Merge branch 'prefetch' of github.com:botbw/ColossalAI into botbw-prefetch

pull/5722/head
genghaozhe 2024-05-16 07:23:40 +00:00
commit 1f6b57099c
7 changed files with 93 additions and 22 deletions

View File

@ -329,6 +329,7 @@ class GeminiPlugin(DPPluginBase):
chunk_init_device: Optional[torch.device] = None, chunk_init_device: Optional[torch.device] = None,
placement_policy: str = "static", placement_policy: str = "static",
enable_gradient_accumulation: bool = False, enable_gradient_accumulation: bool = False,
max_prefetch: int = 0,
shard_param_frac: float = 1.0, # only for static placement shard_param_frac: float = 1.0, # only for static placement
offload_optim_frac: float = 0.0, # only for static placement offload_optim_frac: float = 0.0, # only for static placement
offload_param_frac: float = 0.0, # only for static placement offload_param_frac: float = 0.0, # only for static placement
@ -386,6 +387,7 @@ class GeminiPlugin(DPPluginBase):
memstats=memstats, memstats=memstats,
mixed_precision=PRECISION_STR_TO_DTYPE[precision], mixed_precision=PRECISION_STR_TO_DTYPE[precision],
master_weights=master_weights, master_weights=master_weights,
max_prefetch=max_prefetch,
) )
self.zero_optim_config = dict( self.zero_optim_config = dict(
gpu_margin_mem_ratio=gpu_margin_mem_ratio, gpu_margin_mem_ratio=gpu_margin_mem_ratio,

View File

@ -357,14 +357,14 @@ class Chunk:
else: else:
raise NotImplementedError raise NotImplementedError
def access_chunk(self): def access_chunk(self, async_access: bool = False) -> Optional[dist.Work]:
"""Make the chunk usable for the parameters inside it. It's an operation done in CUDA.""" """Make the chunk usable for the parameters inside it. It's an operation done in CUDA."""
# sanity check # sanity check
assert self.chunk_temp is None assert self.chunk_temp is None
if not self.is_gathered: if not self.is_gathered:
self.__gather() return self.__gather(async_op=async_access)
self.__update_tensors_ptr() self.__update_tensors_ptr()
return None
def release_chunk(self): def release_chunk(self):
"""Release the usable chunk. It's an operation done in CUDA.""" """Release the usable chunk. It's an operation done in CUDA."""
@ -498,17 +498,19 @@ class Chunk:
def get_tensors(self) -> List[torch.Tensor]: def get_tensors(self) -> List[torch.Tensor]:
return list(self.tensors_info.keys()) return list(self.tensors_info.keys())
def __gather(self): def __gather(self, async_op: bool = False) -> Optional[dist.Work]:
if not self.is_gathered: if not self.is_gathered:
# sanity check # sanity check
assert self.cuda_shard is not None assert self.cuda_shard is not None
alloc_storage(self.cuda_global_chunk) alloc_storage(self.cuda_global_chunk)
gather_list = list(torch.chunk(input=self.cuda_global_chunk, chunks=self.pg_size, dim=0)) gather_list = list(torch.chunk(input=self.cuda_global_chunk, chunks=self.pg_size, dim=0))
dist.all_gather(gather_list, self.cuda_shard, self.torch_pg) work = dist.all_gather(gather_list, self.cuda_shard, self.torch_pg, async_op=async_op)
self.cuda_shard = None self.cuda_shard = None
self.is_gathered = True self.is_gathered = True
return work
return None
def __scatter(self): def __scatter(self):
if self.keep_gathered: if self.keep_gathered:

