[fp8] hotfix backward hook (#6053)

* [fp8] hotfix backward hook

* [fp8] hotfix pipeline loss accumulation

colossalai/booster/plugin/hybrid_parallel_plugin.py

@@ -216,7 +216,7 @@ class HybridParallelModule(ModelWrapper, AMPModelMixin):
         if self.convert_fn is not None:
             args = tree_map(self.convert_fn, args)
             kwargs = tree_map(self.convert_fn, kwargs)
-        with self._wait_all_gather():
+        with self._hook_context():
             return super().forward(*args, **kwargs)
 
     def unwrap(self):
@@ -229,12 +229,8 @@ class HybridParallelModule(ModelWrapper, AMPModelMixin):
         for p in self.module.parameters():
             wait_all_gather_handle(p)
 
-    def _wait_all_gather(self):
-        return (
-            ColoParamOpHookManager.use_hooks(*self.op_hooks)
-            if (self.overlap_allgather or self.use_fp8)
-            else nullcontext()
-        )
+    def _hook_context(self):
+        return ColoParamOpHookManager.use_hooks(*self.op_hooks) if len(self.op_hooks) > 0 else nullcontext()
 
 
 def get_param_info(optim: Optimizer):
@@ -317,7 +313,8 @@ class HybridParallelNaiveOptimizer(OptimizerWrapper):
         """
 
         # Call the superclass backward method to compute gradients.
-        super().backward(loss, *args, **kwargs)
+        with self.model._hook_context():
+            super().backward(loss, *args, **kwargs)
 
         if self.model.require_grad_sync:
             # If gradient synchronization is required, sync sequence parallelism gradients.
@@ -540,7 +537,8 @@ class HybridParallelAMPOptimizer(MixedPrecisionOptimizer):
             None
         """
         # Call the superclass backward method to compute gradients.
-        super().backward(loss, *args, **kwargs)
+        with self.model._hook_context():
+            super().backward(loss, *args, **kwargs)
 
         if self.model.require_grad_sync:
             # If gradient synchronization is required, sync sequence parallelism gradients.
@@ -683,6 +681,7 @@ class HybridParallelZeroOptimizer(LowLevelZeroOptimizer):
         pp_process_group: Optional[ProcessGroup] = None,  # if using pp
         forced_dtype: Optional[torch.dtype] = None,
         overlap_allgather: bool = False,
+        fp8_communication: bool = False,
     ):
         self.model = model
         self.param_info = param_info
@@ -712,6 +711,8 @@ class HybridParallelZeroOptimizer(LowLevelZeroOptimizer):
             dp_process_group=dp_process_group,
             forced_dtype=forced_dtype,
             overlap_allgather=overlap_allgather,
+            fp8_communication=fp8_communication,
+            backward_context=model._hook_context,
         )
 
     def sync_dp_grads(self):
@@ -1206,6 +1207,7 @@ class HybridParallelPlugin(PipelinePluginBase):
             partition_grad=(self.zero_stage == 2),
             forced_dtype=PRECISION_TORCH_TYPE[precision],
             overlap_allgather=overlap_allgather,
+            fp8_communication=fp8_communication,
         )
 
         self.max_norm = max_norm
@@ -1371,7 +1373,7 @@ class HybridParallelPlugin(PipelinePluginBase):
         # so we disable it, performing manual reduction instead.
         ctx = optimizer.no_sync() if isinstance(optimizer, HybridParallelZeroOptimizer) else model.no_sync()
 
-        with ctx, model._wait_all_gather():
+        with ctx, model._hook_context():
             outputs = self.schedule.forward_backward_step(
                 model, data_iter, criterion, optimizer, return_loss, return_outputs
             )

colossalai/booster/plugin/low_level_zero_plugin.py

@@ -100,14 +100,16 @@ class LowLevelZeroModel(ModelWrapper, AMPModelMixin):
         if self.convert_fn is not None:
             args = tree_map(self.convert_fn, args)
             kwargs = tree_map(self.convert_fn, kwargs)
-        ctx = ColoParamOpHookManager.use_hooks(*self.op_hooks) if self.overlap_allgather else nullcontext()
-        with ctx:
+        with self._hook_context():
             return super().forward(*args, **kwargs)
 
     def _force_wait_all_gather(self):
         for p in self.module.parameters():
             wait_all_gather_handle(p)
 
