github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-08 20:40:34 +00:00
Code Issues Packages Projects Releases Wiki Activity

[inference] Refactor inference architecture (#5057)

Browse Source 
* [inference] support only TP (#4998)

* support only tp

* enable tp

* add support for bloom (#5008)

* [refactor] refactor gptq and smoothquant llama (#5012)

* refactor gptq and smoothquant llama

* fix import error

* fix linear import torch-int

* fix smoothquant llama import error

* fix import accelerate error

* fix bug

* fix import smooth cuda

* fix smoothcuda

* [Inference Refactor] Merge chatglm2 with pp and tp (#5023)

merge chatglm with pp and tp

* [Refactor] remove useless inference code (#5022)

* remove useless code

* fix quant model

* fix test import bug

* mv original inference legacy

* fix chatglm2

* [Refactor] refactor policy search and quant type controlling in inference (#5035)

* [Refactor] refactor policy search and quant type controling in inference

* [inference] update readme (#5051)

* update readme

* update readme

* fix architecture

* fix table

* fix table

* [inference] udpate example (#5053)

* udpate example

* fix run.sh

* fix rebase bug

* fix some errors

* update readme

* add some features

* update interface

* update readme

* update benchmark

* add requirements-infer

---------

Co-authored-by: Bin Jia <45593998+FoolPlayer@users.noreply.github.com>
Co-authored-by: Zhongkai Zhao <kanezz620@gmail.com>
This commit is contained in:
Xu Kai
2023-11-19 21:05:05 +08:00
committed by GitHub
parent bc09b95f50
commit fd6482ad8c
115 changed files with 6027 additions and 1431 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
colossalai/legacy/inference/dynamic_batching/__init__.py Normal file
View File

40
colossalai/legacy/inference/dynamic_batching/get_tokenizer.py Normal file
View File

@@ -0,0 +1,40 @@
"""
Motivated by VllM (https://github.com/vllm-project/vllm), This module is trying to resolve the tokenizer issue.
license: MIT, see LICENSE for more details.
"""
from transformers import AutoTokenizer
_FAST_LLAMA_TOKENIZER = "hf-internal-testing/llama-tokenizer"
def get_tokenizer(
tokenizer=None,
tokenizer_name: str = "",
trust_remote_code: bool = False,
use_fast: bool = True,
):
if tokenizer is not None:
tokenizer = tokenizer
else:
if "llama" in tokenizer_name.lower() and use_fast == True:
print(
"For some LLaMA-based models, initializing the fast tokenizer may "
"take a long time. To eliminate the initialization time, consider "
f"using '{_FAST_LLAMA_TOKENIZER}' instead of the original "
"tokenizer. This is done automatically in Colossalai."
)
tokenizer_name = _FAST_LLAMA_TOKENIZER
try:
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name, use_fast=use_fast, trust_remote_code=trust_remote_code
)
except TypeError:
use_fast = False
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name, use_fast=use_fast, trust_remote_code=trust_remote_code
)
return tokenizer

346
colossalai/legacy/inference/dynamic_batching/infer_batch.py Normal file
View File

