lmdeploy serve is more than two times slower than normal transformers code

[paniabhisek]

I am using phi-3-vision model. When I run the example snippet from huggingface, it takes around 5 seconds for the response.

However if I serve the model using lmdeploy serve api_server microsoft/Phi-3-vision-128k-instruct --server-port 23333. Then use the following code to get response

import requests
import base64
import time

def encode_image(image_path):
  with open(image_path, "rb") as image_file:
    return base64.b64encode(image_file.read()).decode('utf-8')

image_path = "wisconsin-madison.jpg"
base64_image = encode_image(image_path)

start = time.time()

r = requests.post('http://<ip>:23333/v1/chat/completions', json={
                             "model": "/home/useradmin/.cache/huggingface/hub/models--microsoft--Phi-3-vision-128k-instruct/snapshots/6065b7a1a412feff7ac023149f65358b71334984",
                             "messages": [
                                 {
                                     "role": "user",
                                     "content": [
                                         {
                                             "type": "text",
                                             "text": "What is in this image?"
                                         },
                                         {
                                             "type": "image_url",
                                             "image_url": {
                                                 "url": f"data:image/jpeg;base64,{base64_image}"
                                             }
                                         }
                                     ]
                                 }
                             ],
                             "temperature": 0.7,
                             "top_p": 1,
                             "tools": None,
                             "tool_choice": "none",
                             "logprobs": False,
                             "top_logprobs": 0,
                             "n": 1,
                             "max_tokens": None,
                             "stop": None,
                             "stream": False,
                             "presence_penalty": 0,
                             "frequency_penalty": 0,
                             "user": "string",
                             "repetition_penalty": 1,
                             "session_id": -1,
                             "ignore_eos": False,
                             "skip_special_tokens": True,
                             "top_k": 40
                         })

end = time.time()

It takes around 10 to 12 seconds. Why is there so much delay.
In both of the cases, the same image and query is being used. And I have disabled flash_attention and using eager mode in the config.json

[lvhan028]

@RunningLeon

[RunningLeon]

@paniabhisek hi, pls post your env info by running lmdeploy check_env. BTW, can you post what you changed in config.json? For image embedding part, it surely needs flash attn as in here https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/blob/6065b7a1a412feff7ac023149f65358b71334984/image_embedding_phi3_v.py#L24

[paniabhisek]

sys.platform: linux
Python: 3.10.12 (main, Jul 19 2024, 14:39:24) [GCC 11.4.0]
CUDA available: True
MUSA available: False
numpy_random_seed: 2147483648
GPU 0,1,2,3: Tesla T4
CUDA_HOME: /usr/local/cuda
NVCC: Cuda compilation tools, release 12.4, V12.4.99
GCC: gcc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
PyTorch: 2.2.2+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 v3.3.2 (Git Hash 2dc95a2ad0841e29db8b22fbccaf3e5da7992b01)
  - OpenMP 201511 (a.k.a. OpenMP 4.5)
  - LAPACK is enabled (usually provided by MKL)
  - NNPACK is enabled
  - CPU capability usage: AVX2
  - 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 8.9.2
  - Magma 2.6.1
  - Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=12.1, CUDNN_VERSION=8.9.2, 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_QNNPACK -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=2.2.2, USE_CUDA=ON, USE_CUDNN=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, 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: 0.17.2+cu121
LMDeploy: 0.5.2.post1+
transformers: 4.42.4
gradio: Not Found
fastapi: 0.111.1
pydantic: 2.8.2
triton: 2.1.0
NVIDIA Topology:
        GPU0    GPU1    GPU2    GPU3    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      SYS     SYS     SYS     16-31   1               N/A
GPU1    SYS      X      SYS     SYS     16-31   1               N/A
GPU2    SYS     SYS      X      SYS     32-47   2               N/A
GPU3    SYS     SYS     SYS      X      32-47   2               N/A

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

[paniabhisek]

This is the change.

I am also disabling flash_attention in transformer code too which gives 5 seconds.

[RunningLeon]

hi, changing "_attn_implementation" will not affect the speed. Are you using --tp > 1 ? Is T4+16G able to load the model? Anyway, we don't have T4, but this is what I got on A100+80G. It seems reasonable.

5.06273889541626
0.6262824535369873
0.9383771419525146
0.7238662242889404
0.7138667106628418
0.6676077842712402
0.7242634296417236
0.8624002933502197

[paniabhisek]

It's T4+64G with each 16G. I was only using one card previously. Now when I ran using --tp 4, I got

Does that mean if I use more cards, the inference time will decrease? I do not understand what is happening here.