+    def _hook_context(self):
+        return ColoParamOpHookManager.use_hooks(*self.op_hooks) if len(self.op_hooks) > 0 else nullcontext()
+
 
 class LowLevelZeroCheckpointIO(TorchDDPCheckpointIO):
     def save_unsharded_optimizer(self, optimizer: OptimizerWrapper, checkpoint: str, gather_dtensor: bool = False):
@@ -520,7 +522,7 @@ class LowLevelZeroPlugin(DPPluginBase):
 
         if optimizer is not None and not isinstance(optimizer, OptimizerWrapper):
             optimizer: LowLevelZeroOptimizer = LowLevelZeroOptimizer(
-                optimizer, **zero_optim_kwargs, verbose=self.verbose
+                optimizer, **zero_optim_kwargs, verbose=self.verbose, backward_context=model._hook_context
             )
             # inject update_master_params
             model.update_master_params = MethodType(optimizer.update_master_params, model)

colossalai/initialize.py

@@ -9,6 +9,7 @@ import os
 # https://forums.developer.nvidia.com/t/how-many-streams-maximum-number-of-streams/6571/16
 os.environ["CUDA_DEVICE_MAX_CONNECTIONS"] = "1"
 
+import torch
 import torch.distributed as dist
 
 from colossalai.accelerator import get_accelerator
@@ -64,6 +65,11 @@ def launch(
 
     set_seed(seed)
 
+    try:
+        torch._dynamo.config.optimize_ddp = world_size > 1
+    except AttributeError:
+        pass
+
     if verbose:
         logger = get_dist_logger()
         logger.info(f"Distributed environment is initialized, world size: {dist.get_world_size()}", ranks=[0])

colossalai/pipeline/schedule/interleaved_pp.py

@@ -318,7 +318,7 @@ class InterleavedSchedule(PipelineSchedule):
             if self.stage_manager.is_last_stage():
                 loss = criterion(output_obj, micro_batch) / self.num_microbatch
                 if accum_loss is not None:
-                    accum_loss.add_(loss.detach())
+                    accum_loss.add_(loss.data)
                 if outputs is not None:
                     outputs.append(tree_map(detach, output_obj))
                 return loss

colossalai/pipeline/schedule/one_f_one_b.py

@@ -273,7 +273,7 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
             loss = criterion(output_obj, micro_batch) / self.num_microbatches
 
             if accum_loss is not None:
-                accum_loss.add_(loss.detach())
+                accum_loss.add_(loss.data)
             if outputs is not None:
                 outputs.append(tree_map_hf(detach, output_obj))
             return loss

colossalai/zero/low_level/low_level_optim.py

@@ -1,6 +1,6 @@
 # this code is inspired by the DeepSpeed library and implemented with our own design from scratch
 import copy
-from contextlib import contextmanager
+from contextlib import contextmanager, nullcontext
 from functools import partial
 from typing import Dict, Iterator, List, Optional, Tuple
 from weakref import proxy
@@ -88,6 +88,7 @@ class LowLevelZeroOptimizer(OptimizerWrapper):
         master_weights: bool = True,  # master weights
         overlap_allgather: bool = False,
         fp8_communication: bool = False,
+        backward_context=None,
     ):
         super(LowLevelZeroOptimizer, self).__init__(optim=optimizer)
 
@@ -130,6 +131,7 @@ class LowLevelZeroOptimizer(OptimizerWrapper):
         self._reduce_bucket_size = reduce_bucket_size
         self._communication_dtype = communication_dtype
         self._fp8_communication = fp8_communication
+        self._backward_context = backward_context
 
         # gradient clipping
         self._clip_grad_norm = clip_grad_norm
@@ -429,7 +431,9 @@ class LowLevelZeroOptimizer(OptimizerWrapper):
         if self.mixed_precision_mixin is not None:
             loss = self.mixed_precision_mixin.pre_backward(loss)
 
-        loss.backward(retain_graph=retain_graph)
+        ctx = nullcontext() if self._backward_context is None else self._backward_context()
+        with ctx:
+            loss.backward(retain_graph=retain_graph)
 
         if not self.require_grad_sync:
             return