[Zerobubble] merge main.

colossalai/amp/naive_amp/mixed_precision_mixin/base.py

@@ -43,7 +43,7 @@ def zero_grad(self):
     dtype: torch.dtype
 
     @abstractmethod
-    def pre_backward(self, loss: Tensor) -> Tensor:
+    def pre_backward(self, loss: Tensor, *args, **kwargs) -> Tensor:
         """Called before backward.
 
         Args:

colossalai/amp/naive_amp/mixed_precision_optimizer.py

@@ -86,13 +86,18 @@ def __init__(
             group["params"] = master_params
         self._current_grad_norm: Optional[float] = None
 
-    def backward(self, loss: Tensor, *args, **kwargs):
+    def backward(self, loss: Tensor, inputs=None, retain_graph=False, **kwargs):
         loss = self.mixed_precision.pre_backward(loss)
-        loss.backward(*args, **kwargs)
+        loss.backward(inputs=inputs, retain_graph=retain_graph, **kwargs)
 
-    def backward_by_grad(self, tensor: Tensor, grad: Tensor):
+    def backward_by_grad(self, tensor: Tensor, grad: Tensor, inputs: Tensor = None, retain_graph: bool = False):
         grad = self.mixed_precision.pre_backward_by_grad(tensor, grad)
-        tensor.backward(grad)
+        torch.autograd.backward(
+            tensors=tensor,
+            grad_tensors=grad,
+            inputs=inputs,
+            retain_graph=retain_graph,
+        )
 
     def zero_grad(self, *args, **kwargs):
         for p in self.working_to_master_map.keys():

colossalai/booster/mixed_precision/fp16_torch.py

@@ -46,9 +46,9 @@ def __init__(
             growth_interval=growth_interval,
         )
 
-    def backward(self, loss: Tensor, *args, **kwargs) -> None:
+    def backward(self, loss: Tensor, inputs=None, retain_graph=False, **kwargs) -> None:
         scaled_loss = self.scale_loss(loss)
-        scaled_loss.backward(*args, **kwargs)
+        scaled_loss.backward(inputs=inputs, retain_graph=retain_graph, **kwargs)
 
     def step(self, *args, **kwargs) -> Optional[float]:
         out = self.scaler.step(self.optim, *args, **kwargs)

colossalai/booster/plugin/hybrid_parallel_plugin.py

@@ -28,7 +28,7 @@
 from colossalai.interface.optimizer import DistributedOptim
 from colossalai.logging import get_dist_logger
 from colossalai.nn.optimizer import DistGaloreAwamW, cast_to_distributed
-from colossalai.pipeline.schedule import InterleavedSchedule, OneForwardOneBackwardSchedule
+from colossalai.pipeline.schedule import InterleavedSchedule, OneForwardOneBackwardSchedule, ZeroBubbleVPipeScheduler
 from colossalai.pipeline.stage_manager import PipelineStageManager
 from colossalai.quantization import BnbQuantizationConfig, quantize_model
 from colossalai.quantization.fp8_hook import FP8Hook
@@ -296,7 +296,7 @@ def __init__(
         self._current_grad_norm: Optional[float] = None
         super().__init__(optim)
 
-    def backward(self, loss: Tensor, *args, **kwargs):
+    def backward(self, loss: Tensor, inputs=None, retain_graph=False, **kwargs):
         r"""
         Backpropagate gradients through the model and optionally synchronize sequence parallelism gradients.
 
@@ -315,7 +315,7 @@ def backward(self, loss: Tensor, *args, **kwargs):
 
         # Call the superclass backward method to compute gradients.
         with self.model._hook_context():
-            super().backward(loss, *args, **kwargs)
+            super().backward(loss, inputs=inputs, retain_graph=retain_graph, **kwargs)
 
         if self.model.require_grad_sync:
             # If gradient synchronization is required, sync sequence parallelism gradients.
@@ -324,7 +324,7 @@ def backward(self, loss: Tensor, *args, **kwargs):
             # If gradient synchronization is is not required, return.
             return
 
-    def backward_by_grad(self, tensor: Tensor, grad: Tensor):
+    def backward_by_grad(self, tensor: Tensor, grad: Tensor, inputs: Tensor = None, retain_graph: bool = False):
         """
         Backpropagate gradients through the model using a precomputed gradient and optionally synchronize sequence parallelism gradients.
 
