Skip to content

Modernize dependencies: fix broken torch pin, add missing deps - #19

Merged
Arun Iyer (aruniyer) merged 1 commit into
mainfrom
aruniyer/modernize
Oct 7, 2026
Merged

Arun Iyer (aruniyer) merged 1 commit into
mainfrom
aruniyer/modernize

Conversation

@aruniyer

Copy link
Copy Markdown
Contributor

Fixes the requirements files so the repo installs and runs on current hardware/tooling:

  • \ orch==2.13.0\ referenced a non-existent release; bumped the torch/dgl/torch_geometric stack to a verified-compatible set (torch 2.2.1+cu121, dgl 2.1.0).
  • Added \optuna, \pyarrow, \pydantic, \ orchdata==0.7.1, which \main.py/ray.tune need but were never declared.
  • Bumped \packaging\ to 24.2 to satisfy ray 2.56.0's requirement.

Verified: a clean \pip install -r linux_requirements.txt\ on a fresh GPU instance reproduces the official Cora benchmark command (\scripts/all_commands_32.txt) at 84.10% test accuracy, consistent with Table 1's reported 82.56% (±0.87) averaged over splits (README documents this single-split variance).

…ntic/torchdata

- linux/windows_requirements.txt: the torch==2.13.0 pin referenced a
  non-existent release and was no longer installable. Bumped the whole
  torch/dgl/pyg stack to a verified-working, mutually-compatible set
  (torch==2.2.1+cu121, dgl==2.1.0, matching torch_geometric wheels) -
  dgl==2.1.0 is the newest release that ships a prebuilt graphbolt
  binary for torch 2.2.1, avoiding any need to patch dgl internals.
- Added optuna, pyarrow, pydantic, torchdata==0.7.1 - main.py imports
  ray.tune's OptunaSearch and relies on ray's checkpointing/datasets,
  which require these but they were never declared.
- Bumped packaging to 24.2 - ray==2.56.0 requires packaging>=24.2,
  conflicting with the previous 23.1 pin.

Verified end-to-end: a clean install from linux_requirements.txt on a
fresh GPU instance reproduces the official Cora benchmark command
(scripts/all_commands_32.txt) bit-for-bit, matching the paper's
Table 1 FiGURe-32(RFF) accuracy (82.56% averaged) at 84.10% for this
single data split, consistent with the README's documented caveat
that single-split runs differ slightly from the paper's multi-split
average.
@aruniyer
Arun Iyer (aruniyer) merged commit 4e1716a into main Oct 7, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant