Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions docs/proposals/index.rst
Original file line number Diff line number Diff line change
Expand Up @@ -7,3 +7,4 @@ Fromager Enhancement Proposals
new-patcher-config
new-resolver-config
release-cooldown
wheel-build-tag-hook
221 changes: 221 additions & 0 deletions docs/proposals/wheel-build-tag-hook.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,221 @@
# Unique wheel file names with a `wheel_build_tag` hook

- Author: Christian Heimes
- Created: 2026-04-16
- Status: Open
- GitHub issue: [#1059](https://github.com/python-wheel-build/fromager/issues/1059)

## What

This enhancement proposes a new stevedore hook point, `wheel_build_tag`, in
the existing `fromager.hooks` namespace. The hook lets downstream plugin
packages inject custom suffixes into the
[wheel build tag](https://packaging.python.org/en/latest/specifications/binary-distribution-format/),
producing unique wheel file names that encode platform, accelerator stack,
and dependency ABI information.

## Why

Fromager-built wheels are platform-specific and may depend on:

- **OS / distribution version**, e.g. Fedora 43, RHEL 9.6, RHEL 10.1
- **AI accelerator stack**, e.g. CUDA 13.1 vs CUDA 12.9, ROCm 7.1
- **Torch ABI**, which is unstable across versions; a wheel compiled for
Torch 2.10.0 may have a different ABI than one compiled for Torch 2.11.0

Currently, wheel filenames carry none of this information. A rebuild for a
new Torch version does not produce a new filename, making it difficult to:

1. Invalidate caches when the underlying platform or dependency stack changes.
2. Distinguish wheels built for different accelerator stacks or OS versions.
3. Replace outdated wheels with correctly-targeted rebuilds.
4. Share wheels between indexes. Downstream maintains a separate index for
each accelerator version, but only a couple of dozen packages out of over
1,200 are CUDA/ROCm/Torch-specific. Wheels like `pillow` or `fromager`
are identical across accelerator indexes and could be shared if their
filenames clearly indicate they have no accelerator dependency. Sharing is
out of scope for this proposal but is a possibility for future
improvements in downstream.

## Goals

- introduce a `wheel_build_tag` hook point in the `fromager.hooks` stevedore
namespace
- allow hooks to contribute ordered suffix segments to the wheel build tag
- produce unique, deterministic wheel file names that reflect the build
environment

## Non-goals

- Filtering or selecting wheels by build tag at install time. `pip install`
and `uv pip install` only use the build tag for sorting, not for filtering.
- Sharing wheels across indexes. While unique file names enable this in
principle, the mechanics of cross-index sharing are out of scope.
- Accessing wheel content, the build environment, or ELF dependency info from
within the hook. The hook must work identically whether a wheel is freshly
built or retrieved from cache.
- Validation of annotations such as "depends on libtorch". A validation
system for `build_wheel` may be added in the future.
- Per-package plugin hooks (`overrides.find_and_invoke()`). A per-package
override can be added later when the need arises.

## How

### Wheel spec background

The wheel filename format is:

```
{distribution}-{version}(-{build tag})?-{python tag}-{abi tag}-{platform tag}.whl
```

The build tag is optional, must start with a digit, and must not contain `-`.
It is parsed as `tuple[int, str]`. Fromager already fills the `int` part from
variant + package changelog and sets the `str` suffix to `""`. The build tag
is also stored in the `{name}-{version}.dist-info/WHEEL` metadata file and
shown in `pip list` output.

This proposal extends that suffix with hook-provided segments (e.g.
`_el9.6_rocm7.1_torch2.10.0`).

### Hook signature

```python
def build_tag_hook(
*,
ctx: context.WorkContext,
req: Requirement,
version: Version,
wheel_tags: frozenset[Tag],
) -> typing.Sequence[tuple[int, str]]:
...
```

Each registered hook returns `typing.Sequence[tuple[int, str]]`, a sequence
Comment thread
rd4398 marked this conversation as resolved.
of (sort-order, suffix) pairs. This allows a single hook to contribute
multiple suffix segments (e.g. both an OS tag and an accelerator tag). The
runner collects all pairs from all hooks, sorts them in ascending order, and
returns the suffix parts as a sequence. Neither `sort-order` nor `suffix`
have to be unique. The first appearance of a `suffix` is used. Suffixes are
are joined with `_` and appended to the existing build tag.
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What is the use case for having multiple hooks?

