Skip to content

Week 2 Day 1: KV cache and bounded capacity - #317

Merged
skyzh merged 5 commits into
mainfrom
atlas/week2-day1-kv-capacity
Sep 24, 2026
Merged

skyzh merged 5 commits into
mainfrom
atlas/week2-day1-kv-capacity

Conversation

@skyzh

@skyzh skyzh commented Sep 24, 2026 •

Copy link
Copy Markdown
Owner

Scope

Extract Week 2 Day 1 from the reviewed five-day Draft #316. The active route now teaches request-owned KV cache and bounded capacity together. Starter and reference checkpoints expose only kv-cache and capacity-cache; the book, CLI, supplied tests, and shipped-day CI use the same two-checkpoint contract.

The former Week 2 lesson URLs remain available as clearly labeled historical pages. The M1–M4 bandwidth and roofline example is preserved, with the Mac Studio pairing dated to Apple's 2025 specifications. W4, SIMD, fused operators, and tiled attention are reserved for later day-scoped PRs.

Local evidence

  • The Day 1 author gate on the parent candidate passed 329 reference tests with 1 skip; the focused Day 1 reference file passed 6/6. The supplied starter tests collect and fail at named learner TODOs, including the capacity logical-prefix and rewind seams.
  • The learner extension builds; mdBook, sitemap, rendered local links, changed-file checks, and documented command parsing pass. The final one-page historical-source successor passed fresh factual review without changing code, tests, or CI.
  • Local factual, correctness, learner, and static accessibility reviews were completed on their stated exact commits. Those verdicts do not approve this public PR identity or its hosted checks.

Review and merge gates

This PR is Draft. The macOS reference check passed at this exact head in run 35941542257; fresh public-head reviews and all other live gates remain required before Ready or merge. The inherited reference-native Metal build failure is explicitly deferred to Day 2; Day 1 does not claim that build passed. The separate Qwen3-4B screen on integrated Draft #316 found no selected-versus-all-off regression under a frozen replay, but three rows reached only 66%–77% of full MLX. This Day 1 PR makes no speed claim.

The existing main-branch workflow deploys GitHub Pages on push. Merging this PR would therefore trigger publication; merge must wait for an explicit publication decision or a separately reviewed workflow change, in addition to the course review and repository gates.

@skyzh
skyzh marked this pull request as ready for review September 24, 2026 02:12
@skyzh
skyzh merged commit 071eee1 into main Sep 24, 2026
1 check passed
@skyzh
skyzh deleted the atlas/week2-day1-kv-capacity branch September 24, 2026 02:14
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