@@ -0,0 +1,346 @@
# Adapted from https://github.com/ModelTC/lightllm
import collections
from dataclasses import dataclass
from typing import Dict, List, Tuple
import numpy as np
import torch
from colossalai.inference.tensor_parallel import MemoryManager
# make batch infer state an attr of InferBatch
class InferSamplingParams:
def __init__(
self,
do_sample: bool = False,
presence_penalty: float = 0.0,
frequency_penalty: float = 0.0,
temperature: float = 1.0,
top_p: float = 1.0,
top_k: int = -1,
vocab_size: int = -1,
) -> None:
self.do_sample = do_sample
self.presence_penalty = presence_penalty
self.frequency_penalty = frequency_penalty
self.temperature = temperature
self.top_p = top_p
self.top_k = top_k
if self.top_k == -1:
self.top_k = vocab_size
return
@dataclass
class InferBatch:
batch_id: int
requests: List
requests_idx_mapping: Dict[int, int]
input_ids: torch.Tensor
all_input_ids: List[List[int]]
input_lengths: List[int]
out_token_id_counts: List
sampling_param_list: List[InferSamplingParams]
nopad_total_token_num: int
nopad_max_len_in_batch: int
nopad_b_loc: torch.Tensor
nopad_b_start_loc: torch.Tensor
nopad_b_seq_len: torch.Tensor
cache_manager: MemoryManager
max_total_len: int
@classmethod
@torch.no_grad()
def init_batch(
cls,
batch_id,
requests,
dtype: torch.dtype,
device: torch.device,
cache_manager: MemoryManager,
vocab_size: int,
max_total_len: int,
) -> "InferBatch":
input_lengths = []
all_input_ids = []
requests_idx_mapping = {}
out_token_id_counts = []
sampling_param_list = []
nopad_total_token_num = 0
nopad_max_len_in_batch = 0
nopad_b_loc = torch.empty((len(requests), max_total_len + 12), dtype=torch.long, device="cuda")
# to avoid memory leak , we pre-allocate 12 more space for each batch.
nopad_b_start_loc = torch.zeros(len(requests), dtype=torch.int32, device="cuda")
for i, r in enumerate(requests):
# request id -> idx in list mapping
requests_idx_mapping[r["request_id"]] = i
tokenized_input = r["input_id"]
input_length = len(tokenized_input)
input_lengths.append(input_length)
all_input_ids.append(tokenized_input)
out_token_id_counts.append(collections.defaultdict(int))
# postprocessor
sampling_param = r["sampling_param"]
sampling_param["vocab_size"] = vocab_size
sampling_param_list.append(InferSamplingParams(**sampling_param))
nopad_total_token_num += input_length
nopad_max_len_in_batch = max(nopad_max_len_in_batch, input_length)
nopad_b_seq_len = torch.tensor(input_lengths, dtype=torch.int32, device="cuda")
nopad_b_start_loc[1:] = torch.cumsum(nopad_b_seq_len, dim=0, dtype=torch.int32)[0:-1]
if len(requests) > 1:
input_ids = np.concatenate(all_input_ids, dtype=np.int64)
else:
input_ids = all_input_ids[0]
# Create tensors on device
input_ids = torch.tensor(input_ids, dtype=torch.int64, device=device)
return cls(
batch_id=batch_id,
requests=requests,
requests_idx_mapping=requests_idx_mapping,
input_ids=input_ids,
input_lengths=input_lengths,
all_input_ids=all_input_ids,
nopad_total_token_num=nopad_total_token_num,
nopad_max_len_in_batch=nopad_max_len_in_batch,
nopad_b_loc=nopad_b_loc,
nopad_b_start_loc=nopad_b_start_loc,
nopad_b_seq_len=nopad_b_seq_len,
out_token_id_counts=out_token_id_counts,
sampling_param_list=sampling_param_list,
cache_manager=cache_manager,
max_total_len=max_total_len,
)
@torch.no_grad()
def free_self(self) -> None:
"""
Free the memory of the InferBatch itself
"""
remove_index = []
for idx in range(len(self)):
remove_index.append(
self.nopad_b_loc[
idx,
(self.nopad_max_len_in_batch - 1)
- (self.nopad_b_seq_len[idx] - 1) : (self.nopad_max_len_in_batch - 1),
