[Auto Parallel] Add zero h1 pipeline scheduling for paddle (#62865)

AndSonder · web-flow · commit adf8689c97cb · 2024-04-18T20:12:44.000+08:00
* reconstruct_pipeline_scheduler_pass

* add pipeline_scheduler_pass into __all__

* update __init__.py

* recover __init__.py

* extract split matmul_grad_op to pass_utils

* fix

* add paddle.distributed.passes.pipeline_scheduler_pass' to setup.py

* add paddle.distributed.passes.pipeline_scheduler_pass' to setup.py.in

* apply suggestions from code review

* update

* fix

* change func name

* update

* update

* add zero bubble pipeline

* fix bug

* fix

* update

* fix error micro step id

* add zero bubble unittest

* update comment

* add zb to __init__.py

* fix

* fix

* fix codestyle

* add enable_send_recv_overlap

* fix
diff --git a/python/paddle/distributed/passes/__init__.py b/python/paddle/distributed/passes/__init__.py
@@ -102,6 +102,7 @@
     Pipeline1F1BPass,
     PipelineEager1F1BPass,
     PipelineVirtualPipelinePass,
+    PipelineZeroBubblePipelinePass,
     apply_pass,
 )
 from .ps_trainer_pass import (  # noqa: F401
diff --git a/python/paddle/distributed/passes/pass_utils.py b/python/paddle/distributed/passes/pass_utils.py
@@ -276,10 +276,10 @@ def set_skip_gc_vars(num_micro_batches, job_types, sub_programs, jobs):
             f"Skip gc vars for {job_type}-({micro_batch_id}): {skip_gc_vars}"
         )
 
-        if job_type == "backward":
+        if job_type in ["backward", "backward_w"]:
             assert (
                 len(skip_gc_vars) == 0
-            ), f"When enabling pipeline parallelism strategy, the skip_gc_vars for backward subprogram must be empty, but it is {skip_gc_vars}."
+            ), f"When enabling pipeline parallelism strategy, the skip_gc_vars for {job_type} subprogram must be empty, but it is {skip_gc_vars}."
 
         job.set_skip_gc_vars(skip_gc_vars)
         suffixed_required_vars[micro_batch_id] |= required_vars
@@ -778,6 +778,143 @@ def _split_ops(block):
     return list(type_to_program.keys()), list(type_to_program.values())
 
