add torchcodec to pytorch builds by lamikr · Pull Request #2590 · ROCm/TheRock

lamikr · 2025-12-16T22:23:24Z

Add support for checking out and building torchcodec during the pytorch build phase

version of torchcodec depends from the pytorch version and torchcodecs/README.md lists which torchcodec and pytorch versions are compatible with each others. Because pytorch/.github/ci_commit_pins directory does not mention the torchcodec, we will check the pytorch version in pytorch_torchcodec_repo.py and then use that information to determine which torchcodec version to checkout.
torchcodec source code checkout and build has been enabled in the linux CI job
pytorch_torchcodec_repo.py will first install a torchcodec compatible version of setuptools before launching the torchcodec build command (setuptools > 0.80)

Todo:

Verify windows builds
Add smoketests

DevDotMathias · 2026-01-03T09:50:53Z

I was able to successfully build torchcodec (with torch + torchaudio 2.9) on Windows (gfx103X, 7.11.0a20251230 from nightlies/v2 just in case the CI failures when building pytorch wheels w/Python 3.12 and torch 2.9 were related to the newer rocm wheels) after cherry-picking 5ab4742 and patching torchcodec's setup.py to:

fetch the pybind11 cmake directory, because despite using pybind11[global] and the link in (venv root)/share/cmake populating correctly, it still could not locate the config
- this is a known issue for some, I am not the only one experiencing it
per torchaudio 2.9 tools/setup_helpers/extensions.py:
- default to Ninja generator, instead of Visual Studio
- get CXX and CC from env and add -DCMAKE_C_COMPILER=... and -DCMAKE_CXX_COMPILER=..., so rocm_sdk_devel clang is used

Patch here: patch-2.9-fix-windows-builds.patch. torch+audio+torchcodec from this build worked as expected, as long as the ffmpeg DLL directory was added via os.add_dll_directory() or the DLLs copied to the torchcodec package root in my venv. I used ffmpeg 8.0.1 from gyan.dev via winget. The symbol issue from #1490 did not occur, as expected.

Hope to see this get added to CI soon, ideally with Windows builds, and even more ideally with gfx103X builds (Linux + Windows).

lamikr · 2026-01-08T07:18:20Z

@DevDotMathias Thanks for your input for torchcodec. Not sure would it be easier to add first the Linux support on first PR and then add the windows support on second PR little later.

At least for me the pybind11 worked by installing it with pip to python venv that is used by the build on ci-image.

This PR includes the docker changes I needed to make to our CI build image so that python versions under /opt/python directory contains the libpython libraries required by the torchcodec. For now I have temporarily pushed the modified CI image to docker io but before this patch is applied we need to push it back to github docker server and then re-change the download url to point to there.

With these changes the torchcodec builded now ok for pytorch 2.9 and pytorch nighly.
(I have disabled torchcodec builds for pytorch 2.7 and pytorch 2.8 because with those I got some errors that would require patching the torchcodec to get the build pass)

CI Test build with logs is here: https://github.com/ROCm/TheRock/actions/runs/20797194747

Pytorch tests where failing but those tests should not have anything torchcodec specific yet, so I am not sure why.

stellaraccident · 2026-01-08T18:38:15Z

Assigning @amd-shiraz for visibility.

lamikr · 2026-01-08T21:39:27Z

@stellaraccident I just rebased my patches to latest rock version and noticed that you had also added
2 days ago a support for building python from sources to CI images on master branch 2 days ago.

I did not investigate yet whether these new versions already contains python changes I need, so I will still include those changes in first push to get them saved somewhere...

But if your ci image changes for python are enough for torchcodec build, I can drop ci-image related changes from my patch. I had a little different approach and added the build commands directly to dockerfile for different python version while you call external shell script.

lamikr · 2026-01-09T18:42:28Z

I was now able to drop the docker changes as new CI images from January 7 contained python versions with devel libraries. Patch handling is also removed in this version as support for applying patches for pytorch projects was also dropped on January 7.