]
)
remove_index = torch.cat(remove_index, dim=-1)
self.cache_manager.free(remove_index)
@torch.no_grad()
def filter(self, request_ids: List[int]) -> "InferBatch":
"""
Filter finished batch and return a new InferBatch with left ones.
"""
if len(request_ids) == 0:
raise ValueError("Batch must have at least one request")
if len(request_ids) == len(self):
return self
requests_idx_mapping = {}
indices = []
requests = []
all_input_ids = []
input_lengths = []
nopad_total_token_num = 0
nopad_max_len_in_batch = 0
nopad_b_loc = torch.empty((len(request_ids), self.max_total_len + 12), dtype=torch.long, device="cuda")
nopad_b_start_loc = torch.zeros(len(request_ids), dtype=torch.int32, device="cuda")
nopad_b_seq_len = torch.zeros(len(request_ids), dtype=torch.int32, device="cuda")
left_idx = []
for i, request_id in enumerate(request_ids):
idx = self.requests_idx_mapping[request_id]
left_idx.append(idx)
left_idx_set = set(left_idx)
remove_index = []
for idx in range(len(self)):
if idx not in left_idx_set:
remove_index.append(
self.nopad_b_loc[
idx,
(self.nopad_max_len_in_batch - 1)
- (self.nopad_b_seq_len[idx] - 1) : (self.nopad_max_len_in_batch - 1),
]
)
remove_index = torch.cat(remove_index, dim=-1)
self.cache_manager.free(remove_index)
nopad_max_len_in_batch = 0
for i, request_id in enumerate(request_ids):
idx = self.requests_idx_mapping[request_id]
indices.append(idx)
nopad_b_seq_len[:] = self.nopad_b_seq_len[indices]
nopad_max_len_in_batch = torch.max(nopad_b_seq_len).item()
nopad_b_start_loc[1:] = torch.cumsum(nopad_b_seq_len, dim=0, dtype=torch.int32)[0:-1]
nopad_total_token_num = torch.sum(nopad_b_seq_len).item()
nopad_b_loc[:, 0 : (nopad_max_len_in_batch - 1)] = self.nopad_b_loc[
indices,
(self.nopad_max_len_in_batch - 1) - (nopad_max_len_in_batch - 1) : (self.nopad_max_len_in_batch - 1),
]
for i, request_id in enumerate(request_ids):
idx = self.requests_idx_mapping[request_id]
requests_idx_mapping[request_id] = i
requests.append(self.requests[idx])
all_input_ids.append(self.all_input_ids[idx])
input_lengths.append(self.input_lengths[idx])
input_ids = self.input_ids[indices]
return InferBatch(
batch_id=self.batch_id,
requests=requests,
requests_idx_mapping=requests_idx_mapping,
input_ids=input_ids,
input_lengths=input_lengths,
all_input_ids=all_input_ids,
nopad_total_token_num=nopad_total_token_num,
nopad_max_len_in_batch=nopad_max_len_in_batch,
nopad_b_loc=nopad_b_loc,
nopad_b_start_loc=nopad_b_start_loc,
nopad_b_seq_len=nopad_b_seq_len,
out_token_id_counts=[self.out_token_id_counts[_i] for _i in indices],
sampling_param_list=[self.sampling_param_list[_i] for _i in indices],
cache_manager=self.cache_manager,
max_total_len=self.max_total_len,
)
@classmethod
@torch.no_grad()
def merge(cls, batch1, batch2) -> "InferBatch":
"""
Return megerd new InferBatch
"""
requests = batch1.requests + batch2.requests
requests_idx_mapping = {}
new_batch_size = len(batch1) + len(batch2)
input_ids = batch1.input_ids.new_empty(new_batch_size)
all_input_ids = []
input_lengths = []
out_token_id_counts = []
sampling_param_list = []
cumulative_batch_size = 0
nopad_total_token_num = batch1.nopad_total_token_num + batch2.nopad_total_token_num
nopad_max_len_in_batch = max(batch1.nopad_max_len_in_batch, batch2.nopad_max_len_in_batch)
max_total_len = max(batch1.max_total_len, batch2.max_total_len)
nopad_b_loc = torch.empty((new_batch_size, batch1.max_total_len + 12), dtype=torch.long, device="cuda")
nopad_b_start_loc = torch.zeros(new_batch_size, dtype=torch.int32, device="cuda")
