[Feature] Zigzag Ring attention

.pre-commit-config.yaml

@@ -12,6 +12,7 @@ repos:
     hooks:
       - id: isort
         name: sort all imports (python)
+        args: ["--profile", "black"] # avoid comflict with black
 
   - repo: https://github.com/psf/black-pre-commit-mirror
     rev: 24.4.2

colossalai/booster/plugin/hybrid_parallel_plugin.py

@@ -32,7 +32,7 @@
 from colossalai.pipeline.stage_manager import PipelineStageManager
 from colossalai.quantization import BnbQuantizationConfig, quantize_model
 from colossalai.shardformer import GradientCheckpointConfig, ShardConfig, ShardFormer
-from colossalai.shardformer.layer.utils import SeqParallelUtils
+from colossalai.shardformer.layer.utils import SeqParallelUtils, is_share_sp_tp
 from colossalai.shardformer.policies.base_policy import Policy
 from colossalai.tensor.colo_parameter import ColoParameter
 from colossalai.tensor.d_tensor.api import is_distributed_tensor
@@ -42,7 +42,7 @@
 
 from .pp_plugin_base import PipelinePluginBase
 
-SUPPORT_SP_MODE = ["split_gather", "ring", "all_to_all"]
+SUPPORT_SP_MODE = ["split_gather", "ring", "all_to_all", "ring_attn"]
 
 PRECISION_TORCH_TYPE = {"fp16": torch.float16, "fp32": torch.float32, "bf16": torch.bfloat16}
 
@@ -72,7 +72,7 @@ def __init__(
         self.dp_group = dp_group
         self.tp_group = tp_group
         self.sp_group = sp_group
-        self.use_dpp = use_ddp
+        self.use_ddp = use_ddp
         self.require_grad_sync = True
         self.overlap_allgather = overlap_allgather
 
@@ -139,8 +139,8 @@ def no_sync(self):
         # Disable automatic gradient synchronization.
         self.require_grad_sync = False
         try:
-            if self.use_dpp:
-                # If using data parallel processing (use_dpp), disable synchronization too.
+            if self.use_ddp:
+                # If using data parallel processing (use_ddp), disable synchronization too.
                 with self.module.no_sync():
                     yield
             else:
@@ -188,7 +188,7 @@ def sync_sp_grads(self, grads: Optional[List[torch.Tensor]] = None):
         """
 
         if self.shard_config.enable_sequence_parallelism:
-            if self.shard_config.sequence_parallelism_mode == "all_to_all":
+            if self.shard_config.sequence_parallelism_mode in ["all_to_all", "ring_attn"]:
                 return
 
             if self.shard_config.sequence_parallelism_mode in ["split_gather", "ring"]:
@@ -1041,9 +1041,11 @@ def __init__(
                     )
                 self.sp_size = 1
                 self.dp_size = dist.get_world_size() // (tp_size * pp_size)
-            elif self.sequence_parallelism_mode in ["all_to_all"]:
+            elif self.sequence_parallelism_mode in ["all_to_all", "ring_attn"]:
                 self.sp_size = 1 if sp_size is None else sp_size
                 self.dp_size = dist.get_world_size() // (self.sp_size * pp_size * tp_size)
+                if self.sequence_parallelism_mode == "ring_attn":
+                    enable_flash_attention = True
         else:
             self.dp_size = dist.get_world_size() // (tp_size * pp_size)
             assert (
@@ -1116,6 +1118,14 @@ def __init__(
             self.sp_group = self.pg_mesh.get_group_along_axis(self.tp_axis)
         else:
             self.sp_group = self.pg_mesh.get_group_along_axis(self.sp_axis)
+        # According to https://github.com/InternLM/InternEvo/blob/a53a4ff4fc45761f80d7fe8e9188bc2e02d487fc/internlm/model/ops/ring_flash_attn/zigzag_ring_flash_attn_with_sliding_window.py#L405
+        # and https://zhuanlan.zhihu.com/p/706805407
+        # using a different proc group may put p2p comm on a new
+        # NCCL stream :)
+        dkv_group = None
+        if sequence_parallelism_mode == "ring_attn":
+            sp_ranks = dist.get_process_group_ranks(self.sp_group)
+            dkv_group = dist.new_group(ranks=sp_ranks)
 