@@ -341,7 +341,7 @@ def backward_by_grad(self, tensor: Tensor, grad: Tensor):
         """
 
         # Call the superclass backward method to compute gradients.
-        super().backward_by_grad(tensor, grad)
+        super().backward_by_grad(tensor, grad, inputs=inputs, retain_graph=retain_graph)
 
         if self.model.require_grad_sync:
             # If gradient synchronization is required, sync sequence parallelism gradients.
@@ -525,7 +525,7 @@ def __init__(
             max_norm=max_norm,
         )
 
-    def backward(self, loss: Tensor, *args, **kwargs):
+    def backward(self, loss: Tensor, inputs=None, retain_graph=False, **kwargs):
         r"""
         Backpropagate gradients through the model and optionally synchronize sequence parallelism gradients.
 
@@ -543,7 +543,7 @@ def backward(self, loss: Tensor, *args, **kwargs):
         """
         # Call the superclass backward method to compute gradients.
         with self.model._hook_context():
-            super().backward(loss, *args, **kwargs)
+            super().backward(loss, inputs=inputs, retain_graph=retain_graph, **kwargs)
 
         if self.model.require_grad_sync:
             # If gradient synchronization is required, sync sequence parallelism gradients.
@@ -552,7 +552,7 @@ def backward(self, loss: Tensor, *args, **kwargs):
             # If gradient synchronization is is not required, return.
             return
 
-    def backward_by_grad(self, tensor: Tensor, grad: Tensor):
+    def backward_by_grad(self, tensor: Tensor, grad: Tensor, inputs: Tensor = None, retain_graph: bool = False):
         """
         Backpropagate gradients through the model using a precomputed gradient and optionally synchronize sequence parallelism gradients.
 
@@ -568,7 +568,7 @@ def backward_by_grad(self, tensor: Tensor, grad: Tensor):
             None
         """
         # Call the superclass backward method to compute gradients.
-        super().backward_by_grad(tensor, grad)
+        super().backward_by_grad(tensor, grad, inputs=inputs, retain_graph=retain_graph)
 
         if self.model.require_grad_sync:
             # If gradient synchronization is required, sync sequence parallelism gradients.
@@ -785,7 +785,7 @@ def _get_grads_to_sync(all_working_grads) -> Union[List[Tensor], None]:
         else:
             return
 
-    def backward(self, loss, retain_graph=False):
+    def backward(self, loss, inputs=None, retain_graph=False):
         """
         Backpropagate gradients through the model and optionally synchronize sequence parallelism gradients.
 
@@ -801,7 +801,7 @@ def backward(self, loss, retain_graph=False):
             None
         """
         # Call the superclass backward method to compute gradients.
-        super().backward(loss, retain_graph)
+        super().backward(loss, inputs=inputs, retain_graph=retain_graph)
 
         if self.require_grad_sync and self.model.shard_config.enable_sequence_parallelism:
             # If gradient synchronization is required, sync sequence parallelism gradients.
@@ -810,7 +810,7 @@ def backward(self, loss, retain_graph=False):
             # If gradient synchronization is is not required, return.
             return
 
-    def backward_by_grad(self, tensor, grad):
+    def backward_by_grad(self, tensor, grad, inputs: Tensor = None, retain_graph: bool = False):
         """
         Backpropagate gradients through the model using a precomputed gradient and optionally synchronize sequence parallelism gradients.
 
