtual=False):
+    """Return true if current rank is in embedding group, False otherwise."""
+    rank = torch.distributed.get_rank()
+    global _EMBEDDING_GLOBAL_RANKS
+    if ignore_virtual:
+        return rank in _EMBEDDING_GLOBAL_RANKS
+    if rank in _EMBEDDING_GLOBAL_RANKS:
+        if rank == _EMBEDDING_GLOBAL_RANKS[0]:
+            return is_pipeline_first_stage(ignore_virtual=False)
+        elif rank == _EMBEDDING_GLOBAL_RANKS[-1]:
+            return is_pipeline_last_stage(ignore_virtual=False)
+        else:
+            return True
+    return False
+
+
+def is_rank_in_position_embedding_group():
+    """Return true if current rank is in position embedding group, False otherwise."""
+    rank = torch.distributed.get_rank()
+    global _POSITION_EMBEDDING_GLOBAL_RANKS
+    return rank in _POSITION_EMBEDDING_GLOBAL_RANKS
+
+
+def is_pipeline_stage_before_split(rank=None):
+    """Return True if pipeline stage executes encoder block for a model
+    with both encoder and decoder."""
+    if get_pipeline_model_parallel_world_size() == 1:
+        return True
+    if rank is None:
+        rank = get_pipeline_model_parallel_rank()
+    global _PIPELINE_MODEL_PARALLEL_SPLIT_RANK
+    if _PIPELINE_MODEL_PARALLEL_SPLIT_RANK is None:
+        return True
+    if rank < _PIPELINE_MODEL_PARALLEL_SPLIT_RANK:
+        return True
+    return False
+
+
+def is_pipeline_stage_after_split(rank=None):
+    """Return True if pipeline stage executes decoder block for a model
+    with both encoder and decoder."""
+    if get_pipeline_model_parallel_world_size() == 1:
+        return True
+    if rank is None:
+        rank = get_pipeline_model_parallel_rank()
+    global _PIPELINE_MODEL_PARALLEL_SPLIT_RANK
+    if _PIPELINE_MODEL_PARALLEL_SPLIT_RANK is None:
+        return True
+    if rank >= _PIPELINE_MODEL_PARALLEL_SPLIT_RANK:
+        return True
+    return False
+
+
+def is_pipeline_stage_at_split():
+    """Return true if pipeline stage executes decoder block and next
+    stage executes encoder block for a model with both encoder and
+    decoder."""
+    rank = get_pipeline_model_parallel_rank()
+    return is_pipeline_stage_before_split(rank) and \
+            is_pipeline_stage_after_split(rank+1)
+
+
+def get_virtual_pipeline_model_parallel_rank():
+    """Return the virtual pipeline-parallel rank."""
+    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
+    return _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
+
+
+def set_virtual_pipeline_model_parallel_rank(rank):
+    """Set the virtual pipeline-parallel rank."""
+    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
+    _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = rank
+
+
+def get_virtual_pipeline_model_parallel_world_size():
+    """Return the virtual pipeline-parallel world size."""
+    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+    return _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+
+
+def get_tensor_model_parallel_src_rank():
+    """Calculate the global rank corresponding to the first local rank
+    in the tensor model parallel group."""
+    global_rank = torch.distributed.get_rank()
+    local_world_size = get_tensor_model_parallel_world_size()
+    return (global_rank // local_world_size) * local_world_size
+
+
+def get_data_parallel_src_rank():
+    """Calculate the global rank corresponding to the first local rank
+    in the data parallel group."""
+    assert _DATA_PARALLEL_GLOBAL_RANKS is not None, \
+        "Data parallel group is not initialized"
+    return _DATA_PARALLEL_GLOBAL_RANKS[0]
+
+
+def get_pipeline_model_parallel_first_rank():
+    """Return the global rank of the first process in the pipeline for the
+    current tensor parallel group"""
+    assert _PIPELINE_GLOBAL_RANKS is not None, \
+        "Pipeline parallel group is not initialized"
+    return _PIPELINE_GLOBAL_RANKS[0]
+
+
+def get_pipeline_model_parallel_last_rank():
+    """Return the global rank of the last process in the pipeline for the
+    current tensor parallel group"""
+    assert _PIPELINE_GLOBAL_RANKS is not None, \
+        "Pipeline parallel group is not initialized"
+    last_rank_local = get_pipeline_model_parallel_world_size() - 1
+    return _PIPELINE_GLOBAL_RANKS[last_rank_local]
+
+def get_pipeline_model_parallel_next_rank():
+    """Return the global rank that follows the caller in the pipeline"""
+    assert _PIPELINE_GLOBAL_RANKS is not None, \
+        "Pipeline parallel group is not initialized"
+    rank_in_pipeline = get_pipeline_model_parallel_rank()
+    world_size = get_pipeline_model_parallel_world_size()
+    return _PIPELINE_GLOBAL_RANKS[(rank_in_pipeline + 1) % world_size]
+
+
+def get_pipeline_model_parallel_prev_rank():
+    """Return the global rank that preceeds the caller in the pipeline"""
+    assert _PIPELINE_GLOBAL_RANKS is not None, \
+        "Pipeline parallel group is not initialized"
+    rank_in_pipeline = get_pipeline_model_parallel_rank()
+    world_size = get_pipeline_model_parallel_world_size()
+    return _PIPELINE_GLOBAL_RANKS[(rank_in_pipeline - 1) % world_size]
+
+
+def get_data_parallel_world_size():
+    """Return world size for the data parallel group."""
+    return torch.distributed.get_world_size(group=get_data_parallel_group())
+
+
+def get_data_parallel_rank():
+    """Return my rank for the data parallel group."""
+    return torch.distributed.get_rank(group=get_data_parallel_group())
+
+def _set_global_memory_buffer():
+    """Initialize global buffer"""
+    global _GLOBAL_MEMORY_BUFFER
+    assert _GLOBAL_MEMORY_BUFFER is None, 'global memory buffer is already initialized'
+    _GLOBAL_MEMORY_BUFFER = GlobalMemoryBuffer()
+
+def get_global_memory_buffer():
+    """Return the global GlobalMemoryBuffer object"""
+    assert _GLOBAL_MEMORY_BUFFER is not None, 'global memory buffer is not initialized'
+    return _GLOBAL_MEMORY_BUFFER
+
+def get_all_reduce_launcher() -> 'GraphAllReduce':
+    assert _ALL_REDUCE_LAUNCHER is not None, 'all reduce launcher is not initialized'
+    return _ALL_REDUCE_LAUNCHER
+
+def destroy_model_parallel():
+    """Set the groups to none."""