         self.shard_config = ShardConfig(
             tensor_parallel_process_group=self.tp_group,
@@ -1132,6 +1142,12 @@ def __init__(
             parallel_output=parallel_output,
             make_vocab_size_divisible_by=make_vocab_size_divisible_by,
             gradient_checkpoint_config=gradient_checkpoint_config,
+            sp_stream=(
+                torch.cuda.Stream()
+                if enable_sequence_parallelism and sequence_parallelism_mode == "ring_attn"
+                else None
+            ),
+            dkv_group=dkv_group,
         )
         self.amp_config = dict(
             initial_scale=initial_scale,
@@ -1216,14 +1232,13 @@ def configure(
             zero_stage = 0
 
         if not isinstance(model, ModelWrapper):
+            # Shouldn't use pp (frequent grad accumulation) with torch ddp
             use_ddp = (self.dp_size > 1 and self.pp_size == 1 and self.zero_stage == 0) or (
-                self.dp_size == 1
-                and self.pp_size == 1
-                and self.enable_sequence_parallelism
-                and self.sequence_parallelism_mode == "all_to_all"
+                self.dp_size == 1 and self.pp_size == 1
             )
-            # sync gradients across DP * SP ranks
-            if self.enable_sequence_parallelism and self.sequence_parallelism_mode == "all_to_all":
+
+            # 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])
             else:
                 dp_group = self.dp_group

colossalai/checkpoint_io/hybrid_parallel_checkpoint_io.py

@@ -203,7 +203,6 @@ def save_sharded_model(
             return
 
         Path(checkpoint).mkdir(parents=True, exist_ok=True)
-
         # Devices along the same dp_group share the same copies of model.
         # So only let the device with dp_rank == 0 save the model.
         if self.dp_rank != 0:
@@ -643,14 +642,12 @@ def save_unsharded_model(self, model: ModelWrapper, checkpoint: str, gather_dten
         assert isinstance(model, ModelWrapper), "Please boost the model before saving!"
         model._force_wait_all_gather()
         model = model.unwrap()
-
         if self.dp_rank != 0:
             return
 
         # The logic of collecting parameter shards along tp degree
         # has been implemented by _save_to_state_dict method of ParallelModule in Shardformer.
         state_dict = model.state_dict()
-
         if self.pp_size == 1:
             # When pipeline is not used, let master rank directly save the collected state_dict.
             if self.tp_rank == 0:
@@ -660,7 +657,6 @@ def save_unsharded_model(self, model: ModelWrapper, checkpoint: str, gather_dten
             state_dict_list = [None for _ in range(self.pp_size)]
             dist.barrier(self.pp_group)
             dist.all_gather_object(state_dict_list, state_dict, self.pp_group)
-
             # Only the master rank do the saving.
             if self.coordinator.is_master():
                 complete_state_dict = dict()

colossalai/lazy/pretrained.py

@@ -62,7 +62,6 @@ def new_from_pretrained(
     config = kwargs.pop("config", None)
     cache_dir = kwargs.pop("cache_dir", None)
     force_download = kwargs.pop("force_download", False)
-    resume_download = kwargs.pop("resume_download", False)
     proxies = kwargs.pop("proxies", None)
     local_files_only = kwargs.pop("local_files_only", False)
     use_auth_token = kwargs.pop("use_auth_token", None)
@@ -116,7 +115,6 @@ def new_from_pretrained(
             cache_dir=cache_dir,
             return_unused_kwargs=True,
             force_download=force_download,
-            resume_download=resume_download,
             proxies=proxies,
             local_files_only=local_files_only,
             use_auth_token=use_auth_token,
@@ -195,7 +193,6 @@ def new_from_pretrained(
                     "cache_dir": cache_dir,
                     "force_download": force_download,
                     "proxies": proxies,
-                    "resume_download": resume_download,
                     "local_files_only": local_files_only,
                     "use_auth_token": use_auth_token,
                     "user_agent": user_agent,
@@ -312,7 +309,6 @@ def new_from_pretrained(
                 pretrained_model_name_or_path,
                 cache_dir=cache_dir,
                 force_download=force_download,
-                resume_download=resume_download,
                 proxies=proxies,
                 local_files_only=local_files_only,
                 use_auth_token=use_auth_token,

