conda packaging for holoscan latest version (3.x) #28598

billysuh7 · 2024-12-16T16:09:45Z

Checklist

The following dependencies have been whitelisted to have the build going, but we need to provide them (or call them out as not supported) in conda. Only the checked items may be actionable.

"libcuda.so" # driver
"libgdrapi.so" # part of gdrcopy (https://github.com/NVIDIA/gdrcopy) which operates as kernel module
"libnvidia-ml.so" # driver
"libnvinfer_plugin.so" # NVIDIA's TensorRT (https://github.com/NVIDIA/TensorRT), not on conda-forge
"libnvonnxparser.so" # NVIDIA's TensorRT (https://github.com/NVIDIA/TensorRT), not on conda-forge
"*libonnxruntime.so.1.18.1" # onnxruntime conda package is there, but this particular version does not have this DSO. Higher version onnruntime package has this lib, but holoscan must use that.
"libtorchvision.so" # not available even in torchvision package in conda-forge. The github repo (https://github.com/pytorch/vision) discourages using libtorchvision.so from C++ code due to instability, in favor of Python APIs. Conda does provide pytorch, which torchvision depends on though.
"libv4l2.so" # a wrapper for video4linux2, Video for Linux 2, which is a kernel framework on the Linux operating system designed to support video device. http://git.linuxtv.org/v4l-utils.git
"libxpmem.so" # XPMEM, Linux Cross-Memory Attach kernel module, not available

Also, this package does not depend on MOFED because there is no MOFED available as conda package. Ref on dependencies: https://docs.nvidia.com/holoscan/sdk-user-guide/sdk_installation.html

Xref: #28623
This PR still very much WIP.

github-actions · 2024-12-16T16:11:07Z

Hi! This is the staged-recipes linter and your PR looks excellent! 🚀

conda-forge-admin · 2024-12-16T16:11:39Z

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipes/holoscan/meta.yaml) and found it was in an excellent condition.

I do have some suggestions for making it better though...

For recipes/holoscan/meta.yaml:

ℹ️ The recipe is not parsable by parser conda-souschef (grayskull). This parser is not currently used by conda-forge, but may be in the future. We are collecting information to see which recipes are compatible with grayskull.

_{This message was generated by GitHub Actions workflow run https://github.com/conda-forge/conda-forge-webservices/actions/runs/13403187480. Examine the logs at this URL for more detail.}

isuruf · 2024-12-17T16:49:02Z

Is this the same as https://github.com/nvidia-holoscan/holoscan-sdk? If so, why re-package it when it is open source?

billysuh7 · 2024-12-18T16:20:26Z

Is this the same as https://github.com/nvidia-holoscan/holoscan-sdk? If so, why re-package it when it is open source?

Hi @isuruf, yes holoscan-sdk is the same thing. The main reason for repackaging being difficulty associated with building from source in terms of meeting build dependencies. In the recipe we have white-listed a few dependencies that are kernel modules (libxpmem.so, libv4l2.so), cuda drivers (libcuda.so, libnvidia-ml.so) for which there is no conda package and no docker image prebuilt w/ them AFAIK.

Even with binary repackaging we have other hurdles to overcome before holoscan package can be released: conda packages that happen to be missing specific DSO files (libgdrapi.so missing in gds-tools, libonnxruntime.so.1.18.1 missing in onnxruntime 1.18.1, libtorchvision.so missing in torchvision), and then there are DSOs from a product that we don't have (yet)in conda-forge (TensorRT).

For gds-tools, onnxruntime and torchvision I will see when we can get respective owners to provide missing DSO, and for TensorRT we'll have to find out a plan for it with respect to conda.

… order to easily exclude host/runtime dependencies from getting into the package

… by python

github-actions · 2025-02-17T06:59:43Z

Hi! This is the staged-recipes linter and your PR looks excellent but I have some suggestions.

File-specific lints and/or hints:

recipes/holoscan/meta.yaml:
- hints:
  - It looks like you are submitting a multi-output recipe. In these cases, the correct name for the feedstock is ambiguous, and our infrastructure defaults to the top-level package.name field. Please add a feedstock-name entry in the extra section.

github-actions · 2025-02-19T00:29:41Z

Hi! This is the staged-recipes linter and your PR looks excellent! 🚀

billysuh7 · 2025-02-19T01:02:20Z

Please ignore "XYZ in in requirements/run but it is not used" => the indicated packages are ones that holoscan team wants to declare as explicit dependencies. They are:
libnpp-dev, libcusolver, libcublas, libnvjitlink, nccl, libcusparse, libcufft, cuda-nvtx, cudnn and libcurand

carterbox · 2025-02-19T18:01:37Z

recipes/holoscan/build.sh

+cp -rv examples $PREFIX/
+cp -rv lib $PREFIX/
+cp -rv include $PREFIX/
+cp -rv python $PREFIX/