 
+def _get_backward_op_type(block, op):
+    # For the op doesn't have output such as 'send_v2', it should be backward_b.
+    if len(op.output_arg_names) == 0:
+        return "backward_b"
+    for name in op.output_arg_names:
+        name = name.split("@")[0]
+        if not block._find_var_recursive(name):
+            return "backward_b"
+        var = block._find_var_recursive(name)
+        if not var.is_parameter:
+            return "backward_b"
+
+    return "backward_w"
+
+
+def _program_for_zero_bubble(program, enable_send_recv_overlap=False):
+    if enable_send_recv_overlap:
+        _overlap_send_recv(program)
+    else:
+        _insert_sync_for_fthenb_1f1b(program)
+
+    oprole_type = {
+        0: "forward",
+        1: "backward",
+        2: "backward_b",
+        3: 'backward_w',
+        4: "optimizer",
+    }
+
+    def _split_ops(block):
+        # split the program based on the op_role
+        type_to_ops = OrderedDict()
+        for type in oprole_type.values():
+            type_to_ops[type] = []
+        type_to_ops["fetch"] = []
+
+        for op in block.ops:
+            if _is_fetch_op(op):
+                type_to_ops["fetch"].append(op)
+            elif is_forward_op(op):
+                type_to_ops["forward"].append(op)
+            elif is_backward_op(op):
+                type = _get_backward_op_type(block, op)
+                type_to_ops[type].append(op)
+                type_to_ops["backward"].append(op)
+            elif is_optimize_op(op):
+                type_to_ops["optimizer"].append(op)
+            else:
+                raise ValueError(
+                    "The op role: "
+                    + str(op.attr('op_role'))
+                    + " isn't one of Forward, Backward or Optimizer."
+                )
+        return type_to_ops
+
+    type_to_program = OrderedDict()
+    for type in oprole_type.values():
+        type_to_program[type] = Program()
+
+    for idx, src_block in enumerate(program.blocks):
+        type_to_ops = _split_ops(src_block)
+        fwd_ops, bwd_ops, bwd_b_ops, bwd_w_ops, opt_ops, fetch_ops = (
+            type_to_ops["forward"],
+            type_to_ops["backward"],
+            type_to_ops["backward_b"],
+            type_to_ops["backward_w"],
+            type_to_ops["optimizer"],
+            type_to_ops["fetch"],
+        )
+        if idx == 0:
+            fwd_block = type_to_program["forward"].block(0)
+            _add_ops_into_block(src_block, fwd_block, fwd_ops)
+
+            bwd_block = type_to_program["backward"].block(0)
+            _add_ops_into_block(src_block, bwd_block, bwd_ops)
+
+            bwd_block_b = type_to_program["backward_b"].block(0)
+            _add_ops_into_block(src_block, bwd_block_b, bwd_b_ops)
+
+            bwd_block_w = type_to_program["backward_w"].block(0)
+            _add_ops_into_block(src_block, bwd_block_w, bwd_w_ops)
+
+            opt_block = type_to_program["optimizer"].block(0)
+            _add_ops_into_block(src_block, opt_block, opt_ops)
+        else:
+            if len(fwd_ops):
+                fwd_block = type_to_program["forward"]._create_block(
+                    parent_idx=src_block.parent_idx
+                )
+                fwd_block._set_forward_block_idx(src_block.forward_block_idx)
+                _add_ops_into_block(src_block, fwd_block, fwd_ops)
+
+            if len(bwd_ops):
+                bwd_block = type_to_program["backward"]._create_block(
+                    parent_idx=src_block.parent_idx
+                )
+                bwd_block._set_forward_block_idx(src_block.forward_block_idx)
+                _add_ops_into_block(src_block, bwd_block, bwd_ops)
+
+            if len(bwd_b_ops):
+                bwd_block_b = type_to_program["backward_b"]._create_block(
+                    parent_idx=src_block.parent_idx
+                )
+                bwd_block_b._set_forward_block_idx(src_block.forward_block_idx)
+                _add_ops_into_block(src_block, bwd_block_b, bwd_b_ops)
+
+            if len(bwd_w_ops):
+                bwd_block_w = type_to_program["backward_w"]._create_block(
+                    parent_idx=src_block.parent_idx
+                )
+                bwd_block_w._set_forward_block_idx(src_block.forward_block_idx)
+                _add_ops_into_block(src_block, bwd_block_w, bwd_w_ops)
+
+            if len(opt_ops):
+                opt_block = type_to_program["optimizer"]._create_block(
+                    parent_idx=src_block.parent_idx
+                )
+                opt_block._set_forward_block_idx(src_block.forward_block_idx)
+                _add_ops_into_block(src_block, opt_block, opt_ops)
+
+        for fetch_op in fetch_ops:
+            in_name = fetch_op.input_arg_names[0]
+            dst_block = None
+            for block in [fwd_block, bwd_block_b, bwd_block_w, opt_block]:
+                if block._find_var_recursive(in_name):
+                    dst_block = block
+                    break
+            if dst_block:
+                _create_program(src_block, dst_block, fetch_op)
+
+    for prog in type_to_program.values():
+        prog._sync_with_cpp()
+        prog._roll_to_global_block()
+
+    return list(type_to_program.keys()), list(type_to_program.values())
+
+
 def _add_event_dependency(recorder_op, waiter_op):
     '''
     Add the extra event dependency of the two operators.
diff --git a/python/paddle/distributed/passes/pipeline_scheduler_pass/__init__.py b/python/paddle/distributed/passes/pipeline_scheduler_pass/__init__.py
@@ -19,6 +19,7 @@
 from .pipeline_eager_1f1b import PipelineEager1F1BPass  # noqa: F401
 from .pipeline_fthenb import PipelineFThenBPass  # noqa: F401
 from .pipeline_vpp import PipelineVirtualPipelinePass  # noqa: F401
+from .pipeline_zero_bubble import PipelineZeroBubblePipelinePass  # noqa: F401
 
