[Bug] Use triton to deploy minicpm-v-2_6 GPU memory keeps increasing until it overflows

[LinJianping]

Checklist

• 1. I have searched related issues but cannot get the expected help.
• 2. The bug has not been fixed in the latest version.
• 3. Please note that if the bug-related issue you submitted lacks corresponding environment info and a minimal reproducible demo, it will be challenging for us to reproduce and resolve the issue, reducing the likelihood of receiving feedback.

Describe the bug

When I use lmdeploy to deploy minicpm-V-2_6 in triton, I find that the GPU memory keeps increasing until an exception occurs due to GPU memory issues, but this does not happen when I using a python script for loop inference. I experimented with the official model weights（https://huggingface.co/openbmb/MiniCPM-V-2_6） and still had the same problem. In order to align with the training environment, Torch 2.4.0 was used instead of Torch 2.3.1, which is compatible with lmdeploy. I don't know if this is the reason. When I use torch.cuda.empty_cache() to free unused GPU memory in each request, I can solve this problem, but it will cause the inference time to increase.

I have downgrade the Torch version to 2.3.1, the problem still exists.

Reproduction

The triton service startup script is as follows：

#!/bin/bash
cd ~
export OMP_NUM_THREADS=1
tritonserver --log-dir ./serving-logs \
             --model-control-mode "explicit" \
             --load-model minicpmv_2_6_awq_4bit \
             --whale-rpc-port 26381 \
             --grpc-port 26384 \
             --mport 26382 \
             --whale-rpc-use-async 1 \
             --log-info false \
             --allow-grpc true \
             --allow-whale-rpc true \
             --allow-metrics 0 \
             --model-repository /xx/model_repos/minicpmv_2_6_awq_4bit/2 \
             --exit-timeout-secs 0 \
             --log-verbose 0 \
             --server-appkey  xxxx

The python backend inference script is as follows:

import os
import numpy as np
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:32"
import triton_python_backend_utils as pb_utils
import json
import time
import torch

from lmdeploy import pipeline, TurbomindEngineConfig
from lmdeploy.vl import load_image


class TritonPythonModel:
    """Your Python model must use the same class name. Every Python model
    that is created must have "TritonPythonModel" as the class name.
    """

    def initialize(self, args):
        """`initialize` is called only once when the model is being loaded.
        Implementing `initialize` function is optional. This function allows
        the model to intialize any state associated with this model.

        Parameters
        ----------
        args : dict
          Both keys and values are strings. The dictionary keys and values are:
          * model_config: A JSON string containing the model configuration
          * model_instance_kind: A string containing model instance kind
          * model_instance_device_id: A string containing model instance device ID
          * model_repository: Model repository path
          * model_version: Model version
          * model_name: Model name
        """
        # You must parse model_config. JSON string is not parsed here
        # print("args:", args)
        pwd = os.path.abspath(os.path.dirname(__file__))
        self.ckpt_path = f'{pwd}/MiniCPM-V-2_6/'
        print('self.ckpt_path', self.ckpt_path)
        engine_config = TurbomindEngineConfig(session_len=8192, max_batch_size=1)
        self.pipeline = pipeline(self.ckpt_path, backend_config=engine_config)
        print(os.environ['LD_LIBRARY_PATH'])

    def execute(self, requests):
        """`execute` must be implemented in every Python model. `execute`
        function receives a list of pb_utils.InferenceRequest as the only
        argument. This function is called when an inference request is made
        for this model. Depending on the batching configuration (e.g. Dynamic
        Batching) used, `requests` may contain multiple requests. Every
        Python model, must create one pb_utils.InferenceResponse for every
        pb_utils.InferenceRequest in `requests`. If there is an error, you can
        set the error argument when creating a pb_utils.InferenceResponse

        Parameters
        ----------
        requests : list
          A list of pb_utils.InferenceRequest

        Returns
        -------
        list
          A list of pb_utils.InferenceResponse. The length of this list must
          be the same as `requests`
        """
        responses = []

        # 后处理模型输出，并为每个请求创建响应
        for request in requests:
            query = pb_utils.get_input_tensor_by_name(request, "query")
            query = query.as_numpy()

            image = pb_utils.get_input_tensor_by_name(request, "image")
            image = image.as_numpy()

            result = self.pipeline((query[0][0].decode('utf-8'), image[0][0].decode('utf-8')))

            output = np.array(result.text.encode('utf-8'),  dtype=np.string_)

            out_tensor_0 = pb_utils.Tensor("response", output.astype(np.string_))
            response = pb_utils.InferenceResponse(output_tensors=[out_tensor_0])
            responses.append(response)


        # print("len req:", len(requests))
        # print("len resp:", len(responses))
        #torch.cuda.empty_cache()

        return responses

    def finalize(self):
        """`finalize` is called only once when the model is being unloaded.
        Implementing `finalize` function is OPTIONAL. This function allows
        the model to perform any necessary clean ups before exit.
        """
        print('Cleaning up...')

