[shardformer] support tp+zero for shardformer (#4472)

* support tp+zero/input type cast for hybridplugin

* add tp+zero tests

* fix bucket arguments

colossalai/booster/plugin/hybrid_parallel_plugin.py

@@ -1,5 +1,6 @@
 import random
 from contextlib import nullcontext
+from functools import partial
 from typing import Any, Callable, Iterator, List, Optional, Tuple, Union
 
 import numpy as np
@@ -10,6 +11,7 @@ from torch.nn import Module, SyncBatchNorm
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import _LRScheduler as LRScheduler
+from torch.utils._pytree import tree_map
 from torch.utils.data import DataLoader
 from torch.utils.data.distributed import DistributedSampler
 
@@ -27,32 +29,49 @@ from .pp_plugin_base import PipelinePluginBase
 DP_AXIS, PP_AXIS, TP_AXIS = 0, 1, 2
 
 
+def _convert_floating_point(x, dtype: torch.dtype = torch.float16):
+    if isinstance(x, torch.Tensor) and torch.is_floating_point(x):
+        return x.to(dtype)
+    return x
+
+
 class HybridParallelModule(ModelWrapper):
 
     def __init__(self, module: Module, precision: str, shard_config: ShardConfig, dp_group: ProcessGroup, use_ddp: bool,
                  ddp_config: dict) -> None:
+
         self.stage_manager = shard_config.pipeline_stage_manager
         self.dp_group = dp_group
+
         shardformer = ShardFormer(shard_config)
         module, self.shared_params = shardformer.optimize(module)
-        # TODO(ver217): add input type cast
+
+        # setting process groups for shared parameters
         self.shared_param_process_groups = []
         for shared_param in self.shared_params:
             if len(shared_param) > 0:
                 self.shared_param_process_groups.append(
                     self.stage_manager.init_process_group_by_stages(list(shared_param.keys())))
+
+        # setting mixed_precision
+        self.mixed_precision = None
         if precision == 'fp16':
-            module = module.half().cuda()
+            self.mixed_precision = torch.float16
         elif precision == 'bf16':
-            module = module.to(dtype=torch.bfloat16).cuda()
-        else:
-            module = module.cuda()    # train without AMP
+            self.mixed_precision = torch.bfloat16
+        if self.mixed_precision is not None:
+            module = module.to(self.mixed_precision)
+        module = module.cuda()
 
+        # setting input type cast when using mixed precision
+        self.convert_fn = None
+        if self.mixed_precision is not None:
+            self.convert_fn = partial(_convert_floating_point, dtype=self.mixed_precision)
+
+        # setting ddp configs
         if use_ddp:
-
             # convert model to sync bn
             module = SyncBatchNorm.convert_sync_batchnorm(module, dp_group)
-
             # wrap the model with PyTorch DDP
             module = DDP(module, process_group=dp_group, **ddp_config)
 
@@ -78,6 +97,12 @@ class HybridParallelModule(ModelWrapper):
                 dist.all_reduce(p.grad, group=self.dp_group)
                 p.grad.div_(self.dp_group.size())
 
