Merge pull request #6077 from duanjunwen/dev/zero_bubble

[Feature] ZeroBubble support MoeHybridplugin;

colossalai/booster/plugin/moe_hybrid_parallel_plugin.py

@@ -295,8 +295,11 @@ class MoeHybridParallelPlugin(HybridParallelPlugin):
         if self.pp_size > 1:
             assert pp_style in ["1f1b", "interleaved", "zbv"], "Unsupported pipeline parallelism style"
             assert (
-                pp_style == "interleaved" or pp_style == "zbv"
+                pp_style in ["interleaved", "zbv"] or num_model_chunks == 1
-            ) or num_model_chunks == 1, "num_model_chunks must be 1 when using 1f1b"
+            ), "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"
@@ -309,6 +312,7 @@ class MoeHybridParallelPlugin(HybridParallelPlugin):
                 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":
@@ -329,7 +333,8 @@ class MoeHybridParallelPlugin(HybridParallelPlugin):
                     enable_metadata_cache=enable_metadata_cache,
                 )
             elif pp_style == "zbv":
-                self.schedule = ZeroBubbleVPipeScheduler(
+                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,

colossalai/shardformer/policies/mixtral.py

@@ -258,14 +258,30 @@ class MixtralPolicy(Policy):
         stage_manager = self.pipeline_stage_manager
 
         held_layers = []
-        layers_per_stage = stage_manager.distribute_layers(len(module.layers))
-        if stage_manager.is_first_stage():
-            held_layers.append(module.embed_tokens)
-        start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage)
-        held_layers.extend(module.layers[start_idx:end_idx])
-        if stage_manager.is_last_stage():
-            held_layers.append(module.norm)
 
+        if stage_manager.is_interleave:
+            assert stage_manager.num_model_chunks is not None
+            layers_per_stage = stage_manager.distribute_layers(len(module.layers))
+            stage_indices = stage_manager.get_stage_index(layers_per_stage)
+            stage_manager.stage_indices = stage_indices
+            if stage_manager.is_first_stage(ignore_chunk=True):
+                held_layers.append(module.embed_tokens)
+            for start_idx, end_idx in stage_indices:
+                held_layers.extend(module.layers[start_idx:end_idx])
+            if (stage_manager.use_zbv and stage_manager.is_first_stage(ignore_chunk=True)) or (
+                not stage_manager.use_zbv and stage_manager.is_last_stage(ignore_chunk=True)
+            ):
+                # for zbv, when is_first_stage (last fwd), we append norm
+                # for interleaved, when is_last_stage (last fwd), we also append norm
+                held_layers.append(module.norm)
+        else:
+            layers_per_stage = stage_manager.distribute_layers(len(module.layers))
+            if stage_manager.is_first_stage():
+                held_layers.append(module.embed_tokens)
+            start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage)
+            held_layers.extend(module.layers[start_idx:end_idx])
+            if stage_manager.is_last_stage():
+                held_layers.append(module.norm)
         return held_layers
 
 

tests/test_pipeline/test_schedule/test_zerobubble_pp.py

@@ -7,17 +7,30 @@ import torch
 import torch.distributed as dist
 import torch.nn as nn
 from torch.testing import assert_close
+from transformers.models.mixtral.configuration_mixtral import MixtralConfig
+from transformers.models.mixtral.modeling_mixtral import MixtralModel
 
 import colossalai
+from colossalai.booster.booster import Booster
+from colossalai.booster.plugin.moe_hybrid_parallel_plugin import MoeHybridParallelPlugin
 from colossalai.cluster import ProcessGroupMesh
 from colossalai.interface import OptimizerWrapper
 from colossalai.logging import disable_existing_loggers
 from colossalai.pipeline.schedule.v_schedule import PipelineGraph, ScheduledNode
 from colossalai.pipeline.schedule.zero_bubble_pp import ZeroBubbleVPipeScheduler
 from colossalai.pipeline.stage_manager import PipelineStageManager
+from colossalai.shardformer.layer.utils import Randomizer
 from colossalai.tensor.d_tensor.api import clear_layout_converter
 from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
-from tests.kit.model_zoo import model_zoo
+from colossalai.testing.random import seed_all
+from tests.test_moe.moe_utils import assert_loose_close
+
+NUM_BATCH = 8
+NUM_TOK_PER_BATCH, NUM_EXPERTS = 4, 4
+NUM_LAYERS = 8
+HIDDEN_SIZE_PER_HEAD = 4
+NUM_HEADS = 4
+TOP_K = 1
 
 
 class MlpModel(nn.Module):
@@ -730,127 +743,176 @@ def run_fwd_bwd_vschedule_with_optim(test_config):
     assert_optim_param_groups(optim_base_param_groups, optim_pp_param_groups)
 
 
-# TODO:4) support Hybrid base 3)
+# TODO:3) support booster & Hybrid base 2)
 def run_with_hybridplugin(test_config):
     pass
 