colossalai/legacy/moe/openmoe/model/openmoe_policy.py

@@ -171,7 +171,7 @@ def module_policy(self):
         policy = super().module_policy()
 
         if self.shard_config.enable_tensor_parallelism:
-            # add a new item for casual lm
+            # add a new item for causal lm
             # TODO: recursively assign ep group foe all modules
             new_item = {
                 OpenMoeForCausalLM: ModulePolicyDescription(

colossalai/legacy/nn/layer/parallel_1d/_operation.py

@@ -81,6 +81,9 @@ def backward(ctx, grad_output):
             handle = dist.all_reduce(grad_input, group=gpc.get_group(ctx.parallel_mode), async_op=True)
             # Delay the start of weight gradient computation shortly (3us) to have
             # all-reduce scheduled first and have GPU resources allocated
+            # TODO: This seems to only work if you add torch.cuda.Event.wait()
+
+            # _ = torch.zeros(1, device=grad_output.device)
 
         grad_weight = grad_output.t().matmul(total_input)
         grad_bias = grad_output.sum(dim=0) if use_bias else None

colossalai/pipeline/schedule/interleaved_pp.py

@@ -286,7 +286,6 @@ def forward_step(
         # for the first stage, input_obj is None
         # for other stages, input_obj is the output of the previous stage containing hidden_states etc.
         # Only attention_mask from micro_batch is used
-
         with self.stage_manager.switch_model_chunk_id(model_chunk_id):
             if isinstance(model_chunk, ModuleList):
                 output_obj = model_forward(model_chunk[model_chunk_id], micro_batch, input_obj)

colossalai/pipeline/schedule/one_f_one_b.py

@@ -244,6 +244,7 @@ def forward_step(
         output_obj = model_forward(model, micro_batch, input_obj)
         if self.stage_manager.is_last_stage():
             loss = criterion(output_obj, micro_batch) / self.num_microbatches
+
             if accum_loss is not None:
                 accum_loss.add_(loss.detach())
             if outputs is not None:

colossalai/shardformer/layer/__init__.py

@@ -1,5 +1,5 @@
 from ._operation import all_to_all_comm
-from .attn import AttnMaskType, ColoAttention
+from .attn import AttnMaskType, ColoAttention, RingAttention, get_pad_info
 from .dropout import DropoutForParallelInput, DropoutForReplicatedInput
 from .embedding import Embedding1D, PaddingEmbedding, VocabParallelEmbedding1D
 from .linear import Linear1D_Col, Linear1D_Row, PaddingLMHead, VocabParallelLMHead1D
@@ -31,5 +31,7 @@
     "VocabParallelLMHead1D",
     "AttnMaskType",
     "ColoAttention",
+    "RingAttention",
+    "get_pad_info",
     "all_to_all_comm",
 ]

colossalai/shardformer/layer/_operation.py

@@ -2,6 +2,8 @@
 import torch.distributed as dist
 import torch.nn.functional as F
 
+from .utils import is_share_sp_tp
+
 try:
     import fused_mix_prec_layer_norm_cuda
 except:
@@ -93,7 +95,7 @@ def backward(ctx, grad_output):
         if ctx.async_grad_allreduce:
             # Asynchronous all-reduce
             handle = dist.all_reduce(grad_input, group=ctx.process_group, async_op=True)
-            # Relay on CUDA_DEVICE_MAX_CONNECTIONS=1 to have
+            # Rely on CUDA_DEVICE_MAX_CONNECTIONS=1 to have
             # all-reduce scheduled first and have GPU resources allocated, CUDA_DEVICE_MAX_CONNECTIONS=1 is set in shardformer.py
 