 __all__ = []
 
@@ -29,7 +30,8 @@ def apply_pass(main_program, startup_program, pass_name, pass_attr={}):
         "1F1B",
         "Eager1F1B",
         "VPP",
-    ], f"pipeline scheduler only support FThenB, 1F1B, Eager1F1B and VPP, but receive {pass_name}"
+        "ZBH1",
+    ], f"pipeline scheduler only support FThenB, 1F1B, Eager1F1B, VPP and ZBH1, but receive {pass_name}"
 
     if pass_name == "1F1B":
         # TODO(Ruibiao): Move FLAGS_1f1b_backward_forward_overlap and
diff --git a/python/paddle/distributed/passes/pipeline_scheduler_pass/pipeline_zero_bubble.py b/python/paddle/distributed/passes/pipeline_scheduler_pass/pipeline_zero_bubble.py
@@ -0,0 +1,135 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+
+from paddle.base import core
+
+from ...utils.log_utils import get_logger
+from ..pass_base import register_pass
+from ..pass_utils import _program_for_zero_bubble, split_matmul_grad_to_matmul
+from .pipeline_pass_base import PipelinePassBase
+
+FORWARD = "forward"
+BACKWARD = "backward"
+OPT = "optimizer"
+
+logger = get_logger(logging.INFO)
+
+
+@register_pass("pipeline_scheduler_ZBH1")
+class PipelineZeroBubblePipelinePass(PipelinePassBase):
+    def __init__(self):
+        super().__init__()
+        self.set_attr("enable_optimizer_post_validation", 0)
+
+    def _create_job_list(self):
+        num_micro_batches = self.get_attr("num_micro_batches")
+        pp_stage = self.get_attr("pp_stage")
+        pp_degree = self.get_attr("pp_degree")
+
+        job_list = []
+        assert (
+            pp_degree <= num_micro_batches
+        ), "Num of micro batches should larger than or equal to pp degree."
+
+        micro_batch_in_warmup = pp_degree - pp_stage
+        micro_batch_in_zero_bubble = num_micro_batches - pp_degree
+
+        forward_micro_batch_id = 0
+        for _ in range(micro_batch_in_warmup):
+            forward_job = core.Job(FORWARD)
+            forward_job.set_micro_batch_id(forward_micro_batch_id)
+            job_list.append(forward_job)
+            forward_micro_batch_id += 1
+
+        backward_micro_batch_id = 0
+        for _ in range(pp_stage):
+            backward_b_job = core.Job(BACKWARD + '_b')
+            backward_b_job.set_micro_batch_id(backward_micro_batch_id)
+            job_list.append(backward_b_job)
+            backward_micro_batch_id += 1
+
+            forward_job = core.Job(FORWARD)
+            forward_job.set_micro_batch_id(forward_micro_batch_id)
+            job_list.append(forward_job)
+            forward_micro_batch_id += 1
+
+        for _ in range(micro_batch_in_zero_bubble):
+            backward_job = core.Job(BACKWARD)
+            backward_job.set_micro_batch_id(backward_micro_batch_id)
+            job_list.append(backward_job)
+
+            forward_job = core.Job(FORWARD)
+            forward_job.set_micro_batch_id(forward_micro_batch_id)
+            job_list.append(forward_job)
+
+            forward_micro_batch_id += 1
+            backward_micro_batch_id += 1
+
+        for _ in range(micro_batch_in_warmup - 1):
+            backward_job = core.Job(BACKWARD)
+            backward_job.set_micro_batch_id(backward_micro_batch_id)
+            job_list.append(backward_job)
+            backward_micro_batch_id += 1
+
