support Pipeline parallelism

mlora.py

@@ -26,6 +26,7 @@
 
 # Command Line Arguments
 parser = argparse.ArgumentParser(description='m-LoRA main program')
+# the argument about model and tokenizer
 parser.add_argument('--base_model', type=str,
                     help='Path to or name of base model')
 parser.add_argument('--tokenizer', type=str,
@@ -34,25 +35,31 @@
                     help='The model type, support: llama, chatglm')
 parser.add_argument('--device', type=str, default='cuda:0',
                     help='Specify which GPU to be used, default is cuda:0')
-
-parser.add_argument('--load_8bit', action="store_true",
-                    help='Load model in 8bit mode')
-parser.add_argument('--load_4bit', action="store_true",
-                    help='Load model in 4bit mode')
-
 parser.add_argument('--inference', action="store_true",
                     help='The inference mode (just for test)')
-
+# the argument about lora
 parser.add_argument('--load_lora', action="store_true",
                     help="Load lora from file instead of init randomly")
 parser.add_argument('--disable_lora', action="store_true",
                     help="Disable the lora modules")
-
+# the argument about pipeline
+parser.add_argument('--pipeline', action="store_true",
+                    help="Train the LoRA model use the pipeline parallelism")
+parser.add_argument('--rank', type=int, default=-1,
+                    help="The device's rank number")
+parser.add_argument('--balance', type=int, default=-1,
+                    help="The device's rank number")
+# the argument about quant
+parser.add_argument('--load_8bit', action="store_true",
+                    help='Load model in 8bit mode')
+parser.add_argument('--load_4bit', action="store_true",
+                    help='Load model in 4bit mode')
+# the argument about config
 parser.add_argument('--config', type=str,
                     help='Path to finetune configuration')
 parser.add_argument('--seed', type=int, default=42,
                     help='Random seed in integer, default is 42')
-
+# the argument about log
 parser.add_argument('--log_level', type=str, default="INFO",
                     help="Set the log level.")
 parser.add_argument('--log_file', type=str,
@@ -294,6 +301,13 @@ def inference(config: Dict[str, any],
     tokenizer, model = load_base_model(config)
     init_lora_model(config, model)
 
+    if args.pipeline:
+        pipe = mlora.Pipe(model,
+                          device=torch.device(args.device),
+                          rank=args.rank,
+                          balance=[9, 8, 8, 10])
+        exit(pipe.run())
+
     torch.cuda.empty_cache()
 
     if args.inference:

mlora/__init__.py

@@ -5,6 +5,7 @@
 from mlora.model_chatglm import ChatGLMModel
 from mlora.modelargs import LLMModelArgs, MultiLoraBatchData, LoraBatchDataConfig
 from mlora.dispatcher import TrainTask, Dispatcher
+from mlora.pipeline.pipe import Pipe
 
 __all__ = [
     "Tokenizer",
@@ -17,5 +18,6 @@
     "convert_hf_to_pth",
     "save_lora_model",
     "TrainTask",
-    "Dispatcher"
+    "Dispatcher",
+    "Pipe"
 ]

mlora/model.py

@@ -112,6 +112,9 @@ def forward(self, data: torch.Tensor) -> torch.Tensor:
 
 
 class LLMModel(metaclass=ABCMeta):
+    n_heads_: int = -1
+    rope_angle_: torch.tensor = None
+
     @abstractclassmethod
     def forward(self, input: MultiLoraBatchData):
         pass

mlora/pipeline/function.py

@@ -0,0 +1,18 @@
+import torch
+
+
+class AutogradWithoutActivations(torch.autograd.Function):
+    # helper to reduce the activation memory
+    @staticmethod
+    def forward(ctx, *x):
+        return torch.tensor(1, 0)
+
+    @staticmethod
+    def backward(ctx, grad):
+        assert ctx.grad_from_next_worker is not None
+        return ctx.grad_from_next_worker
+
+
+class RecvOperator(torch.autograd.Function):
+    # backward will auto send the grad to pre worker
+    pass

