Repository navigation
feat(installer): add an opt-in cu124 PEAR runtime for RTX 30/40 - #141
Open
AWildRShakes wants to merge 2 commits into
Open
AWildRShakes wants to merge 2 commits into
AWildRShakes wants to merge 2 commits into
Conversation
ADR-0016's cu128 matrix costs pre-Blackwell GPUs about a third of their live throughput (RTX 3080: 30.25 -> 20.53 NET FPS), and today the only recovery is installing v1.0.6-win.4, which drops every fix since. Task 0041 already names a temporary cu124 path as a legitimate fallback. ADR-0019 proposes shipping the frozen cu124 matrix as a user-selected runtime inside the same PEAR payload, reusing the hash-pinned PyTorch3D wheel from the v1.0.6-win.4 payload. It partially reverses ADR-0016's rejection of per-architecture payloads, so it stays proposed until the maintainer decides. Task 0044 tracks the implementation and records the RTX 3080 end-to-end evidence. Signed-off-by: RShakes <uoreyn@gmail.com>
RTX 30/40 owners lose about a third of live throughput on the default cu128 matrix and could only recover it by downgrading to v1.0.6-win.4. The two matrices differ only in torch, torchvision and the PyTorch3D build (the PyPI lock is identical), so the PEAR payload now carries the frozen cu124 matrix next to the default and the user picks one. A PyTorch3D build only loads under the torch it was compiled against, so the payload build stages the hash-pinned cu124 wheel from the v1.0.6-win.4 payload under wheels/cu124/ together with its torch lock, and the packer refuses one without the other. The choice persists in pear_runtime.json so install, repair and upgrade honour it. The switch script, exposed as two Start-menu shortcuts, refuses cu124 when a Blackwell GPU is present and restores the previous choice if the reinstall fails; the PEAR handler falls back to cu128 with a warning if a Blackwell GPU appears later. The release notice now points RTX 30/40 owners at the shortcut instead of telling them to weigh the upgrade. Verified end to end on an RTX 3080: default install, switch to cu124 with a PyTorch3D CUDA kernel running, and switch back. Per ADR-0019. Refs task 0044. Signed-off-by: RShakes <uoreyn@gmail.com>
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Test plan
uv run ruff check ./uv run ruff format --check .uv run pyright --pythonplatform Windows/--pythonplatform Linuxuv run lint-importsuv run pytest(646 passed). New behavioral tests intests/test_pear_runtime_switch.pyrun the switch script under Windows PowerShell 5.1 with a stubnvidia-smi.build_pear_payload.ps1runknn_pointsran on the GPULinks
🤖 Generated with Claude Code