+    global _MODEL_PARALLEL_GROUP
+    _MODEL_PARALLEL_GROUP = None
+    global _TENSOR_MODEL_PARALLEL_GROUP
+    _TENSOR_MODEL_PARALLEL_GROUP = None
+    global _PIPELINE_MODEL_PARALLEL_GROUP
+    _PIPELINE_MODEL_PARALLEL_GROUP = None
+    global _DATA_PARALLEL_GROUP
+    _DATA_PARALLEL_GROUP = None
+    global _EMBEDDING_GROUP
+    _EMBEDDING_GROUP = None
+    global _POSITION_EMBEDDING_GROUP
+    _POSITION_EMBEDDING_GROUP = None
+    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
+    _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = None
+    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+    _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
+    global _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE
+    _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE = None
+    global _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+    _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
+    global _MPU_TENSOR_MODEL_PARALLEL_RANK
+    _MPU_TENSOR_MODEL_PARALLEL_RANK = None
+    global _MPU_PIPELINE_MODEL_PARALLEL_RANK
+    _MPU_PIPELINE_MODEL_PARALLEL_RANK = None
+    global _GLOBAL_MEMORY_BUFFER
+    _GLOBAL_MEMORY_BUFFER = None
+
+
+class GraphAllReduce:
+
+    def __init__(
+        self,
+        max_num_tokens: int,
+        hidden_size: int,
+        dtype: torch.dtype,
+        disable_graph: bool = False,
+    ) -> None:
+        self.max_num_tokens = max_num_tokens
+        self.hidden_size = hidden_size
+        self.disable_graph = disable_graph
+
+        tp_world_size = get_tensor_model_parallel_world_size()
+        if tp_world_size == 1:
+            return
+
+        self.group = get_tensor_model_parallel_group()
+        self.buffer = torch.empty(
+            size=(max_num_tokens, hidden_size),
+            dtype=dtype,
+            device='cuda',
+        )
+
+        # Build graphs for different number of tokens.
+        if not self.disable_graph:
+            self.graphs = {}
+            for num_tokens in range(8, max_num_tokens + 1, 8):
+                self.graphs[num_tokens] = self._build_graph(num_tokens)
+
+    def _build_graph(self, num_tokens: int) -> torch.cuda.CUDAGraph:
+        # Warm up.
+        torch.distributed.all_reduce(self.buffer[:num_tokens], group=self.group)
+        torch.cuda.synchronize()
+
+        # Build graph.
+        graph = torch.cuda.CUDAGraph()
+        with torch.cuda.graph(graph):
+            torch.distributed.all_reduce(
+                self.buffer[:num_tokens], group=self.group)
+        torch.cuda.synchronize()
+        return graph
+
+    def launch(self, x: torch.Tensor) -> torch.Tensor:
+        # NOTE: x must be a slice of self.buffer.
+        num_tokens = x.shape[0]
+        if self.disable_graph:
+            torch.distributed.all_reduce(x, group=self.group)
+        else:
+            self.graphs[num_tokens].replay()
+        return x

cacheflow/model_executor/parallel_utils/tensor_parallel/__init__.py

@@ -0,0 +1,55 @@
+from .layers import (
+    ColumnParallelLinear,
+    RowParallelLinear,
+    VocabParallelEmbedding,
+    set_tensor_model_parallel_attributes,
+    set_defaults_if_not_set_tensor_model_parallel_attributes,
+    copy_tensor_model_parallel_attributes,
+    param_is_not_tensor_parallel_duplicate,
+)
+
+from .mappings import (
+    copy_to_tensor_model_parallel_region,
+    gather_from_tensor_model_parallel_region,
+    gather_from_sequence_parallel_region,
+    scatter_to_tensor_model_parallel_region,
+    scatter_to_sequence_parallel_region,
+)
+
+from .random import (
+    checkpoint,
+    get_cuda_rng_tracker,
+    model_parallel_cuda_manual_seed,
+)
+
+from .utils import (
+    split_tensor_along_last_dim,
+    split_tensor_into_1d_equal_chunks,
+    gather_split_1d_tensor,
+)
+
+__all__ = [
+    #layers.py
+    "ColumnParallelLinear",
+    "RowParallelLinear",
+    "VocabParallelEmbedding",
+    "set_tensor_model_parallel_attributes",
+    "set_defaults_if_not_set_tensor_model_parallel_attributes",
+    "copy_tensor_model_parallel_attributes",
+    "param_is_not_tensor_parallel_duplicate",
+    # mappings.py
+    "copy_to_tensor_model_parallel_region",
+    "gather_from_tensor_model_parallel_region",
+    "gather_from_sequence_parallel_region",
+#    "reduce_from_tensor_model_parallel_region",
+    "scatter_to_tensor_model_parallel_region",
+    "scatter_to_sequence_parallel_region",
+    # random.py
+    "checkpoint",
+    "get_cuda_rng_tracker",
+    "model_parallel_cuda_manual_seed",
+    # utils.py
+    "split_tensor_along_last_dim",
+    "split_tensor_into_1d_equal_chunks",
+    "gather_split_1d_tensor",
+]

cacheflow/model_executor/parallel_utils/tensor_parallel/layers.py

@@ -0,0 +1,446 @@
+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+# Parts of the code here are adapted from PyTorch
+# repo: https://github.com/pytorch/pytorch
+
+
+import torch
+import torch.nn.functional as F
+import torch.nn.init as init
+from torch.nn.parameter import Parameter
+
+from cacheflow.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+    get_all_reduce_launcher,
+)
+from .mappings import (
+    copy_to_tensor_model_parallel_region,
+    gather_from_tensor_model_parallel_region,
+    reduce_from_tensor_model_parallel_region,
+    scatter_to_tensor_model_parallel_region,
+)
+
+from .random import get_cuda_rng_tracker
+from .utils import (
+    divide,
+    VocabUtility,
+)
+
+_MODEL_PARALLEL_ATTRIBUTE_DEFAULTS = {'tensor_model_parallel': False,
+                                      'partition_dim': -1,
+                                      'partition_stride': 1}
+
+def param_is_not_tensor_parallel_duplicate(param):
+    return (hasattr(param, 'tensor_model_parallel') and
+            param.tensor_model_parallel) or (
+                get_tensor_model_parallel_rank() == 0)
+
+
+def set_tensor_model_parallel_attributes(tensor, is_parallel, dim, stride):
+    # Make sure the attributes are not set.
+    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
+        assert not hasattr(tensor, attribute)
+    # Set the attributes.
+    setattr(tensor, 'tensor_model_parallel', is_parallel)
+    setattr(tensor, 'partition_dim', dim)
+    setattr(tensor, 'partition_stride', stride)
+
+
+def set_defaults_if_not_set_tensor_model_parallel_attributes(tensor):
+    def maybe_set(attribute, value):
+        if not hasattr(tensor, attribute):
+            setattr(tensor, attribute, value)
+    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
+        maybe_set(attribute, _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS[attribute])
+
+
+def copy_tensor_model_parallel_attributes(destination_tensor, source_tensor):
+    def maybe_copy(attribute):
+        if hasattr(source_tensor, attribute):
+            setattr(destination_tensor, attribute,
+                    getattr(source_tensor, attribute))
+    for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
+        maybe_copy(attribute)
+
+
+def _initialize_affine_weight_gpu(weight, init_method,
+                                  partition_dim, stride=1):
+    """Initialize affine weight for model parallel on GPU."""