nopad_b_seq_len = torch.zeros(new_batch_size, dtype=torch.int32, device="cuda")
nopad_start_loc_len_temp = 0
batches = [batch1, batch2]
for i, batch in enumerate(batches):
if i == 0:
requests_idx_mapping = batch.requests_idx_mapping
else:
for k, v in batch.requests_idx_mapping.items():
requests_idx_mapping[k] = v + cumulative_batch_size
start_index = cumulative_batch_size
end_index = cumulative_batch_size + len(batch)
input_ids[start_index:end_index] = batch.input_ids
nopad_b_seq_len[start_index:end_index] = batch.nopad_b_seq_len
nopad_b_start_loc[start_index:end_index] = batch.nopad_b_start_loc + nopad_start_loc_len_temp
nopad_start_loc_len_temp = nopad_b_start_loc[end_index - 1] + nopad_b_seq_len[end_index - 1]
nopad_b_loc[
start_index:end_index,
nopad_max_len_in_batch - batch.nopad_max_len_in_batch : nopad_max_len_in_batch - 1,
] = batch.nopad_b_loc[:, : batch.nopad_max_len_in_batch - 1]
all_input_ids.extend(batch.all_input_ids)
input_lengths.extend(batch.input_lengths)
out_token_id_counts.extend(batch.out_token_id_counts)
sampling_param_list.extend(batch.sampling_param_list)
# Update
cumulative_batch_size += len(batch)
nopad_b_loc[:, nopad_max_len_in_batch - 1] = (
nopad_total_token_num - new_batch_size + torch.arange(0, new_batch_size, dtype=torch.int32, device="cuda")
)
return InferBatch(
batch_id=batches[0].batch_id,
requests=requests,
requests_idx_mapping=requests_idx_mapping,
input_ids=input_ids,
input_lengths=input_lengths,
all_input_ids=all_input_ids,
nopad_total_token_num=nopad_total_token_num,
nopad_max_len_in_batch=nopad_max_len_in_batch,
nopad_b_loc=nopad_b_loc,
nopad_b_start_loc=nopad_b_start_loc,
nopad_b_seq_len=nopad_b_seq_len,
out_token_id_counts=out_token_id_counts,
sampling_param_list=sampling_param_list,
cache_manager=batches[0].cache_manager,
max_total_len=max_total_len,
)
def __len__(self):
return len(self.requests)
def get_post_sample_tensors(self) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
presence_penalties: List[float] = []
frequency_penalties: List[float] = []
temperatures: List[float] = []
top_ps: List[float] = []
top_ks: List[int] = []
p_token_ids: List[int] = []
p_token_counts: List[int] = []
p_seq_len: List[int] = [
0,
]
p_max_len_in_batch: int = 0
for i, id_to_count in enumerate(self.out_token_id_counts):
sample_param = self.sampling_param_list[i]
presence_penalties.append(sample_param.presence_penalty)
frequency_penalties.append(sample_param.frequency_penalty)
temperatures.append(sample_param.temperature)
top_ps.append(sample_param.top_p)
top_ks.append(sample_param.top_k)
for token_id, count in id_to_count.items():
p_token_ids.append(token_id)
p_token_counts.append(count)
p_seq_len.append(len(id_to_count))
p_max_len_in_batch = max(p_max_len_in_batch, len(id_to_count))
presence_penalties = torch.tensor(presence_penalties, dtype=torch.float, device="cuda")
frequency_penalties = torch.tensor(frequency_penalties, dtype=torch.float, device="cuda")
temperatures = torch.tensor(temperatures, dtype=torch.float, device="cuda")
top_ps = torch.tensor(top_ps, dtype=torch.float, device="cuda")
top_ks = torch.tensor(top_ks, dtype=torch.int32, device="cuda")
p_token_ids = torch.tensor(p_token_ids, dtype=torch.int32, device="cuda")
p_token_counts = torch.tensor(p_token_counts, dtype=torch.int32, device="cuda")
p_seq_len = torch.tensor(p_seq_len, dtype=torch.int32, device="cuda")
p_cumsum_seq_len = torch.cumsum(p_seq_len, dim=0, dtype=torch.int32)
return (
presence_penalties,
frequency_penalties,
temperatures,
top_ps,
top_ks,
p_token_ids,
p_token_counts,
p_cumsum_seq_len,
p_max_len_in_batch,
)