@@ -826,7 +826,7 @@ def backward_by_grad(self, tensor, grad):
             None
         """
         # Call the superclass backward_by_grad method to compute gradients.
-        super().backward_by_grad(tensor, grad)
+        super().backward_by_grad(tensor, grad, inputs=inputs, retain_graph=retain_graph)
 
         if self.require_grad_sync and self.model.shard_config.enable_sequence_parallelism:
             # If gradient synchronization is required, sync sequence parallelism gradients.
@@ -1030,6 +1030,7 @@ def __init__(
         custom_policy: Policy = None,
         pp_style: str = "1f1b",
         num_model_chunks: int = 1,
+        scheduler_nodes: List = None,
         num_layers_per_stage: Optional[List[int]] = None,
         gradient_checkpoint_config: Optional[GradientCheckpointConfig] = None,
         enable_metadata_cache: bool = True,
@@ -1048,6 +1049,9 @@ def __init__(
             dist.get_world_size() % (tp_size * pp_size) == 0
         ), f"World size {dist.get_world_size()} is not divisible by tp_size {tp_size} * pp_size {pp_size}"
 
+        assert (
+            not pp_style == "zbv" or scheduler_nodes is not None
+        ), f"scheduler_nodes must not be None when using zero bubble pipeline."
         if enable_sequence_parallelism:
             self.sequence_parallelism_mode = (
                 sequence_parallelism_mode if sequence_parallelism_mode is not None else "all_to_all"
@@ -1109,29 +1113,39 @@ def __init__(
                 self.pg_mesh = ProcessGroupMesh(self.pp_size, self.dp_size, self.tp_size, self.sp_size)
 
         self.stage_manager = None
-        self.schedule = None
+        self.scheduler = None
         self.custom_policy = custom_policy
         assert zero_stage in (0, 1, 2)
         if self.pp_size > 1:
-            assert pp_style in ["1f1b", "interleaved"], "Unsupported pipeline parallelism style"
-            assert pp_style == "interleaved" or num_model_chunks == 1, "num_model_chunks must be 1 when using 1f1b"
+            assert pp_style in ["1f1b", "interleaved", "zbv"], "Unsupported pipeline parallelism style"
+            assert (
+                pp_style in ["interleaved", "zbv"] or num_model_chunks == 1
+            ), "num_model_chunks must be 1 when using 1f1b"
+            assert (
+                pp_style in ["1f1b", "interleaved"] or num_model_chunks == 2
+            ), "num_model_chunks must be 2 when using zero bubble pipeline"
             assert (
                 num_microbatches is not None or microbatch_size is not None
             ), "num_microbatches or microbatch_size must be specified when using pipeline parallelism"
             assert (
                 self.zero_stage <= 1
             ), "To avoid prohibitive gradient synchronization costs, zero stage must be 0 or 1 when using pipeline parallelism"
+            if pp_style == "zbv":
+                self.logger.warning(
+                    """the enable_gradient_checkpointing function must set the use_reentrant to False, such as  model.gradient_checkpointing_enable(gradient_checkpointing_kwargs={'use_reentrant':False})"""
+                )
             self.stage_manager = PipelineStageManager(
                 self.pg_mesh,
                 pipeline_axis=self.pp_axis,
-                enable_interleave=pp_style == "interleaved",
+                enable_interleave=(pp_style == "interleaved" or pp_style == "zbv"),
+                use_zbv=(pp_style == "zbv"),
                 num_model_chunks=num_model_chunks,
                 num_layers_per_stage=num_layers_per_stage,
             )
 
             if pp_style == "interleaved":
                 assert num_model_chunks > 1, "number of model chunks must be > 1 when using interleaved"
-                self.schedule = InterleavedSchedule(
+                self.scheduler = InterleavedSchedule(
                     stage_manager=self.stage_manager,
                     num_model_chunks=num_model_chunks,
                     num_microbatch=num_microbatches,
@@ -1141,13 +1155,21 @@ def __init__(
                     fp8_communication=fp8_communication,
                 )
             elif pp_style == "1f1b":
-                self.schedule = OneForwardOneBackwardSchedule(
+                self.scheduler = OneForwardOneBackwardSchedule(
                     stage_manager=self.stage_manager,
                     num_microbatches=num_microbatches,
                     microbatch_size=microbatch_size,
                     enable_metadata_cache=enable_metadata_cache,
                     fp8_communication=fp8_communication,
                 )
+            elif pp_style == "zbv":
+                self.scheduler = ZeroBubbleVPipeScheduler(
+                    stage_manager=self.stage_manager,
+                    schedule=scheduler_nodes,
+                    num_model_chunks=num_model_chunks,
+                    num_microbatch=num_microbatches,
+                    microbatch_size=microbatch_size,
+                )
             else:
                 raise NotImplementedError()
         if sequence_parallelism_mode == "ring_attn":
@@ -1263,7 +1285,6 @@ def configure(
 
