dist runtime opt source

colossalai/inference/batch_bucket.py

@@ -521,3 +521,67 @@ class BatchBucket:
 
     def __repr__(self) -> str:
         return f"(sequences_dict={self._sequences_dict}, is_prompts={self.is_prompts})"
+
+
+class RPCBatchBucket(BatchBucket):
+    def __init__(self, *args, **argv):
+        self.is_rpc = True
+        super().__init__(*args, **argv)
+
+    # For compatibility
+    def get_1D_inputs(self) -> List[int]:
+        assert len(self._sequences_dict) > 0, "No sequence in the batch"
+        first_seq = next(iter(self._sequences_dict.values()))  # not exactly the first sequence
+        if first_seq.output_len == 0:
+            # Assume prefill stage
+            assert all(
+                seq.output_len == 0 for seq in self._sequences_dict.values()
+            ), "Sequence stage (Prefill/Decoding) must be the same in the batch"
+            out_li = []
+            seq_ids = sorted(self._sequences_indexes.keys(), key=lambda x: self._sequences_indexes[x])
+            for seq_id in seq_ids:
+                seq: Sequence = self._sequences_dict[seq_id]
+                out_li.extend(seq.input_token_id)
+            return out_li
+        else:
+            # Assume decoding stage
+            if self.use_spec_dec:
+                # For Speculative Decoding
+                # the number of tokens to be verified in parallel plus the correct token in the last step
+                return self.get_1D_inputs_spec_dec(self.num_tokens_to_verify + 1)
+            assert all(
+                seq.output_len > 0 for seq in self._sequences_dict.values()
+            ), "Sequence stage (Prefill/Decoding) must be the same in the batch"
+            assert self.is_compact, "BatchBucket is not compact"
+            out = [0] * self.current_batch_size
+            for seq_id, index_in_b in self._sequences_indexes.items():
+                seq: Sequence = self._sequences_dict[seq_id]
+                out[index_in_b] = seq.output_token_id[-1]
+            return out
+
+    # For compatibility
+    def get_sequence_lengths(self) -> List[int]:
+        assert self.is_compact  # Debug usage
+        sequence_lengths = self.seq_lengths[: self.current_batch_size]
+        return sequence_lengths
+
+    def get_1D_inputs_spec_dec(self, n: int) -> List[int]:
+        # Used for main model verification in **Decoding Stage**
+        # `n` is the number of tokens to be verified,
+        # and so that prepare the last `n` tokens of each sequence as the inputs
+        assert len(self._sequences_dict) > 0, "No sequence in the batch"
+        assert all(
+            seq.output_len >= n for seq in self._sequences_dict.values()
+        ), "Sequence output tokens must be greater than or equal to the number of tokens to be verified."
+        out_li = []
+        seq_ids = sorted(self._sequences_indexes.keys(), key=lambda x: self._sequences_indexes[x])
+        for seq_id in seq_ids:
+            seq: Sequence = self._sequences_dict[seq_id]
+            out_li.extend(seq.output_token_id[-n:])
+        return out_li
+
+    # For compatibility
+    def get_block_table_tensor(self) -> torch.Tensor:
+        assert self.is_compact  # Debug usage
+        block_table = self.block_tables[: self.current_batch_size]
+        return block_table

colossalai/inference/config.py

@@ -87,8 +87,12 @@ class InputMetaData(RPC_PARAM):
 
     def to_rpc_param(self) -> Dict[str, any]:
         return {
-            "block_tables": self.block_tables.tolist(),
-            "sequence_lengths": self.sequence_lengths.tolist(),
+            "block_tables": self.block_tables.tolist()
+            if isinstance(self.block_tables, torch.Tensor)
+            else self.block_tables,
+            "sequence_lengths": self.sequence_lengths.tolist()
+            if isinstance(self.block_tables, torch.Tensor)
+            else self.sequence_lengths,
             "batch_size": self.batch_size,
             "is_prompts": self.is_prompts,
             "use_cuda_kernel": self.use_cuda_kernel,
@@ -113,10 +117,14 @@ class InputMetaData(RPC_PARAM):
         return InputMetaData(
             block_tables=torch.tensor(
                 rpc_dict["block_tables"], dtype=torch.int, device=get_accelerator().get_current_device()
-            ),
+            )
+            if isinstance(rpc_dict["block_tables"], list)
+            else rpc_dict["block_tables"],
             sequence_lengths=torch.tensor(
                 rpc_dict["sequence_lengths"], dtype=torch.int, device=get_accelerator().get_current_device()
-            ),
+            )
+            if isinstance(rpc_dict["sequence_lengths"], list)
+            else rpc_dict["sequence_lengths"],
             batch_size=rpc_dict["batch_size"],
             is_prompts=rpc_dict["is_prompts"],
             use_cuda_kernel=rpc_dict["use_cuda_kernel"],