View File

@ -111,15 +111,16 @@ class ChunkManager:
for group_name in self.chunk_groups: for group_name in self.chunk_groups:
self.__close_one_chunk(self.chunk_groups[group_name][-1]) self.__close_one_chunk(self.chunk_groups[group_name][-1])
def access_chunk(self, chunk: Chunk) -> None: def access_chunk(self, chunk: Chunk, async_access: bool = False) -> Optional[dist.Work]:
"""Make the chunk can be used for calculation.""" """Make the chunk can be used for calculation."""
if chunk in self.accessed_chunks: if chunk in self.accessed_chunks:
return return None
self.__sub_memory_usage(chunk.memory_usage) self.__sub_memory_usage(chunk.memory_usage)
if chunk.device_type == "cpu": if chunk.device_type == "cpu":
chunk.shard_move(get_accelerator().get_current_device()) chunk.shard_move(get_accelerator().get_current_device())
self.__add_accessed_chunk(chunk) maybe_work = self.__add_accessed_chunk(chunk, async_access=async_access)
self.__add_memory_usage(chunk.memory_usage) self.__add_memory_usage(chunk.memory_usage)
return maybe_work
def release_chunk(self, chunk: Chunk) -> None: def release_chunk(self, chunk: Chunk) -> None:
"""Scatter the chunk in CUDA.""" """Scatter the chunk in CUDA."""
@ -251,10 +252,11 @@ class ChunkManager:
for k, v in usage.items(): for k, v in usage.items():
self.total_mem[k] += v self.total_mem[k] += v
def __add_accessed_chunk(self, chunk: Chunk): def __add_accessed_chunk(self, chunk: Chunk, async_access: bool = False) -> Optional[dist.Work]:
chunk.access_chunk() maybe_work = chunk.access_chunk(async_access=async_access)
self.accessed_chunks.add(chunk) self.accessed_chunks.add(chunk)
self.accessed_mem += chunk.chunk_mem self.accessed_mem += chunk.chunk_mem
return maybe_work
def __sub_accessed_chunk(self, chunk: Chunk): def __sub_accessed_chunk(self, chunk: Chunk):
chunk.release_chunk() chunk.release_chunk()

View File

@ -78,6 +78,7 @@ class GeminiDDP(ModelWrapper):
chunk_init_device: torch.device = torch.device("cpu"), chunk_init_device: torch.device = torch.device("cpu"),
placement_policy: str = "static", placement_policy: str = "static",
enable_gradient_accumulation: bool = False, enable_gradient_accumulation: bool = False,
max_prefetch: int = 0,
shard_param_frac: float = 1.0, # only for static placement shard_param_frac: float = 1.0, # only for static placement
offload_optim_frac: float = 0.0, # only for static placement offload_optim_frac: float = 0.0, # only for static placement
offload_param_frac: float = 0.0, # only for static placement offload_param_frac: float = 0.0, # only for static placement
@ -132,7 +133,7 @@ class GeminiDDP(ModelWrapper):
steady_cuda_cap_ratio=steady_cuda_cap_ratio, steady_cuda_cap_ratio=steady_cuda_cap_ratio,
) )
self.force_outputs_fp32 = force_outputs_fp32 self.force_outputs_fp32 = force_outputs_fp32
self.param_op_hook = GeminiZeROHook(self.gemini_manager) self.param_op_hook = GeminiZeROHook(self.gemini_manager, max_prefetch=max_prefetch)
self.fp32_params: List[torch.Tensor] = list() self.fp32_params: List[torch.Tensor] = list()
self.fp16_params: List[ColoParameter] = list() self.fp16_params: List[ColoParameter] = list()
self.grads_device: Dict[torch.Tensor, torch.device] = dict() self.grads_device: Dict[torch.Tensor, torch.device] = dict()

