🚧 Week 3: Build a Mini vLLM

🚧 This overview and chapters carrying the same marker are under review and may change.

Week 3 turns the optimized single-request model into a multi-request serving engine. Students add scheduling, request-owned cache state, shared page pools, and the runtime metadata needed to read noncontiguous K/V directly. The final model uses one page-aware attention interface. A correct direct page-walking implementation may serve every query shape; the completed reference adds a tiled schedule for the supported BF16 long-prefill hot case.

Week 2’s course-owned quantized projections remain the inspectable endpoint of that week’s kernel lessons. Week 3 deliberately switches dense-model projections to mx.quantized_matmul at model construction, while retaining the course-owned normalization, activation, cache, attention, paging, and scheduler paths. This keeps Week 3 focused on serving-system mechanisms rather than carrying the teaching kernel’s projection cost through every benchmark.

What We’ll Cover

• Continuous batching and request-slot reuse
• Chunked prefill and scheduler fairness
• Paged KV storage and page-walking attention
• Paged FlashAttention for long prefill
• Optional Day 6 Mixture-of-Experts model support
• Optional Day 7 speculative decoding over rewindable caches

Day 1 introduces that projection seam and batches independent request states. Day 2 splits long prefills so they cannot monopolize the scheduler. Day 3 replaces a growing dense cache with fixed-size pages while retaining a dense-gather compatibility path. Day 4 removes that gather by teaching attention to walk the page table directly with a correctness-first implementation; one kernel may serve all query shapes. Day 5 then replaces the supported BF16 long-prefill hot case with a tiled schedule. Page translation is therefore introduced before it is optimized.

These five days form the required path in your solution. The final model in your solution runs paged FlashAttention for supported BF16 long prefill and retains correct direct fallbacks for short queries and generic shapes. Every schedule reads the same page pool through the same block-table interface; none rebuilds dense K/V.

Paged attention is not an automatic single-request latency win. The checked trace measures lower KV storage, page reuse, incremental growth, and batching; page-table indirection can make one request slower. It does not establish an admission-capacity gain without a memory-capped sweep. Each chapter ends with a focused measurement, while the performance appendix records the matched chapter-by-chapter results.

Optional Day 6 adds MoE model support independently of the cache and scheduler. Optional Day 7 then adds speculative decoding, whose rejection path needs a precise cache rewind operation and whose multi-token verification needs the page-aware long-query path. Neither extension is required to complete Week 3.

Your feedback is greatly appreciated. Join our Discord community.
Found an issue? Open an issue or pull request at github.com/skyzh/tiny-llm.
tiny-llm-book © 2025 by Alex Chi Z is licensed under CC BY-NC-SA 4.0.