 
-# TODO:5) support MoEHybrid base 3)
+# TODO:4) support booster & MoEHybrid base 2)
 @parameterize(
-    "test_config",
+    "config",
     [
-        {
+        (0, 1, 4, 1, 1),
-            "pp_style": "zbv",
+        (1, 2, 2, 1, 1),
-            "tp_size": 1,
+        (1, 2, 1, 2, 1),
-            "ep_size": 1,
+        (1, 2, 1, 1, 2),
-            "pp_size": 4,
-            "num_microbatches": 4,
-            "zero_stage": 1,
-            "precision": "bf16",
-            "num_model_chunks": 2,
-        },
     ],
 )
-def run_with_moehybridplugin(test_config):
+def run_with_booster_moehybridplugin(config: Tuple[int, ...]):
-    sub_model_zoo = model_zoo.get_sub_registry("transformers_bert")
+    test_config = config
-    # test_config["use_lazy_init"] = False
+    stage, ep_size, pp_size, tp_size, sp_size = config
-    test_config["initial_scale"] = 2**16
+    num_microbatches = pp_size
-    model_list = [
+    dist.get_world_size()
-        "transformers_bert",
+    rank = dist.get_rank()
-    ]
+    dtype, precision = torch.float16, "fp16"
-    clear_layout_converter()
+    torch.cuda.set_device(dist.get_rank())
 
-    for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():
+    ########
-        if name in model_list:
+    # init base model
-            # base param
+    ########
-            model = model_fn()
+    assert pp_size <= NUM_LAYERS, "pp_size should be less than or equal to NUM_LAYERS"
-            data = data_gen_fn()
+    config = MixtralConfig(
-            print(f"data {data}")
+        hidden_size=HIDDEN_SIZE_PER_HEAD * NUM_HEADS,
-            criterion = loss_fn
+        intermediate_size=HIDDEN_SIZE_PER_HEAD * NUM_HEADS * 2,
-            optimizer = torch.optim.SGD(model.parameters(), momentum=0.1, lr=1e-5)
+        num_hidden_layers=NUM_LAYERS,
+        num_attention_heads=NUM_HEADS,
+        num_key_value_heads=NUM_HEADS,
+        num_local_experts=NUM_EXPERTS,
+        num_experts_per_tok=TOP_K,
+        attn_implementation="flash_attention_2",
+    )
 
-            output = model(**data)
+    # init model with the same seed
-            loss = criterion(output)
+    seed_all(10086)
-            loss.backward()
-            optimizer.step()
-            print(f"output {output}")
 
-            # # pp param
+    torch_model = MixtralModel(config).to(dtype).cuda()
-            # model_pp = deepcopy(model)
+    torch_optimizer = torch.optim.SGD(torch_model.parameters(), lr=1)
-            # data_pp = deepcopy(data)
+    # init schedule
-            # optimizer_pp = OptimizerWrapper(torch.optim.SGD(model_pp.parameters(), momentum=0.1, lr=1e-5))
+    h, a, s = config.hidden_size, config.num_attention_heads, 1024
+    mem_f = 34 * h + 5 * a * s
+    mem_w = -32 * h
+    mem_b = -mem_w - mem_f
+    graph = PipelineGraph(
+        n_stage=pp_size,
+        n_micro=num_microbatches,
+        f_cost=1,
+        b_cost=1,
+        w_cost=1,
+        c_cost=1,
+        f_mem=mem_f,
+        b_mem=mem_b,
+        w_mem=mem_w,
+        # max_mem=mem_f * (p * 2 + m_offset),
+    )
 
-            # # init pipeline graph
+    zbv_schedule = graph.get_v_schedule()
-            # h, a, s = model.config.hidden_size, model.config.num_attention_heads, 1024
-            # mem_f = 34 * h + 5 * a * s
-            # mem_w = -32 * h
-            # mem_b = -mem_w - mem_f
-            # graph = PipelineGraph(
-            #     n_stage=test_config["pp_size"],
-            #     n_micro=test_config["num_microbatches"],
-            #     f_cost=1,
-            #     b_cost=1,
-            #     w_cost=1,
-            #     c_cost=1,
-            #     f_mem=mem_f,
-            #     b_mem=mem_b,
-            #     w_mem=mem_w,
-            #     # max_mem=mem_f * (p * 2 + m_offset),
-            # )
 
-            # zbv_schedule = graph.get_v_schedule()
+    # init MoeHybridPlugin
+    plugin = MoeHybridParallelPlugin(
+        pp_size=pp_size,
+        num_microbatches=pp_size,
+        tp_size=tp_size,
+        sp_size=sp_size,
+        ep_size=ep_size,
+        zero_stage=stage,
+        enable_sequence_parallelism=sp_size > 1,
+        sequence_parallelism_mode="all_to_all" if sp_size > 1 else None,
+        overlap_communication=False,
+        initial_scale=1,
+        precision=precision,
+        find_unused_parameters=True,
+        pp_style="zbv",
+        scheduler_nodes=zbv_schedule,
+        num_model_chunks=2,
+    )
 