colossalai/inference/core/request_handler.py

@@ -4,7 +4,7 @@ import torch
 from transformers.configuration_utils import PretrainedConfig
 from transformers.generation import GenerationConfig
 
-from colossalai.inference.batch_bucket import BatchBucket
+from colossalai.inference.batch_bucket import BatchBucket, RPCBatchBucket
 from colossalai.inference.config import InferenceConfig
 from colossalai.inference.flash_decoding_utils import FDIntermTensors
 from colossalai.inference.kv_cache import KVCacheManager, RPCKVCacheManager
@@ -376,7 +376,7 @@ class RPCRequestHandler(RequestHandler):
 
         # TODO In the continuous batching scenario, the batch size may be greater than max_batch_size,
         # which may cause bugs and this issue should be fixed later.
-        self.running_bb = BatchBucket(
+        self.running_bb = RPCBatchBucket(
             num_heads=model_config.num_attention_heads // inference_config.tp_size,
             head_dim=head_dim,
             max_batch_size=self.max_batch_size,
@@ -386,7 +386,7 @@ class RPCRequestHandler(RequestHandler):
             fd_interm_tensor=None,
             dtype=self.dtype,
         )
-        self.prefill_bb = BatchBucket(
+        self.prefill_bb = RPCBatchBucket(
             num_heads=model_config.num_attention_heads // inference_config.tp_size,
             head_dim=head_dim,
             max_batch_size=self.max_batch_size,

colossalai/inference/core/rpc_engine.py

@@ -11,7 +11,7 @@ from torch import multiprocessing as mp
 from transformers import AutoConfig, PreTrainedTokenizer, PreTrainedTokenizerFast
 from transformers.configuration_utils import PretrainedConfig
 
-from colossalai.inference.batch_bucket import BatchBucket
+from colossalai.inference.batch_bucket import RPCBatchBucket
 from colossalai.inference.config import InferenceConfig, InputMetaData
 from colossalai.inference.executor.rpc_worker import rpcWorkerService
 from colossalai.inference.utils import find_available_ports
@@ -161,8 +161,16 @@ class RPCInferenceEngine(InferenceEngine):
                 raise Exception("conn error!")
         self.logger.info(f"Build RPC Connection Success! Begin to load model...")
         asyncio.run(self.init_worker_env())
+        self._init_worker_forward()
         self.logger.info(f"init dist env over")
 
+    def _init_worker_forward(self):
+        """
+        Async wrappers for forward, because it will be invoked many times.
+        """
+        assert len(self.workers) == self.tp_size, "init workers first"
+        self.worker_forwards = [rpyc.async_(worker.execute_model_forward) for worker in self.workers]
+
     async def async_parallel_wrapper(self, f, *args, **kwargs):
         async_res = rpyc.async_(f)(*args, **kwargs)
         await asyncio.to_thread(async_res.wait)
@@ -209,7 +217,8 @@ class RPCInferenceEngine(InferenceEngine):
     def init_device_cache(self, alloc_shape: Tuple[Tuple[int, ...], Tuple[int, ...]]):
         asyncio.run(self._init_device_cache(alloc_shape))
 
-    def prepare_input(self, batch: BatchBucket) -> Tuple[List[int], InputMetaData]:
+    def prepare_input(self, batch: RPCBatchBucket) -> Tuple[List[int], InputMetaData]:
+        assert batch.is_rpc, "the batch must be RPCBatchBucket"
         input_ids = batch.get_1D_inputs()
         sequence_lengths = batch.get_sequence_lengths()
 
