pytorch/builder

Investigate size increase in pytorch pypi binary

atalman opened this issue · 2 comments

ptrblck commented

I see a reduction between 1.13.1 to 2.0.0:

Collecting torch==1.13.1
  Using cached torch-1.13.1-cp310-cp310-manylinux1_x86_64.whl (887.5 MB)
Collecting nvidia-cuda-nvrtc-cu11==11.7.99
  Using cached nvidia_cuda_nvrtc_cu11-11.7.99-2-py3-none-manylinux1_x86_64.whl (21.0 MB)
Collecting typing-extensions
  Using cached typing_extensions-4.5.0-py3-none-any.whl (27 kB)
Collecting nvidia-cuda-runtime-cu11==11.7.99
  Using cached nvidia_cuda_runtime_cu11-11.7.99-py3-none-manylinux1_x86_64.whl (849 kB)
Collecting nvidia-cudnn-cu11==8.5.0.96
  Using cached nvidia_cudnn_cu11-8.5.0.96-2-py3-none-manylinux1_x86_64.whl (557.1 MB)
Collecting nvidia-cublas-cu11==11.10.3.66
  Using cached nvidia_cublas_cu11-11.10.3.66-py3-none-manylinux1_x86_64.whl (317.1 MB)
Collecting setuptools
  Using cached setuptools-67.6.1-py3-none-any.whl (1.1 MB)
Collecting wheel
  Using cached wheel-0.40.0-py3-none-any.whl (64 kB)

vs.

Collecting torch==2.0.0
  Using cached torch-2.0.0-cp310-cp310-manylinux1_x86_64.whl (619.9 MB)
Collecting jinja2
  Using cached Jinja2-3.1.2-py3-none-any.whl (133 kB)
Collecting nvidia-cusolver-cu11==11.4.0.1
  Using cached nvidia_cusolver_cu11-11.4.0.1-2-py3-none-manylinux1_x86_64.whl (102.6 MB)
Collecting nvidia-cuda-runtime-cu11==11.7.99
  Using cached nvidia_cuda_runtime_cu11-11.7.99-py3-none-manylinux1_x86_64.whl (849 kB)
Collecting triton==2.0.0
  Using cached triton-2.0.0-1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl (63.3 MB)
Collecting sympy
  Using cached sympy-1.11.1-py3-none-any.whl (6.5 MB)
Collecting nvidia-cuda-cupti-cu11==11.7.101
  Using cached nvidia_cuda_cupti_cu11-11.7.101-py3-none-manylinux1_x86_64.whl (11.8 MB)
Collecting nvidia-cufft-cu11==10.9.0.58
  Using cached nvidia_cufft_cu11-10.9.0.58-py3-none-manylinux1_x86_64.whl (168.4 MB)
Collecting nvidia-curand-cu11==10.2.10.91
  Using cached nvidia_curand_cu11-10.2.10.91-py3-none-manylinux1_x86_64.whl (54.6 MB)
Collecting filelock
  Using cached filelock-3.10.7-py3-none-any.whl (10 kB)
Collecting networkx
  Using cached networkx-3.0-py3-none-any.whl (2.0 MB)
Collecting nvidia-cuda-nvrtc-cu11==11.7.99
  Using cached nvidia_cuda_nvrtc_cu11-11.7.99-2-py3-none-manylinux1_x86_64.whl (21.0 MB)
Collecting nvidia-nccl-cu11==2.14.3
  Using cached nvidia_nccl_cu11-2.14.3-py3-none-manylinux1_x86_64.whl (177.1 MB)
Collecting nvidia-nvtx-cu11==11.7.91
  Using cached nvidia_nvtx_cu11-11.7.91-py3-none-manylinux1_x86_64.whl (98 kB)
Collecting nvidia-cudnn-cu11==8.5.0.96
  Using cached nvidia_cudnn_cu11-8.5.0.96-2-py3-none-manylinux1_x86_64.whl (557.1 MB)
Collecting typing-extensions
  Using cached typing_extensions-4.5.0-py3-none-any.whl (27 kB)
Collecting nvidia-cusparse-cu11==11.7.4.91
  Using cached nvidia_cusparse_cu11-11.7.4.91-py3-none-manylinux1_x86_64.whl (173.2 MB)
Collecting nvidia-cublas-cu11==11.10.3.66
  Using cached nvidia_cublas_cu11-11.10.3.66-py3-none-manylinux1_x86_64.whl (317.1 MB)
Collecting setuptools
  Using cached setuptools-67.6.1-py3-none-any.whl (1.1 MB)
Collecting wheel
  Using cached wheel-0.40.0-py3-none-any.whl (64 kB)
Collecting lit
  Using cached lit-16.0.0.tar.gz (144 kB)
  Preparing metadata (setup.py) ... done
Collecting cmake
  Using cached cmake-3.26.1-py2.py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl (24.0 MB)
Collecting MarkupSafe>=2.0
  Using cached MarkupSafe-2.1.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (25 kB)
Collecting mpmath>=0.19
  Using cached mpmath-1.3.0-py3-none-any.whl (536 kB)

in the actual torch wheel as more CUDA pip wheel dependencies were added. Is the total size the concern here?

msaroufim commented

Yeah so the core problem was when upgrading from torchserve 0.6.1 to 0.7 we upgraded CUDA versions to 11 and torch to 1.13 and that almost doubled our docker image size and one of our customers was concerned about this

Check out layer 53 for the problem

I also checked for Ubuntu package sizes and indeed CUDA seems to be a big contributor to this increase (about 800MB difference from CUDA 10.2 to 11.7)

And compared all the pip package size differences (a 1.6GB difference for torch)

I'm not sure how pip computes download sizes but my script is doing this https://gist.github.com/msaroufim/098c0478bd2c629312acaa59b535fa9c#file-download_size_torch-py-L16

image

cc @agunapal