mlora/pipeline/messages.py

@@ -0,0 +1,77 @@
+import torch
+import torch.distributed.rpc
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from queue import Queue
+from typing import Dict
+from enum import Enum
+
+
+class PipeMessageType(Enum):
+    ACTIVATIONS = "ACTIVATIONS"
+    GRADIENTS = "GRADIENTS"
+
+
+@dataclass()
+class PipeMessage:
+    src_: str
+    dst_: str
+
+    msg_type_: PipeMessageType
+    tensors_: torch.Tensor
+
+
+# save by different message type
+RPCMessageQueues: Dict[str, Queue] = {
+    PipeMessageType.ACTIVATIONS.value: Queue(),
+    PipeMessageType.GRADIENTS.value: Queue()
+}
+
+
+def rpc_push_queue(msg: PipeMessage) -> None:
+    assert msg.msg_type_ in globals(
+    )["RPCMessageQueues"], f"No this message type: {msg.msg_type_.value}"
+
+    globals()["RPCMessageQueues"][msg.msg_type_.value].put(msg)
+
+
+class Transport(ABC):
+    device_: torch.device
+
+    @abstractmethod
+    def recv_message(self, msg_type: PipeMessageType, block: bool = False) -> PipeMessage:
+        pass
+
+    @abstractmethod
+    def send_message(self, msg: PipeMessage, sync: bool = False) -> None:
+        pass
+
+
+class RpcTransport(Transport):
+    def __init__(self, worker_device: torch.device) -> None:
+        super().__init__()
+        self.device_ = worker_device
+
+    # recv message will be blocked by convert the cpu memory to gpu memory
+    def recv_message(self, msg_type: PipeMessageType, block: bool = True) -> PipeMessage:
+        assert msg_type.value in globals(
+        )["RPCMessageQueues"], f"No this message type: {msg_type.value}"
+
+        queue = globals()["RPCMessageQueues"][msg_type.value]
+        if queue.empty():
+            return None
+
+        result = queue.get(block=block)
+        result.tensors_ = tuple(t.to(self.device_) for t in result.tensors_)
+        return result
+
+    # send message will be blocked by convert the gpu memory to cpu memory
+    def send_message(self, msg: PipeMessage, sync: bool = False) -> None:
+        msg.tensors_ = tuple(t.to("cpu") for t in msg.tensors_)
+        if sync:
+            send_fn = torch.distributed.rpc.rpc_sync
+        else:
+            send_fn = torch.distributed.rpc.rpc_async
+
+        send_fn(msg.dst_, rpc_push_queue, args=(msg,))