+
+    set_tensor_model_parallel_attributes(tensor=weight,
+                                         is_parallel=True,
+                                         dim=partition_dim,
+                                         stride=stride)
+
+    with get_cuda_rng_tracker().fork():
+        init_method(weight)
+
+
+def _initialize_affine_weight_cpu(weight, output_size, input_size,
+                                  per_partition_size, partition_dim,
+                                  init_method, stride=1,
+                                  return_master_weight=False,
+                                  *, params_dtype=None):
+    """Initialize affine weight for model parallel.
+
+    Build the master weight on all processes and scatter
+    the relevant chunk."""
+
+    set_tensor_model_parallel_attributes(tensor=weight,
+                                         is_parallel=True,
+                                         dim=partition_dim,
+                                         stride=stride)
+
+    if params_dtype is None:
+        params_dtype = torch.get_default_dtype()
+
+    # Initialize master weight
+    master_weight = torch.empty(output_size, input_size,
+                                dtype=torch.float,
+                                requires_grad=False)
+    init_method(master_weight)
+    master_weight = master_weight.to(dtype=params_dtype)
+
+    # Split and copy
+    per_partition_per_stride_size = divide(per_partition_size, stride)
+    weight_list = torch.split(master_weight, per_partition_per_stride_size,
+                              dim=partition_dim)
+    rank = get_tensor_model_parallel_rank()
+    world_size = get_tensor_model_parallel_world_size()
+    my_weight_list = weight_list[rank::world_size]
+
+    with torch.no_grad():
+        torch.cat(my_weight_list, dim=partition_dim, out=weight)
+    if return_master_weight:
+        return master_weight
+    return None
+
+
+class VocabParallelEmbedding(torch.nn.Module):
+    """Embedding parallelized in the vocabulary dimension.
+
+    This is mainly adapted from torch.nn.Embedding and all the default
+    values are kept.
+    Arguments:
+        num_embeddings: vocabulary size.
+        embedding_dim: size of hidden state.
+
+    Keyword Arguments:
+        init_method: method to initialize weights.
+        params_dtype
+        use_cpu_initialization
+        perform_initialization
+    """
+
+    def __init__(self, num_embeddings: int, embedding_dim: int, *,
+                 init_method=init.xavier_normal_,
+                 params_dtype: torch.dtype=None,
+                 use_cpu_initialization: bool=False,
+                 perform_initialization: bool=True):
+        super(VocabParallelEmbedding, self).__init__()
+        # Keep the input dimensions.
+        self.num_embeddings = num_embeddings
+        self.embedding_dim = embedding_dim
+        if params_dtype is None:
+            params_dtype = torch.get_default_dtype()
+
+        # Set the defaults for compatibility.
+        self.padding_idx = None
+        self.max_norm = None
+        self.norm_type = 2.
+        self.scale_grad_by_freq = False
+        self.sparse = False
+        self._weight = None
+        self.tensor_model_parallel_size = get_tensor_model_parallel_world_size()
+        # Divide the weight matrix along the vocaburaly dimension.
+        self.vocab_start_index, self.vocab_end_index = \
+            VocabUtility.vocab_range_from_global_vocab_size(
+                self.num_embeddings, get_tensor_model_parallel_rank(),
+                self.tensor_model_parallel_size)
+        self.num_embeddings_per_partition = self.vocab_end_index - \
+            self.vocab_start_index
+
+        # Allocate weights and initialize.
+        if use_cpu_initialization:
+            self.weight = Parameter(torch.empty(
+                self.num_embeddings_per_partition, self.embedding_dim,
+                dtype=params_dtype))
+            if perform_initialization:
+                _initialize_affine_weight_cpu(
+                    self.weight, self.num_embeddings, self.embedding_dim,
+                    self.num_embeddings_per_partition, 0, init_method,
+                    params_dtype=params_dtype)
+        else:
+            self.weight = Parameter(torch.empty(
+                self.num_embeddings_per_partition, self.embedding_dim,
+                device=torch.cuda.current_device(), dtype=params_dtype))
+            if perform_initialization:
+                _initialize_affine_weight_gpu(self.weight, init_method,
+                                              partition_dim=0, stride=1)
+
+    def forward(self, input_):
+        if self.tensor_model_parallel_size > 1:
+            # Build the mask.
+            input_mask = (input_ < self.vocab_start_index) | \
+                         (input_ >= self.vocab_end_index)
+            # Mask the input.
+            masked_input = input_.clone() - self.vocab_start_index
+            masked_input[input_mask] = 0
+        else:
+            masked_input = input_
+            # Get the embeddings.
+        output_parallel = F.embedding(masked_input, self.weight,
+                                      self.padding_idx, self.max_norm,
+                                      self.norm_type, self.scale_grad_by_freq,
+                                      self.sparse)
+        # Mask the output embedding.
+        if self.tensor_model_parallel_size > 1:
+            output_parallel[input_mask, :] = 0.0
+        # Reduce across all the model parallel GPUs.
+        output = reduce_from_tensor_model_parallel_region(output_parallel)
+        return output
+
+
+class ColumnParallelLinear(torch.nn.Module):
+    """Linear layer with column parallelism.
+
+    The linear layer is defined as Y = XA + b. A is parallelized along
+    its second dimension as A = [A_1, ..., A_p].
+
+    Arguments:
+        input_size: first dimension of matrix A.
+        output_size: second dimension of matrix A.
+
+    Keyword Arguments
+        bias: If true, add bias
+        gather_output: If true, call all-gather on output and make Y available
+                       to all GPUs, otherwise, every GPU will have its output
+                       which is Y_i = XA_i
+        init_method: method to initialize weights. Note that bias is always set
+                     to zero.
+        stride: For the strided linear layers.
+        keep_master_weight_for_test: This was added for testing and should be
+                                     set to False. It returns the master weights
+                                     used for initialization.
+        skip_bias_add: This was added to enable performance optimations where bias
+                       can be fused with other elementwise operations. we skip
+                       adding bias but instead return it.
+        params_dtype:
+        use_cpu_initialization:
+    """
+
+    def __init__(self, input_size, output_size, *,
+                 bias=True, gather_output=True,
+                 init_method=init.xavier_normal_, stride=1,
+                 keep_master_weight_for_test=False,
+                 skip_bias_add=False,
+                 params_dtype=None,
+                 use_cpu_initialization=False,
+                 perform_initialization=True,
+                 ):
+        super(ColumnParallelLinear, self).__init__()
+
+        # Keep input parameters
+        self.input_size = input_size
+        self.output_size = output_size
+        self.gather_output = gather_output
+        # Divide the weight matrix along the last dimension.
