Merge pull request #5722 from botbw/prefetch

[gemini] prefetch chunks

colossalai/booster/plugin/gemini_plugin.py

@@ -329,6 +329,7 @@ class GeminiPlugin(DPPluginBase):
         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
@@ -386,6 +387,7 @@ class GeminiPlugin(DPPluginBase):
             memstats=memstats,
             mixed_precision=PRECISION_STR_TO_DTYPE[precision],
             master_weights=master_weights,
+            max_prefetch=max_prefetch,
         )
         self.zero_optim_config = dict(
             gpu_margin_mem_ratio=gpu_margin_mem_ratio,

colossalai/zero/gemini/chunk/chunk.py

@@ -357,14 +357,14 @@ class Chunk:
         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
-
         if not self.is_gathered:
-            self.__gather()
+            return self.__gather(async_op=async_access)
         self.__update_tensors_ptr()
+        return None
 
     def release_chunk(self):
         """Release the usable chunk. It's an operation done in CUDA."""
@@ -498,17 +498,19 @@ class Chunk:
     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 @@ class ChunkManager:
         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 @@ class ChunkManager:
         for k, v in usage.items():
             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_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 @@ class GeminiDDP(ModelWrapper):
         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
@@ -130,6 +131,7 @@ class GeminiDDP(ModelWrapper):
             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

@@ -5,6 +5,7 @@ from typing import List
 
 import torch
 
+from colossalai.logging import DistributedLogger
 from colossalai.tensor.param_op_hook import ColoParamOpHook
 from colossalai.utils import is_ddp_ignored
 from colossalai.zero.gemini import TensorState
@@ -16,6 +17,9 @@ class TrainingPhase(Enum):
     BACKWARD = 1
 
 
+logger = DistributedLogger("gemini_hook")
+
+
 class GeminiZeROHook(ColoParamOpHook):
     def __init__(self, gemini_manager: GeminiManager) -> None:
         super().__init__()
@@ -24,16 +28,37 @@ class GeminiZeROHook(ColoParamOpHook):
         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()
+
+        # 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 +42,10 @@ class GeminiManager:
         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(self, chunk_manager, 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 +93,20 @@ class GeminiManager:
         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)
+        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, 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 +118,21 @@ class GeminiManager:
         # 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 +152,9 @@ class GeminiManager:
         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 +175,18 @@ class GeminiManager:
             return self._mem_stats_collector.cuda_margin_mem
         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
     def is_cuda_margin_mem_avail(self) -> bool:
         return self._placement_policy.need_mem_stats

colossalai/zero/gemini/placement_policy.py

@@ -18,10 +18,17 @@ class PlacementPolicy(ABC):
     need_mem_stats: bool = False
 
     def __init__(
-        self, chunk_manager: ChunkManager, mem_stats_collector: Optional[ChunkMemStatsCollector] = None, **kwargs
+        self,
+        gemini_manager: "GeminiManager",
+        chunk_manager: ChunkManager,
+        mem_stats_collector: Optional[ChunkMemStatsCollector] = None,
+        max_prefetch: int = 0,
+        **kwargs,
     ) -> None:
+        self.gemini_manager = gemini_manager
         self.chunk_manager = chunk_manager
         self.mem_stats_collector: Optional[ChunkMemStatsCollector] = mem_stats_collector
+        self.max_prefetch = max_prefetch
 
     @abstractmethod
     def evict_tensors(self, can_evict_chunks: List[Chunk], **kwargs) -> Tuple[int, float]:
@@ -33,18 +40,26 @@ class PlacementPolicy(ABC):
     ) -> None:
         raise NotImplementedError
 
+    @abstractmethod
+    def get_prefetch_chunks(self) -> List[Chunk]:
+        raise NotImplementedError
+
 
 class StaticPlacementPolicy(PlacementPolicy):
     def __init__(
         self,
+        gemini_manager: "GeminiManager",
         chunk_manager: ChunkManager,
         mem_stats_collector: Optional[ChunkMemStatsCollector] = None,
+        max_prefetch: int = 0,
         shard_param_frac: float = 1.0,
         offload_optim_frac: float = 0.0,
         offload_param_frac: float = 0.0,
         **kwargs,
     ) -> None:
-        super().__init__(chunk_manager, mem_stats_collector=mem_stats_collector)
+        super().__init__(
+            gemini_manager, chunk_manager, mem_stats_collector=mem_stats_collector, max_prefetch=max_prefetch
+        )
         if offload_param_frac > 0.0 and (shard_param_frac != 1.0 or offload_optim_frac != 1.0):
             warnings.warn("offload_param_frac is ignored when shard_param_frac != 1.0 or offload_optim_frac != 1.0")
             offload_param_frac = 0.0
@@ -95,19 +110,38 @@ class StaticPlacementPolicy(PlacementPolicy):
         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)
 