View File

@ -1,39 +1,67 @@
from contextlib import contextmanager from contextlib import contextmanager
from enum import Enum from enum import Enum
from functools import partial from functools import partial
from typing import List from typing import Dict, List
import torch import torch
import torch.distributed as dist
from colossalai.logging import DistributedLogger
from colossalai.tensor.param_op_hook import ColoParamOpHook from colossalai.tensor.param_op_hook import ColoParamOpHook
from colossalai.utils import is_ddp_ignored from colossalai.utils import is_ddp_ignored
from colossalai.zero.gemini import TensorState from colossalai.zero.gemini import TensorState
from colossalai.zero.gemini.gemini_mgr import GeminiManager from colossalai.zero.gemini.gemini_mgr import GeminiManager
from .chunk import Chunk
class TrainingPhase(Enum): class TrainingPhase(Enum):
FORWARD = 0 FORWARD = 0
BACKWARD = 1 BACKWARD = 1
logger = DistributedLogger("gemini_hook")
class GeminiZeROHook(ColoParamOpHook): class GeminiZeROHook(ColoParamOpHook):
def __init__(self, gemini_manager: GeminiManager) -> None: def __init__(self, gemini_manager: GeminiManager, max_prefetch: int = 0) -> None:
super().__init__() super().__init__()
self._gemini_manager = gemini_manager self._gemini_manager = gemini_manager
self._chunk_manager = gemini_manager.chunk_manager self._chunk_manager = gemini_manager.chunk_manager
self._training_phase = TrainingPhase.FORWARD self._training_phase = TrainingPhase.FORWARD
self._max_prefetch = max_prefetch
self._async_works: Dict[Chunk, dist.work] = {}
def wait_chunks(self, chunks: List[Chunk]) -> List[Chunk]:
non_prefetched_chunks = []
for chunk in chunks:
if chunk in self._async_works:
print(f"prefetched {chunk.count_id}")
self._async_works[chunk].wait()
del self._async_works[chunk]
else:
non_prefetched_chunks.append(chunk)
return non_prefetched_chunks
def pre_op(self, params): def pre_op(self, params):
params = [p for p in params if not is_ddp_ignored(p)] params = [p for p in params if not is_ddp_ignored(p)]
chunks = self._chunk_manager.get_chunks(params) all_chunks = self._chunk_manager.get_chunks(params)
# wait for prefetched chunks, filter those are not prefetched
chunks_fetch_sync = tuple(self.wait_chunks(all_chunks))
for p in params: for p in params:
self._chunk_manager.trans_tensor_state(p, TensorState.COMPUTE) self._chunk_manager.trans_tensor_state(p, TensorState.COMPUTE)
self._gemini_manager.sample_overall_data() self._gemini_manager.sample_overall_data()
self._gemini_manager.adjust_layout(chunks) self._gemini_manager.adjust_layout(all_chunks, record_anyway=self._max_prefetch > 0)
for chunk in chunks: # fetch the rest chunks synchronously
for chunk in chunks_fetch_sync:
self._chunk_manager.access_chunk(chunk) self._chunk_manager.access_chunk(chunk)
chunks_fetch_async = self._gemini_manager.placement_policy.get_prefetch_chunks(max_prefetch=self._max_prefetch)
for chunk in chunks_fetch_async:
maybe_work = self._chunk_manager.access_chunk(chunk, async_access=True)
if maybe_work is not None:
self._async_works[chunk] = maybe_work
# record cuda model data of the current OP # record cuda model data of the current OP, including memory for prefetched chunks
self._gemini_manager.record_model_data_volume() self._gemini_manager.record_model_data_volume()
def post_op(self, params): def post_op(self, params):
@ -60,6 +88,11 @@ class GeminiZeROHook(ColoParamOpHook):
@contextmanager @contextmanager
def switch_training_phase(self, training_phase: TrainingPhase = TrainingPhase.BACKWARD): def switch_training_phase(self, training_phase: TrainingPhase = TrainingPhase.BACKWARD):
if training_phase == TrainingPhase.FORWARD:
self._cur_param_idx = 0
else:
self._cur_param_idx = len(self._param_visited_order) - 1
old_training_phase = self._training_phase old_training_phase = self._training_phase
try: try:
self._training_phase = training_phase self._training_phase = training_phase