+        world_size = get_tensor_model_parallel_world_size()
+        self.output_size_per_partition = divide(output_size, world_size)
+        self.skip_bias_add = skip_bias_add
+
+        if params_dtype is None:
+            params_dtype = torch.get_default_dtype()
+
+        # Parameters.
+        # Note: torch.nn.functional.linear performs XA^T + b and as a result
+        # we allocate the transpose.
+        # Initialize weight.
+        if use_cpu_initialization:
+            self.weight = Parameter(torch.empty(self.output_size_per_partition,
+                                                self.input_size,
+                                                dtype=params_dtype))
+            if perform_initialization:
+                self.master_weight = _initialize_affine_weight_cpu(
+                    self.weight, self.output_size, self.input_size,
+                    self.output_size_per_partition, 0, init_method,
+                    stride=stride, return_master_weight=keep_master_weight_for_test)
+        else:
+            self.weight = Parameter(torch.empty(
+                self.output_size_per_partition, self.input_size,
+                device=torch.cuda.current_device(), dtype=params_dtype))
+            if perform_initialization:
+                _initialize_affine_weight_gpu(self.weight, init_method,
+                                              partition_dim=0, stride=stride)
+
+        if bias:
+            if use_cpu_initialization:
+                self.bias = Parameter(torch.empty(
+                    self.output_size_per_partition, dtype=params_dtype))
+            else:
+                self.bias = Parameter(torch.empty(
+                    self.output_size_per_partition,
+                    device=torch.cuda.current_device(),
+                    dtype=params_dtype))
+            set_tensor_model_parallel_attributes(self.bias, True, 0, stride)
+            # Always initialize bias to zero.
+            with torch.no_grad():
+                self.bias.zero_()
+        else:
+            self.register_parameter('bias', None)
+
+
+    def forward(self, input_):
+        """Forward of ColumnParallelLinear
+
+        Args:
+            input_: 3D tensor whose order of dimension is [sequence, batch, hidden]
+
+        Returns:
+            - output
+            - bias
+        """
+        bias = self.bias if not self.skip_bias_add else None
+
+        input_parallel = copy_to_tensor_model_parallel_region(input_)
+        # Matrix multiply.
+        output_parallel = F.linear(input_parallel, self.weight, bias)
+        if self.gather_output:
+            # All-gather across the partitions.
+            output = gather_from_tensor_model_parallel_region(output_parallel)
+        else:
+            output = output_parallel
+        output_bias = self.bias if self.skip_bias_add else None
+        return output, output_bias
+
+
+class RowParallelLinear(torch.nn.Module):
+    """Linear layer with row parallelism.
+
+    The linear layer is defined as Y = XA + b. A is parallelized along
+    its first dimension and X along its second dimension as:
+               -   -
+              | A_1 |
+              | .   |
+          A = | .   |        X = [X_1, ..., X_p]
+              | .   |
+              | A_p |
+               -   -
+    Arguments:
+        input_size: first dimension of matrix A.
+        output_size: second dimension of matrix A.
+
+    Keyword Arguments:
+        bias: If true, add bias. Note that bias is not parallelized.
+        input_is_parallel: If true, we assume that the input is already
+                           split across the GPUs and we do not split
+                           again.
+        init_method: method to initialize weights. Note that bias is always set
+                     to zero.
+        stride: For the strided linear layers.
+        keep_master_weight_for_test: This was added for testing and should be
+                                     set to False. It returns the master weights
+                                     used for initialization.
+        skip_bias_add: This was added to enable performance optimization where bias
+                       can be fused with other elementwise operations. We skip
+                       adding bias but instead return it.
+        params_dtype:
+        use_cpu_initialization:
+        perform_initialization:
+    """
+
+    def __init__(self, input_size, output_size, *,
+                 bias=True, input_is_parallel=False,
+                 init_method=init.xavier_normal_, stride=1,
+                 keep_master_weight_for_test=False,
+                 skip_bias_add=False,
+                 params_dtype=None,
+                 use_cpu_initialization=False,
+                 perform_initialization=True,
+                 ):
+        super(RowParallelLinear, self).__init__()
+
+        # Keep input parameters
+        self.input_size = input_size
+        self.output_size = output_size
+        self.input_is_parallel = input_is_parallel
+        if params_dtype is None:
+            params_dtype = torch.get_default_dtype()
+
+        # Divide the weight matrix along the last dimension.
+        world_size = get_tensor_model_parallel_world_size()
+        self.input_size_per_partition = divide(input_size, world_size)
+        self.skip_bias_add = skip_bias_add
+
+        # Parameters.
+        # Note: torch.nn.functional.linear performs XA^T + b and as a result
+        # we allocate the transpose.
+        # Initialize weight.
+        if use_cpu_initialization:
+            self.weight = Parameter(torch.empty(self.output_size,
+                                                self.input_size_per_partition,
+                                                dtype=params_dtype))
+            if perform_initialization:
+                self.master_weight = _initialize_affine_weight_cpu(
+                    self.weight, self.output_size, self.input_size,
+                    self.input_size_per_partition, 1, init_method,
+                    stride=stride, return_master_weight=keep_master_weight_for_test,
+                    params_dtype=params_dtype)
+        else:
+            self.weight = Parameter(torch.empty(
+                self.output_size, self.input_size_per_partition,
+                device=torch.cuda.current_device(), dtype=params_dtype))
+            if perform_initialization:
+                _initialize_affine_weight_gpu(self.weight, init_method,
+                                              partition_dim=1, stride=stride)
+        if bias:
+            if use_cpu_initialization:
+                self.bias = Parameter(torch.empty(self.output_size,
+                                                  dtype=params_dtype))
+            else:
+                self.bias = Parameter(torch.empty(
+                    self.output_size, device=torch.cuda.current_device(),
+                    dtype=params_dtype))
+
+            # Always initialize bias to zero.
+            with torch.no_grad():
+                self.bias.zero_()
+        else:
+            self.register_parameter('bias', None)
+        self.weight_t = self.weight.t()
+
+    def forward(self, input_):
+        """Forward of RowParallelLinear
+
+        Args:
+            input_: 3D tensor whose order of dimension is [sequence, batch, hidden]
+
+        Returns:
+            - output
+            - bias
+        """
+        # Set up backprop all-reduce.
+        if self.input_is_parallel:
+            input_parallel = input_
+        else:
+            input_parallel = scatter_to_tensor_model_parallel_region(input_)
+        if get_tensor_model_parallel_world_size() == 1:
+            # Matrix multiply.
