Skip to content

Implement a vendored nvidia variant provider - #22690

Merged
konstin merged 26 commits into
mainfrom
konsti/nvidia-variant-provider
Oct 22, 2026
Merged

konstin merged 26 commits into
mainfrom
konsti/nvidia-variant-provider

Conversation

@konstin

@konstin konstin commented Oct 18, 2026 •

Copy link
Copy Markdown
Member

This PR connects the wheel variant support with real hardware. It queries the CUDA version and compute capability. The CUDA version detection is shared with torch-backend, the compute capability is added on top. This gives us end-to-end wheel variant support.

While variant markers (required for torch) and more/generic backends are still missing, this should be enough to select the first, simple ML wheels with for GPU support. A smoke test translating torch wheel from the older wheel variant index to the new format passed on a real GPU.

The shape of this provider will evolve, right now it's important that it works as an end to end a demo provider with real hardware. The provider is a vendored Rust translation of https://github.com/wheelnext/nvidia-variant-provider. The more interesting part here is the internal provider interface.

@konstin
konstin requested a review from zsol October 18, 2026 18:21
@konsti-openai konsti-openai changed the title Support the NVIDIA wheel variant provider Implement a vendored nvidia variant provider Oct 18, 2026
@konsti-openai konsti-openai added the preview Experimental behavior label Oct 18, 2026
@konstin
konstin force-pushed the konsti/nvidia-variant-provider branch from ecafd9a to 75886b9 Compare October 18, 2026 18:31
@konstin
konstin force-pushed the konsti/nvidia-variant-provider branch from 75886b9 to 5ce425e Compare October 19, 2026 10:21
@konstin
konstin force-pushed the konsti/nvidia-variant-provider branch from 5ce425e to 870762e Compare October 19, 2026 10:45
@codspeed

codspeed Bot commented Oct 19, 2026 •

Copy link
Copy Markdown

Merging this PR will not alter performance

1 warning

⚠️ 6 benchmarks are dominated by system calls

System calls cannot be simulated consistently and are left out of the measure. The reported value understates the real cost. Switch to the Walltime instrument to measure them.

Learn more about system calls

See affected benchmarks

✅ 117 untouched benchmarks
⏩ 54 skipped benchmarks1


Comparing konsti/nvidia-variant-provider (b3448a5) with main (02cad35)

Open in CodSpeed

Footnotes

  1. 54 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

@konstin
konstin force-pushed the konsti/nvidia-variant-provider branch from 870762e to 3732975 Compare October 19, 2026 11:36
@konstin
konstin force-pushed the konsti/nvidia-variant-provider branch from 3732975 to 9b442bb Compare October 19, 2026 13:14
@konstin
konstin force-pushed the konsti/nvidia-variant-provider branch from 9b442bb to 121fb3b Compare October 19, 2026 15:02
@konstin
konstin merged commit 8b7edb8 into main Oct 22, 2026
120 checks passed
@konstin
konstin deleted the konsti/nvidia-variant-provider branch October 22, 2026 22:14
konstin added a commit that referenced this pull request Oct 23, 2026
I forgot to list the NVIDIA provider's variant properties in #22690.
hbjydev pushed a commit to hbjydev/phoebe that referenced this pull request Oct 27, 2026
This PR contains the following updates:

| Package | Type | Update | Change |
|---|---|---|---|
| [uv](https://github.com/astral-sh/uv) | tools | patch | `0.13.5` → `0.13.6` |

---

### Release Notes

<details>
<summary>astral-sh/uv (uv)</summary>

### [`v0.13.6`](https://github.com/astral-sh/uv/blob/HEAD/CHANGELOG.md#0136)

[Compare Source](astral-sh/uv@0.13.5...0.13.6)

Released on 2026-10-26.

##### Enhancements

- Accept an empty `UV_RUN_RECURSION_DEPTH` as zero and report invalid values consistently ([#&#8203;22814](astral-sh/uv#22814))
- Explain how to add project metadata when `uv init` finds a `pyproject.toml` without a `[project]` table ([#&#8203;22916](astral-sh/uv#22916))

##### Python

- Add CPython 3.11.18, 3.12.16, 3.13.18, 3.14.10, and 3.15.2 ([#&#8203;22938](astral-sh/uv#22938))

##### Preview features

- Detect NVIDIA GPU capabilities when selecting compatible wheel variants with `wheel-variants` ([#&#8203;22690](astral-sh/uv#22690))
- Nest each module's owners in an object in `uv workspace metadata` output ([#&#8203;22638](astral-sh/uv#22638))
- Report the unsupported properties required by an incompatible wheel variant ([#&#8203;22719](astral-sh/uv#22719))