+    def get_prefetch_chunks(self) -> List[Chunk]:
+        if self.gemini_manager.is_warmup():  # no prefetch during warmup since we need compute_list
+            return []
+        can_prefetch = self.max_prefetch - len(self.gemini_manager._async_works)
+        prefetch = []
+        for i in range(self.gemini_manager.compute_idx + 1, len(self.gemini_manager.compute_list)):
+            for chunk in self.gemini_manager.compute_list[i]:
+                if len(prefetch) >= can_prefetch:
+                    break
+                if chunk not in prefetch and chunk not in self.chunk_manager.accessed_chunks:
+                    prefetch.append(chunk)
+            if len(prefetch) >= can_prefetch:
+                break
+        return prefetch
+
 
 class AutoPlacementPolicy(PlacementPolicy):
     need_mem_stats: bool = True
 
     def __init__(
         self,
+        gemini_manager: "GeminiManager",
         chunk_manager: ChunkManager,
         mem_stats_collector: Optional[ChunkMemStatsCollector] = None,
+        max_prefetch: int = 0,
         warmup_non_model_data_ratio: float = 0.8,
         steady_cuda_cap_ratio: float = 0.9,
         **kwargs,
     ) -> None:
-        super().__init__(chunk_manager, mem_stats_collector=mem_stats_collector)
+        super().__init__(
+            gemini_manager, chunk_manager, mem_stats_collector=mem_stats_collector, max_prefetch=max_prefetch
+        )
         # model data will use 1-_warmup_non_model_data_ratio CUDA memory in warmup phase
         # you can set them by AutoPlacementPolicy.set_warmup_non_model_data_ratio()
         # and AutoPlacementPolicy.set_steady_cuda_cap_ratio()
@@ -198,6 +232,9 @@ class AutoPlacementPolicy(PlacementPolicy):
             else:
                 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:
     policies: Dict[str, Type[PlacementPolicy]] = {

examples/language/gpt/gemini/commons/utils.py

@@ -30,8 +30,9 @@ def get_profile_context(enable_flag, warmup_steps, active_steps, save_dir):
             activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
             schedule=schedule(wait=0, warmup=warmup_steps, active=active_steps),
             on_trace_ready=tensorboard_trace_handler(save_dir),
-            record_shapes=True,
+            # record_shapes=True,
-            profile_memory=True,
+            # profile_memory=True,
+            with_stack=True,
         )
     else:
         return nullcontext(DummyProfiler())

examples/language/gpt/gemini/train_gpt_demo.py

@@ -129,7 +129,7 @@ def main():
     WARMUP_STEPS = 1
     assert WARMUP_STEPS < NUM_STEPS, "warmup steps should smaller than the total steps"
     assert (NUM_STEPS - WARMUP_STEPS) % 2 == 1, "the number of valid steps should be odd to take the median"
-    PROF_FLAG = False  # The flag of profiling, False by default
+    PROF_FLAG = True  # The flag of profiling, False by default
 
     disable_existing_loggers()
     colossalai.launch_from_torch()
@@ -166,7 +166,7 @@ def main():
                 stage=zero_stage, reduce_bucket_size_in_m=12, overlap_communication=True, verbose=True
             )
         elif args.distplan == "CAI_Gemini":
-            plugin = GeminiPlugin(search_range_m=128, hidden_dim=model.config.n_embd)
+            plugin = GeminiPlugin(search_range_m=128, hidden_dim=model.config.n_embd, max_prefetch=1)
         else:
             raise RuntimeError
 
@@ -248,7 +248,7 @@ def main():
             prof.step()
 
     tflops_list.sort()
-    median_index = ((NUM_STEPS - WARMUP_STEPS) >> 1) + WARMUP_STEPS
+    median_index = min(((NUM_STEPS - WARMUP_STEPS) >> 1) + WARMUP_STEPS, len(tflops_list) - 1)
     logger.info(f"Median TFLOPS is {tflops_list[median_index]:.3f}")
     torch.cuda.synchronize()
 

tests/test_zero/test_gemini/test_optim.py

@@ -40,9 +40,7 @@ EXAMPLE_MODELS = [
 ]
 
 # bfloat16 cannot represent them exactly
-BF16_IGNORED_KEYS = [
+BF16_IGNORED_KEYS = ["masked_bias"]
-    "masked_bias",
-]
 
 
 def check_param(model: GeminiDDP, torch_model: torch.nn.Module, dtype: torch.dtype):