DevDotMathias · 2026-01-10T06:47:25Z

Thanks for your input for torchcodec. Not sure would it be easier to add first the Linux support on first PR and then add the windows support on second PR little later.

Whatever works best, I just noticed you had Windows builds listed in the description of this PR. If you have to get upstream to accept a PR (I'd submit, but I think it'll carry more weight coming from an AMD dev than from me) for switching to Ninja and reading env vars, I can see holding off until they merge.

At least for me the pybind11 worked by installing it with pip to python venv that is used by the build on ci-image.

I expect that to always be the case for Linux. If it works straight off for you in Windows CI images, all the better.

lamikr · 2026-01-12T18:50:03Z

I am now trying testing the images build by ci with gfx1100 on Ubuntu 24.04. I used the rocm-sdk from 20260108 for building and install is still little tricky as I need to pickup packages from v2 and v2-staging dirs.
These steps worked at least:

Install ffmpeg to ubuntu as torchcodec requires it but does not ship it in itself

sudo apt install ffmpeg
--> will install on ffmpeg 6.1 on ubuntu 24.04. Torchcodec will support ffmpeg 4,5,6,7 and 8
--> libtorchcodec_core6.so will be used by the torchcodec for ffmpeg version 6

Create python venv

cd mkdir therock_gfx1100 cd therock_gfx1100 python3 -m venv .venv_sdk2 && source .venv_sdk2/bin/activate

Install pytorch and rocm-sdk

pip install --index-url https://rocm.devreleases.amd.com/v2-staging/gfx110X-all torch==2.9.1+rocm7.11.0a20251208

Install torchcodec

pip install --index-url https://rocm.devreleases.amd.com/v2-staging/gfx110X-all torchcodec pip install numpy

Execute test app to print torchcodec version:

python3 -c "import torchcodec; 
try: print('version: ' + torchcodec.__version__)
except Exception as ex: print(f'An error occurred: {ex}')"

which should printout:

version: 0.9.1+rocm7.11.0a20260108

marbre · 2026-01-13T18:06:53Z

We had an issue with a history rewrite which confused the GH UI and the commits it is showing on your PR. Please make sure to update your branch e.g. in the GH UI via

Afterwards you need update your local branch via git pull.

Alternative git rebase solution

Alternatively you can manually update your branch locally by following the below instructions:

git checkout <yourbranch>
git fetch origin main
git rebase origin/main
git push --force-with-lease origin <yourbranch>

- add support for checking out and building torchcodec as a part of the pytorch build process - enable the torchcodec source code checkout and build in the linux CI job Note: - Version of torchcodec we checkout and build depends from the pytorch version and torchcodecs/README.md lists which torchcodec and pytorch versions are compatible with each others. - Pinmapping file used by many other pytorch related subprojects does not yet exist for the torchcodec in directory pytorch/.github/ci_commit_pins and therefore we will check the pytorch version in pytorch_torchcodec_repo.py and then use that information to determine which torchcodec version to checkout. - At least for now we will only support the torchcodec with pytorch 2.9 and never releases and nightlies. - Different pytorch components may require a different versions of python setuptools. Torchcodec requires that setup tools version needs to be never than 80.0. Because of that the build scripts install install a torchcodec compatible version of setuptools before launching the torchcodec build itself. Todo: - Add smoketests - Add windows build support fixes: #1490 Signed-off-by: Mika Laitio <mika.laitio@amd.com>

lamikr · 2026-01-13T21:26:04Z

Rebased latest test build for gfx110x gpus passed ok for all python versions:

https://github.com/ROCm/TheRock/actions/runs/20970122325

amd-shiraz · 2026-01-15T22:25:46Z

external-builds/pytorch/build_prod_wheels.py

+    except ValueError:
+        print("Error, Skipping pytorch torchcodec build")
+        print(
+            f"Could not convert version string major and minor to int. Check the format of '{version_string}'."


is this a typo ? should this be version_string or build_version ?

I'm assuming it should be build_version. "version_string" is not defined, only "version_string_arr".