@@ -219,7 +228,7 @@ class RPCInferenceEngine(InferenceEngine):
             n_tokens = batch.current_batch_size
             if batch.use_spec_dec:
                 n_tokens = batch.num_tokens_to_verify + 1
-                assert n_tokens == input_ids.size(0)
+                assert n_tokens == len(input_ids)
                 n_tokens = n_tokens * batch.current_batch_size
 
         batch_token_ids = None
@@ -251,20 +260,38 @@ class RPCInferenceEngine(InferenceEngine):
             batch_token_ids=batch_token_ids,
         )
 
-        return input_ids.tolist(), input_meta_data
+        return input_ids, input_meta_data
+
+    async def async_parallel_forward(self, async_f, *args, **kwargs):
+        async_res = async_f(*args, **kwargs)
+        await asyncio.to_thread(async_res.wait)
+        assert async_res.ready
+        return async_res.value
 
     async def step_(self, input_token_ids, input_meta_data: InputMetaData):
         assert len(self.workers) == self.tp_size, "init workers first"
 
-        init_tasks = [
-            self.async_parallel_wrapper(
-                worker.execute_model_forward,
-                input_token_ids,
-                input_meta_data.to_rpc_param(),
-                self.generation_config_dict,
-            )
-            for worker in self.workers
-        ]
+        init_tasks = []
+        for rank, async_forward in enumerate(self.worker_forwards):
+            if rank == 0:
+                init_tasks.append(
+                    self.async_parallel_forward(
+                        async_forward,
+                        input_token_ids,
+                        input_meta_data.to_rpc_param(),
+                        self.generation_config_dict,
+                    )
+                )
+            else:
+                init_tasks.append(
+                    self.async_parallel_forward(
+                        async_forward,
+                        None,
+                        None,
+                        None,
+                    )
+                )
+
         ret = await asyncio.gather(*init_tasks)
 
         return ret[0]
@@ -277,7 +304,6 @@ class RPCInferenceEngine(InferenceEngine):
         next_tokens = asyncio.run(self.step_(input_token_ids, input_meta_data))
 
         # update the request_handler
-        next_tokens = torch.tensor(next_tokens, dtype=torch.int)
         self.request_handler.append_next_tokens(next_tokens)
         finished_sequences = self.request_handler.update()
         return finished_sequences

colossalai/inference/executor/rpc_worker.py

@@ -1,4 +1,4 @@
-from typing import List, Tuple, Union
+from typing import List, Optional, Tuple, Union
 
 import rpyc
 import torch
@@ -18,7 +18,7 @@ from colossalai.inference.modeling.policy import (
     model_policy_map,
 )
 from colossalai.inference.sampler import search_tokens
-from colossalai.inference.utils import get_model_size
+from colossalai.inference.utils import Timer, get_model_size
 from colossalai.interface import ModelWrapper
 from colossalai.logging import get_dist_logger
 from colossalai.pipeline.stage_manager import PipelineStageManager
@@ -51,6 +51,12 @@ class rpcWorkerService(rpyc.Service):
     def exposed_init_dist_env(self, rank, world_size, master_address, master_port):
         logger.info(f"init process group for rank {rank}")
         colossalai.launch(rank=rank, world_size=world_size, port=master_port, host=master_address)
+        self.rank = rank
+
+        # profiling only, remove later
+        self.t_prepare = Timer("[Timer] prepare the data")
+        self.t_exe = Timer("[Timer] execute the model forward")
+        self.t_sampler = Timer("[Timer] sampler time")
         logger.info(f"init process group done for rank {rank}")
 
     def exposed_init_model(
@@ -98,38 +104,50 @@ class rpcWorkerService(rpyc.Service):
         logger.info("physical cache init over")
 
     def exposed_execute_model_forward(
-        self, input_token_ids_param: List[int], input_meta_data_param: dict, generation_config_param: dict
+        self,
+        input_token_ids_param: Optional[List[int]] = None,
+        input_meta_data_param: Optional[dict] = None,
+        generation_config_param: Optional[dict] = None,
     ):
         # prepare the data for model forward
-        input_meta_data = InputMetaData.from_rpc_param(input_meta_data_param)
-        input_meta_data.fd_inter_tensor = self.fd_inter_tensor
+        with self.t_prepare:
+            input_token_ids, input_meta_data, generation_config = self._broadcast_param_to_all_workers(
+                input_token_ids_param=input_token_ids_param,
+                input_meta_data_param=input_meta_data_param,
+                generation_config_param=generation_config_param,
+            )
+
         if input_meta_data.is_prompts:
             n_tokens = input_meta_data.sequence_lengths.sum().item()
         else:
             n_tokens = input_meta_data.batch_size
-        input_token_ids = torch.tensor(input_token_ids_param, dtype=torch.int, device=self.device)
 