+            output_ = F.linear(input_parallel, self.weight)
+        else:
+            # Matrix multiply.
+            all_reduce_launcher = get_all_reduce_launcher()
+            num_tokens = input_parallel.shape[0]
+            output_buffer = all_reduce_launcher.buffer[:num_tokens]
+            torch.matmul(input_parallel, self.weight_t, out=output_buffer)
+            # All-reduce across all the partitions.
+            output_ = all_reduce_launcher.launch(output_buffer)
+
+        if not self.skip_bias_add:
+            output = output_ + self.bias if self.bias is not None else output_
+            output_bias = None
+        else:
+            output = output_
+            output_bias = self.bias
+        return output, output_bias

cacheflow/model_executor/parallel_utils/tensor_parallel/mappings.py

@@ -0,0 +1,279 @@
+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+import torch
+
+from cacheflow.model_executor.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+    get_tensor_model_parallel_group,
+)
+from .utils import split_tensor_along_last_dim
+
+
+def _reduce(input_):
+    """All-reduce the input tensor across model parallel group."""
+
+    # Bypass the function if we are using only 1 GPU.
+    if get_tensor_model_parallel_world_size()==1:
+        return input_
+
+    # All-reduce.
+    torch.distributed.all_reduce(input_, group=get_tensor_model_parallel_group())
+
+    return input_
+
+
+def _split_along_last_dim(input_):
+    """Split the tensor along its last dimension and keep the
+    corresponding slice."""
+
+    world_size = get_tensor_model_parallel_world_size()
+    # Bypass the function if we are using only 1 GPU.
+    if world_size == 1:
+        return input_
+
+    # Split along last dimension.
+    input_list = split_tensor_along_last_dim(input_, world_size)
+
+    # Note: torch.split does not create contiguous tensors by default.
+    rank = get_tensor_model_parallel_rank()
+    output = input_list[rank].contiguous()
+
+    return output
+
+
+def _split_along_first_dim(input_):
+    """Split the tensor along its first dimension and keep the
+    corresponding slice."""
+
+    world_size = get_tensor_model_parallel_world_size()
+    # Bypass the function if we are using only 1 GPU.
+    if world_size == 1:
+        return input_
+
+    # Split along first dimension.
+    dim_size = input_.size()[0]
+    assert dim_size % world_size == 0, \
+        "First dimension of the tensor should be divisible by tensor parallel size"
+    local_dim_size = dim_size // world_size
+    rank = get_tensor_model_parallel_rank()
+    dim_offset = rank * local_dim_size
+
+    output = input_[dim_offset:dim_offset+local_dim_size].contiguous()
+
+    return output
+
+
+def _gather_along_last_dim(input_):
+    """Gather tensors and concatinate along the last dimension."""
+
+    world_size = get_tensor_model_parallel_world_size()
+    # Bypass the function if we are using only 1 GPU.
+    if world_size == 1:
+        return input_
+
+    # Size and dimension.
+    last_dim = input_.dim() - 1
+    rank = get_tensor_model_parallel_rank()
+
+    tensor_list = [torch.empty_like(input_) for _ in range(world_size)]
+    tensor_list[rank] = input_
+    torch.distributed.all_gather(tensor_list, input_, group=get_tensor_model_parallel_group())
+
+    # Note: torch.cat already creates a contiguous tensor.
+    output = torch.cat(tensor_list, dim=last_dim).contiguous()
+
+    return output
+
+
+def _gather_along_first_dim(input_):
+    """Gather tensors and concatinate along the first dimension."""
+
+    world_size = get_tensor_model_parallel_world_size()
+    # Bypass the function if we are using only 1 GPU.
+    if world_size == 1:
+        return input_
+
+    dim_size = list(input_.size())
+    dim_size[0] = dim_size[0] * world_size
+
+    output = torch.empty(dim_size, dtype=input_.dtype,
+                         device=torch.cuda.current_device())
+    torch.distributed._all_gather_base(output, input_.contiguous(),
+                                       group=get_tensor_model_parallel_group())
+
+    return output
+
+def _reduce_scatter_along_first_dim(input_):
+    """Reduce-scatter the input tensor across model parallel group."""
+    world_size = get_tensor_model_parallel_world_size()
+    # Bypass the function if we are using only 1 GPU.
+    if world_size == 1:
+        return input_
+
+    dim_size = list(input_.size())
+    assert dim_size[0] % world_size == 0, \
+        "First dimension of the tensor should be divisible by tensor parallel size"
+
+    dim_size[0] = dim_size[0] // world_size
+
+    output = torch.empty(dim_size, dtype=input_.dtype,
+                         device=torch.cuda.current_device())
+    torch.distributed._reduce_scatter_base(output, input_.contiguous(),
+                                           group=get_tensor_model_parallel_group())
+    return output
+
+
+class _CopyToModelParallelRegion(torch.autograd.Function):
+    """Pass the input to the model parallel region."""
+
+    @staticmethod
+    def symbolic(graph, input_):
+        return input_
+
+    @staticmethod
+    def forward(ctx, input_):
+        return input_
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return _reduce(grad_output)
+
+
+class _ReduceFromModelParallelRegion(torch.autograd.Function):
+    """All-reduce the input from the model parallel region."""
+
+    @staticmethod
+    def symbolic(graph, input_):
+        return _reduce(input_)
+
+    @staticmethod
+    def forward(ctx, input_):
+        return _reduce(input_)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return grad_output
+
+
+class _ScatterToModelParallelRegion(torch.autograd.Function):
+    """Split the input and keep only the corresponding chuck to the rank."""
+
+    @staticmethod
+    def symbolic(graph, input_):
+        return _split_along_last_dim(input_)
+
+    @staticmethod
+    def forward(ctx, input_):
+        return _split_along_last_dim(input_)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return _gather_along_last_dim(grad_output)
+
+
+class _GatherFromModelParallelRegion(torch.autograd.Function):
+    """Gather the input from model parallel region and concatinate."""
+
+    @staticmethod
+    def symbolic(graph, input_):
+        return _gather_along_last_dim(input_)
+
+    @staticmethod
+    def forward(ctx, input_):
+        return _gather_along_last_dim(input_)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return _split_along_last_dim(grad_output)
+
+
+class _ScatterToSequenceParallelRegion(torch.autograd.Function):
+    """Split the input and keep only the corresponding chuck to the rank."""
+
+    @staticmethod
+    def symbolic(graph, input_):
+        return _split_along_first_dim(input_)
+
+    @staticmethod
+    def forward(ctx, input_):
+        return _split_along_first_dim(input_)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return _gather_along_first_dim(grad_output)
+
+
+class _GatherFromSequenceParallelRegion(torch.autograd.Function):
+    """Gather the input from sequence parallel region and concatinate."""