Why are the pythons unpacked here instead of into site-package?

carterbox · 2025-02-19T18:25:25Z

recipes/holoscan/meta.yaml

@@ -0,0 +1,362 @@
+{% set version = "3.0.0.1" %}
+{% set cuda_version = "12.6" %} # min version holoscan works with


If this is the minimum version, then why do you implement CEC such that this package can be installed with CTK 12.0 packages?

carterbox · 2025-02-19T18:31:59Z

recipes/holoscan/compile_examples.sh

+cmake -S $PREFIX/examples \
+  -DCMAKE_LIBRARY_PATH=${GCC_DIR} \
+  -DCMAKE_C_COMPILER=$CC \
+  -DCMAKE_CXX_COMPILER=$CXX \
+  -DCUDAToolkit_INCLUDE_DIRECTORIES="$PREFIX/include;$PREFIX/${targetsDir}/include" \
+  -DCMAKE_PREFIX_PATH=$CONDA_PREFIX \
+  -DCMAKE_CUDA_COMPILER=$PREFIX/bin/nvcc \
+  -DCUDAToolkit_ROOT=$PREFIX


Did you try using $CMAKE_ARGS instead of manually setting all of these variables?

Also, this test takes a long time to build all the examples. Something shorter like

cmake -S $PREFIX/examples/hello_world/cpp $CMAKE_ARGS

might be more appropriate.

carterbox · 2025-02-19T18:36:22Z

recipes/holoscan/meta.yaml

+      run:
+        - {{ pin_subpackage("libholoscan-dev", exact=True) }}


Pin in both host and run when using exact pinnings, so that the packages are built in the correct order.

carterbox · 2025-02-19T18:36:36Z

recipes/holoscan/meta.yaml

+        - cuda-version {{ cuda_compiler_version }}  # [cuda_compiler_version != "None"]
+      run:
+        - {{ pin_compatible('cuda-version', max_pin='x', min_pin='x') }}  # [cuda_compiler_version != "None"]
+        - {{ pin_subpackage("libholoscan" ~ soname, exact=True) }}


Pin in both host and run when using exact pinnings.

carterbox · 2025-02-19T18:39:08Z

recipes/holoscan/meta.yaml

+        - python =3.10
+        - rdma-core


Can we split the python artifacts into a separate output or are they part of the core functionality even when you are only using the C-API? Python is a large dependency for programs that don't use it.

carterbox · 2025-02-19T18:48:48Z

recipes/holoscan/meta.yaml

+        - test -L $PREFIX/lib/libucm.so
+        - test -L $PREFIX/lib/libucp.so
+        - test -L $PREFIX/lib/libucs.so
+        - test -L $PREFIX/lib/libucs_signal.so
+        - test -L $PREFIX/lib/libuct.so
+        - test -L $PREFIX/lib/libyaml-cpp.so


Are these third-party libraries that are being revended?

carterbox · 2025-02-19T18:50:43Z

recipes/holoscan/meta.yaml

+        - test -f $PREFIX/bin/video_validation.py
+        - test -d $PREFIX/lib/cmake/holoscan
+        - test -d $PREFIX/examples
+        - test -d $PREFIX/include/3rdparty


We shouldn't be revending third party headers

carterbox · 2025-02-19T18:57:36Z

recipes/holoscan/meta.yaml

+        - test -L $PREFIX/lib/ucx/libuct_xpmem.so
+        - test -L $PREFIX/lib/ucx/libucx_perftest_cuda.so
+
+  - name: libholoscan-static


The static library packages are meant to separate out optional static libraries when users have the option of linking to dynamic instead. For example, cudart vs cudart_static. In this case, this package only contains libholoviz_imgui.a and there is not an alternative shared library option? So a separate static library package is not needed.

carterbox · 2025-02-19T19:06:24Z

recipes/holoscan/build.sh

+check-glibc bin/* lib/* lib/ucx/* lib/gxf_extensions/*
+find python/ -name "*.so*" | xargs -I"{}" check-glibc "{}"
+
+cp -rv bin $PREFIX/


Why are we unpacking a redist instead of building from source when the source is available?

https://github.com/nvidia-holoscan/holoscan-sdk

I see that we already discussed this above back in December 2024.

For gds-tools, onnxruntime and torchvision I will see when we can get respective owners to provide missing DSO, and for TensorRT we'll have to find out a plan for it with respect to conda.

Have you opened issues on the feedstocks for these packages asking for the relevant libraries?

conda-forge-admin · 2025-02-22T08:02:39Z

Hi! This is the friendly automated conda-forge-linting service.

I wanted to let you know that I linted all conda-recipes in your PR (recipes/holoscan/meta.yaml) and found some lint.

Here's what I've got...

For recipes/holoscan/meta.yaml:

❌ This recipe is using a compiler, which now requires adding a build dependence on {{ stdlib("c") }} as well. Note that this rule applies to each output of the recipe using a compiler. For further details, please see META: {{ stdlib("c") }} migration conda-forge.github.io#2102.