mlora/pipeline/pipe.py

@@ -0,0 +1,159 @@
+from mlora.modelargs import MultiLoraBatchData, LoraBatchDataConfig
+from mlora.model import LLMModel, precompute_mask
+from mlora.pipeline.messages import Transport, RpcTransport
+from mlora.pipeline.messages import PipeMessage, PipeMessageType
+
+import os
+import torch
+import time
+import logging
+
+import torch.distributed.rpc
+from typing import List
+from enum import Enum, auto
+from dataclasses import dataclass
+
+
+class WorkerRole(Enum):
+    HEAD = auto()
+    MID = auto()
+    TAIL = auto()
+
+
+@dataclass(frozen=True)
+class PipeModelLocInfo():
+    n_heads_: int = -1
+    rope_angle_: torch.Tensor = None
+
+
+class Pipe():
+    world_size_: int = -1
+    balance_: List[int] = None
+
+    rank_: int = -1
+    device_: torch.device = None
+
+    model_partition_: torch.nn.Sequential = torch.nn.Sequential()
+    model_info_: PipeModelLocInfo = None
+    role_: WorkerRole = None
+
+    transport_: Transport = None
+
+    def __init__(self,
+                 model: LLMModel,
+                 device: torch.device,
+                 rank: int,
+                 balance: List[int]) -> None:
+
+        world_size = torch.cuda.device_count()
+        assert world_size == len(balance)
+
+        self.rank_ = rank
+        self.world_size_ = world_size
+        self.balance_ = balance
+        self.device_ = device
+        self.model_info_ = PipeModelLocInfo(
+            n_heads_=model.n_heads_,
+            rope_angle_=model.rope_angle_
+        )
+
+        if rank == 0:
+            self.role_ = WorkerRole.HEAD
+        elif rank == len(balance) - 1:
+            self.role_ = WorkerRole.TAIL
+        else:
+            self.role_ = WorkerRole.MID
+
+        self.transport_ = RpcTransport(self.device_)
+        self.init_partition(model)
+        self.init_rpc()
+
+    @property
+    def next_worker_name(self) -> str:
+        return f"worker-{self.rank_ + 1}"
+
+    @property
+    def prev_worker_name(self) -> str:
+        return f"worker-{self.rank_ - 1}"
+
+    # process the backward
+    # process the forward
+    def run(self):
+        while True:
+            self.process_backward()
+            self.process_forward()
+            self.process_input()
+            time.sleep(5)
+
+    def process_input(self) -> None:
+        if self.role_ != WorkerRole.HEAD:
+            return
+        logging.debug("Train input data.")
+        import random
+        test_data_len = 96
+        tokens_list = [[random.random() for _ in range(test_data_len)]]
+        input = MultiLoraBatchData(
+            batch_tokens_=tokens_list,
+            additional_mask_=None,
+            lora_batch_data_config_=[LoraBatchDataConfig(
+                adapter_name_="lora_0",
+                batch_start_idx_=0,
+                batch_end_idx_=1
+            )],
+            inference_model_=False
+        )
+
+        tokens_batch = torch.tensor(
+            input.batch_tokens_, dtype=torch.int64).to(self.device_)
+        mask = precompute_mask(
+            tokens_batch, self.model_info_.n_heads_, self.device_)
+
+        dummy_data = (tokens_batch, mask,
+                      self.model_info_.rope_angle_, input, True)
+
+        for seq_model in self.model_partition_:
+            dummy_data = seq_model.forward(dummy_data)
+
+    def process_forward(self) -> None:
+        if self.role_ == WorkerRole.HEAD:
+            return
+
+    def process_backward(self) -> None:
+        logging.debug("To recv the gradients message.")
+        message = self.transport_.recv_message(
+            PipeMessageType.GRADIENTS, block=False)
+        if message is None:
+            logging.debug("No gradients to deal.")
+            return
+        logging.debug(f"Have already recv the gradients from {message.src_}")
+
+    def init_partition(self, model: LLMModel) -> None:
+        balance = self.balance_[self.rank_]
+        start_module_idx = sum(self.balance_[:self.rank_])
+        logging.info(
+            f"RANK-{self.rank_} in device {self.device_} to load module layers from {start_module_idx} to {start_module_idx + balance}.")
+
+        seq_model = model.sequential_module()
+        del seq_model[:start_module_idx]
+        del seq_model[balance:]
+        for idx in range(0, len(seq_model)):
+            self.model_partition_.append(seq_model[idx])
+
+        assert len(self.model_partition_) == balance
+
+        del model
+
+        torch.cuda.empty_cache()
+
+    def init_rpc(self) -> None:
+        if "MASTER_ADDR" not in os.environ:
+            os.environ["MASTER_ADDR"] = "localhost"
+        if "MASTER_PORT" not in os.environ:
+            os.environ["MASTER_PORT"] = "12355"
+
+        # will be block when all world size's gpu join the group
+        torch.distributed.rpc.init_rpc(
+            f"worker-{self.rank_}", rank=self.rank_, world_size=self.world_size_)
+
+        logging.info(
+            f"Init rpc with rank {self.rank_} world_size: {self.world_size_}")