         # execute the model
-        logits = self.model(
-            input_token_ids,
-            self.output_tensor[:n_tokens],
-            input_meta_data,
-            self.k_cache,
-            self.v_cache,
-        )
+        with self.t_exe:
+            logits = self.model(
+                input_token_ids,
+                self.output_tensor[:n_tokens],
+                input_meta_data,
+                self.k_cache,
+                self.v_cache,
+            )
 
-        # sampler
-        if self.inference_config.pad_input:
-            logits = logits[:, -1, :]
-        next_tokens = search_tokens(
-            generation_config_param,
-            logits,
-            input_meta_data.is_prompts,
-            input_meta_data.batch_token_ids,
-        )
+        if self.rank == 0:
+            with self.t_sampler:
+                # sampler
+                if self.inference_config.pad_input:
+                    logits = logits[:, -1, :]
+                next_tokens = search_tokens(
+                    generation_config,
+                    logits,
+                    input_meta_data.is_prompts,
+                    input_meta_data.batch_token_ids,
+                )
 
-        # return the tokens generated to scheduler
-        return next_tokens.tolist()
+            # return the tokens generated to scheduler
+            # only rank 0 need to pass the data back
+            # to reduce the overhead of rpc param passing
+            return next_tokens.cpu()
 
     def _init_output_tensor(self):
         alloc_shape = (
@@ -166,6 +184,84 @@ class rpcWorkerService(rpyc.Service):
 
         self.fd_inter_tensor = fd_inter_tensor
 
+    def _broadcast_param_to_all_workers(
+        self,
+        input_token_ids_param: Optional[List[int]] = None,
+        input_meta_data_param: Optional[dict] = None,
+        generation_config_param: Optional[dict] = None,
+    ):
+        if self.rank == 0:
+            input_meta_data = InputMetaData.from_rpc_param(input_meta_data_param)
+            input_meta_data.fd_inter_tensor = self.fd_inter_tensor
+            input_token_ids = torch.tensor(input_token_ids_param, dtype=torch.int, device=self.device)
+            generation_config = generation_config_param
+
+            if dist.get_world_size() > 1:
+                broadcast_list = {}
+                for k, v in input_meta_data_param.items():
+                    if not isinstance(v, List):
+                        broadcast_list[k] = v
+
+                # Pass the tensor shape and type in advance for
+                # other workers to prepare the empty tensor and async transport tensors
+                broadcast_list["block_tables"] = (
+                    input_meta_data.block_tables.size(),
+                    input_meta_data.block_tables.dtype,
+                )
+                broadcast_list["sequence_lengths"] = (
+                    input_meta_data.sequence_lengths.size(),
+                    input_meta_data.sequence_lengths.dtype,
+                )
+                broadcast_list["input_token_ids"] = (input_token_ids.size(), input_token_ids.dtype)
+
+                # Generation Config Param
+                broadcast_list["generation_config"] = generation_config_param
+
+                # send some meta data and some tensor shape
+                torch.distributed.broadcast_object_list([broadcast_list], src=self.rank)
+
+                # send the real tensor
+                torch.distributed.broadcast(input_meta_data.block_tables, src=self.rank)
+                torch.distributed.broadcast(input_meta_data.sequence_lengths, src=self.rank)
+                torch.distributed.broadcast(input_token_ids, src=self.rank)
+
+        else:
+            assert input_meta_data_param is None, "Input Must Be None"
+
+            # recv the meta data
+            recv_list = [None]
+            torch.distributed.broadcast_object_list(recv_list, src=0)
+            input_meta_data_param = recv_list[0]
+
+            generation_config = input_meta_data_param["generation_config"]
+
+            blocktable_shape, blocktable_type = input_meta_data_param["block_tables"]
+            blocktables = torch.empty(blocktable_shape, dtype=blocktable_type, device=self.device)
+            sequence_lengths_shape, sequence_lengths_type = input_meta_data_param["sequence_lengths"]
+            sequence_lengths = torch.empty(sequence_lengths_shape, dtype=sequence_lengths_type, device=self.device)
+            input_token_ids_shape, input_token_ids_type = input_meta_data_param["input_token_ids"]
+            input_token_ids = torch.empty(input_token_ids_shape, dtype=input_token_ids_type, device=self.device)
+
+            # recv the real tensor
+            async1 = torch.distributed.broadcast(blocktables, src=0, async_op=True)
+            async2 = torch.distributed.broadcast(sequence_lengths, src=0, async_op=True)
+            async3 = torch.distributed.broadcast(input_token_ids, src=0, async_op=True)
+
+            input_meta_data_param["sequence_lengths"] = sequence_lengths
+            input_meta_data_param["blocktables"] = blocktables
+
+            input_meta_data = InputMetaData.from_rpc_param(input_meta_data_param)
+            input_meta_data.fd_inter_tensor = self.fd_inter_tensor
+
+            async1.wait()
+            async2.wait()
+            async3.wait()
+
+            input_meta_data.block_tables = blocktables
+            input_meta_data.sequence_lengths = sequence_lengths
+
+        return input_token_ids, input_meta_data, generation_config
+
     def _init_model(self, model_or_path: Union[nn.Module, str], model_policy: Policy = None):
         """
         Shard model or/and Load weight
@@ -304,3 +400,11 @@ class rpcWorkerService(rpyc.Service):
         logger.info(f"Worker rank {dist_rank}: Sum after all_reduce: {data.item()}")
 