166
colossalai/legacy/inference/dynamic_batching/io_struct.py Normal file
View File

@@ -0,0 +1,166 @@
# Adapted from https://github.com/ModelTC/lightllm
from typing import Dict, List, Tuple
from .sampling_params import SamplingParams
class Req:
def __init__(self, request_id, prompt_ids, sample_params: SamplingParams, prompts: str = ""):
self.request_id = request_id
self.prompt_ids = prompt_ids
self.input_len = len(prompt_ids)
self.max_output_len = sample_params.max_new_tokens
self.sample_params = sample_params
self.output_ids = []
self.output_metadata_list = []
self.has_generate_finished = False
self.aborted = False
self.prompts = prompts
def to_rpc_obj(self):
return {
"request_id": self.request_id,
"input_id": self.prompt_ids,
"output_len": self.max_output_len,
"sampling_param": self.sample_params.to_dict(),
}
def stop_sequences_matched(self):
# should we add stpp sequences to the sample params?
if self.sample_params.stop_sequences is not None:
for stop_token_ids in self.sample_params.stop_sequences:
stop_len = len(stop_token_ids)
if (
stop_len > 0
and len(self.output_ids) >= stop_len
and all(self.output_ids[-(stop_len - i)] == stop_token_ids[i] for i in range(stop_len))
):
return True
return False
def __repr__(self):
return f"request_id(n={self.request_id}, " f"prompt_ids={self.prompt_ids}, "
class Batch:
def __init__(self, batch_id, reqs: List[Req]):
self.batch_id = batch_id
self.reqs = reqs
self.id_to_reqs = {req.request_id: req for req in reqs}
def input_tokens(self):
batch_input_tokens = 0
for req in self.reqs:
batch_input_tokens += req.input_len
return batch_input_tokens
def calcu_max_tokens(self):
tokens = 0
for req in self.reqs:
tokens += req.input_len + req.max_output_len
return tokens
def calcu_used_tokens(self):
tokens = 0
for req in self.reqs:
tokens += req.input_len + len(req.output_ids)
return tokens
def mark_finished_req(self, eos_id, engine_max_output_len):
has_new_finish = False
for req in self.reqs:
if req.stop_sequences_matched():
req.has_generate_finished = True
has_new_finish = True
if len(req.output_ids) >= engine_max_output_len:
req.has_generate_finished = True
has_new_finish = True
if req.output_ids[-1] == eos_id and req.sample_params.ignore_eos == False:
req.has_generate_finished = True
has_new_finish = True
if len(req.output_ids) >= req.max_output_len or req.aborted:
req.has_generate_finished = True
has_new_finish = True
return has_new_finish
def filter_finished(self) -> List[Req]:
"""
Filter finished requests from the batch, the finished ones will be removed from 'reqs'.
"""
# TODO: the logic of return should be defined here.
unfinished_req = []
finished_req = []
for req in self.reqs:
if not req.has_generate_finished:
unfinished_req.append(req)
else:
finished_req.append(req)
self.reqs = unfinished_req
self.id_to_reqs = {req.request_id: req for req in self.reqs}
return finished_req
def is_clear(self):
return len(self.reqs) == 0
def merge(self, mini_batch):
for _req in mini_batch.reqs:
self.reqs.append(_req)
self.id_to_reqs = {req.request_id: req for req in self.reqs}
return
def __repr__(self):
return f"batch_id={self.batch_id}, " f"reqs={self.reqs}, "
def __len__(self):
return len(self.reqs)
class BatchTokenIdOut:
def __init__(self):
self.reqs_infs: List[
Tuple[str, int, Dict, bool, bool]
] = [] # [req_id, new_token_id, gen_metadata, finished_state, abort_state]
class BatchStrOut:
def __init__(self):
self.reqs_infs: List[
Tuple[str, str, Dict, bool, bool]
] = [] # [req_id, token_str, gen_metadata, finished_state, abort_state]
class AbortReq:
def __init__(self, req_id):
self.req_id = req_id
class RequestOutput:
"""The output data of a request to the LLM.
Args:
request_id: The unique ID of the request.
prompt: The prompt string of the request.
prompt_token_ids: The token IDs of the prompt.
outputs: The output sequences of the request.
"""
def __init__(
self,
request_id: str,
prompt: str,
prompt_token_ids: List[int],
outputs,
) -> None:
self.request_id = request_id
self.prompt = prompt
self.prompt_token_ids = prompt_token_ids
self.outputs = outputs
def __repr__(self) -> str:
return (
f"RequestOutput(request_id={self.request_id}, "
f"prompt={self.prompt!r}, "
f"prompt_token_ids={self.prompt_token_ids}, "
f"outputs={self.outputs}, "
)