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  • The other hooks also support multiple hook implementations.
  • Downstream can define multiple hooks, e.g. one for platform tag, one for Torch, one for each accelerator stack, and so on.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The way the other hooks are invoked only allows for one entry point, based on the package name. Stevedore should report an error if multiples are registered. For example, it wouldn't make sense to have 2 plugins give resolver providers for the same package.

This case feels different, because the hooks might combine data usefully. I'm just wondering if that's a "maybe useful" thing or a "definitely useful" thing, because the sorting seems complex to get right.

If we have 1 hook and it returns a list, then it can understand the ordering itself without leaking implementation details between hooks. But if we allow mixing hooks from multiple sources, how would we ensure that their sort ordering makes sense? I could see us releasing a bunch of simple hooks upstream, for example, to do platform tags and change log length and maybe other values. But how would we ensure those sort properly without having them all know about each other to specify their sort order values? And how would we add the one for torch into that mix without upstream and downstream knowing about each other in some meaningful way?

If instead of supporting multiple hooks, we just provide helper functions, then the single downstream hook implementation could specify the ordering.

If we want multiple hooks, then maybe we need a config setting to manage the order, instead of having the hooks try to express it directly? Stevedore has a way to invoke plugins based on a list of names, for example, so we could directly convert a list of names in a config file into invocation of the hooks in the expected order.

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It's "maybe useful". We can get by with just a single hook. It would simplify the implementation, too.

I designed the feature with multiple hooks in mind to give us more flexibility and because stevedore may return multiple hooks. We do not need that flexibility in downstream. There seems to be no option to enforce a single wheel_build_tag entry point for fromager.hooks. Multiple packages can provide a wheel_build_tag hook.

Maybe entry points and stevedore extensions are the wrong approach for hooks that are singletons? This feature feels less like a notification hook and more like a setting. We could take another approach like a global config setting of type ImportString.

# overrides/settings.yaml
[wheels]
build_tag_hook = "mysettings.hooks:custom_tag_hook"

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It looks like for overrides we're using a normal ExtensionManager and relying on only having one plugin with a given name. That's a bug, we should at least register a conflict_resolver function. Maybe just using the builtin error_on_conflict is what we want?

For hooks, it may make sense to have multiple values registered. If we use multiple hooks, I recommend building some sort of determinism into the execution order of the hooks themselves, though, rather than relying on the data returned by the hooks to be mixable directly. For example, in this case if you sorted by (name, order-key, value) then even if 2 hooks return the same order-key the results would always come out in the same order.

If we want a singleton hook, using the NamedExtensionManager or DriverExtensionManager may fit that pattern better.

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does it even make sense to use an entry point and stevedore hook for this feature? The wheel build tag hook feels like a runtime configuration option.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That's fair. It breaks the pattern we use elsewhere, but maybe since it's just one it does make sense.


`suffix` must only contain alphanumeric ASCII characters or dot
(`[a-zA-Z0-9.]`). When a hook returns any other character or raises an
exception, the build fails.

Build tag hooks are registered for `fromager.hooks` entry point and name
`wheel_build_tag`.

### Example hook

```python
def example_hook(
*,
ctx: context.WorkContext,
req: Requirement,
version: Version,
wheel_tags: frozenset[Tag],
) -> typing.Sequence[tuple[int, str]]:
result: list[tuple[int, str]] = []
platlib = any(tag.platform != "any" for tag in wheel_tags)
if platlib:
# fc43, el9.6, ...
result.append((1, get_distro_tag()))
Comment thread
rd4398 marked this conversation as resolved.
pbi = ctx.package_build_info(req)

# example how to use anntoations and ctx.variant for custom flags
if pbi.annotations.get("example.accelerator-specific") == "true":
# cpu, cuda13.0, ...
if ctx.variant.startswith("cpu"):
result.append((2, "cpu"))
elif ctx.variant.startswith("cuda"):
cv = Version(os.environ["CUDA_VERSION"])
result.append((2, f"cuda{cv.major}.{cv.minor}"))
else:
Comment thread
coderabbitai[bot] marked this conversation as resolved.
raise NotImplementedError(ctx.variant)
return result
```

```python
@functools.cache
def get_distro_tag() -> str:
info = platform.freedesktop_os_release()
ids = [info["ID"]] # always defined
if "ID_LIKE" in info: # ids in precedence order
ids.extend(info["ID_LIKE"].split())
version_id = info.get("VERSION_ID", "")
for ident in ids:
if ident == "rhel": # RHEL and CentOS
return f"el{version_id}"
elif ident == "fedora":
return f"fc{version_id}"
# other distros
return f"{ids[0]}{version_id}".replace("_", "").replace("-", "")
```