+    def forward(self, *args, **kwargs):
+        if self.convert_fn is not None:
+            args = tree_map(self.convert_fn, args)
+            kwargs = tree_map(self.convert_fn, kwargs)
+        return super().forward(*args, **kwargs)
+
     def unwrap(self):
         module = super().unwrap()
         if isinstance(module, DDP):
@@ -180,7 +205,6 @@ class HybridParallelPlugin(PipelinePluginBase):
                                     Defaults to 'fp16'.
         zero_stage (int, optional): The stage of ZeRO for data parallelism. Can only be choosed from [0, 1, 2].
                                         When set to 0, ZeRO will not be used. Defaults to 0.
-        cpu_offload (bool, optional): Whether to open cpu_offload when using ZeRO. Defaults to False.
         enable_all_optimization (bool, optional): Whether to switch on all the optimizations supported by Shardformer.
                                                     Currently all the optimization methods include fused normalization, flash attention and JIT.
                                                     Defaults to False.
@@ -196,12 +220,16 @@ class HybridParallelPlugin(PipelinePluginBase):
         hysteresis (int, optional):  The number of overflows before decreasing loss scale when using AMP. Defaults to 2.
         max_scale (float, optional): The maximum loss scale of AMP. Defaults to 2**32.
         max_norm (float, optional): Maximum norm for gradient clipping. Defaults to 0.
-        broadcast_buffers (bool, optional): Whether to broadcast buffers in the beginning of training. Only for usage of DDP. Defaults to True.
-        bucket_cap_mb (int, optional): The bucket size in MB. Only for usage of DDP. Defaults to 25.
-        find_unused_parameters (bool, optional): Whether to find unused parameters. Only for usage of DDP. Defaults to False.
-        check_reduction (bool, optional): Whether to check reduction. Only for usage of DDP. Defaults to False.
-        gradient_as_bucket_view (bool, optional): Whether to use gradient as bucket view. Only for usage of DDP. Defaults to False.
-        static_graph (bool, optional): Whether to use static graph. Only for usage of DDP. Defaults to False.
+        broadcast_buffers (bool, optional): Whether to broadcast buffers in the beginning of training when using DDP. Defaults to True.
+        ddp_bucket_cap_mb (int, optional): The bucket size in MB when using DDP. Defaults to 25.
+        find_unused_parameters (bool, optional): Whether to find unused parameters when using DDP. Defaults to False.
+        check_reduction (bool, optional): Whether to check reduction when using DDP. Defaults to False.
+        gradient_as_bucket_view (bool, optional): Whether to use gradient as bucket view when using DDP. Defaults to False.
+        static_graph (bool, optional): Whether to use static graph when using DDP. Defaults to False.
+        zero_bucket_size_in_m (int, optional): Gradient reduce bucket size in million elements when using ZeRO. Defaults to 12.
+        cpu_offload (bool, optional): Whether to open cpu_offload when using ZeRO. Defaults to False.
+        communication_dtype (torch.dtype, optional): Communication dtype when using ZeRO. If not specified, the dtype of param will be used. Defaults to None.
+        overlap_communication (bool, optional): Whether to overlap communication and computation when using ZeRO. Defaults to True.
     """
 
     def __init__(self,
@@ -209,7 +237,6 @@ class HybridParallelPlugin(PipelinePluginBase):
                  pp_size: int,
                  precision: str = 'fp16',
                  zero_stage: int = 0,
-                 cpu_offload: bool = False,
                  enable_all_optimization: bool = False,
                  enable_fused_normalization: bool = False,
                  enable_flash_attention: bool = False,
@@ -224,12 +251,16 @@ class HybridParallelPlugin(PipelinePluginBase):
                  hysteresis: int = 2,
                  max_scale: float = 2**32,
                  max_norm: float = 0,
-                 broadcast_buffers=True,
-                 bucket_cap_mb=25,
-                 find_unused_parameters=False,
-                 check_reduction=False,
-                 gradient_as_bucket_view=False,
-                 static_graph=False) -> None:
+                 broadcast_buffers: bool = True,
+                 ddp_bucket_cap_mb: int = 25,
+                 find_unused_parameters: bool = False,
+                 check_reduction: bool = False,
+                 gradient_as_bucket_view: bool = False,
+                 static_graph: bool = False,
+                 zero_bucket_size_in_m: int = 12,
+                 cpu_offload: bool = False,
+                 communication_dtype: Optional[torch.dtype] = None,
+                 overlap_communication: bool = True) -> None:
 
         super().__init__()
         assert dist.get_world_size() % (
@@ -239,8 +270,6 @@ class HybridParallelPlugin(PipelinePluginBase):
         if enable_sequence_parallelism:
             assert tp_size > 1, 'Sequence parallelism must be enabled when using tensor parallelism'
 
-        # TODO(ver217): support zero
-        assert zero_stage == 0, 'zero is not support yet'
         self.tp_size = tp_size
         self.pp_size = pp_size
         self.dp_size = dist.get_world_size() // (tp_size * pp_size)
@@ -282,11 +311,18 @@ class HybridParallelPlugin(PipelinePluginBase):
         )
 
         self.ddp_config = dict(broadcast_buffers=broadcast_buffers,
-                               bucket_cap_mb=bucket_cap_mb,
+                               bucket_cap_mb=ddp_bucket_cap_mb,
                                find_unused_parameters=find_unused_parameters,
                                check_reduction=check_reduction,
                                gradient_as_bucket_view=gradient_as_bucket_view,
                                static_graph=static_graph)
+
+        self.zero_config = dict(reduce_bucket_size=zero_bucket_size_in_m * 1024 * 1024,
+                                communication_dtype=communication_dtype,
+                                overlap_communication=overlap_communication,
+                                cpu_offload=cpu_offload,
+                                partition_grad=(self.zero_stage == 2))
+
         self.max_norm = max_norm
 