Fixed it was typo, needs to be build_version.

amd-shiraz · 2026-01-15T22:27:36Z

external-builds/pytorch/build_prod_wheels.py

+
+    # verify that installed setuptools > 80.0 (80.9.0 tested)
+    exec(
+        [sys.executable, "-m", "pip", "install", "setuptools>80.0"],


Prefer installing deps once in the workflow (pinned versions / constraints file), not ad-hoc inside the build script. unless there is a string reason to. if there is pls add that here as comment.

I agree. I would replace this with a version check of setuptools and return an error if setputools is not a sufficient version.

Similarly for the other deps, return an error if they aren't already installed.

I noticed torchcodec build errors if the setuptools version was 0.7.x while pytorch, audio and visio had build fine with that version. This is the reason why I have added the pip install commands for setuptools and pybind.

Triton and pytorch have requirements.txt and when we build those components we install their dependencies in build_prod_wheels.py with following commands:

python -m pip install -r triton/python/requirements.txt (line 583)
and
python -m pip install -r pytorch/requirements.txt (line 827)

Problem is that the torchcodec does not have requirements.txt file. Because we have dropped support for patching the pytorch related projects, I can not use the patch file to create a requirements.txt for torchcodec. This is the reason I have installed packages I have noticed to be need one by one.

I did local builds for pytorch 2.9, 2.10 and nightly for python 3.12
and Iwas not able to reproduce anymore the error related to setuptools version.

Depending from the pytorch version build, setuptools version ended up being in my build in the end either 80.9.0 or 80.10.2.

I launched one more test build with CI for 9070/gfx120x target.

https://github.com/ROCm/TheRock/actions/runs/21376750243

CI build failed after I removed the dependency install at least for the pybinds cmake files missing.
It worked on my local Ubuntu build because I had there the pybind-dev installed.

Error:

2026-01-26T23:17:04.1824662Z CMake Error at src/torchcodec/_core/CMakeLists.txt:7 (find_package): 2026-01-26T23:17:04.1825183Z -- Configuring incomplete, errors occurred! 2026-01-26T23:17:04.1826085Z By not providing "Findpybind11.cmake" in CMAKE_MODULE_PATH this project has 2026-01-26T23:17:04.1826924Z asked CMake to find a package configuration file provided by "pybind11", 2026-01-26T23:17:04.1827417Z but CMake did not find one. 2026-01-26T23:17:04.1827621Z 2026-01-26T23:17:04.1827875Z Could not find a package configuration file provided by "pybind11" with any 2026-01-26T23:17:04.1828350Z of the following names: 2026-01-26T23:17:04.1828521Z 2026-01-26T23:17:04.1828625Z pybind11Config.cmake 2026-01-26T23:17:04.1828906Z pybind11-config.cmake 2026-01-26T23:17:04.1829075Z 2026-01-26T23:17:04.1829309Z Add the installation prefix of "pybind11" to CMAKE_PREFIX_PATH or set 2026-01-26T23:17:04.1829893Z "pybind11_DIR" to a directory containing one of the above files. If 2026-01-26T23:17:04.1830584Z "pybind11" provides a separate development package or SDK, be sure it has 2026-01-26T23:17:04.1830928Z been installed.

Either we need to install the pybind to our manylinux install natively or we need to install it via pip install dynamically. For now I test the pip install way as it is most similar to "pip install -r requirements.txt" that we do for torch and triton.

We discussed this in the morning meeting because pybind-devel headers needs to still get installed even with the latest patch version somehow to build image.

We could pre-install pybind-devel rpm to manylinux-docker image from rpm package. Marius did not like from that idea.

We could install pybind-devel rpm during the build time. This idea was not good. One reason is that in our docker image we have multiple python versions under /opt/python directory and install of pybind-devel rpm would install one more as a dependency under /usr.