+
+    @staticmethod
+    def symbolic(graph, input_, tensor_parallel_output_grad=True):
+        return _gather_along_first_dim(input_)
+
+    @staticmethod
+    def forward(ctx, input_, tensor_parallel_output_grad=True):
+        ctx.tensor_parallel_output_grad = tensor_parallel_output_grad
+        return _gather_along_first_dim(input_)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        tensor_parallel_output_grad = ctx.tensor_parallel_output_grad
+
+        # If the computation graph after the gather operation is
+        # in the tensor parallel mode, output gradients need to reduce
+        # scattered and whereas if the computation is duplicated,
+        # output gradients need to be scattered.
+        if tensor_parallel_output_grad:
+            return _reduce_scatter_along_first_dim(grad_output), None
+        else:
+            return _split_along_first_dim(grad_output), None
+
+
+class _ReduceScatterToSequenceParallelRegion(torch.autograd.Function):
+    """Reduce scatter the input from the model parallel region."""
+
+    @staticmethod
+    def symbolic(graph, input_):
+        return _reduce_scatter_along_first_dim(input_)
+
+    @staticmethod
+    def forward(ctx, input_):
+        return _reduce_scatter_along_first_dim(input_)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return _gather_along_first_dim(grad_output)
+
+
+# -----------------
+# Helper functions.
+# -----------------
+
+def copy_to_tensor_model_parallel_region(input_):
+    return _CopyToModelParallelRegion.apply(input_)
+
+
+def reduce_from_tensor_model_parallel_region(input_):
+    return _ReduceFromModelParallelRegion.apply(input_)
+
+
+def scatter_to_tensor_model_parallel_region(input_):
+    return _ScatterToModelParallelRegion.apply(input_)
+
+
+def gather_from_tensor_model_parallel_region(input_):
+    return _GatherFromModelParallelRegion.apply(input_)
+
+
+def scatter_to_sequence_parallel_region(input_):
+    return _ScatterToSequenceParallelRegion.apply(input_)
+
+
+def gather_from_sequence_parallel_region(input_, tensor_parallel_output_grad=True):
+    return _GatherFromSequenceParallelRegion.apply(input_, tensor_parallel_output_grad)
+
+
+def reduce_scatter_to_sequence_parallel_region(input_):
+    return _ReduceScatterToSequenceParallelRegion.apply(input_)
+

cacheflow/model_executor/parallel_utils/tensor_parallel/random.py

@@ -0,0 +1,253 @@
+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+# Parts of the code here are adapted from PyTorch
+# repo: https://github.com/pytorch/pytorch
+
+import contextlib
+
+import torch
+from torch import _C
+from torch.cuda import _lazy_call, device as device_ctx_manager
+from torch.utils.checkpoint import detach_variable
+
+from cacheflow.model_executor.parallel_utils.parallel_state import (
+    get_data_parallel_rank,
+    get_tensor_model_parallel_group,
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+)
+
+from .utils import (
+    split_tensor_into_1d_equal_chunks,
+    gather_split_1d_tensor,
+)
+
+from cacheflow.model_executor.parallel_utils.utils import safely_set_viewless_tensor_data
+
+# Default name for the model parallel rng tracker.
+_MODEL_PARALLEL_RNG_TRACKER_NAME = 'model-parallel-rng'
+
+
+def _set_cuda_rng_state(new_state, device=-1):
+    """Sets the random number generator state of the current GPU.
+
+    Argumentss:
+        new_state (torch.ByteTensor): The desired state
+    This function is adapted from PyTorch repo (torch.cuda.set_rng_state)
+    with a single change: the input state is not cloned. Cloning caused
+    major performance issues for +4 GPU cases.
+    """
+    if hasattr(_C, '_cuda_setRNGState') and callable(_C._cuda_setRNGState):
+        # older PyTorch
+        def cb():
+            with device_ctx_manager(device):
+                _C._cuda_setRNGState(new_state)
+    else:
+        # newer PyTorch
+        if device == -1:
+            device = torch.device('cuda')
+        elif isinstance(device, str):
+            device = torch.device(device)
+        elif isinstance(device, int):
+            device = torch.device('cuda', device)
+
+        def cb():
+            idx = device.index
+            if idx is None:
+                idx = torch.cuda.current_device()
+            default_generator = torch.cuda.default_generators[idx]
+            default_generator.set_state(new_state)
+
+    _lazy_call(cb)
+
+
+
+class CudaRNGStatesTracker:
+    """Tracker for the cuda RNG states.
+
+    Using the `add` method, a cuda rng state is initialized based on
+    the input `seed` and is assigned to `name`. Later, by forking the
+    rng state, we can perform operations and return to our starting
+    cuda state.
+    """
+
+    def __init__(self):
+        # Map from a string name to the cuda rng state.
+        self.states_ = {}
+        # Seeds are just for book keeping and ensure no seed is set twice.
+        self.seeds_ = set()
+
+    def reset(self):
+        """Set to the initial state (no tracker)."""
+        self.states_ = {}
+        self.seeds_ = set()
+
+    def get_states(self):
+        """Get rng states. Copy the dictionary so we have direct
+        pointers to the states, not just a pointer to the dictionary."""
+        states = {}
+        for name in self.states_:
+            states[name] = self.states_[name]
+        return states
+
+    def set_states(self, states):
+        """Set the rng states. For efficiency purposes, we do not check
+        the size of seed for compatibility."""
+        self.states_ = states
+
+    def add(self, name, seed):
+        """Track the rng state."""
+        # Check seed is not already used.
+        if seed in self.seeds_:
+            raise Exception('seed {} already exists'.format(seed))
+        self.seeds_.add(seed)
+        # Check that state is not already defined.
+        if name in self.states_:
+            raise Exception('cuda rng state {} already exists'.format(name))
+        # Get the current rng state.
+        orig_rng_state = torch.cuda.get_rng_state()
+        # Set the new state and store it.
+        torch.cuda.manual_seed(seed)
+        self.states_[name] = torch.cuda.get_rng_state()
+        # Reset rng state to what it was.
+        _set_cuda_rng_state(orig_rng_state)
+
+    @contextlib.contextmanager
+    def fork(self, name=_MODEL_PARALLEL_RNG_TRACKER_NAME):
+        """Fork the cuda rng state, perform operations, and exit with
+        the original state."""
+        # Check if we have added the state
+        if name not in self.states_:
+            raise Exception('cuda rng state {} is not added'.format(name))
+        # Store current rng state.
+        orig_cuda_rng_state = torch.cuda.get_rng_state()
+        # Set rng state to the desired one
+        _set_cuda_rng_state(self.states_[name])
+        # Do the stuff we wanted to do.
+        try:
+            yield
+        finally:
+            # Update the current rng state for later use.
+            self.states_[name] = torch.cuda.get_rng_state()
+            # And set the state to the original state we started with.