Registration:

```toml
[project.entry-points."fromager.hooks"]
wheel_build_tag = "mypackage.hooks:example_hook"
```

### What hooks can access

- **`ctx` + `req`**: for `ctx.variant` and package configuration like annotations
- **`wheel_tags`**: detect whether a wheel is purelib or platlib
(platform/arch-specific).

Hooks can also use other information like **`os.environ`** or the
**`platform`** stdlib module. They should only use information that
is immutable and identifies the build context, e.g.
**`platform.freedesktop_os_release()`** to read distribution name and
version from `/etc/os-release`.

### What hooks cannot access

The hook **does not** have access to wheel content, the `build_env`, or
ELF dependency info. While this information exists during the build, it is
not available when wheels are retrieved from cache servers or local cache. The
hook must work identically in both paths.

### Examples

#### RHEL 9.6, ROCm 7.1, Torch 2.10.0

| Wheel | Build tag |
| -- | -- |
| `flash_attn-2.8.3-8_el9.6_rocm7.1_torch2.10.0-cp312-cp312-linux_x86_64.whl` | `8_el9.6_rocm7.1_torch2.10.0` |
| `torch-2.10.0-7_el9.6_rocm7.1-cp312-cp312-linux_x86_64.whl` | `7_el9.6_rocm7.1` |
| `pillow-12.2.0-2_el9.6-cp312-cp312-linux_x86_64.whl` | `2_el9.6` |
| `fromager-0.79.0-2-py3-none-any.whl` | `2` (pure-python, no suffix) |

#### Fedora 43, CUDA 13.0, Torch 2.9.1

| Wheel | Build tag |
| -- | -- |
| `flash_attn-2.8.3-8_fc43_cuda13.0_torch2.9.1-cp312-cp312-linux_x86_64.whl` | `8_fc43_cuda13.0_torch2.9.1` |
| `torch-2.9.1-8_fc43_cuda13.0-cp312-cp312-linux_x86_64.whl` | `8_fc43_cuda13.0` |
| `pillow-12.2.0-2_fc43-cp312-cp312-linux_x86_64.whl` | `2_fc43` |
| `fromager-0.79.0-2-py3-none-any.whl` | `2` (pure-python, no suffix) |

Note how pure-python wheels (`py3-none-any`) receive no suffix, while
platlib wheels get progressively more specific tags based on their actual
dependencies.

## Limitations

The current wheel standard does not support multiple accelerator variants in
a single package index. A single index cannot contain both CUDA and ROCm
builds of the same package because `pip install` and `uv pip install` only
use the build tag for sorting, not for filtering. An index with both CUDA
and ROCm wheels would result in the installer picking whichever has the
highest build tag, not the correct accelerator.

The upcoming [Wheel.Next](https://wheelnext.dev/) initiative and
[PEP 817](https://peps.python.org/pep-0817/) /
[PEP 825](https://peps.python.org/pep-0825/) aim to address this by
introducing wheel variants, which will enable CUDA and ROCm wheels to
coexist in the same index. Logic for Torch and CUDA/ROCm dependency
selection in this proposal may be reused by wheel variants when the new
standard becomes available.
2 changes: 2 additions & 0 deletions docs/spelling_wordlist.txt
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,7 @@ iteratively
json
lexicographically
libcurl
libtorch
linter
localhost
matcher
Expand All @@ -57,6 +58,7 @@ setuptools
statelessly
stderr
stdin
stdlib
stdout
subcommands
subdirectory
Expand Down
Loading