154
colossalai/legacy/inference/dynamic_batching/ray_dist_init.py Normal file
View File

@@ -0,0 +1,154 @@
import logging
import os
from typing import List
import ray
import ray.util.collective as collective
import torch
from transformers import AutoModelForCausalLM
import colossalai
from colossalai.inference.async_manager import start_dynamic_batching
from colossalai.inference.dynamic_batching.get_tokenizer import get_tokenizer
from colossalai.inference.dynamic_batching.io_struct import RequestOutput
from colossalai.inference.dynamic_batching.ray_init_config import EngineArgsClass, RooterArgsClass
from colossalai.inference.dynamic_batching.sampling_params import SamplingParams
from colossalai.inference.tensor_parallel.engine import TPInferEngine
from colossalai.shardformer import ShardConfig
from colossalai.testing import free_port
ray_serve_logger = logging.getLogger("ray.serve")
def log_cuda_info(scope_name: str):
ray_serve_logger.info(f" {scope_name}: ray.get_gpu_ids(): {ray.get_gpu_ids()}")
ray_serve_logger.info(
f" {scope_name}: CUDA_VISIBLE_DEVICES: {os.getenv('CUDA_VISIBLE_DEVICES', 'NO DEVICES FOUND!')}"
)
if torch.cuda.is_available():
ray_serve_logger.info(
f" {scope_name}: cuda current_device: {torch.cuda.current_device()}, cuda device count: {torch.cuda.device_count()}"
)
else:
ray_serve_logger.info(f" {scope_name}: cuda is not available!")
@ray.remote(num_gpus=1)
class Worker:
def __init__(
self,
model_path: str,
tensor_parallel_size: int,
max_batch_size: int,
max_input_len: int,
max_output_len: int,
router_config: RooterArgsClass,
):
log_cuda_info("Worker.init")
self.tensor_parallel_size = tensor_parallel_size
self.model_path = model_path
self.max_batch_size = max_batch_size
self.max_input_len = max_input_len
self.max_output_len = max_output_len
self.router_config = router_config
def setup(self, world_size, rank, port):
# initialize a ray collective group, otherwise colossalai distributed env won't be built successfully
collective.init_collective_group(world_size, rank, "nccl", "default")
# initialize and set distributed environment
colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
ray_serve_logger.info(f"Worker with rank {rank} (world size {world_size}) setting up..")
log_cuda_info("Worker.setup")
# Load model
self.tokenizer = get_tokenizer(tokenizer_name=self.model_path)
if self.tokenizer.pad_token is None:
self.tokenizer.pad_token = self.tokenizer.eos_token
self.model = AutoModelForCausalLM.from_pretrained(
self.model_path, pad_token_id=self.tokenizer.pad_token_id, torch_dtype=torch.float16
)
shard_config = ShardConfig(
enable_tensor_parallelism=True if world_size > 1 else False, extra_kwargs={"inference_only": True}
)
self.infer_engine = TPInferEngine(
self.model, shard_config, self.max_batch_size, self.max_input_len, self.max_output_len
)
self.start_dynamic_batching = start_dynamic_batching(self.router_config, self.infer_engine, [])
return True
# def generate(self, request_id: str, prompt: str, sampling_params: SamplingParams) -> List[str]:
# ray_serve_logger.info(f"text: {prompt}")
# final_outputs = self.start_dynamic_batching.generate(prompt, sampling_params, request_id)
# return final_outputs
def add_input(self, request_id: str, prompt: str, sampling_params: SamplingParams):
self.start_dynamic_batching.add_input(request_id, prompt, sampling_params)
def abort(self, request_id: str):
self.start_dynamic_batching.abort(request_id)
def step(self) -> List[RequestOutput]:
return self.start_dynamic_batching._step()
def add_req(self, prompt_ids: List[int], sampling_params: SamplingParams, request_id: str, prompt: str):
self.start_dynamic_batching.add_req(prompt_ids, sampling_params, request_id, prompt)
def is_running(self):
return self.start_dynamic_batching.is_running()
class Driver:
def __init__(self, router_config: RooterArgsClass, engine_config: EngineArgsClass):
log_cuda_info("Driver:init")
model_path = engine_config.model
tensor_parallel_size = engine_config.tensor_parallel_size
self.num_workers = tensor_parallel_size
self.workers = []
init_rets = []
# Just grab a free port on localhost
# NOTE workers in this communication group listen to the same port
available_port = free_port()
for i in range(self.num_workers):
worker_name = "worker_idx_{}".format(i)
w = Worker.options(name=worker_name).remote(
model_path,
self.num_workers,
engine_config.max_batch_size,
engine_config.max_input_len,
engine_config.max_output_len,
router_config,
)
self.workers.append(w)
init_rets.append(w.setup.remote(self.num_workers, i, available_port))
_options = {
"group_name": "default_driver",
"world_size": self.num_workers,
"ranks": [i for i in range(self.num_workers)],
"backend": "nccl",
}
collective.create_collective_group(self.workers, **_options)
_ = ray.get(init_rets)
def add_input(self, request_id: str, prompt: str, sampling_params: SamplingParams):
ray.get([w.add_input.remote(request_id, prompt, sampling_params) for w in self.workers])
def abort(self, request_id: str):
ray.get([w.abort.remote(request_id) for w in self.workers])
def step(self):
results = ray.get([w.step.remote() for w in self.workers])
outputs = results[0] # get any one of the copies
return outputs
def add_req(self, request_id: str, prompt_ids: List[int], sampling_params: SamplingParams, prompt: str):
ray.get([w.add_req.remote(prompt_ids, sampling_params, request_id, prompt) for w in self.workers])
def is_running(self):
results = ray.get([w.is_running.remote() for w in self.workers])
return any(results)

