87ebaef4e4
### What this PR does / why we need it? Based on the design of dual-batch overlap proposed by Deepseek team and also the implementation of fused moe in VLLM project, we implement the multi-stream(also known as dual-batch) overlap for deepseek+mla on Ascend NPU. We split the input batch of model into two microbatches and then overlap the comp/comm ops in attention and moe layers using two streams to improve the performance. Our approach can be easily extended when adding dispatch/combine communications for moe layer. Compared with the previously proposed [draft](https://github.com/vllm-project/vllm-ascend/pull/842), we use one stream for computation ops and the other for communication ops, separately. In out opinions, it is beneficial for arranging the order of executing different ops and thus avoiding the contention of computation/communication resources. ref: [overlap for llama](https://github.com/vllm-project/vllm/pull/15787/files) ref: [dbo in sglang](https://github.com/sgl-project/sglang/pull/4068/files#diff-b4937569fc71f6ad215181b633b2f89c7183a2b4ac39e41fc22635599a9be7de) ### Does this PR introduce _any_ user-facing change? Adding an env variable "VLLM_ENABLE_DBO". Users can enable dbo by setting "VLLM_ASCEND_ENABLE_DBO=1" See /examples/offline_dualbatch_overlap_npu.py for more info. ### How was this patch tested? This patch can be tested with vllm-0.9.0 using its online service with benchmark tests. We have decoupled the func of dbo from vllm and it should be able to run without any modification to the code of vllm(some modifications is better to implement in vllm though). Any advice/discussion is welcome. ### Performance Benchmark We have ran the benchmark_serving script of vllm to test the performance after using dual-batch overlap. `python -m vllm.entrypoints.openai.api_server \ --model=DeepSeek-R1-W8A8 \ --trust-remote-code \ --distributed-executor-backend=mp \ -tp=16 \ --port 8006 \ --max-num-seqs 390 \ --max-model-len 32768 \ --max-num-batched-tokens 65536 \ --block-size 128 \ --compilation_config 0 \ --gpu-memory-utilization 0.90 \ --disable-log-requests \ --additional-config '{"expert_tensor_parallel_size":1,"enable_inter_dp_scheduling":true,"init_torchair_graph_batch_sizes":true,"trace_recompiles":true,"ascend_scheduler_config":{},"enable_graph_mode":false}'` and run benchmark with the parameters of : `--dataset-name random --random-input-len 4096 --random-output-len 1 --num-prompts 200 --max-concurrency 8 --request-rate 5 --metric-percentiles 90` 1. test with the version using allgather+allreduce in Ascend 910B (tp16 ep16 + deepseek r1 w8a8) 2. test with the version using alltoall: prefill qps: 0.90 -> 1.01 Mean TTFT：8226->7432ms The overlap approach when using alltoall communication can be further optimized by overlapping micro-batch1's moe comp with micro-batch2's dispatch a2a comm --------- Signed-off-by: zhuohuan <zxdu1997@gmail.com>
98 lines
3.3 KiB
Python
98 lines
3.3 KiB
Python
#
|
|
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
|
|
# Copyright 2023 The vLLM team.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
# This file is a part of the vllm-ascend project.
|
|
# Adapted from vllm/tests/basic_correctness/test_basic_correctness.py
|
|
#
|
|
"""Compare the short outputs of HF and vLLM when using greedy sampling.
|
|
|
|
Run `pytest tests/test_offline_inference.py`.
|
|
"""
|
|
import os
|
|
from unittest.mock import patch
|
|
|
|
import vllm # noqa: F401
|
|
from vllm import SamplingParams
|
|
|
|
from tests.conftest import VllmRunner
|
|
|
|
os.environ["PYTORCH_NPU_ALLOC_CONF"] = "max_split_size_mb:256"
|
|
|
|
|
|
def test_models_distributed_QwQ():
|
|
example_prompts = [
|
|
"Hello, my name is",
|
|
]
|
|
dtype = "half"
|
|
max_tokens = 5
|
|
with VllmRunner(
|
|
"Qwen/QwQ-32B",
|
|
dtype=dtype,
|
|
tensor_parallel_size=4,
|
|
distributed_executor_backend="mp",
|
|
) as vllm_model:
|
|
vllm_model.generate_greedy(example_prompts, max_tokens)
|
|
|
|
|
|
def test_models_distributed_DeepSeek():
|
|
example_prompts = [
|
|
"Hello, my name is",
|
|
]
|
|
dtype = "half"
|
|
max_tokens = 5
|
|
with VllmRunner(
|
|
"deepseek-ai/DeepSeek-V2-Lite",
|
|
dtype=dtype,
|
|
tensor_parallel_size=4,
|
|
distributed_executor_backend="mp",
|
|
) as vllm_model:
|
|
vllm_model.generate_greedy(example_prompts, max_tokens)
|
|
|
|
|
|
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_TOPK_OPTIMZE": "1"})
|
|
def test_models_distributed_topk() -> None:
|
|
example_prompts = [
|
|
"vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs.",
|
|
"Briefly describe the major milestones in the development of artificial intelligence from 1950 to 2020.",
|
|
"Compare and contrast artificial intelligence with human intelligence in terms of processing information.",
|
|
]
|
|
dtype = "half"
|
|
sampling_params = SamplingParams(max_tokens=5,
|
|
temperature=0.0,
|
|
top_k=50,
|
|
top_p=0.9)
|
|
|
|
with VllmRunner(
|
|
"deepseek-ai/DeepSeek-V2-Lite",
|
|
dtype=dtype,
|
|
tensor_parallel_size=4,
|
|
distributed_executor_backend="mp",
|
|
) as vllm_model:
|
|
vllm_model.generate(example_prompts, sampling_params)
|
|
|
|
|
|
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_DBO": "1"})
|
|
def test_models_distributed_DeepSeek_dbo():
|
|
example_prompts = ["The president of the United States is"] * 41
|
|
dtype = "half"
|
|
sampling_params = SamplingParams(max_tokens=100, temperature=0.0)
|
|
with VllmRunner(
|
|
"deepseek-ai/DeepSeek-V2-Lite",
|
|
dtype=dtype,
|
|
tensor_parallel_size=4,
|
|
distributed_executor_backend="mp",
|
|
) as vllm_model:
|
|
vllm_model.generate(example_prompts, sampling_params)
|