         return data.item()
+
+    def __del__(self):
+        """
+        profiling only, remove later
+        """
+        del self.t_prepare
+        del self.t_exe
+        del self.t_sampler

colossalai/inference/utils.py

@@ -113,3 +113,70 @@ def find_available_ports(num: int):
         print(f"An OS error occurred: {e}")
         raise RuntimeError("Error finding available ports")
     return free_ports
+
+
+"""
+below just for profiling temporarily, will removed before merge
+"""
+import time
+from contextlib import asynccontextmanager, contextmanager
+
+
+@contextmanager
+def timer(name=""):
+    # (@lry89757) will remove later
+    start_time = time.time()
+    try:
+        yield
+    finally:
+        end_time = time.time()
+        elapsed_time = end_time - start_time
+        print(f"{name} took {elapsed_time:.6f} seconds")
+
+
+class Timer:
+    # (@lry89757) will remove later
+    def __init__(self, name=""):
+        print(f"init timer, {name}")
+        self.name = name
+        self.times = []
+
+    def __enter__(self):
+        self.start_time = time.time()
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        end_time = time.time()
+        elapsed_time = end_time - self.start_time
+        self.times.append(elapsed_time)
+        print(f"{self.name} took {elapsed_time:.6f} seconds")
+        self.print_info()
+
+    def print_info(self):
+        average_prefill_time = self.times[0]
+        print(f"{self.name} prefill average time: {average_prefill_time:.6f} seconds")
+        if len(self.times) > 1:
+            average_decoding_time = sum(self.times[1:]) / len(self.times[1:])
+            print(f"{self.name} decoding average time: {average_decoding_time:.6f} seconds")
+
+    def __del__(self):
+        if self.times:
+            average_prefill_time = self.times[0]
+            print(f"{self.name} prefill average time: {average_prefill_time:.6f} seconds")
+            if len(self.times) > 1:
+                average_decoding_time = sum(self.times[1:]) / len(self.times[1:])
+                print(f"{self.name} decoding average time: {average_decoding_time:.6f} seconds")
+        else:
+            print(f"{self.name} no timings recorded")
+
+
+@asynccontextmanager
+async def async_timer(name=""):
+    # (@lry89757) will remove later
+    start_time = time.time()
+    try:
+        yield
+    finally:
+        end_time = time.time()
+        elapsed_time = end_time - start_time
+        print(f"{name} took {elapsed_time:.6f} seconds")