+            _set_cuda_rng_state(orig_cuda_rng_state)
+
+
+# RNG tracker object.
+_CUDA_RNG_STATE_TRACKER = CudaRNGStatesTracker()
+
+
+def get_cuda_rng_tracker():
+    """Get cuda rng tracker."""
+    return _CUDA_RNG_STATE_TRACKER
+
+
+def model_parallel_cuda_manual_seed(seed):
+    """Initialize model parallel cuda seed.
+
+    This function should be called after the model parallel is
+    initialized. Also, no torch.cuda.manual_seed should be called
+    after this function. Basically, this is replacement for that
+    function.
+    Two set of RNG states are tracked:
+        default state: This is for data parallelism and is the same among a
+                       set of model parallel GPUs but different across
+                       different model paralle groups. This is used for
+                       example for dropout in the non-tensor-model-parallel regions.
+        tensor-model-parallel state: This state is different among a set of model
+                              parallel GPUs, but the same across data parallel
+                              groups. This is used for example for dropout in
+                              model parallel regions.
+    """
+    # 2718 is just for fun and any POSITIVE value will work.
+    offset = seed + 2718
+    tensor_model_parallel_seed = offset + get_tensor_model_parallel_rank()
+    # Data parallel gets the original seed.
+    data_parallel_seed = seed
+
+    _CUDA_RNG_STATE_TRACKER.reset()
+    # Set the default state.
+    torch.cuda.manual_seed(data_parallel_seed)
+    # and model parallel state.
+    _CUDA_RNG_STATE_TRACKER.add(_MODEL_PARALLEL_RNG_TRACKER_NAME,
+                                tensor_model_parallel_seed)
+
+
+class CheckpointFunction(torch.autograd.Function):
+    """This function is adapted from torch.utils.checkpoint with
+       two main changes:
+           1) torch.cuda.set_rng_state is replaced with `_set_cuda_rng_state`
+           2) the states in the model parallel tracker are also properly
+              tracked/set/reset.
+    """
+    @staticmethod
+    def forward(ctx, run_function, distribute_saved_activations, *args):
+        ctx.run_function = run_function
+        ctx.distribute_saved_activations \
+            = distribute_saved_activations
+
+        # Copy the rng states.
+        ctx.fwd_cpu_rng_state = torch.get_rng_state()
+        ctx.fwd_cuda_rng_state = torch.cuda.get_rng_state()
+        ctx.fwd_cuda_rng_state_tracker = get_cuda_rng_tracker().get_states()
+
+        with torch.no_grad():
+            outputs = run_function(*args)
+
+        # Divide hidden states across model parallel group and only keep
+        # the chunk corresponding to the current rank.
+        if distribute_saved_activations:
+            ctx.input_0_shape = args[0].data.shape
+            safely_set_viewless_tensor_data(
+                args[0],
+                split_tensor_into_1d_equal_chunks(args[0].data, new_buffer=True))
+
+        # Store everything.
+        ctx.save_for_backward(*args)
+
+        return outputs
+
+    @staticmethod
+    def backward(ctx, *args):
+        if not torch.autograd._is_checkpoint_valid():
+            raise RuntimeError("Checkpointing is not compatible with .grad(), "
+                               "please use .backward() if possible")
+        inputs = ctx.saved_tensors
+        if ctx.distribute_saved_activations:
+            safely_set_viewless_tensor_data(
+                inputs[0],
+                gather_split_1d_tensor(inputs[0].data).view(ctx.input_0_shape))
+
+        # Store the current states.
+        bwd_cpu_rng_state = torch.get_rng_state()
+        bwd_cuda_rng_state = torch.cuda.get_rng_state()
+        bwd_cuda_rng_state_tracker = get_cuda_rng_tracker().get_states()
+
+        # Set the states to what it used to be before the forward pass.
+        torch.set_rng_state(ctx.fwd_cpu_rng_state)
+        _set_cuda_rng_state(ctx.fwd_cuda_rng_state)
+        get_cuda_rng_tracker().set_states(ctx.fwd_cuda_rng_state_tracker)
+
+        # Compute the forward pass.
+        detached_inputs = detach_variable(inputs)
+        with torch.enable_grad():
+            outputs = ctx.run_function(*detached_inputs)
+
+        # Set the states back to what it was at the start of this function.
+        torch.set_rng_state(bwd_cpu_rng_state)
+        _set_cuda_rng_state(bwd_cuda_rng_state)
+        get_cuda_rng_tracker().set_states(bwd_cuda_rng_state_tracker)
+
+        if isinstance(outputs, torch.Tensor):
+            outputs = (outputs,)
+        torch.autograd.backward(outputs, args)
+        grads = tuple(inp.grad if isinstance(inp, torch.Tensor) else inp
+                      for inp in detached_inputs)
+        return (None, None) + grads
+
+
+def checkpoint(function, distribute_saved_activations, *args):
+    """Checkpoint a model or part of the model.
+    This has been directly copied from torch.utils.checkpoint."""
+    return CheckpointFunction.apply(function,
+                                    distribute_saved_activations, *args)

cacheflow/model_executor/parallel_utils/tensor_parallel/utils.py

@@ -0,0 +1,108 @@
+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+import torch
+from typing import List, Sequence
+
+from cacheflow.model_executor.parallel_utils.utils import divide
+from cacheflow.model_executor.parallel_utils import parallel_state
+
+def split_tensor_along_last_dim(
+    tensor: torch.Tensor,
+    num_partitions: int,
+    contiguous_split_chunks: bool = False,
+) -> List[torch.Tensor]:
+    """ Split a tensor along its last dimension.
+
+        Arguments:
+            tensor: input tensor.
+            num_partitions: number of partitions to split the tensor
+            contiguous_split_chunks: If True, make each chunk contiguous
+                                     in memory.
+
+        Returns:
+            A list of Tensors
+    """
+    # Get the size and dimension.
+    last_dim = tensor.dim() - 1
+    last_dim_size = divide(tensor.size()[last_dim], num_partitions)
+    # Split.
+    tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
+    # Note: torch.split does not create contiguous tensors by default.
+    if contiguous_split_chunks:
+        return tuple(chunk.contiguous() for chunk in tensor_list)
+
+    return tensor_list
+
+def split_tensor_into_1d_equal_chunks(tensor, new_buffer=False):
+    """ Break a tensor into equal 1D chunks across tensor parallel ranks.
+
+        Returns a Tensor or View with this rank's portion of the data.
+
+        Arguments:
+            tensor: The tensor to split
+
+        Keyword Arguments:
+            new_buffer (bool): If True, returns a new Tensor.
+                               If False, returns a view into the existing Tensor.
