Declare the torch version in release wheels by shoumikhin · Pull Request #23311 · pytorch/executorch

shoumikhin · 2026-10-01T14:11:54Z

The ExecuTorch wheel only works with a PyTorch (torch) version close to the one it was built with. But since 1.5, nothing tells pip which version that is. So pip can install any torch next to it, and the problem only shows up later, at runtime:

pip install executorch==1.5.1 torch==2.12.0   # installs fine
# then running a model fails:
RuntimeError: tensor does not have a device    # 1.5.1 was built on torch 2.14

Releases 0.2 to 1.4 declared torch, because each release branch added the requirement by hand. For 1.5 it was not added.

This change makes a release declare its torch version automatically, so pip picks a torch that works:

Requires-Dist: torch<2.15,>=2.14.0a0

How it works:

The version comes from the torch the wheel is built against, so nobody has to type it in.
The upper limit (<2.15) matters as much as the lower one. Without it, a later torch can break a release that already shipped, as executorch==1.3.1 with torch 2.14.1 does today.
The a0 in the lower limit lets a torch built from source (version 2.14.0a0+git...) count as 2.14. A build on a nightly torch snapshot (version 2.14.0.dev...) uses that snapshot as the lower limit, because it sorts below a0.
It does not choose between the CPU and CUDA builds of torch. The package index you install from still does that.

Only releases declare it. Nightly builds, local builds, and the small export-only wheel still declare no torch. The release workflows mark a build as a release by setting BUILD_VERSION to a plain version like 1.6.0+cu132.

The wheel tests now also check that a release declares torch and that the installed torch fits the range. Nightly wheels skip this, so on main the new unit tests cover the rule instead.

Test plan:

New unit tests. Each fails if a part of the rule is removed, including the line in setup.py that adds it.
Built the package metadata from the real setup.py for release, nightly, local, and export-only builds, and checked the torch line in each.
Checked what pip does with the new range. A fresh install gets torch 2.14.1. An installed 2.14.1 CUDA build or source build is kept, and 2.13 or 2.15 is replaced.
The new wheel test fails on the published 1.5.1 wheel, which shipped without the requirement.

pytorch-bot · 2026-10-01T14:11:58Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/23311

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

⏳ No Failures, 1 Pending

As of commit 511909b with merge base fdd5140 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

Copilot

Copilot review overview

🟡 Changes recommended

Release metadata generation fails under standard isolated PEP 517 builds because PyTorch is unavailable to the build backend.

Review effort: Balanced
Findings: 1

Open (1)

PEP 517 build fails when BUILD_VERSION requires unavailable torch · New

What changed in this PR

Adds PyTorch minor-version constraints to release wheels to prevent ABI-incompatible installations.

Changes:

Derives release-wheel PyTorch bounds from the build environment.
Adds unit and wheel smoke-test coverage.
Updates installation and packaging documentation.

File	Description
`setup.py`	Adds release-only PyTorch dependency metadata.
`install_utils.py`	Derives the compatible PyTorch range.
`docs/source/getting-started.md`	Documents release and nightly behavior.
`.ci/scripts/wheel/test_shared_libraries.py`	Updates dependency commentary.
`.ci/scripts/wheel/test_cuda_linux.py`	Checks CUDA release metadata.
`.ci/scripts/wheel/test_clean_install.py`	Validates the declared PyTorch range.
`.ci/scripts/wheel/cuda_arch_list.sh`	Clarifies PyTorch build selection.
`.ci/scripts/tests/test_release_torch_requirement.py`	Tests dependency generation and wheel scope.

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

+    public = (build_version or "").strip().split("+", 1)[0]
+    if not re.fullmatch(r"\d+(\.\d+)*", public):
+        return None
+    torch_version = importlib.metadata.version("torch").split("+", 1)[0]


Copilot

Copilot review overview

🟡 Changes recommended

The current reusable Linux and macOS wheel workflows reject the new requirement format before repository smoke tests run.

Review effort: Balanced
Findings: 2

Open (2)

Update wheel workflows to support version ranges and nightly builds · New
PEP 517 build fails when BUILD_VERSION requires unavailable torch

+    major, minor = (int(part) for part in torch_version.split(".")[:2])
+    snapshot = re.fullmatch(rf"{major}\.{minor}\.0\.dev\d+", torch_version)
+    floor = snapshot.group(0) if snapshot else f"{major}.{minor}.0a0"
+    return f"torch>={floor},<{major}.{minor + 1}"