+        if pp_stage > 0:
+            backward_b_job = core.Job(BACKWARD + '_b')
+            backward_b_job.set_micro_batch_id(backward_micro_batch_id)
+            job_list.append(backward_b_job)
+
+            backward_w_job = core.Job(BACKWARD + '_w')
+            backward_w_job.set_micro_batch_id(backward_micro_batch_id)
+            job_list.append(backward_w_job)
+        else:
+            backward_job = core.Job(BACKWARD)
+            backward_job.set_micro_batch_id(backward_micro_batch_id)
+            job_list.append(backward_job)
+        backward_micro_batch_id += 1
+
+        for i in range(pp_stage):
+            backward_w_job = core.Job(BACKWARD + '_w')
+            backward_w_job.set_micro_batch_id(i)
+            job_list.append(backward_w_job)
+
+        opt_job = core.Job(OPT)
+        opt_job.set_micro_batch_id(0)
+        job_list.append(opt_job)
+        return job_list
+
+    def _split_matmul_grad_ops_to_matmul(self, program, dist_context):
+        for block in program.blocks:
+            matmul_grad_op_idx = []
+            ops = block.ops
+            for i, op_i in enumerate(ops):
+                if (
+                    op_i.type == "matmul_v2_grad"
+                    and not op_i.attr("trans_x")
+                    and not op_i.attr("trans_y")
+                ):
+                    matmul_grad_op_idx.append(i)
+
+            for matmul_grad_id in reversed(matmul_grad_op_idx):
+                split_matmul_grad_to_matmul(
+                    block, matmul_grad_id, dist_context=dist_context
+                )
+
+    def _partial_programs(self, program):
+        dist_context = self.get_attr("dist_context")
+        self._split_matmul_grad_ops_to_matmul(program, dist_context)
+        enable_send_recv_overlap = self.get_attr("enable_send_recv_overlap")
+        types, sub_program_list = _program_for_zero_bubble(
+            program, enable_send_recv_overlap
+        )
+        return types, sub_program_list
diff --git a/test/auto_parallel/CMakeLists.txt b/test/auto_parallel/CMakeLists.txt
@@ -87,6 +87,9 @@ if(WITH_DISTRIBUTE AND WITH_GPU)
                   test_pipeline_scheduler_vpp)
   set_tests_properties(test_pipeline_scheduler_vpp
                        PROPERTIES LABELS "RUN_TYPE=EXCLUSIVE" TIMEOUT 50)
+  py_test_modules(test_pipeline_scheduler_zb MODULES test_pipeline_scheduler_zb)
+  set_tests_properties(test_pipeline_scheduler_zb
+                       PROPERTIES LABELS "RUN_TYPE=EXCLUSIVE" TIMEOUT 50)
   py_test_modules(test_auto_tuner_compare MODULES test_auto_tuner_compare)
   set_tests_properties(test_auto_tuner_compare
                        PROPERTIES LABELS "RUN_TYPE=EXCLUSIVE" TIMEOUT 100)
diff --git a/test/auto_parallel/pipeline_scheduler_zb_unittest.py b/test/auto_parallel/pipeline_scheduler_zb_unittest.py
diff --git a/test/auto_parallel/test_pipeline_scheduler_zb.py b/test/auto_parallel/test_pipeline_scheduler_zb.py

Original file line number	Diff line number	Diff line change
`@@ -102,6 +102,7 @@`
`102`	`102`	`Pipeline1F1BPass,`
`103`	`103`	`PipelineEager1F1BPass,`
`104`	`104`	`PipelineVirtualPipelinePass,`
	`105`	`+ PipelineZeroBubblePipelinePass,`
`105`	`106`	`apply_pass,`
`106`	`107`	`)`
`107`	`108`	`from .ps_trainer_pass import ( # noqa: F401`