         # Replace with distributed implementation if exists
         optimizer = cast_to_distributed(optimizer)
-
         if isinstance(optimizer, DistGaloreAwamW) and zero_stage > 0 and self.dp_size > 0:
             self.logger.warning(
                 "Galore is only supported for Tensor Parallel and vanilla Data Parallel yet. Disabling ZeRO.",
@@ -1278,6 +1299,7 @@ def configure(
                 self.dp_size == 1 and self.pp_size == 1
             )
             # sync gradients across DP * SP ranks
+            # sync gradients across DP * SP ranks
             # Apply Hybrid ZeRO across DP * SP ranks
             if self.enable_sequence_parallelism and not is_share_sp_tp(self.sequence_parallelism_mode):
                 dp_group = self.pg_mesh.create_group_along_axis([self.dp_axis, self.sp_axis])
@@ -1380,7 +1402,7 @@ def execute_pipeline(
         ctx = optimizer.no_sync() if isinstance(optimizer, HybridParallelZeroOptimizer) else model.no_sync()
 
         with ctx, model._hook_context():
-            outputs = self.schedule.forward_backward_step(
+            outputs = self.scheduler.forward_backward_step(
                 model, data_iter, criterion, optimizer, return_loss, return_outputs
             )
 

colossalai/booster/plugin/moe_hybrid_parallel_plugin.py

@@ -29,6 +29,7 @@
 from colossalai.nn.optimizer import cast_to_distributed
 from colossalai.pipeline.schedule.interleaved_pp import InterleavedSchedule
 from colossalai.pipeline.schedule.one_f_one_b import OneForwardOneBackwardSchedule
+from colossalai.pipeline.schedule.zero_bubble_pp import ZeroBubbleVPipeScheduler
 from colossalai.pipeline.stage_manager import PipelineStageManager
 from colossalai.shardformer.policies.base_policy import Policy
 from colossalai.shardformer.shard.grad_ckpt_config import GradientCheckpointConfig
@@ -212,6 +213,7 @@ def __init__(
         custom_policy: Policy = None,
         pp_style: str = "1f1b",
         num_model_chunks: int = 1,
+        scheduler_nodes: List = None,
         num_layers_per_stage: Optional[List[int]] = None,
         gradient_checkpoint_config: Optional[GradientCheckpointConfig] = None,
         enable_metadata_cache: bool = True,
@@ -285,12 +287,17 @@ def __init__(
             self.pg_mesh = ProcessGroupMesh(self.pp_size, self.moe_dp_size, self.ep_size, self.tp_size, self.sp_size)
 
         self.stage_manager = None
-        self.schedule = None
+        self.scheduler = None
         self.custom_policy = custom_policy
         assert zero_stage in (0, 1, 2)
         if self.pp_size > 1:
-            assert pp_style in ["1f1b", "interleaved"], "Unsupported pipeline parallelism style"
-            assert pp_style == "interleaved" or num_model_chunks == 1, "num_model_chunks must be 1 when using 1f1b"
+            assert pp_style in ["1f1b", "interleaved", "zbv"], "Unsupported pipeline parallelism style"
+            assert (
+                pp_style in ["interleaved", "zbv"] or num_model_chunks == 1
+            ), "num_model_chunks must be 1 when using 1f1b"
+            assert (
+                pp_style in ["1f1b", "interleaved"] or num_model_chunks == 2
+            ), "num_model_chunks must be 2 when using zero bubble pipeline"
             assert (
                 num_microbatches is not None or microbatch_size is not None
             ), "num_microbatches or microbatch_size must be specified when using pipeline parallelism"
@@ -300,14 +307,15 @@ def __init__(
             self.stage_manager = PipelineStageManager(
                 self.pg_mesh,
                 pipeline_axis=self.pp_axis,
-                enable_interleave=pp_style == "interleaved",
+                enable_interleave=(pp_style == "interleaved" or pp_style == "zbv"),
                 num_model_chunks=num_model_chunks,
                 num_layers_per_stage=num_layers_per_stage,
+                use_zbv=(pp_style == "zbv"),
             )
 
