How mikamika06/mlsys-lab works
mlsys-lab is an offline, auto-graded low-level ML systems learning platform. Learners choose a task or multi-file project, copy its starter into a work directory, run it, and submit it for deterministic grading against computed oracles and metric gates. The bank spans Python, C++, CUDA, quantization, inference, memory, and performance topics; a VS Code extension is also documented. The supplied tree and excerpts omit the main package implementation, so CLI, grader, and simulator internals are not mapped; documented workflow links are cited to the README.
Main entry points: cache.py Hybrid KV cache · ngram/engine.py N-gram lookup · cache_engine.py Prefix hashing · runner/engine.py Constrained output
Components 20
Learning bank
- Task bank
- Project bank
- Project starter
projects/p-hybrid-sliding-window-cache/skeleton/kvcache/cache.py - Run task
- Grade submission
- Computed oracle
- Metric gates
Inference projects
- Batching simulator
- Hybrid KV cache
projects/p-hybrid-sliding-window-cache/skeleton/kvcache/cache.py - Cache configuration
projects/p-hybrid-sliding-window-cache/skeleton/kvcache/config.py - N-gram lookup
projects/p-ngram-prompt-lookup/reference/ngram/engine.py - Prefix hashing
projects/m-implement-vllm-s-prefix-block-hashing-scheme/reference/prefix_cache/cache_engine.py - Constrained output
projects/p-structured-output-contract-breaks/reference/runner/engine.py - JSON grammar
projects/p-structured-output-contract-breaks/skeleton/runner/grammar.py
Systems analysis
- Quantization parity
projects/m-ovquantizer-one-shot-vs-manual-nncf-quantize-output-/reference/quantizer/pipeline.py - Opset validation
projects/m-ovquantizer-one-shot-vs-manual-nncf-quantize-output-/skeleton/quantizer/opset.py - Layout analysis
projects/m-transpose-overhead-from-layout-mismatch/reference/edgeio/pipeline.py - Layout transforms
projects/m-transpose-overhead-from-layout-mismatch/skeleton/edgeio/layout.py
Editor integration
- VS Code extension
editor/extension.js
Other components
- Learner
Connections 16
- Learner → Task bank: chooses task
README.md - Learner → Project bank: chooses project
README.md - Project bank → Project starter: provides starter
- Learner → Run task: runs work
README.md - Learner → Grade submission: submits work
README.md - Run task → Grade submission: grades separately
README.md - Grade submission → Computed oracle: computes reference
README.md - Grade submission → Metric gates: checks metrics
README.md - Batching simulator → Hybrid KV cache: models KV cache
- Hybrid KV cache → Cache configuration: uses configuration
projects/p-hybrid-sliding-window-cache/skeleton/kvcache/cache.py - N-gram lookup → Task bank: represents feature
- Prefix hashing → Task bank: represents feature
- Constrained output → JSON grammar: applies mask
projects/p-structured-output-contract-breaks/reference/runner/engine.py - Quantization parity → Opset validation: validates opset
projects/m-ovquantizer-one-shot-vs-manual-nncf-quantize-output-/reference/quantizer/pipeline.py - Layout analysis → Layout transforms: converts layout
projects/m-transpose-overhead-from-layout-mismatch/reference/edgeio/pipeline.py - Learner → VS Code extension: opens workspace
README.md