Merge pull request #5749 from hpcaitech/prefetch

[Gemini] Prefetch next chunk before each op

colossalai/booster/plugin/gemini_plugin.py

@@ -329,6 +329,7 @@ def __init__(
         chunk_init_device: Optional[torch.device] = None,
         placement_policy: str = "static",
         enable_gradient_accumulation: bool = False,
+        max_prefetch: int = 0,
         shard_param_frac: float = 1.0,  # only for static placement
         offload_optim_frac: float = 0.0,  # only for static placement
         offload_param_frac: float = 0.0,  # only for static placement
@@ -387,6 +388,7 @@ def __init__(
             memstats=memstats,
             mixed_precision=PRECISION_STR_TO_DTYPE[precision],
             master_weights=master_weights,
+            max_prefetch=max_prefetch,
             enable_async_reduce=enable_async_reduce,
         )
         self.zero_optim_config = dict(

colossalai/zero/gemini/chunk/chunk.py

@@ -359,14 +359,15 @@ def shard_move(self, device: torch.device, force_copy: bool = False):
         else:
             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."""
         # sanity check
         assert self.chunk_temp is None
-
+        maybe_work = None
         if not self.is_gathered:
-            self.__gather()
+            maybe_work = self.__gather(async_op=async_access)
         self.__update_tensors_ptr()
+        return maybe_work
 
     def release_chunk(self):
         """Release the usable chunk. It's an operation done in CUDA."""
@@ -512,17 +513,19 @@ def optim_update(self) -> None:
     def get_tensors(self) -> List[torch.Tensor]:
         return list(self.tensors_info.keys())
 
-    def __gather(self):
+    def __gather(self, async_op: bool = False) -> Optional[dist.Work]:
         if not self.is_gathered:
             # sanity check
             assert self.cuda_shard is not None
 
             alloc_storage(self.cuda_global_chunk)
             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.is_gathered = True
+            return work
+        return None
 
     def __scatter(self):
         if self.keep_gathered:

colossalai/zero/gemini/chunk/manager.py

@@ -111,15 +111,16 @@ def close_all_groups(self):
         for group_name in self.chunk_groups:
             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."""
         if chunk in self.accessed_chunks:
-            return
+            return None
         self.__sub_memory_usage(chunk.memory_usage)
         if chunk.device_type == "cpu":
             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)
+        return maybe_work
 
     def release_chunk(self, chunk: Chunk) -> None:
         """Scatter the chunk in CUDA."""
@@ -251,10 +252,11 @@ def __add_memory_usage(self, usage: Dict[str, int]):
         for k, v in usage.items():
             self.total_mem[k] += v
 
-    def __add_accessed_chunk(self, chunk: Chunk):
-        chunk.access_chunk()
+    def __add_accessed_chunk(self, chunk: Chunk, async_access: bool = False) -> Optional[dist.Work]:
+        maybe_work = chunk.access_chunk(async_access=async_access)
         self.accessed_chunks.add(chunk)
         self.accessed_mem += chunk.chunk_mem
+        return maybe_work
 
     def __sub_accessed_chunk(self, chunk: Chunk):
         chunk.release_chunk()

colossalai/zero/gemini/gemini_ddp.py

@@ -78,6 +78,7 @@ def __init__(
         chunk_init_device: torch.device = torch.device("cpu"),
         placement_policy: str = "static",
         enable_gradient_accumulation: bool = False,
+        max_prefetch: int = 0,
         shard_param_frac: float = 1.0,  # only for static placement
         offload_optim_frac: float = 0.0,  # only for static placement
         offload_param_frac: float = 0.0,  # only for static placement
@@ -131,6 +132,7 @@ def __init__(
             offload_param_frac=offload_param_frac,
             warmup_non_model_data_ratio=warmup_non_model_data_ratio,
             steady_cuda_cap_ratio=steady_cuda_cap_ratio,
+            max_prefetch=max_prefetch,
         )
         self.force_outputs_fp32 = force_outputs_fp32
         self.param_op_hook = GeminiZeROHook(self.gemini_manager)

colossalai/zero/gemini/gemini_hook.py

@@ -24,16 +24,42 @@ def __init__(self, gemini_manager: GeminiManager) -> None:
         self._training_phase = TrainingPhase.FORWARD
 
     def pre_op(self, params):
+        # map params to chunks
         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 = self._gemini_manager.wait_chunks(all_chunks)
+
+        # transfer state
         for p in params:
             self._chunk_manager.trans_tensor_state(p, TensorState.COMPUTE)
         self._gemini_manager.sample_overall_data()
-        self._gemini_manager.adjust_layout(chunks)
-        for chunk in chunks:
+
+        # evit chunks, aware of async fetched
+        self._gemini_manager.adjust_layout(
+            all_chunks, record_anyway=self._gemini_manager.placement_policy.max_prefetch > 0
+        )
+
+        # fetch the rest synchronously
+        for chunk in chunks_fetch_sync:
             self._chunk_manager.access_chunk(chunk)
 
