Flashinfer Submission Tagger

Flashinfer Submission Tagger

Overview

Create a new submission-vN tag for the current commit, push that tag to the remote repository, and verify the exact tag and commit that FlashInfer contest evaluation should consume.

Use the bundled script to compute the next tag deterministically instead of re-implementing version selection by hand. Use the bundled topology checker to block invalid config.toml layouts before commit, tag, or push.

Workflow

1. Check repository state before tagging.

• Read the current branch, commit SHA, existing submission-* tags, and remotes.
• Assume origin is the push remote unless the user explicitly asks for another remote or origin is missing.

1. Choose the tag name.

• If the user provides a concrete tag, use it exactly.
• Otherwise, select the next numeric version after the highest existing submission-vN.
• Never reuse or retarget an existing tag unless the user explicitly asks for that destructive action.

1. Validate config.toml topology before touching metadata. This is a hard gate.

• Run:

python3 /Users/yue/.codex/skills/flashinfer-submission-tagger/scripts/check_config_topology.py --repo /path/to/repo

• If the checker exits non-zero, STOP. Do not commit, tag, or push until the layout is fixed.
• The checker only evaluates the submission surface: the repo root, immediate definition subdirectories, and the solution/ trees attached to them. It intentionally ignores archived artifacts under directories such as optimize_ops/ or reference/.
• Accepted layouts:
• Single-definition root layout: root config.toml + root solution//...
• Definition-subdir layout: one or more /config.toml files, each paired with /solution//...
• Forbidden layouts:
• root config.toml in a multi-definition repo
• both root config.toml and definition-subdir configs in the same repo
• any config.toml nested under solution/
• a config whose entry_point file does not exist under the resolved source dir
• a resolved source dir with zero real source files

1. Discover the solution layout from the validated config topology. Do NOT hardcode paths like solution/triton/ or solution/cuda/.

• For a single-definition root layout, read root config.toml and derive solution//.
• For a definition-subdir layout, read the target /config.toml and derive /solution//.
• When the repo contains multiple definition subdirectories, operate only on the target definition the user asked to submit. Do not infer metadata from unrelated definitions.

1. Update all submission metadata files before committing. Every tagged commit must have consistent, up-to-date metadata. Paths are relative to the solution directory discovered in step 4.

• config.toml (root for a single-definition root layout, or /config.toml for a definition-subdir layout): update name field to match the new submission version.
• /summary.json: update with the best benchmark run data — date, candidate name, previoussubmission, per-run results (avglatency_ms, median, p95, min, speedup, passed), mean across repeats, and improvement vs previous. Use avg_latency_ms as the primary metric.
• /summary.md: update with a human-readable comparison table (latency-primary), description of changes, decision rationale, and selected retained run.
• /retained_run.json: if a new benchmark run was performed, update with the retained run's full JSON data.
• README.md: update the Performance table with the best single-run result, update Implementation Notes if kernel behavior changed, update the solution name.
• Kernel file (e.g. kernel.py or kernel.cu in the solution directory): ensure the docstring/header comment reflects the new version number and lists key changes.
• Move any backup/exploration kernels and benchmark artifacts out of the solution directory into optimize_ops/ so the submission directory stays clean.
• Do NOT modify .gitignore.

1. Commit all updated files.

• Stage only the specific files that changed (avoid git add -A).
• Write a descriptive commit message with the key metric (avglatencyms), changes summary, and Co-Authored-By if applicable.

1. Create the tag on HEAD.

• Tag the current commit, not the working tree.
• If the repo has uncommitted changes after step 5, call that out — something was missed.

1. Push the branch and tag.

• Push both origin and origin .
• After push, verify the tag exists and report the commit SHA it points to.

Scripts

Use scripts/tag_submission.py when you need deterministic tag selection. Use scripts/check_config_topology.py as a mandatory submission gate.

Common commands:

python3 scripts/tag_submission.py --repo /path/to/repo --print-next
python3 scripts/tag_submission.py --repo /path/to/repo --create
python3 scripts/tag_submission.py --repo /path/to/repo --create --push
python3 scripts/tag_submission.py --repo /path/to/repo --tag submission-v7 --create --push

Behavior:

• --print-next: print the next available submission-vN without changing git state.
• --create: create the tag on the current HEAD.
• --push: push the chosen tag to the configured remote. Use with --create for the common contest workflow.
• --tag: override auto-increment and use an explicit tag name.

Topology gate:

python3 /Users/yue/.codex/skills/flashinfer-submission-tagger/scripts/check_config_topology.py --repo /path/to/repo

Behavior:

• exit 0: topology is valid for submission work
• exit non-zero: invalid layout; fix the reported config.toml issue before commit/tag/push
• prints the detected layout, active config(s), resolved source dir(s), and the first blocking error when invalid

Output Expectations

When using this skill, report:

• the final tag name
• the commit SHA it points to
• whether the tag was pushed successfully

If anything blocks the workflow, report the exact git condition:

• no git repository
• tag already exists
• remote missing
• push failed

If the topology gate blocks the workflow, report the exact config issue:

• multi-definition repo has a root config.toml
• root and definition-subdir configs coexist
• config.toml is nested under solution/
• resolved source dir is missing or empty
• entry_point file is missing relative to the resolved source dir

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: syhya
• Source: syhya/mlsys26-flashinfer-contest
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.