-            # test_config["scheduler_nodes"] = zbv_schedule
+    dp_size = plugin.dp_size
-            # plugin = MoeHybridParallelPlugin(
-            #     **test_config
-            # )
-            # model_pp, optimizer_pp, criterion, data_pp, _ = plugin.configure(
-            #     model = model_pp,
-            #     optimizer = optimizer_pp,
-            #     criterion = criterion,
-            #     dataloader = data_pp,
-            # )
 
-            # output_pp = plugin.execute_pipeline(
+    booster = Booster(plugin=plugin)
-            #     data_iter=iter(data),
-            #     model=model,
-            #     criterion=criterion,
-            #     optimizer=optimizer,
-            #     return_loss = True,
-            #     return_outputs = True,
-            # )
 
+    ########
+    # init pp model
+    ########
 
-# TODO:6) support booster & Hybrid base 4)
+    parallel_model = deepcopy(torch_model)
+    parallel_optimizer = torch.optim.SGD(parallel_model.parameters(), lr=1)
+    parallel_model, parallel_optimizer, _, _, _ = booster.boost(parallel_model, parallel_optimizer)
+    # create different input along dp axis
+    seed_all(1453 + rank)
 
+    torch_model.train()
+    parallel_model.train()
+    for _ in range(2):
+        # gen random input
+        input_embeddings = torch.rand(
+            NUM_BATCH, NUM_TOK_PER_BATCH, HIDDEN_SIZE_PER_HEAD * NUM_HEADS, requires_grad=True
+        ).cuda()
+        dist.all_reduce(
+            input_embeddings, group=plugin.pp_group
+        )  # pp inputs except the first stage doesn't matter, but need to be replicate for torch model check
 
-# TODO:7) support booster & MoEHybrid base 4)
+        dist.all_reduce(input_embeddings, group=plugin.tp_group)  # tp group duplicate input
-@parameterize(
+        dist.all_reduce(input_embeddings, group=plugin.sp_group)  # sp group duplicate input
-    "test_config",
+
-    [
+        # run the model with hybrid parallel
-        {
+        if booster.plugin.stage_manager is not None:
-            "pp_style": "zbv",
+            # for test with pp
-            "tp_size": 1,
+            data_iter = iter([{"inputs_embeds": input_embeddings}])
-            "ep_size": 1,
+            sharded_output = booster.execute_pipeline(
-            "pp_size": 4,
+                data_iter,
-            "num_microbatches": 4,
+                parallel_model,
-            "zero_stage": 1,
+                lambda x, y: x.last_hidden_state.mean(),
-            "precision": "bf16",
+                parallel_optimizer,
-            "num_model_chunks": 2,
+                return_loss=True,
-        },
+                return_outputs=True,
-    ],
+            )
-)
+            # stage 0 chunk 0
-def run_with_booster_moehybridplugin(test_config):
+            parallel_output = None
-    pass
+            if (
+                booster.plugin.stage_manager.is_first_stage(ignore_chunk=True)
+                and rank == dist.get_process_group_ranks(plugin.pp_group)[0]
+            ):
+                parallel_output = sharded_output["loss"]
+            else:
+                parallel_output = torch.tensor(12345.0, device="cuda")
+            # broadcast along pp axis
+            dist.broadcast(parallel_output, src=dist.get_process_group_ranks(plugin.pp_group)[0], group=plugin.pp_group)
+
+        else:
+            # for test without pp
+            parallel_output = parallel_model(inputs_embeds=input_embeddings.to(dtype)).last_hidden_state.mean()
+            parallel_optimizer.backward(parallel_output)
+        parallel_optimizer.step()
+        parallel_optimizer.zero_grad()
+        dist.all_reduce(parallel_output, group=plugin.dp_group)
+
+        # ===================================================================================
+        # run normal model with all dp(different) inputs
+        all_inputs = [torch.empty_like(input_embeddings) for _ in range(dp_size)]
+        dist.all_gather(all_inputs, input_embeddings, group=plugin.dp_group)
+        torch_output_sum = 0
+        for input_data_ in all_inputs:
+            torch_output = torch_model(inputs_embeds=input_data_.to(dtype)).last_hidden_state.mean()
+            torch_output.backward()
+            torch_output_sum += torch_output.detach()
+        # avg dp grads follows zero optimizer
+        for p in torch_model.parameters():
+            if p.grad is not None:
+                p.grad /= dp_size
+        torch_optimizer.step()
+        torch_optimizer.zero_grad()
+        assert_loose_close(parallel_output, torch_output_sum, dtype=dtype)
+        print(f"rank {dist.get_rank()} config {test_config}  test passed")
+        clear_layout_converter()
+        Randomizer.reset_index()
+        torch.cuda.empty_cache()
 
 
 def run_dist(rank, world_size, port):
     disable_existing_loggers()
     colossalai.launch(rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
-    # run_fwd_bwd_iter_input()
+    # run_fwd_bwd_vschedule_with_optim()
-    run_fwd_bwd_vschedule_with_optim()
+    run_with_booster_moehybridplugin()
-    # run_with_moehybridplugin()
-    # run_with_booster_moehybridplugin()
 
 
 @pytest.mark.dist