     @property
@@ -337,15 +373,16 @@ class HybridParallelPlugin(PipelinePluginBase):
                                                              model,
                                                              use_pipeline=self.enable_pipeline_parallelism)
             else:
+                assert self.dp_size > 1, "Please use Zero when data parallel size is greater than 1."
+                assert self.precision != 'fp32', "Please set precision to 'fp16' or 'bf16' when using ZeRO."
                 optimizer = HybridParallelZeroOptimizer(optimizer,
                                                         model,
                                                         use_pipeline=self.enable_pipeline_parallelism,
-                                                        partition_grad=(self.zero_stage == 2),
-                                                        cpu_offload=self.cpu_offload,
                                                         dp_process_group=self.dp_group,
                                                         tp_process_group=self.tp_group,
                                                         verbose=True,
                                                         clip_grad_norm=self.max_norm,
+                                                        **self.zero_config,
                                                         **self.amp_config)
         return model, optimizer, criterion, dataloader, lr_scheduler
 

tests/test_shardformer/test_model/test_shard_bert.py

@@ -56,9 +56,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
         atol, rtol = 1e-4, 1e-3
     else:
         atol, rtol = 5e-3, 5e-3
-    if stage_manager is None or stage_manager.is_first_stage():
-        #check_weight(bert.embeddings.word_embeddings, sharded_bert.embeddings.word_embeddings, tp_group, atol=1e-5, rtol=1e-3)
-        #check_weight(bert.encoder.layer[0].attention.self.query, sharded_bert.encoder.layer[0].attention.self.query, tp_group, atol=5e-3, rtol=1e-3)
+    if (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
         check_grad(bert, sharded_bert, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False)
         check_grad(bert, sharded_bert, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0, verbose=False)
 
@@ -101,6 +99,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     'enable_all_optimization': True,
     'use_lazy_init': False,
     'precision': 'fp32'
+}, {
+    'tp_size': 2,
+    'pp_size': 1,
+    'enable_all_optimization': True,
+    'use_lazy_init': True,
+    'zero_stage': 2,
+    'precision': 'fp16',
+    'initial_scale': 1
 }])
 def run_bert_test(test_config):
 

tests/test_shardformer/test_model/test_shard_bloom.py

@@ -53,7 +53,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     # check grad
     row_layer_for_check = ['h[0].self_attention.query_key_value', 'word_embeddings']
     col_layer_for_check = ['h[0].self_attention.dense']
-    if stage_manager is None or stage_manager.is_first_stage():
+    if (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
         if test_config['precision'] == 'fp32':
             atol, rtol = 1e-6, 1e-5
         else:
@@ -101,6 +101,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     'enable_all_optimization': True,
     'use_lazy_init': False,
     'precision': 'fp32'
+}, {
+    'tp_size': 2,
+    'pp_size': 1,
+    'enable_all_optimization': True,
+    'use_lazy_init': True,
+    'zero_stage': 2,
+    'precision': 'fp16',
+    'initial_scale': 1
 }])
 def run_bloom_test(test_config):
 

tests/test_shardformer/test_model/test_shard_chatglm.py

@@ -55,7 +55,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     # check grad
     row_layer_for_check = ['encoder.layers[0].self_attention.query_key_value', 'embedding.word_embeddings']
     col_layer_for_check = ['encoder.layers[0].self_attention.dense']
-    if stage_manager is None or stage_manager.is_first_stage():
+    if (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
         if test_config['precision'] == 'fp32':
             atol, rtol = 1e-6, 1e-3
         else:
@@ -125,6 +125,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     'enable_all_optimization': True,
     'use_lazy_init': False,
     'precision': 'fp32'
+}, {
+    'tp_size': 2,
+    'pp_size': 1,
+    'enable_all_optimization': True,
+    'use_lazy_init': True,
+    'zero_stage': 2,
+    'precision': 'fp16',
+    'initial_scale': 1
 }])
 def run_chatglm_test(test_config):
 

tests/test_shardformer/test_model/test_shard_gpt2.py

@@ -56,7 +56,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     row_layer_for_check = ['wte', 'h[0].mlp.c_proj']
 