##### Performance

- Reduce peak memory use when resolving large workspaces ([#&#8203;22860](astral-sh/uv#22860), [#&#8203;22863](astral-sh/uv#22863), [#&#8203;22864](astral-sh/uv#22864), [#&#8203;22866](astral-sh/uv#22866), [#&#8203;22870](astral-sh/uv#22870))
- Reduce peak memory use when locking or upgrading large projects ([#&#8203;22833](astral-sh/uv#22833), [#&#8203;22834](astral-sh/uv#22834))
- Speed up redirected output from `uv export` and `uv pip compile` ([#&#8203;22840](astral-sh/uv#22840))
- Reuse cached or active environments when running PEP 723 scripts without dependencies ([#&#8203;21458](astral-sh/uv#21458))
- Reduce parsing overhead for build dependency and build backend settings ([#&#8203;22848](astral-sh/uv#22848), [#&#8203;22871](astral-sh/uv#22871), [#&#8203;22879](astral-sh/uv#22879), [#&#8203;22884](astral-sh/uv#22884))
- Reduce parsing overhead for project sources and workspace references ([#&#8203;22880](astral-sh/uv#22880), [#&#8203;22883](astral-sh/uv#22883))
- Reduce parsing overhead for cache, exclusion, cutoff, and preview settings ([#&#8203;22876](astral-sh/uv#22876), [#&#8203;22877](astral-sh/uv#22877), [#&#8203;22878](astral-sh/uv#22878), [#&#8203;22887](astral-sh/uv#22887))
- Reduce PyPI metadata parsing overhead ([#&#8203;22874](astral-sh/uv#22874), [#&#8203;22875](astral-sh/uv#22875))
- Speed up parsing tool receipt requirements ([#&#8203;22885](astral-sh/uv#22885))

##### Bug fixes

- Prevent `uv run --active --script` and `uv sync --active --script` from deleting non-virtual-environment directories ([#&#8203;21464](astral-sh/uv#21464))
- Use `AMD64` for `platform_machine` when resolving for x86-64 Windows ([#&#8203;22935](astral-sh/uv#22935))
- Preserve free-threaded and other non-default variants in `uv python upgrade` ([#&#8203;22836](astral-sh/uv#22836))
- Allow `uv init --no-package` and `uv init --script` when `UV_INIT_BUILD_BACKEND` is set ([#&#8203;22907](astral-sh/uv#22907))
- Prevent panics when rendering containment markers with local versions ([#&#8203;22815](astral-sh/uv#22815))

##### Documentation

- Update the Docker guide and reproduction examples to use uv 0.13.5 ([#&#8203;22867](astral-sh/uv#22867))

</details>

---

### Configuration

📅 **Schedule**: (in timezone Europe/London)

- Branch creation
  - At any time (no schedule defined)
- Automerge
  - At any time (no schedule defined)

🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied.

♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox.

🔕 **Ignore**: Close this PR and you won't be reminded about this update again.

---

 - [ ] <!-- rebase-check -->If you want to rebase/retry this PR, check this box

---

This PR has been generated by [Mend Renovate CLI](https://github.com/renovatebot/renovate).
<!--renovate-debug:eyJjcmVhdGVkSW5WZXIiOiI0NC4xODEuNyIsInVwZGF0ZWRJblZlciI6IjQ0LjE4MS43IiwidGFyZ2V0QnJhbmNoIjoibWFpbiIsImxhYmVscyI6WyJyZW5vdmF0ZS9naXRodWItcmVsZWFzZSIsInR5cGUvcGF0Y2giXX0=-->

Reviewed-on: https://git.hayden.moe/hayden/phoebe/pulls/1017
luketainton pushed a commit to luketainton/luke_rsu that referenced this pull request Oct 27, 2026
This PR contains the following updates:

| Package | Type | Update | Change |
|---|---|---|---|
| [ghcr.io/astral-sh/uv](https://github.com/astral-sh/uv) | stage | patch | `0.13.5` → `0.13.6` |

---

### Release Notes

<details>
<summary>astral-sh/uv (ghcr.io/astral-sh/uv)</summary>

### [`v0.13.6`](https://github.com/astral-sh/uv/blob/HEAD/CHANGELOG.md#0136)

[Compare Source](astral-sh/uv@0.13.5...0.13.6)

Released on 2026-10-26.