-        # record cuda model data of the current OP
+        # get possible chunks to prefetch
+        chunks_fetch_async = self._gemini_manager.placement_policy.get_prefetch_chunks(
+            is_warmup=self._gemini_manager.is_warmup(),
+            compute_list=self._gemini_manager.compute_list,
+            compute_idx=self._gemini_manager.compute_idx,
+            async_works=self._gemini_manager.async_works,
+        )
+
+        # 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._gemini_manager.add_work(chunk, maybe_work)
+
+        # record cuda model data of the current OP, including memory for prefetched chunks
         self._gemini_manager.record_model_data_volume()
 
     def post_op(self, params):

colossalai/zero/gemini/gemini_mgr.py

@@ -1,12 +1,13 @@
 import functools
 from time import time
-from typing import Dict, List, Optional, Tuple
+from typing import Dict, Iterable, List, Optional, Tuple
 
 import torch
+import torch.distributed as dist
 
 from .chunk import Chunk, ChunkManager
 from .memory_tracer import ChunkMemStatsCollector, MemStats
-from .placement_policy import PlacementPolicyFactory
+from .placement_policy import PlacementPolicy, PlacementPolicyFactory
 
 
 class GeminiManager:
@@ -41,9 +42,12 @@ def __init__(
         self._mem_stats_collector = (
             ChunkMemStatsCollector(chunk_manager, self._memstats) if policy_cls.need_mem_stats else None
         )
-        self._placement_policy = policy_cls(chunk_manager, self._mem_stats_collector, **placement_kwargs)
+        self._placement_policy = policy_cls(
+            chunk_manager=chunk_manager, mem_stats_collector=self._mem_stats_collector, **placement_kwargs
+        )
         self._compute_list: List[Tuple[Chunk, ...]] = []
         self._compute_idx: int = -1
+        self._async_works: Dict[Chunk, dist.Work] = {}
 
         self._h2d_volume = 0
         self._d2h_volume = 0
@@ -91,18 +95,20 @@ def post_iter(self):
         self._warmup = False
         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
         by mem_stats_collector, which should belongs to a Sharded Model.
         """
         # find stateful tensor in state COMPUTE
         start = time()
-        self._record_chunks_order(chunks)
-        cuda_demand, hold_cuda_tensor_list = self._get_layout_info(self._compute_idx, self._warmup, chunks)
+        self._record_warmup_chunks_order(chunks, record_anyway=record_anyway)
+        cuda_demand, can_evict_chunks = self._get_layout_info(self._compute_idx, self._warmup, chunks)
+        # don't evict chunks that are asynchronously fetched
+        can_evict_chunks = [chunk for chunk in can_evict_chunks if chunk not in self._async_works]
         self._layout_time += time() - start
 
         vol, evict_time = self._placement_policy.evict_tensors(
-            can_evict_chunks=hold_cuda_tensor_list,
+            can_evict_chunks=can_evict_chunks,
             cuda_demand=cuda_demand,
             warmup=self._warmup,
             compute_list=self._compute_list,
@@ -114,6 +120,21 @@ def adjust_layout(self, chunks: Tuple[Chunk, ...]) -> None:
         # move COMPUTE tensors to CUDA
         self._h2d_volume += cuda_demand
 
+    def wait_chunks(self, chunks: Iterable[Chunk]) -> Tuple[Chunk]:
+        non_prefetched_chunks = []
+        for chunk in chunks:
+            if chunk in self._async_works:
+                self._async_works[chunk].wait()
+                del self._async_works[chunk]
+            else:
+                non_prefetched_chunks.append(chunk)
+        return tuple(non_prefetched_chunks)
+
+    def add_work(self, chunk: Chunk, work: dist.Work):
+        assert work is not None
+        assert chunk not in self._async_works
+        self._async_works[chunk] = work
+
     @functools.lru_cache(maxsize=None)
     def _get_layout_info(self, compute_idx: int, warmup: bool, chunks: Tuple[Chunk, ...]):
         start = time()
@@ -133,9 +154,9 @@ def _get_layout_info(self, compute_idx: int, warmup: bool, chunks: Tuple[Chunk,
         can_evict_chunks = self._chunk_manager.get_cuda_movable_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
-        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)
 
     def sample_overall_data(self):
@@ -156,6 +177,22 @@ def cuda_margin_mem(self) -> Optional[float]:
             return self._mem_stats_collector.cuda_margin_mem
         return None
 
+    @property
+    def placement_policy(self) -> PlacementPolicy:
+        return self._placement_policy
+
+    @property
+    def compute_list(self) -> List[Tuple[Chunk, ...]]:
+        return self._compute_list
+
+    @property
+    def compute_idx(self) -> int:
+        return self._compute_idx
+
+    @property
+    def async_works(self) -> Dict[Chunk, dist.Work]:
+        return self._async_works
+
     @property
     def is_cuda_margin_mem_avail(self) -> bool:
         return self._placement_policy.need_mem_stats