     # check grad
-    if stage_manager is None or stage_manager.is_first_stage():
+    if (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
         if test_config['precision'] == 'fp32':
             atol, rtol = 1e-4, 1e-3
         else:
@@ -120,6 +120,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     'use_lazy_init': True,
     'enable_sequence_parallelism': True,
     'precision': 'fp32',
+}, {
+    'tp_size': 2,
+    'pp_size': 1,
+    'enable_all_optimization': True,
+    'use_lazy_init': True,
+    'zero_stage': 2,
+    'precision': 'fp16',
+    'initial_scale': 1
 }])
 @clear_cache_before_run()
 def run_gpt2_test(test_config):

tests/test_shardformer/test_model/test_shard_llama.py

@@ -60,7 +60,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     # check grad
     row_layer_for_check = ['layers[0].self_attn.q_proj', 'embed_tokens']
     col_layer_for_check = ['layers[0].self_attn.o_proj']
-    if stage_manager is None or stage_manager.is_first_stage():
+    if (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
         if test_config['precision'] == 'fp32':
             atol, rtol = 1e-6, 1e-4
         else:
@@ -135,6 +135,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     'enable_all_optimization': True,
     'use_lazy_init': False,
     'precision': 'fp32'
+}, {
+    'tp_size': 2,
+    'pp_size': 1,
+    'enable_all_optimization': True,
+    'use_lazy_init': True,
+    'zero_stage': 2,
+    'precision': 'fp16',
+    'initial_scale': 1
 }])
 def run_llama_test(test_config):
 

tests/test_shardformer/test_model/test_shard_opt.py

@@ -58,7 +58,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     # check grad
     row_layer_for_check = ['decoder.layers[0].self_attn.q_proj', 'decoder.embed_tokens']    # 'decoder.embed_tokens'
     col_layer_for_check = ['decoder.layers[0].self_attn.out_proj']
-    if stage_manager is None or stage_manager.is_first_stage():
+    if (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
         if test_config['precision'] == 'fp32':
             atol, rtol = 1e-6, 1e-3
         else:
@@ -127,6 +127,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     'enable_all_optimization': True,
     'use_lazy_init': False,
     'precision': 'fp32'
+}, {
+    'tp_size': 2,
+    'pp_size': 1,
+    'enable_all_optimization': True,
+    'use_lazy_init': True,
+    'zero_stage': 2,
+    'precision': 'fp16',
+    'initial_scale': 1
 }])
 def run_opt_test(test_config):
 

tests/test_shardformer/test_model/test_shard_t5.py

@@ -55,12 +55,12 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
 
     row_layer_for_check = ['shared', 'encoder.block[0].layer[0].SelfAttention.q']
 
-    # check weights and gradients
+    # check grad
     if test_config['precision'] == 'fp32':
         atol, rtol = 1e-5, 1e-3
     else:
         atol, rtol = 5e-3, 5e-3
-    if stage_manager is None or stage_manager.is_first_stage():
+    if (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
         check_grad(t5, sharded_t5, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0)
 
     # check weights after optimizer.step()
@@ -110,6 +110,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     'enable_all_optimization': True,
     'use_lazy_init': False,
     'precision': 'fp32'
+}, {
+    'tp_size': 2,
+    'pp_size': 1,
+    'enable_all_optimization': True,
+    'use_lazy_init': True,
+    'zero_stage': 2,
+    'precision': 'fp16',
+    'initial_scale': 1
 }])
 @clear_cache_before_run()
 def run_t5_test(test_config):

tests/test_shardformer/test_model/test_shard_vit.py

@@ -55,7 +55,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     # check grad
     row_layer_for_check = ['encoder.layer[0].attention.attention.query', 'embeddings.patch_embeddings.projection']
     col_layer_for_check = ['encoder.layer[0].attention.output.dense']
-    if stage_manager is None or stage_manager.is_first_stage():
+    if (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
         if test_config['precision'] == 'fp32':
             atol, rtol = 1e-5, 1e-3
         else:
@@ -124,6 +124,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     'enable_all_optimization': True,
     'use_lazy_init': False,
     'precision': 'fp32'
+}, {
+    'tp_size': 2,
+    'pp_size': 1,
+    'enable_all_optimization': True,
+    'use_lazy_init': False,
+    'zero_stage': 2,
+    'precision': 'fp16',
+    'initial_scale': 1
 }])
 def run_vit_test(test_config):