Third option is to install with pip during the build time which is now implemented and I got impression that it is preferred. I would not like to tie the exact pybind version to pip-install command because we are doing a builds wit multiple different python versions and I am not sure can the pybind-version be different depending from the python version currently used. (3.10, 3.11, 3.12, 3.13 or 3.14)

amd-shiraz · 2026-01-15T22:29:46Z

external-builds/pytorch/build_prod_wheels.py

+            "Python3_ROOT_DIR": python_home_dir,
+        }
+    )
+    print("do_build_pytorch_torchcodec")


do we need this twice ? also a dup above in line 962

Fixed, removed the dub printout line. Leaving other the printouts for env-variables.

jeffdaily

Same changes requested as @amd-shiraz .

jeffdaily · 2026-01-16T16:42:16Z

external-builds/pytorch/build_prod_wheels.py

+    except ValueError:
+        print("Error, Skipping pytorch torchcodec build")
+        print(
+            f"Could not convert version string major and minor to int. Check the format of '{version_string}'."


I'm assuming it should be build_version. "version_string" is not defined, only "version_string_arr".

jeffdaily · 2026-01-16T16:45:08Z

external-builds/pytorch/build_prod_wheels.py

+
+    # verify that installed setuptools > 80.0 (80.9.0 tested)
+    exec(
+        [sys.executable, "-m", "pip", "install", "setuptools>80.0"],


I agree. I would replace this with a version check of setuptools and return an error if setputools is not a sufficient version.

Similarly for the other deps, return an error if they aren't already installed.

Signed-off-by: Mika Laitio <mika.laitio@amd.com>

- update from v0.10.0-rc1 to v0.10.0 Signed-off-by: Mika Laitio <mika.laitio@amd.com>

I did local builds for pytorch 2.9, 2.10 and nightly and was not able to reproduce anymore the error that I have seen if not making sure before the torchcodec build starts that the right pybind and setuptools versions were installed. It may have been a problem with some older torchcodec version from the time I tested the building also for pytorch 2.7 and pytorch 2.8 Signed-off-by: Mika Laitio <mika.laitio@amd.com>

Fixes following build error on ci-machine: By not providing "Findpybind11.cmake" in CMAKE_MODULE_PATH this project has asked CMake to find a package configuration file provided by "pybind11", but CMake did not find one. Could not find a package configuration file provided by "pybind11" with any of the following names: pybind11Config.cmake pybind11-config.cmake Signed-off-by: Mika Laitio <mika.laitio@amd.com>

lamikr · 2026-01-27T00:33:55Z

Now with pybind-dev dependency inserted back for torchcodec the build passes.

https://github.com/ROCm/TheRock/actions/runs/21378178853/job/61543032071

And short test for this on linux machine with gfx1201/RX 9070 with python 3.12

# add python 3.12 to PATH and LD_LIBRARY_PATH if you are running some other python version by default
source ./init_python_312.sh
# create venv that uses this python 3.12 version
python3 -m venv venv_py312 && source venv_py312/bin/activate
# install triton, torch, audio, vision and torchcodec
python3 -m pip install torchaudio==2.9.0+rocm7.11.0a20260121 torchvision==0.22.1+rocm7.11.0a20260121 torchcodec==0.9.1+rocm7.11.0a20260121 --index-url https://rocm.devreleases.amd.com/v2-staging/gfx120X-all
# run small test to import torchcodec and then printout torchcodec version
python3 -c "import torchcodec; print('torcodec_version: ' + str(torchcodec.__version__))"

This should printout:

torcodec_version: 0.9.1+rocm7.11.0a20260121

lamikr · 2026-02-05T16:57:47Z

As a sidenote there is now PR to remove the direct linking of libpython to custom_ops-libraries in torchcodec.
Because of this linking the torchcodec will at the moment require that libpython.so is also installed on the system which is not actually required.

Upstream PR: meta-pytorch/torchcodec#1224
Upstream bug: meta-pytorch/torchcodec#1164

ScottTodd

The overall structure of the script and workflow changes looks fine. I'm wondering at a high level - is this working? I don't see any evidence of what testing has been done in the PR description and the code seems to be building with CUDA (and also ROCM?) disabled.