58
colossalai/legacy/inference/dynamic_batching/ray_init_config.py Normal file
View File

@@ -0,0 +1,58 @@
import logging
import yaml
from pydantic import BaseModel
logger = logging.getLogger(__name__)
class EngineArgsClass(BaseModel):
"""Config for Engine"""
model: str
tensor_parallel_size: int = 2
max_batch_size: int = 4
max_input_len: int = 128
max_output_len: int = 32
class RooterArgsClass(BaseModel):
"""Config for Rooter"""
max_total_token_num: int = 42
batch_max_tokens: int = 42
eos_id: int = 0
disable_log_stats: bool = False
log_stats_interval: int = 10
model: str
class RayInitConfig(BaseModel):
"""All-together configs without app router config"""
engine_config_data: EngineArgsClass
router_config_data: RooterArgsClass
@classmethod
def from_yaml_path(cls, path: str):
try:
with open(path, "r") as yaml_file:
try:
config = yaml.safe_load(yaml_file)
# serve deployment config
engine_config = config.get("engine_config", {})
router_config = config.get("router_config", {})
return cls(
engine_config_data=engine_config,
router_config_data=router_config,
)
except yaml.YAMLError as e:
logger.error(f"An Error occurred when parsing yaml: {e}")
raise
except FileNotFoundError:
logger.error(f"The file '{path}' does not exist!")
raise
except OSError as e:
logger.error(f"An Error occurred: {e}")
raise

73
colossalai/legacy/inference/dynamic_batching/req_queue.py Normal file
View File

@@ -0,0 +1,73 @@
# Adapted from https://github.com/ModelTC/lightllm
import uuid
from typing import List
import numpy as np
from .io_struct import Batch, Req
class ReqQueue:
def __init__(self, max_total_tokens, batch_max_tokens, running_max_req_size, waiting_req_list=[]) -> None:
self.max_total_tokens = max_total_tokens
assert batch_max_tokens is not None
self.batch_max_tokens = batch_max_tokens
self.running_max_req_size = running_max_req_size
self.waiting_req_list: List[Req] = waiting_req_list
def append(self, req):
self.waiting_req_list.append(req)
return
def _init_cache_list(self, current_batch: Batch):
if current_batch is not None:
self.cache_len_list = [
(req.input_len + len(req.output_ids), req.max_output_len - len(req.output_ids) - 1)
for req in current_batch.reqs
]
else:
self.cache_len_list = []
# @calculate_time(show=True, min_cost_ms=0.1)
def _can_add_new_req(self, req):
self.cache_len_list.append((req.input_len + 1, req.max_output_len - 1)) # hard to analysis
self.cache_len_list.sort(key=lambda x: -x[1])
left_out_len_array = np.array([e[1] for e in self.cache_len_list])
# assert left_out_len_array.min() >= 0
has_run_len_array = np.array([e[0] for e in self.cache_len_list])
cum_run_len_array = np.cumsum(has_run_len_array)
size_array = np.arange(1, len(self.cache_len_list) + 1, 1)
need_max_token_num = (left_out_len_array * size_array + cum_run_len_array).max()
# NOTE: change here < to <=
return need_max_token_num <= self.max_total_tokens and len(self.cache_len_list) <= self.running_max_req_size
def generate_new_batch(self, current_batch: Batch = None):
if current_batch is not None and len(current_batch.reqs) >= self.running_max_req_size:
return None
self._init_cache_list(current_batch)
can_run_list = []
new_batch_total_tokens = 0
aborted_count = 0
for req in self.waiting_req_list:
flag = self._can_add_new_req(req)
if req.aborted:
aborted_count += 1
continue
if flag and new_batch_total_tokens + req.input_len <= self.batch_max_tokens:
can_run_list.append(req)
new_batch_total_tokens += req.input_len
else:
break
if len(can_run_list) != 0:
new_batch = Batch(uuid.uuid4().hex, can_run_list)
self.waiting_req_list = self.waiting_req_list[len(can_run_list) + aborted_count :]
return new_batch
else:
return None
def __len__(self):
return self.waiting_req_list.__len__()

83
colossalai/legacy/inference/dynamic_batching/sampling_params.py Normal file
View File