For recipes/holoscan/meta.yaml:

ℹ️ The recipe is not parsable by parser conda-souschef (grayskull). This parser is not currently used by conda-forge, but may be in the future. We are collecting information to see which recipes are compatible with grayskull.

_{This message was generated by GitHub Actions workflow run https://github.com/conda-forge/conda-forge-webservices/actions/runs/13471290947. Examine the logs at this URL for more detail.}

conda-forge-admin · 2025-02-22T18:15:26Z

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipes/holoscan/meta.yaml) and found it was in an excellent condition.

I do have some suggestions for making it better though...

For recipes/holoscan/meta.yaml:

ℹ️ The recipe is not parsable by parser conda-souschef (grayskull). This parser is not currently used by conda-forge, but may be in the future. We are collecting information to see which recipes are compatible with grayskull.

_{This message was generated by GitHub Actions workflow run https://github.com/conda-forge/conda-forge-webservices/actions/runs/13514215908. Examine the logs at this URL for more detail.}

…python -> holoscan 2) libholoscan3 -> libholoscan

holoscan 2.6.0.1

620cf20

billysuh7 and others added 4 commits December 18, 2024 16:59

do not split package under output section; use "monolithic" recipe in…

e96cce5

… order to easily exclude host/runtime dependencies from getting into the package

add cuda-runtime, and take out many misc dependencies as a result

4feb1a1

Add $PREFIX/python/lib to the sys.path so that holoscan can be loaded…

937bfdb

… by python

add more dependencies

c50ca66

billysuh7 changed the title ~~holoscan 2.6.0.1~~ conda packaging for holoscan latest version (3.x) Feb 5, 2025

billysuh7 added 2 commits February 14, 2025 03:32

3.0.0.1

0a64294

split into base, dev and static packages

6acc840

billysuh7 added 2 commits February 18, 2025 14:54

update source tarball and add cuda-nvrtc-dev

7504e6c

remove lint by specifying feedstock name in the multi-output recipe

90f6a0b

take out _openmp_mutex, not needed as dependency

2a4e1c2

carterbox requested changes Feb 19, 2025

View reviewed changes

Updated per review comment and updated source tarball

825fb1a

billysuh7 added 2 commits February 22, 2025 00:34

be explicit about python version

b54e588

remove lint by adding stdlib for the compilers

0f10a50

change of package names per request by holoscan team. 1) libholoscan-…

d7d8a3c

…python -> holoscan 2) libholoscan3 -> libholoscan

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

conda packaging for holoscan latest version (3.x) #28598

conda packaging for holoscan latest version (3.x) #28598

billysuh7 commented Dec 16, 2024 •

edited

Loading

github-actions bot commented Dec 16, 2024

conda-forge-admin commented Dec 16, 2024 •

edited

Loading

isuruf commented Dec 17, 2024

billysuh7 commented Dec 18, 2024 •

edited

Loading

github-actions bot commented Feb 17, 2025

github-actions bot commented Feb 19, 2025

billysuh7 commented Feb 19, 2025

carterbox Feb 19, 2025

carterbox Feb 19, 2025

carterbox Feb 19, 2025

carterbox Feb 19, 2025

carterbox Feb 19, 2025

carterbox Feb 19, 2025

carterbox Feb 19, 2025

carterbox Feb 19, 2025

carterbox Feb 19, 2025

carterbox Feb 19, 2025

carterbox Feb 19, 2025

carterbox Feb 19, 2025

conda-forge-admin commented Feb 22, 2025 •

edited

Loading

conda-forge-admin commented Feb 22, 2025 •

edited

Loading

		@@ -0,0 +1,362 @@
		{% set version = "3.0.0.1" %}
		{% set cuda_version = "12.6" %} # min version holoscan works with

conda packaging for holoscan latest version (3.x) #28598

Are you sure you want to change the base?

conda packaging for holoscan latest version (3.x) #28598

Conversation

billysuh7 commented Dec 16, 2024 • edited Loading

github-actions bot commented Dec 16, 2024

conda-forge-admin commented Dec 16, 2024 • edited Loading

isuruf commented Dec 17, 2024

billysuh7 commented Dec 18, 2024 • edited Loading

github-actions bot commented Feb 17, 2025

github-actions bot commented Feb 19, 2025

billysuh7 commented Feb 19, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

conda-forge-admin commented Feb 22, 2025 • edited Loading

conda-forge-admin commented Feb 22, 2025 • edited Loading

billysuh7 commented Dec 16, 2024 •

edited

Loading

conda-forge-admin commented Dec 16, 2024 •

edited

Loading

billysuh7 commented Dec 18, 2024 •

edited

Loading

conda-forge-admin commented Feb 22, 2025 •

edited

Loading

conda-forge-admin commented Feb 22, 2025 •

edited

Loading