The ExecuTorch wheel only works with a PyTorch (`torch`) version close to the one it was built with. But since 1.5, nothing tells pip which version that is. So pip can install any torch next to it, and the problem only shows up later, at runtime: ``` pip install executorch==1.5.1 torch==2.12.0 # installs fine # then running a model fails: RuntimeError: tensor does not have a device # 1.5.1 was built on torch 2.14 ``` Releases 0.2 to 1.4 declared torch, because each release branch added the requirement by hand. For 1.5 it was not added. This change makes a release declare its torch version automatically, so pip picks a torch that works: ``` Requires-Dist: torch<2.15,>=2.14.0a0 ``` How it works: - The version comes from the torch the wheel is built against, so nobody has to type it in. - The upper limit (`<2.15`) matters as much as the lower one. Without it, a later torch can break a release that already shipped, as `executorch==1.3.1` with torch 2.14.1 does today. - The `a0` in the lower limit lets a torch built from source (version `2.14.0a0+git...`) count as 2.14. A build on a nightly torch snapshot (version `2.14.0.dev...`) uses that snapshot as the lower limit, because it sorts below `a0`. - It does not choose between the CPU and CUDA builds of torch. The package index you install from still does that. Only releases declare it. Nightly builds, local builds, and the small export-only wheel still declare no torch. The release workflows mark a build as a release by setting `BUILD_VERSION` to a plain version like `1.6.0+cu132`. The wheel tests now also check that a release declares torch and that the installed torch fits the range. Nightly wheels skip this, so on `main` the new unit tests cover the rule instead. Test plan: - New unit tests. Each fails if a part of the rule is removed, including the line in `setup.py` that adds it. - Built the package metadata from the real `setup.py` for release, nightly, local, and export-only builds, and checked the torch line in each. - Checked what pip does with the new range. A fresh install gets torch 2.14.1. An installed 2.14.1 CUDA build or source build is kept, and 2.13 or 2.15 is replaced. - The new wheel test fails on the published 1.5.1 wheel, which shipped without the requirement.

Copilot

Copilot review overview

🔵 Needs a closer look

The shared Linux and macOS wheel workflows reject the new range syntax because they still require an exact torch==... metadata entry.

Review effort: Balanced
Findings: 2

Open (2)

Update wheel workflows to support version ranges and nightly builds
PEP 517 build fails when BUILD_VERSION requires unavailable torch

Copilot

Copilot review overview

🟡 Changes recommended

The generated bounded requirement is rejected by the shared Linux and macOS wheel workflow’s exact-pin check.

Review effort: Balanced
Findings: 3

Open (3)

Smoke test rejects bounded torch dependency range · New
Update wheel workflows to support version ranges and nightly builds
PEP 517 build fails when BUILD_VERSION requires unavailable torch

    # carry it are declared here. A CPU wheel adds nothing.
-    setup_kwargs["install_requires"] = _base_dependencies() + _cuda_dependencies()
+    setup_kwargs["install_requires"] = (
+        _base_dependencies() + _cuda_dependencies() + _torch_dependencies()


shoumikhin added the release notes: build Changes related to build, including dependency upgrades, build flags, optimizations, etc. label Oct 1, 2026

shoumikhin requested a review from mergennachin as a code owner October 1, 2026 14:11

Copilot AI balanced review requested due to automatic review settings October 1, 2026 14:11

meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Oct 1, 2026

Copilot started reviewing on behalf of shoumikhin October 1, 2026 14:12 View session

Copilot AI reviewed Oct 1, 2026

View reviewed changes

Comment thread install_utils.py

public = (build_version or "").strip().split("+", 1)[0]

if not re.fullmatch(r"\d+(\.\d+)*", public):

return None

torch_version = importlib.metadata.version("torch").split("+", 1)[0]

shoumikhin force-pushed the release-wheels-declare-torch branch from 7424449 to 33eab4f Compare October 1, 2026 15:49

Copilot AI balanced review requested due to automatic review settings October 1, 2026 22:48

Copilot started reviewing on behalf of shoumikhin October 1, 2026 22:49 View session

Copilot AI reviewed Oct 1, 2026

View reviewed changes

Comment thread install_utils.py

major, minor = (int(part) for part in torch_version.split(".")[:2])

snapshot = re.fullmatch(rf"{major}\.{minor}\.0\.dev\d+", torch_version)

floor = snapshot.group(0) if snapshot else f"{major}.{minor}.0a0"

return f"torch>={floor},<{major}.{minor + 1}"

Copilot AI balanced review requested due to automatic review settings October 2, 2026 00:03

shoumikhin force-pushed the release-wheels-declare-torch branch from 72ee216 to 74646aa Compare October 2, 2026 00:03

Copilot started reviewing on behalf of shoumikhin October 2, 2026 00:03 View session

Copilot AI reviewed Oct 2, 2026

View reviewed changes

Merge branch 'main' into release-wheels-declare-torch

511909b

Copilot AI balanced review requested due to automatic review settings October 2, 2026 13:52

Copilot started reviewing on behalf of shoumikhin October 2, 2026 13:53 View session

Copilot AI reviewed Oct 2, 2026

View reviewed changes

Comment thread setup.py

# carry it are declared here. A CPU wheel adds nothing.

setup_kwargs["install_requires"] = _base_dependencies() + _cuda_dependencies()

setup_kwargs["install_requires"] = (

_base_dependencies() + _cuda_dependencies() + _torch_dependencies()

Navigation Menu

Declare the torch version in release wheels - #23311

Declare the torch version in release wheels#23311
shoumikhin wants to merge 2 commits into
pytorch:mainfrom
shoumikhin:release-wheels-declare-torch

shoumikhin commented Oct 1, 2026 •

edited

Loading

Uh oh!

pytorch-bot Bot commented Oct 1, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

shoumikhin commented Oct 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot Bot commented Oct 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/23311

⏳ No Failures, 1 Pending

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🔵 Needs a closer look

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

shoumikhin commented Oct 1, 2026 •

edited

Loading

pytorch-bot Bot commented Oct 1, 2026 •

edited

Loading