ScottTodd · 2026-02-10T23:38:22Z

.github/workflows/build_portable_linux_pytorch_wheels.yml

As was just done for adding https://github.com/rocm/apex, please sequence this as at least two PRs:

Add build_prod_wheels.py changes together with new pytorch_torchcodec_repo.py changes

Add code to .github/workflows/build_portable_linux_pytorch_wheels.yml

See the commit sequence for Apex:

49b454a

39d840b (note that there should also be changes to manage.py, write_torch_versions.py)

That way, we can get the build code in place for other developers to start using and then enable in our release builds separately. If we then need to revert for some reason we can revert just the workflow changes as a flag flip while keeping local build support for easier debugging.

ScottTodd · 2026-02-10T23:40:25Z

.github/workflows/build_portable_linux_pytorch_wheels.yml

      torchaudio_version: ${{ steps.build-pytorch-wheels.outputs.torchaudio_version }}
      torchvision_version: ${{ steps.build-pytorch-wheels.outputs.torchvision_version }}
      triton_version: ${{ steps.build-pytorch-wheels.outputs.triton_version }}
+      torchcodec_version: ${{ steps.build-pytorch-wheels.outputs.torchcodec_version }}


This output var won't exist without changes to write_torch_versions.py. See 39d840b for example

ScottTodd · 2026-02-10T23:41:51Z

external-builds/pytorch/build_prod_wheels.py

    else:
        print("--- Not build pytorch-vision (no --pytorch-vision-dir)")
+    # Build pytorch torchcodec
+    if args.build_pytorch_torchcodec or (


nit: add a line break between the "Build pytorch vision" and the "Build pytorch torchcodec" sections

ScottTodd · 2026-02-10T23:45:26Z

external-builds/pytorch/build_prod_wheels.py

+    python_home_dir = get_python_home_dir()
+    # python env executed in ci build is intepreter version only version in /opt/python/cp3xy-cp3xy dir
+    # environment which contains also the libpython development files is under dir /opt/python-shared/cp3xy-cp3xy dir
+    # and we need to force the torchcodec to find it with find_package("python" command
+    if "/opt/python/cp" in python_home_dir:
+        python_home_dir = python_home_dir.replace(
+            "/opt/python/cp", "/opt/python-shared/cp", 1
+        )


This doesn't seem right. Don't your changes in meta-pytorch/torchcodec#1224 remove the need for this?

We can fork torchcodec under https://github.com/ROCm/ and point the torchcodec checkout there as needed for this bringup work, then switch back to using upstream once that PR (and any other ROCm/TheRock-specific fixes are landed). How does that approach sound, @jeffdaily ?

(and if we do need this python_home_dir code there are more portable ways to do this that work inside and outside of virtual environments on Windows and Linux... the path walking and /opt/python code is brittle)

We will still need this even if the patch is applied in upstream that will remove the linking to libpython because of the find package call is used also to find the location of python header files. Basically

find_package(Python3 ${PYTHON_VERSION} EXACT COMPONENTS Development)

when succeeds will setup environment variable

${Python3_INCLUDE_DIRS}

that is then used on this cmake method call:

target_include_directories(${library_name} PRIVATE ./../../../ "${TORCH_INSTALL_PREFIX}/include" ${Python3_INCLUDE_DIRS} )

ScottTodd · 2026-02-10T23:47:08Z

external-builds/pytorch/build_prod_wheels.py

+    env.update(
+        {
+            "ENABLE_CUDA": "0",


Why ENABLE_CUDA=0? Is ROCM enabled or does this produce a CPU-only torchcodec wheel? (If so we could just copy upstream-produced torchcodec CPU wheels into our release index instead of building from source ourselves)

At least for now it's CPU only.

That answers part of the question, but not what is critical here. Are CPU-only torchcodec packages useful at all for us? If so, why wouldn't we just use upstream-produced packages instead of building them ourselves? Is this a stepping stone to building with ROCm support? What still needs to be done for that?

We discussed offline. CPU-only torchcodec packages are useful and a reasonable starting point. We may just be able to use packages that are already built upstream though, by adding an entry to https://github.com/ROCm/TheRock/blob/main/build_tools/third_party/s3_management/update_dependencies.py.