+                               Default is False
+
+    """
+    partition_size = torch.numel(tensor) // \
+        parallel_state.get_tensor_model_parallel_world_size()
+    start_index = partition_size * parallel_state.get_tensor_model_parallel_rank()
+    end_index = start_index + partition_size
+    if new_buffer:
+        data = torch.empty(partition_size, dtype=tensor.dtype,
+                           device=torch.cuda.current_device(),
+                           requires_grad=False)
+        data.copy_(tensor.view(-1)[start_index:end_index])
+    else:
+        data = tensor.view(-1)[start_index:end_index]
+    return data
+
+
+def gather_split_1d_tensor(tensor):
+    """ Opposite of split_tensor_into_1d_equal_chunks. Gather values from tensor
+        model parallel ranks.
+
+        Returns a new Tensor with the gathered data.
+
+        Arguments:
+            tensor: A Tensor or view of this rank's portion of the data.
+    """
+    numel_gathered = torch.numel(tensor) * \
+        parallel_state.get_tensor_model_parallel_world_size()
+    gathered = torch.empty(numel_gathered, dtype=tensor.dtype,
+                           device=torch.cuda.current_device(),
+                           requires_grad=False)
+    # TODO: This API is experimental in pytorch (as of Feb 2022) and
+    # this might break in future pytorch releases. We chose this API
+    # as opposed to torch.distributed.all_gather for efficiency reasons.
+    # This API calls directly NCCL all-gather versus the former does
+    # internal copies and can potentially cause slow down.
+    torch.distributed._all_gather_base(gathered, tensor,
+                                       group=parallel_state.get_tensor_model_parallel_group())
+    return gathered
+
+
+class VocabUtility:
+    """ Split the vocabulary into `world_size` chunks and return the first
+        and last index of the vocabulary belonging to the `rank`
+        partition: Note that indices in [fist, last)
+
+    """
+
+    @staticmethod
+    def vocab_range_from_per_partition_vocab_size(
+        per_partition_vocab_size: int, rank, world_size: int
+    ) -> Sequence[int]:
+        index_f = rank * per_partition_vocab_size
+        index_l = index_f + per_partition_vocab_size
+        return index_f, index_l
+
+    @staticmethod
+    def vocab_range_from_global_vocab_size(global_vocab_size: int, rank: int, world_size: int) -> Sequence[int]:
+        per_partition_vocab_size = divide(global_vocab_size, world_size)
+        return VocabUtility.vocab_range_from_per_partition_vocab_size(
+            per_partition_vocab_size, rank, world_size
+        )

cacheflow/model_executor/parallel_utils/utils.py

@@ -0,0 +1,120 @@
+"""Utility functions used throughout Megatron core"""
+from functools import reduce
+import operator
+
+import torch
+
+from cacheflow.model_executor.parallel_utils import parallel_state
+
+
+def ensure_divisibility(numerator, denominator):
+    """Ensure that numerator is divisible by the denominator."""
+    assert numerator % denominator == 0, "{} is not divisible by {}".format(
+        numerator, denominator
+    )
+
+
+def divide(numerator, denominator):
+    """Ensure that numerator is divisible by the denominator and return
+    the division value."""
+    ensure_divisibility(numerator, denominator)
+    return numerator // denominator
+
+
+class GlobalMemoryBuffer:
+    """Global buffer to avoid dynamic memory allocations.
+    Caller should ensure that buffers of the same name
+    are not used concurrently."""
+
+    def __init__(self):
+        self.buffer = {}
+
+    def get_tensor(self, tensor_shape, dtype, name):
+        required_len = reduce(operator.mul, tensor_shape, 1)
+        if self.buffer.get((name, dtype), None) is None or \
+                self.buffer[(name, dtype)].numel() < required_len:
+            self.buffer[(name, dtype)] = \
+                torch.empty(required_len,
+                            dtype=dtype,
+                            device=torch.cuda.current_device(),
+                            requires_grad=False)
+
+        return self.buffer[(name, dtype)][0:required_len].view(*tensor_shape)
+
+def _kernel_make_viewless_tensor(inp, requires_grad):
+    '''Make a viewless tensor.
+
+    View tensors have the undesirable side-affect of retaining a reference
+    to the originally-viewed tensor, even after manually setting the '.data'
+    field. This method creates a new tensor that links to the old tensor's
+    data, without linking the viewed tensor, referenced via the '._base'
+    field.
+    '''
+    out = torch.empty(
+        (1,),
+        dtype = inp.dtype,
+        device = inp.device,
+        requires_grad = requires_grad,
+    )
+    out.data = inp.data
+    return out
+
+class MakeViewlessTensor(torch.autograd.Function):
+    '''
+    Autograd function to make a viewless tensor.
+
+    This function should be used in cases where the computation graph needs
+    to be propagated, but we only want a viewless tensor (e.g.,
+    ParallelTransformer's hidden_states). Call this function by passing
+    'keep_graph = True' to 'make_viewless_tensor()'.
+    '''
+    @staticmethod
+    def forward(ctx, inp, requires_grad):
+        return _kernel_make_viewless_tensor(inp, requires_grad)
+    @staticmethod
+    def backward(ctx, grad_output):
+        return grad_output, None
+
+def make_viewless_tensor(inp, requires_grad, keep_graph):
+    '''
+    Entry-point for creating viewless tensors.
+
+    This method should be used, rather than calling 'MakeViewlessTensor'
+    or '_kernel_make_viewless_tensor' directly. This method acts as a
+    switch for determining if an autograd function or a regular method
+    should be used to create the tensor.
+    '''
+
+    # return tensor as-is, if not a 'view'
+    if inp._base is None:
+        return inp
+
+    # create viewless tensor
+    if keep_graph:
+        return MakeViewlessTensor.apply(inp, requires_grad)
+    else:
+        return _kernel_make_viewless_tensor(inp, requires_grad)
+
+def assert_viewless_tensor(tensor, extra_msg = None):
+    '''Assert that a tensor is not a view (i.e., its '._base' field is
+    not set).'''
+    if isinstance(tensor, list):
+        [ assert_viewless_tensor(t) for t in tensor ]
+        return tensor
+    if not isinstance(tensor, torch.Tensor):
+        return tensor
+    assert tensor._base is None, (
+        "Ensure tensor._base is None before setting tensor.data or storing "
+        "tensor to memory buffer. Otherwise, a memory leak will occur (and "
+        "likely accumulate over iterations). %s"
+    ) % extra_msg
+    return tensor
+
+def safely_set_viewless_tensor_data(tensor, new_data_tensor):
+    '''Safely set tensor's '.data' field.
+
+    Check first that the tensor is viewless (i.e., '._base' not set). If not,
+    raise an exception.
+    '''
+    assert_viewless_tensor(tensor, extra_msg = "FYI, tensor._base has shape %s, and new_data_tensor has shape %s." % ("--" if tensor._base is None else tensor._base.shape, new_data_tensor.shape))
+    tensor.data = new_data_tensor