             if pp_style == "interleaved":
                 assert num_model_chunks > 1, "number of model chunks must be > 1 when using interleaved"
-                self.schedule = InterleavedSchedule(
+                self.scheduler = InterleavedSchedule(
                     stage_manager=self.stage_manager,
                     num_model_chunks=num_model_chunks,
                     num_microbatch=num_microbatches,
@@ -316,12 +324,21 @@ def __init__(
                     overlap_p2p=overlap_p2p,
                 )
             elif pp_style == "1f1b":
-                self.schedule = OneForwardOneBackwardSchedule(
+                self.scheduler = OneForwardOneBackwardSchedule(
                     stage_manager=self.stage_manager,
                     num_microbatches=num_microbatches,
                     microbatch_size=microbatch_size,
                     enable_metadata_cache=enable_metadata_cache,
                 )
+            elif pp_style == "zbv":
+                assert num_model_chunks > 1, "number of model chunks must be > 1 when using ZerbubbleV"
+                self.scheduler = ZeroBubbleVPipeScheduler(
+                    schedule=scheduler_nodes,
+                    stage_manager=self.stage_manager,
+                    num_model_chunks=num_model_chunks,
+                    num_microbatch=num_microbatches,
+                    overlap_p2p=overlap_p2p,
+                )
             else:
                 raise NotImplementedError()
 

colossalai/interface/optimizer.py

@@ -49,14 +49,31 @@ def zero_grad(self, *args, **kwargs):
         """
         self.optim.zero_grad(*args, **kwargs)
 
-    def backward(self, loss: Tensor, *args, **kwargs):
+    def backward(self, loss: Tensor, inputs=None, retain_graph=False, **kwargs):
         """
         Performs a backward pass on the loss.
         """
-        loss.backward(*args, **kwargs)
+        loss.backward(inputs=inputs, retain_graph=retain_graph, **kwargs)
 
-    def backward_by_grad(self, tensor: Tensor, grad: Tensor):
-        torch.autograd.backward(tensor, grad)
+    def backward_by_grad(self, tensor: Tensor, grad: Tensor, inputs: Tensor = None, retain_graph: bool = False):
+        """
+        Performs a backward pass for dx or dw,
+        for dx, we only calculate dx = w*dy here
+        for dw, we only calculate dw = x*dy here
+
+        Args:
+            tensor (Tensor): y or loss of current chunk;
+            grad_tensors (Tensor): dy of current chunk;
+            input_obj (Tensor): for dx, input_obj is x of current chunk;
+                                for dw, input_obj is w of current chunk;
+            retain_graph (bool): default to be True, we retain graph in backward_b
+        """
+        torch.autograd.backward(
+            tensors=tensor,
+            grad_tensors=grad,
+            inputs=inputs,
+            retain_graph=retain_graph,
+        )
 
     def state_dict(self):
         """

colossalai/pipeline/__init__.py

@@ -1,11 +1,12 @@
 from .p2p import PipelineP2PCommunication
-from .schedule import InterleavedSchedule, OneForwardOneBackwardSchedule, PipelineSchedule
+from .schedule import InterleavedSchedule, OneForwardOneBackwardSchedule, PipelineSchedule, ZeroBubbleVPipeScheduler
 from .stage_manager import PipelineStageManager
 
 __all__ = [
     "PipelineSchedule",
     "OneForwardOneBackwardSchedule",
     "InterleavedSchedule",
+    "ZeroBubbleVPipeScheduler",
     "PipelineP2PCommunication",
     "PipelineStageManager",
 ]