If we aren't setting any custom build options (like ROCm support using the ROCm Python packages), I would prefer that over including all this extra code and spending more compute time on our build runners.

We discussed offline. CPU-only torchcodec packages are useful and a reasonable starting point. We may just be able to use packages that are already built upstream though, by adding an entry to https://github.com/ROCm/TheRock/blob/main/build_tools/third_party/s3_management/update_dependencies.py.

If we aren't setting any custom build options (like ROCm support using the ROCm Python packages), I would prefer that over including all this extra code and spending more compute time on our build runners.

If you guys go this route, please submit a PR to upstream[1] so that their builds with be compatible with yours across all platforms. It makes no sense to me that TorchCodec would be the only one to force MSVC for Windows builds when both PyTorch and TorchAudio use ninja as a generator and read the CC/CXX env vars, and until the bug in clang is fixed (which could be anywhere from next week to never) that will never play nice.

[1] As I've said before, I'm happy to do this, but my PR would carry zero weight, where one from AMD would be treated with significance.

ScottTodd · 2026-02-10T23:48:17Z

external-builds/pytorch/build_prod_wheels.py

+    # cmake support for pybind11 to avoid error:
+    # Could not find a package configuration file provided by "pybind11"
+    exec(
+        [sys.executable, "-m", "pip", "install", "pybind11[global]"],
+        cwd=pytorch_torchcodec_dir,
+        env=env,
+    )
+    # torchcodec build
+    exec(
+        [sys.executable, "-m", "build", "--wheel", "-vvv", "--no-isolation"],
+        cwd=pytorch_torchcodec_dir,
+        env=env,
+    )


Shouldn't build requirements be automatically installed by python -m build --wheel? How is this handled for the other projects?

It seems that setup.py does not handle this in torchcodec as we would get pybind error.
In triton and pytorch we handle this by doing "pip install -r requirements.txt" but torchcodec does not have this file, we call "pip install pybind" instead directly because otherwise we will get following build error on our CI:

By not providing "Findpybind11.cmake" in CMAKE_MODULE_PATH this project has asked CMake to find a package configuration file provided by "pybind11", but CMake did not find one. Could not find a package configuration file provided by "pybind11" with any of the following names: pybind11Config.cmake pybind11-config.cmake

In theory we could also install dependency to CI machine with "rpm" but that version would then install python 3.11 as a dependency to our CI image that we do not want as we have our own build python versions on the image. We previously discussed about this and conclusion was that install via pip is better alternative.

ScottTodd · 2026-02-10T23:50:41Z

external-builds/pytorch/pytorch_torchcodec_repo.py

+THIS_MAIN_REPO_NAME = "pytorch_torchcodec"
+THIS_DIR = Path(__file__).resolve().parent
+
+DEFAULT_ORIGIN = "https://github.com/pytorch/torchcodec.git"


This moved to https://github.com/meta-pytorch/torchcodec (redirects should still work, but we might as well update)

Suggested change

DEFAULT_ORIGIN = "https://github.com/pytorch/torchcodec.git"

DEFAULT_ORIGIN = "https://github.com/meta-pytorch/torchcodec.git"

(longer term I'm not sure about the politics with the PyTorch foundation vs Meta for such projects... I wonder what sorts of lines we should draw for packages that we build and distribute as AMD)

I updated the URL and pushed the change in. I can't comment too much whether we should or should not include this to our distro.

external-builds/pytorch/pytorch_torchcodec_repo.py

New upstream url: https://github.com/meta-pytorch/torchcodec.git Signed-off-by: Mika Laitio <mika.laitio@amd.com>

Change audio related help texts to point to torchcodec. Co-authored-by: Scott Todd <scott.todd0@gmail.com>

github-project-automation bot added this to TheRock Triage Dec 16, 2025

github-project-automation bot moved this to TODO in TheRock Triage Dec 16, 2025

lamikr force-pushed the users/lamikr/wip/pytorch_torchcodec branch 2 times, most recently from f878ea1 to 5ab4742 Compare December 18, 2025 01:23

lamikr mentioned this pull request Dec 18, 2025

[Issue]: As of TorchAudio 2.9, this save_with_torchcodec function relies on TorchCodec's decoding capabilities under the hood.However, there is no accompanying TorchCodec wheel. #1490