         grad_weight = total_input.t().matmul(grad_output)
@@ -143,7 +145,9 @@ def backward(ctx, grad_output):
         if ctx.async_grad_allreduce:
             # Asynchronous all-reduce
             handle = dist.all_reduce(grad_input, group=ctx.process_group, async_op=True)
-            # Relay on CUDA_DEVICE_MAX_CONNECTIONS=1 to have
+            _ = torch.zeros(1, device=grad_input.device)
+
+            # Rely on CUDA_DEVICE_MAX_CONNECTIONS=1 to have
             # all-reduce scheduled first and have GPU resources allocated, CUDA_DEVICE_MAX_CONNECTIONS=1 is set in shardformer.py
 
         if _grad_accum_fusion_available and weight.grad is not None:
@@ -331,7 +335,7 @@ def backward(ctx, grad_output):
                     input_.shape, dtype=input_parallel.dtype, device=input_parallel.device
                 ).contiguous()
                 handle = dist.reduce_scatter(output, input_list, group=process_group, async_op=True)
-                # Relay on CUDA_DEVICE_MAX_CONNECTIONS=1 to have
+                # Rely on CUDA_DEVICE_MAX_CONNECTIONS=1 to have
                 # all-reduce scheduled first and have GPU resources allocated, CUDA_DEVICE_MAX_CONNECTIONS=1 is set in shardformer.py
 
             if _grad_accum_fusion_available and weight.grad is not None:
@@ -646,8 +650,8 @@ def backward(ctx, grad_output):
                     input_.shape, dtype=input_parallel.dtype, device=input_parallel.device
                 ).contiguous()
                 handle = dist.reduce_scatter(output, input_list, group=process_group, async_op=True)
-                # Relay on CUDA_DEVICE_MAX_CONNECTIONS=1 to have
-                # all-reduce scheduled first and have GPU resources allocated, CUDA_DEVICE_MAX_CONNECTIONS=1 is set in shardformer.py
+                # Rely on CUDA_DEVICE_MAX_CONNECTIONS=1 to have
+                # all-reduce scheduled first and have GPU resources allocated
 
             grad_weight = total_input.t().matmul(grad_output)
             grad_bias = grad_output.sum(dim=0) if use_bias else None
@@ -721,16 +725,20 @@ class _ReduceForward(torch.autograd.Function):
 
     Args:
         input_: input matrix.
-        parallel_mode: parallel mode.
+        process_group: communication group.
+
     """
 
     @staticmethod
-    def forward(ctx, input_, process_group):
+    def forward(ctx, input_, process_group, grad_scale=None):
+        ctx.grad_scale = grad_scale
         return _reduce(input_, process_group)
 
     @staticmethod
     def backward(ctx, grad_output):
-        return grad_output, None
+        if ctx.grad_scale is not None:
+            grad_output = grad_output * ctx.grad_scale
+        return grad_output, None, None
 
 
 class _ReduceBackward(torch.autograd.Function):
@@ -979,8 +987,8 @@ def split_forward_gather_backward(input_, dim, process_group, grad_scale=None):
     return _SplitForwardGatherBackward.apply(input_, dim, process_group, grad_scale)
 
 
-def reduce_forward(input_, process_group):
-    return _ReduceForward.apply(input_, process_group)
+def reduce_forward(input_, process_group, grad_scale=None):
+    return _ReduceForward.apply(input_, process_group, grad_scale)
 
 
 def reduce_backward(input_, process_group):
@@ -989,3 +997,13 @@ def reduce_backward(input_, process_group):
 
 def all_to_all_comm(input_, process_group=None, scatter_dim=2, gather_dim=1):
     return _AllToAll.apply(input_, process_group, scatter_dim, gather_dim)
+
+
+def gather_sp_output(hidden_states, sp_group, sp_mode):
+    """
+    Gather the output of the last layer for cross entropy computation
+    """
+    # Rescale grad (HybridParallelPlugin applies ZeRO grad averaging on the DP * SP group)
+    scale = None if is_share_sp_tp(sp_mode) else dist.get_world_size(sp_group)
+    hidden_states = gather_forward_split_backward(hidden_states, 1, sp_group, grad_scale=scale)
+    return hidden_states