View File

@ -6,7 +6,7 @@ import torch
from .chunk import Chunk, ChunkManager from .chunk import Chunk, ChunkManager
from .memory_tracer import ChunkMemStatsCollector, MemStats from .memory_tracer import ChunkMemStatsCollector, MemStats
from .placement_policy import PlacementPolicyFactory from .placement_policy import PlacementPolicy, PlacementPolicyFactory
class GeminiManager: class GeminiManager:
@ -91,13 +91,13 @@ class GeminiManager:
self._warmup = False self._warmup = False
self.reset_attributes() self.reset_attributes()
def adjust_layout(self, chunks: Tuple[Chunk, ...]) -> None: def adjust_layout(self, chunks: Tuple[Chunk, ...], record_anyway: bool = False) -> None:
"""Adjust the layout of stateful tensors according to the information provided """Adjust the layout of stateful tensors according to the information provided
by mem_stats_collector, which should belongs to a Sharded Model. by mem_stats_collector, which should belongs to a Sharded Model.
""" """
# find stateful tensor in state COMPUTE # find stateful tensor in state COMPUTE
start = time() start = time()
self._record_chunks_order(chunks) self._record_warmup_chunks_order(chunks, record_anyway=record_anyway)
cuda_demand, hold_cuda_tensor_list = self._get_layout_info(self._compute_idx, self._warmup, chunks) cuda_demand, hold_cuda_tensor_list = self._get_layout_info(self._compute_idx, self._warmup, chunks)
self._layout_time += time() - start self._layout_time += time() - start
@ -133,9 +133,9 @@ class GeminiManager:
can_evict_chunks = self._chunk_manager.get_cuda_movable_chunks() can_evict_chunks = self._chunk_manager.get_cuda_movable_chunks()
return cuda_demand, can_evict_chunks return cuda_demand, can_evict_chunks
def _record_chunks_order(self, chunks: Tuple[Chunk, ...]) -> None: def _record_warmup_chunks_order(self, chunks: Tuple[Chunk, ...], record_anyway: bool = False) -> None:
self._compute_idx += 1 self._compute_idx += 1
if self._warmup and self._placement_policy.need_mem_stats: if self._warmup and (self._placement_policy.need_mem_stats or record_anyway):
self._compute_list.append(chunks) self._compute_list.append(chunks)
def sample_overall_data(self): def sample_overall_data(self):
@ -156,6 +156,18 @@ class GeminiManager:
return self._mem_stats_collector.cuda_margin_mem return self._mem_stats_collector.cuda_margin_mem
return None return None
@property
def compute_list(self) -> List[Tuple[Chunk, ...]]:
return self._compute_list
@property
def compute_idx(self) -> int:
return self._compute_idx
@property
def placement_policy(self) -> PlacementPolicy:
return self._placement_policy
@property @property
def is_cuda_margin_mem_avail(self) -> bool: def is_cuda_margin_mem_avail(self) -> bool:
return self._placement_policy.need_mem_stats return self._placement_policy.need_mem_stats

View File

@ -33,6 +33,10 @@ class PlacementPolicy(ABC):
) -> None: ) -> None:
raise NotImplementedError raise NotImplementedError
@abstractmethod
def get_prefetch_chunks(self, max_prefetch: int) -> List[Chunk]:
raise NotImplementedError
class StaticPlacementPolicy(PlacementPolicy): class StaticPlacementPolicy(PlacementPolicy):
def __init__( def __init__(
@ -95,6 +99,18 @@ class StaticPlacementPolicy(PlacementPolicy):
self.keep_gathered_chunk_mem = total_chunk_mem * (1 - self.shard_param_frac) self.keep_gathered_chunk_mem = total_chunk_mem * (1 - self.shard_param_frac)
self.keep_cuda_chunk_mem = total_chunk_mem * (1 - self.offload_param_frac) self.keep_cuda_chunk_mem = total_chunk_mem * (1 - self.offload_param_frac)
def get_prefetch_chunks(self, max_prefetch: int) -> List[Chunk]:
prefetch = []
for i in range(self.chunk_manager.compute_idx + 1, len(self.chunk_manager.compute_list)):
for chunk in self.chunk_manager.compute_list[i]:
if len(prefetch) >= max_prefetch:
break
if chunk not in prefetch:
prefetch.append(chunk)
if len(prefetch) >= max_prefetch:
break
return prefetch
class AutoPlacementPolicy(PlacementPolicy): class AutoPlacementPolicy(PlacementPolicy):
need_mem_stats: bool = True need_mem_stats: bool = True
@ -198,6 +214,9 @@ class AutoPlacementPolicy(PlacementPolicy):
else: else:
grads_device_map[p] = torch.device("cpu") grads_device_map[p] = torch.device("cpu")
def get_prefetch_chunks(self, max_prefetch: int) -> List[Chunk]:
return [] # TODO @botbw: implement prefetching for auto
class PlacementPolicyFactory: class PlacementPolicyFactory:
policies: Dict[str, Type[PlacementPolicy]] = { policies: Dict[str, Type[PlacementPolicy]] = {