Open

lamikr force-pushed the users/lamikr/wip/pytorch_torchcodec branch 5 times, most recently from d02804b to 7929a49 Compare December 18, 2025 21:12

stellaraccident requested a review from amd-shiraz January 8, 2026 18:38

lamikr force-pushed the users/lamikr/wip/pytorch_torchcodec branch 2 times, most recently from 3aa7f92 to 4f33771 Compare January 9, 2026 18:40

subodh-dubey-amd force-pushed the main branch from 7e9ba03 to 3077ce8 Compare January 13, 2026 12:08

subodh-dubey-amd requested review from a team, ScottTodd and marbre as code owners January 13, 2026 12:08

lamikr force-pushed the users/lamikr/wip/pytorch_torchcodec branch from 4f33771 to 4681062 Compare January 13, 2026 18:21

lamikr force-pushed the users/lamikr/wip/pytorch_torchcodec branch from 4681062 to 0de0c8f Compare January 13, 2026 19:32

ScottTodd removed the request for review from a team January 15, 2026 21:13

ScottTodd removed the request for review from marbre January 15, 2026 21:13

amd-shiraz requested changes Jan 15, 2026

View reviewed changes

jeffdaily requested changes Jan 16, 2026

View reviewed changes

lamikr added 4 commits January 26, 2026 08:42

fix printouts based on to review

a8888dc

Signed-off-by: Mika Laitio <mika.laitio@amd.com>

build torchcodec v0.10.0 for pytorch 2.10

f026814

- update from v0.10.0-rc1 to v0.10.0 Signed-off-by: Mika Laitio <mika.laitio@amd.com>

lamikr requested review from amd-shiraz, jeffdaily and marbre January 27, 2026 01:25

ScottTodd requested changes Feb 10, 2026

View reviewed changes

lamikr and others added 2 commits February 11, 2026 09:27

Update torchcodec github url

e05921d

New upstream url: https://github.com/meta-pytorch/torchcodec.git Signed-off-by: Mika Laitio <mika.laitio@amd.com>

Apply argument parser suggestions from code review

8ab7ae4

Change audio related help texts to point to torchcodec. Co-authored-by: Scott Todd <scott.todd0@gmail.com>

lamikr requested a review from ScottTodd February 11, 2026 17:51

	DEFAULT_ORIGIN = "https://github.com/pytorch/torchcodec.git"
	DEFAULT_ORIGIN = "https://github.com/meta-pytorch/torchcodec.git"

Conversation

lamikr commented Dec 16, 2025

Uh oh!

DevDotMathias commented Jan 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

lamikr commented Jan 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

stellaraccident commented Jan 8, 2026

Uh oh!

lamikr commented Jan 8, 2026

Uh oh!

lamikr commented Jan 9, 2026

Uh oh!

DevDotMathias commented Jan 10, 2026

Uh oh!

lamikr commented Jan 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

marbre commented Jan 13, 2026

Uh oh!

lamikr commented Jan 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lamikr Jan 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lamikr Jan 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lamikr Jan 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lamikr Jan 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lamikr Jan 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jeffdaily left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lamikr commented Jan 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

lamikr commented Feb 5, 2026

Uh oh!

ScottTodd left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

DevDotMathias commented Jan 3, 2026 •

edited

Loading

lamikr commented Jan 8, 2026 •

edited

Loading

lamikr commented Jan 12, 2026 •

edited

Loading

lamikr commented Jan 13, 2026 •

edited

Loading

lamikr Jan 26, 2026 •

edited

Loading

lamikr Jan 26, 2026 •

edited

Loading

lamikr Jan 26, 2026 •

edited

Loading

lamikr Jan 27, 2026 •

edited

Loading

lamikr Jan 26, 2026 •

edited

Loading

lamikr commented Jan 27, 2026 •

edited

Loading

lamikr Feb 11, 2026 •

edited

Loading