@@ -0,0 +1,83 @@
# Adapted from https://github.com/ModelTC/lightllm
"""Sampling parameters for text generation."""
from typing import List, Optional, Union
_SAMPLING_EPS = 1e-5
class SamplingParams:
def __init__(
self,
do_sample: bool = False,
presence_penalty: float = 0.0,
frequency_penalty: float = 0.0,
temperature: float = 1.0,
top_p: float = 1.0,
top_k: int = -1, # -1 is for all
ignore_eos: bool = False,
max_new_tokens: int = 256,
stop_sequences: Optional[Union[str, List[str]]] = None, # conditions to stop generation
) -> None:
self.do_sample = do_sample
self.presence_penalty = presence_penalty
self.frequency_penalty = frequency_penalty
self.temperature = temperature
self.top_p = top_p
self.top_k = top_k
self.ignore_eos = ignore_eos
self.max_new_tokens = max_new_tokens
self.stop_sequences = stop_sequences
if self.do_sample == False:
self.temperature = 1.0
self.top_p = 1.0
self.top_k = 1
if (
self.temperature >= 0.0 and self.temperature < _SAMPLING_EPS
): # temperature is too slow, change to greedy search
self.temperature = 1.0
self.top_k = 1
return
def verify(self):
if self.presence_penalty < 0.0:
raise ValueError(f"presence_penalty must >= 0.0, got {self.presence_penalty}")
if self.frequency_penalty < 0.0:
raise ValueError(f"frequency_penalty must >= 0.0, got {self.frequency_penalty}")
if self.temperature <= 0.0:
raise ValueError(f"temperature must > 0.0, got {self.temperature}")
if self.top_p <= 0.0 or self.top_p > 1.0:
raise ValueError(f"top_p must in (0.0, 1.0], got {self.top_p}")
if self.top_k < -1 or self.top_k == 0:
raise ValueError(f"top_k must be -1 (disable), or at least 1, got {self.top_k}.")
if self.max_new_tokens < 1:
raise ValueError(f"max_new_tokens must be at least 1 , got {self.max_new_tokens}.")
return
def stop_sentences_to_token_ids(self, tokenizer):
if self.stop_sequences is None:
self.stop_sequences = []
else:
if isinstance(self.stop_sequences, str):
self.stop_sequences = [self.stop_sequences]
new_stop_sequences = []
for stop_str in self.stop_sequences:
stop_str_ids = tokenizer.encode(stop_str)
if stop_str_ids is not None and len(stop_str_ids) >= 1: # remove bos_token_id
stop_str_ids = stop_str_ids[1:]
if len(stop_str_ids) > 0:
new_stop_sequences.append(stop_str_ids)
self.stop_sequences = new_stop_sequences
return
def to_dict(self):
ret = {}
ret["do_sample"] = self.do_sample
ret["presence_penalty"] = self.presence_penalty
ret["frequency_penalty"] = self.frequency_penalty
ret["temperature"] = self.temperature
ret["top_p"] = self.top_p
ret["top_k"] = self.top_k
# if self.ignore_eos is not None:
# ret["ignore_eos"] = self.ignore_eos
return ret

45
colossalai/legacy/inference/dynamic_batching/stats.py Normal file
View File

@@ -0,0 +1,45 @@
# Adapted from https://github.com/ModelTC/lightllm
import time
class Stats:
def __init__(self, log_status, log_stats_interval) -> None:
self.log_stats = log_status
self.log_stats_interval = log_stats_interval
self.last_log_time = time.time()
self.all_tokens = 0
self.output_tokens = 0
self.prompt_tokens = 0
return
def count_prompt_tokens(self, run_batch):
if self.log_stats:
tokens = run_batch.input_tokens()
self.prompt_tokens += tokens
self.all_tokens += tokens
return
def count_output_tokens(self, run_batch):
if self.log_stats:
tokens = len(run_batch.reqs)
self.output_tokens += tokens
self.all_tokens += tokens
return
def print_stats(self):
if not self.log_stats:
return
now = time.time()
if now - self.last_log_time > self.log_stats_interval:
print(
f"Avg tokens(prompt+generate) throughput: {self.all_tokens/(now-self.last_log_time):8.3f} tokens/s\n"
f"Avg prompt tokens throughput: {self.prompt_tokens/(now-self.last_log_time):8.3f} tokens/s\n"
f"Avg generate tokens throughput: {self.output_tokens/(now-self.last_log_time):8.3f} tokens/s"
)
self.all_tokens = 0
self.output_tokens = 0
self.prompt_tokens = 0
self.last_log_time = now
return