Environment

sys.platform: linux
Python: 3.9.16 (main, Apr  2 2024, 20:40:25) [GCC 10.2.1 20210130 (Red Hat 10.2.1-11)]
CUDA available: True
MUSA available: False
numpy_random_seed: 2147483648
GPU 0: NVIDIA L40
CUDA_HOME: /usr/local/cuda
NVCC: Cuda compilation tools, release 12.2, V12.2.91
GCC: gcc (GCC) 4.8.5 20150623 (Red Hat 4.8.5-44)
PyTorch: 2.4.0+cu121
PyTorch compiling details: PyTorch built with:
  - GCC 9.3
  - C++ Version: 201703
  - Intel(R) oneAPI Math Kernel Library Version 2022.2-Product Build 20220804 for Intel(R) 64 architecture applications
  - Intel(R) MKL-DNN v3.4.2 (Git Hash 1137e04ec0b5251ca2b4400a4fd3c667ce843d67)
  - OpenMP 201511 (a.k.a. OpenMP 4.5)
  - LAPACK is enabled (usually provided by MKL)
  - NNPACK is enabled
  - CPU capability usage: AVX512
  - CUDA Runtime 12.1
  - NVCC architecture flags: -gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_86,code=sm_86;-gencode;arch=compute_90,code=sm_90
  - CuDNN 90.1  (built against CUDA 12.4)
  - Magma 2.6.1
  - Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=12.1, CUDNN_VERSION=9.1.0, CXX_COMPILER=/opt/rh/devtoolset-9/root/usr/bin/c++, CXX_FLAGS= -D_GLIBCXX_USE_CXX11_ABI=0 -fabi-version=11 -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -DNDEBUG -DUSE_KINETO -DLIBKINETO_NOROCTRACER -DUSE_FBGEMM -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -O2 -fPIC -Wall -Wextra -Werror=return-type -Werror=non-virtual-dtor -Werror=bool-operation -Wnarrowing -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-unused-parameter -Wno-unused-function -Wno-unused-result -Wno-strict-overflow -Wno-strict-aliasing -Wno-stringop-overflow -Wsuggest-override -Wno-psabi -Wno-error=pedantic -Wno-error=old-style-cast -Wno-missing-braces -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Wno-stringop-overflow, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_VERSION=2.4.0, USE_CUDA=ON, USE_CUDNN=ON, USE_CUSPARSELT=1, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_GLOO=ON, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=1, USE_NNPACK=ON, USE_OPENMP=ON, USE_ROCM=OFF, USE_ROCM_KERNEL_ASSERT=OFF,

TorchVision: 0.19.0+cu121
LMDeploy: 0.6.1+
transformers: 4.45.2
gradio: Not Found
fastapi: 0.115.0
pydantic: 2.9.2
triton: 3.0.0
NVIDIA Topology:
	GPU0	NIC0	NIC1	CPU Affinity	NUMA Affinity	GPU NUMA ID
GPU0	 X 	SYS	SYS				N/A
NIC0	SYS	 X 	SYS
NIC1	SYS	SYS	 X

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

NIC Legend:

  NIC0: mlx5_bond_0
  NIC1: mlx5_bond_1

Error traceback

terminate called after throwing an instance of 'std::runtime_error'
  what():  [TM][ERROR] CUDA runtime error: out of memory /lmdeploy/src/turbomind/utils/allocator.h:246

[RunningLeon]

hi seems you are using quantized model /xx/model_repos/minicpmv_2_6_awq_4bit, you should add model_format='awq' to TurbomindEngineConfig, see here. Could you try again? If the issue still exists, could you provide a dockerfile to reproduce it ? thanks.

--model-repository /xx/model_repos/minicpmv_2_6_awq_4bit/2
...
self.ckpt_path = f'{pwd}/MiniCPM-V-2_6/'
print('self.ckpt_path', self.ckpt_path)
engine_config = TurbomindEngineConfig(session_len=8192, max_batch_size=1)
self.pipeline = pipeline(self.ckpt_path, backend_config=engine_config)