qwenlm / parscale
pythonmaincd6acb41.5K linessynced 1mo ago
Code health
8.3out of 10Excellent
This codebase scores 8.3 out of 10 on defect risk, which we rate excellent. It also scores maintainability 8.2 and static performance risk 10.0 out of 10. The three are scored separately and never blended into one number.
- Lines of code
- 1.5K
Recent activity
All commitsCommits
No commits indexed yet.
Decisions
- This .to() is needed if the model has been moved to a device after being initialized (because the buproposed
- Since we use batch_size * p as the new batch size, the latency for llm-analysis assumes the embeddinproposed
- Use cost analysis tooling with llm-analysis dependencyproposed
- Use parametric fitting code to fit the parallel scaling lawproposed
- Release specific inference and configuration code for ParScaleproposed
- Continual pretraining approach: freeze backbone and fine-tune new parametersproposed
Needs attention
20 open- medium severity. ProposedThis .to() is needed if the model has been moved to a device after being initialized (because the buAuto-proposed decision awaiting review
- medium severity. ProposedSince we use batch_size * p as the new batch size, the latency for llm-analysis assumes the embeddinAuto-proposed decision awaiting review
- medium severity. ProposedUse cost analysis tooling with llm-analysis dependencyAuto-proposed decision awaiting review
- medium severity. ProposedUse parametric fitting code to fit the parallel scaling lawAuto-proposed decision awaiting review
- medium severity. ProposedRelease specific inference and configuration code for ParScaleAuto-proposed decision awaiting review
Composition
Open the graphpython 80%markdown 20%
Explore this codebase
- Docs12 pages across 4 modules
- ChatAsk this codebase a question and get an answer with its sources
- Files5 files with per-file docs, health and history
- ArchitectureDependency graph, layers, and 0 entry points
- Code healthPer-file scores, 12 open findings, coverage and refactoring targets
- Knowledge graphEntities, communities, and the paths between them
- Change couplingFiles that keep changing together, mined from commit history
- CommitsChange-risk ranked history with agent provenance
- ContributorsBus factor, per-file maintainers, and the human/agent split
- StatsSize class, origin, lifetime churn, rhythm and records
- CostsWhat indexing this snapshot cost, by model and by run