##### Enhancements

- Accept an empty `UV_RUN_RECURSION_DEPTH` as zero and report invalid values consistently ([#&#8203;22814](astral-sh/uv#22814))
- Explain how to add project metadata when `uv init` finds a `pyproject.toml` without a `[project]` table ([#&#8203;22916](astral-sh/uv#22916))

##### Python

- Add CPython 3.11.18, 3.12.16, 3.13.18, 3.14.10, and 3.15.2 ([#&#8203;22938](astral-sh/uv#22938))

##### Preview features

- Detect NVIDIA GPU capabilities when selecting compatible wheel variants with `wheel-variants` ([#&#8203;22690](astral-sh/uv#22690))
- Nest each module's owners in an object in `uv workspace metadata` output ([#&#8203;22638](astral-sh/uv#22638))
- Report the unsupported properties required by an incompatible wheel variant ([#&#8203;22719](astral-sh/uv#22719))

##### Performance

- Reduce peak memory use when resolving large workspaces ([#&#8203;22860](astral-sh/uv#22860), [#&#8203;22863](astral-sh/uv#22863), [#&#8203;22864](astral-sh/uv#22864), [#&#8203;22866](astral-sh/uv#22866), [#&#8203;22870](astral-sh/uv#22870))
- Reduce peak memory use when locking or upgrading large projects ([#&#8203;22833](astral-sh/uv#22833), [#&#8203;22834](astral-sh/uv#22834))
- Speed up redirected output from `uv export` and `uv pip compile` ([#&#8203;22840](astral-sh/uv#22840))
- Reuse cached or active environments when running PEP 723 scripts without dependencies ([#&#8203;21458](astral-sh/uv#21458))
- Reduce parsing overhead for build dependency and build backend settings ([#&#8203;22848](astral-sh/uv#22848), [#&#8203;22871](astral-sh/uv#22871), [#&#8203;22879](astral-sh/uv#22879), [#&#8203;22884](astral-sh/uv#22884))
- Reduce parsing overhead for project sources and workspace references ([#&#8203;22880](astral-sh/uv#22880), [#&#8203;22883](astral-sh/uv#22883))
- Reduce parsing overhead for cache, exclusion, cutoff, and preview settings ([#&#8203;22876](astral-sh/uv#22876), [#&#8203;22877](astral-sh/uv#22877), [#&#8203;22878](astral-sh/uv#22878), [#&#8203;22887](astral-sh/uv#22887))
- Reduce PyPI metadata parsing overhead ([#&#8203;22874](astral-sh/uv#22874), [#&#8203;22875](astral-sh/uv#22875))
- Speed up parsing tool receipt requirements ([#&#8203;22885](astral-sh/uv#22885))

##### Bug fixes

- Prevent `uv run --active --script` and `uv sync --active --script` from deleting non-virtual-environment directories ([#&#8203;21464](astral-sh/uv#21464))
- Use `AMD64` for `platform_machine` when resolving for x86-64 Windows ([#&#8203;22935](astral-sh/uv#22935))
- Preserve free-threaded and other non-default variants in `uv python upgrade` ([#&#8203;22836](astral-sh/uv#22836))
- Allow `uv init --no-package` and `uv init --script` when `UV_INIT_BUILD_BACKEND` is set ([#&#8203;22907](astral-sh/uv#22907))
- Prevent panics when rendering containment markers with local versions ([#&#8203;22815](astral-sh/uv#22815))

##### Documentation

- Update the Docker guide and reproduction examples to use uv 0.13.5 ([#&#8203;22867](astral-sh/uv#22867))

</details>

---

### Configuration

📅 **Schedule**: (UTC)

- Branch creation
  - At any time (no schedule defined)
- Automerge
  - At any time (no schedule defined)

🚦 **Automerge**: Enabled.

♻ **Rebasing**: Whenever PR is behind base branch, or you tick the rebase/retry checkbox.

🔕 **Ignore**: Close this PR and you won't be reminded about this update again.

---

 - [ ] <!-- rebase-check -->If you want to rebase/retry this PR, check this box

---

This PR has been generated by [Mend Renovate CLI](https://github.com/renovatebot/renovate).
<!--renovate-debug:eyJjcmVhdGVkSW5WZXIiOiI0NC4xODYuMSIsInVwZGF0ZWRJblZlciI6IjQ0LjE4Ni4xIiwidGFyZ2V0QnJhbmNoIjoibWFpbiIsImxhYmVscyI6WyJ0eXBlL2RlcGVuZGVuY2llcyJdfQ==-->

Reviewed-on: https://git.tainton.uk/repos/rsu/pulls/62
Co-authored-by: renovate[bot] <renovate-bot@git.tainton.uk>
luketainton pushed a commit to luketainton/repos_labmcp that referenced this pull request Oct 27, 2026
This PR contains the following updates:

| Package | Type | Update | Change |
|---|---|---|---|
| [ghcr.io/astral-sh/uv](https://github.com/astral-sh/uv) | final | patch | `0.13.5` → `0.13.6` |

---

### Release Notes

<details>
<summary>astral-sh/uv (ghcr.io/astral-sh/uv)</summary>

### [`v0.13.6`](https://github.com/astral-sh/uv/blob/HEAD/CHANGELOG.md#0136)

[Compare Source](astral-sh/uv@0.13.5...0.13.6)

Released on 2026-10-26.

##### Enhancements

- Accept an empty `UV_RUN_RECURSION_DEPTH` as zero and report invalid values consistently ([#&#8203;22814](astral-sh/uv#22814))
- Explain how to add project metadata when `uv init` finds a `pyproject.toml` without a `[project]` table ([#&#8203;22916](astral-sh/uv#22916))

##### Python

- Add CPython 3.11.18, 3.12.16, 3.13.18, 3.14.10, and 3.15.2 ([#&#8203;22938](astral-sh/uv#22938))

##### Preview features

- Detect NVIDIA GPU capabilities when selecting compatible wheel variants with `wheel-variants` ([#&#8203;22690](astral-sh/uv#22690))
- Nest each module's owners in an object in `uv workspace metadata` output ([#&#8203;22638](astral-sh/uv#22638))
- Report the unsupported properties required by an incompatible wheel variant ([#&#8203;22719](astral-sh/uv#22719))

##### Performance

- Reduce peak memory use when resolving large workspaces ([#&#8203;22860](astral-sh/uv#22860), [#&#8203;22863](astral-sh/uv#22863), [#&#8203;22864](astral-sh/uv#22864), [#&#8203;22866](astral-sh/uv#22866), [#&#8203;22870](astral-sh/uv#22870))
- Reduce peak memory use when locking or upgrading large projects ([#&#8203;22833](astral-sh/uv#22833), [#&#8203;22834](astral-sh/uv#22834))
- Speed up redirected output from `uv export` and `uv pip compile` ([#&#8203;22840](astral-sh/uv#22840))
- Reuse cached or active environments when running PEP 723 scripts without dependencies ([#&#8203;21458](astral-sh/uv#21458))
- Reduce parsing overhead for build dependency and build backend settings ([#&#8203;22848](astral-sh/uv#22848), [#&#8203;22871](astral-sh/uv#22871), [#&#8203;22879](astral-sh/uv#22879), [#&#8203;22884](astral-sh/uv#22884))
- Reduce parsing overhead for project sources and workspace references ([#&#8203;22880](astral-sh/uv#22880), [#&#8203;22883](astral-sh/uv#22883))
- Reduce parsing overhead for cache, exclusion, cutoff, and preview settings ([#&#8203;22876](astral-sh/uv#22876), [#&#8203;22877](astral-sh/uv#22877), [#&#8203;22878](astral-sh/uv#22878), [#&#8203;22887](astral-sh/uv#22887))
- Reduce PyPI metadata parsing overhead ([#&#8203;22874](astral-sh/uv#22874), [#&#8203;22875](astral-sh/uv#22875))
- Speed up parsing tool receipt requirements ([#&#8203;22885](astral-sh/uv#22885))

##### Bug fixes

- Prevent `uv run --active --script` and `uv sync --active --script` from deleting non-virtual-environment directories ([#&#8203;21464](astral-sh/uv#21464))
- Use `AMD64` for `platform_machine` when resolving for x86-64 Windows ([#&#8203;22935](astral-sh/uv#22935))
- Preserve free-threaded and other non-default variants in `uv python upgrade` ([#&#8203;22836](astral-sh/uv#22836))
- Allow `uv init --no-package` and `uv init --script` when `UV_INIT_BUILD_BACKEND` is set ([#&#8203;22907](astral-sh/uv#22907))
- Prevent panics when rendering containment markers with local versions ([#&#8203;22815](astral-sh/uv#22815))

##### Documentation

- Update the Docker guide and reproduction examples to use uv 0.13.5 ([#&#8203;22867](astral-sh/uv#22867))

</details>

---

### Configuration

📅 **Schedule**: (UTC)

- Branch creation
  - At any time (no schedule defined)
- Automerge
  - At any time (no schedule defined)

🚦 **Automerge**: Enabled.

♻ **Rebasing**: Whenever PR is behind base branch, or you tick the rebase/retry checkbox.

🔕 **Ignore**: Close this PR and you won't be reminded about this update again.

---

 - [ ] <!-- rebase-check -->If you want to rebase/retry this PR, check this box

---

This PR has been generated by [Mend Renovate CLI](https://github.com/renovatebot/renovate).
<!--renovate-debug:eyJjcmVhdGVkSW5WZXIiOiI0NC4xODYuMSIsInVwZGF0ZWRJblZlciI6IjQ0LjE4Ni4xIiwidGFyZ2V0QnJhbmNoIjoibWFpbiIsImxhYmVscyI6WyJ0eXBlL2RlcGVuZGVuY2llcyJdfQ==-->

Reviewed-on: https://git.tainton.uk/repos/labmcp/pulls/83
Co-authored-by: renovate[bot] <renovate-bot@git.tainton.uk>

This branch was previously deployed

1 inactive deployment
automations — b3448a53 Deployed Oct 22, 2026 by konstin via review / security review #52283
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

preview Experimental behavior

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants

⌂ 换网址