Month 2 — CSAPP finish, linear algebra, PMPP/CUDA proper, FA through the algorithm

Continues first-30-days.md (Track A, strong SWE / weak ML). Prefer that file over 05-interview/full-prep.md. Delta: full-prep's first 30 days is Karpathy micrograd → CS336; the delivered month 1 is CSAPP + take-home + metal first. Do not restart at micrograd. If you never did Karpathy lectures 1–6, park them on Sunday walks, not Deep 1.

Clock: operating-system.md (12h × 6 days). Deep 1 / Deep 2 / paper / Friday sport as below. Cite, do not re-list: 02-foundations/curriculum.md (CSAPP 7a, LA module 1, FP module 5), 03-kernels/curriculum.md Layer 0–1.

Month goals

Finish CSAPP Ch 4, 7–9 (and any leftover 5/8) with practice problems, not highlights. Cache + malloc labs if you have them.
Linear algebra you can feel: Trefethen & Bau Lectures 1–12 or Axler Ch 1–5 (pick one book; stay). SVD stated, thin SVD by hand for a 4×3.
PMPP 4e Ch 1–6 (or 5e equivalent) + CUDA C++ Programming Guide programming model. Tiled GEMM you profile. Triton GEMM that matches it.
FlashAttention (Dao 2022) through the algorithm box + §2 IO. Online softmax from memory. No FA2/FA3 yet.
Take-home < 3,000 cycles, tests green, written changelog.

If month 1 exit had more than two unchecked boxes, spend week 1 as a patch week. Do not "go on to MoE."

Week 5 — CSAPP backend + LA start

Goal: linking / VM / malloc taste; LA started; take-home not ignored.

Day	Deep 1	Deep 2	Paper / book	Friday sport / Sat
Mon	CSAPP Ch 7 (linking). `objdump` your month-1 C timing program.	Take-home: one new trick or a cleaner scheduler. Measure.	None.
Tue	CSAPP Ch 8 (ECF) 90 min then Ch 9 start (VM, page tables).	LA: Trefethen 1–3 or Axler Ch 1. All exercises in the sitting.	None.
Wed	Profile / debug day: month-1 Triton GEMM or fused RMSNorm. No new papers. Make it correct and measured.	CSAPP malloc chapter problems, or start malloc lab.	Roofline reread (Williams 2009) against your GEMM numbers.
Thu	LA continues (Trefethen 4–6 / Axler Ch 2).	CUDA/PMPP Ch 1–2 if not solid from month 1; else occupancy sweep (`03-kernels` L1 ex. 2).	FlashAttention §1 only (motivation, IO as the problem).
Fri	Take-home sport (5h). Target < 4,000 if you ended month 1 at ~5k; < 3,500 if you were already under 5k.	—	—	Changelog.
Sat	Long build: CSAPP cache lab or a from-scratch bump+free-list allocator (`02-foundations` L0 / module 7a).	—	—	Writeup.
Sun	Off-ish. Anki.

Week 5 exit: CSAPP 7–9 started, LA book chosen and used, take-home moving.

Week 6 — Linear algebra that is not a meme + FA algorithm

Goal: attention is linear algebra + a softmax. You can say SVD without irony.

Day	Deep 1	Deep 2	Paper / book
Mon	Trefethen 7–9 (QR, conditioning) or Axler Ch 3. Problems.	Re-derive attention + online softmax, blank page (25 min camera). Then backprop-through-softmax on paper.	None.
Tue	LA (Trefethen 10–12 / Axler inner products).	Numpy: implement Householder QR or a thin SVD path; residual vs naive GS in f32 vs f64 (`02-foundations` module 1 ready-when).	FlashAttention algorithm box. Copy it by hand. Mark what is exact vs approximate.
Wed	Debug day: your month-1 attention kernel vs `F.scaled_dot_product_attention` on more shapes.	Take-home.	FA §2 IO analysis. Reproduce the IO complexity on paper.
Thu	PMPP Ch 3–4 (memory, tiling).	Tiled GEMM: bank-conflict microbench (`03-kernels` L1 ex. 3).	Goldberg / IEEE-754 reread if Ch 2 is rusty.
Fri	Take-home. Target < 3,200.
Sat	Long build: write `notes/fa1-algorithm.md` — tiling, online m/ℓ, what is stored for backward. No CUDA FA yet.
Sun	Off-ish.

Week 6 exit: FA algorithm from memory, LA through conditioning or inner products, GEMM still getting faster.

Week 7 — PMPP / CUDA proper (not SAXPY tourism)

Goal: Layer 1 start in 03-kernels/curriculum.md. Occupancy, coalescing, SMEM, a profiled tiled GEMM.

Day	Deep 1	Deep 2	Paper / book
Mon	PMPP Ch 5–6 + CUDA programming model (current guide).	Rewrite tiled GEMM from a blank file. Time it. % of peak.	None.
Tue	Shared-memory swizzle / padding. Measure bank conflicts.	Triton GEMM tutorial (`03-kernels` L2 resource 6) — type it, do not paste.	Tillet et al. 2019 Triton paper as BG; current docs are the language.
Wed	`ncu` or `torch.profiler` on naive vs tiled vs Triton. Write the limiter in one sentence.	CSAPP leftover chapter you skipped.	None.
Thu	LA catch-up (eigen / SVD intro).	Take-home.	Kaplan 2020 or Chinchilla — the figure and the claim, 90 min.
Fri	Take-home. Hold < 3,000 or write the next 1.3× on paper if you miss.
Sat	Long build: occupancy vs performance sweep (`03-kernels` L1 ex. 2). Plot.
Sun	Off-ish.

Week 7 exit: one GEMM you can defend with counters; Triton sibling; PMPP 1–6 actually worked.

Week 8 — Close month 2

Goal: ship. CSAPP problems exist. FA algorithm exists. Take-home < 3k.

Day	Deep 1	Deep 2	Paper / book
Mon	CSAPP leftover problems (Ch 5 optimizing performance if thin).	Fuse RMSNorm+residual if month 1's is ugly; else start a correct standard attention in Triton (not Flash).	None.
Tue	Blank-page FA algorithm + IO.	Tiny GPT train continues (`from-scratch/gpt.py`). Loss must still go down.	Vaswani §3 reread if shapes slipped.
Wed	Debug / measure day.	Take-home polish.	None.
Thu	LA problem set you owe.	Interview drill (`05-interview/drills.md`): 60 min incremental system.	CAI paper: flowchart from memory (preview of month 3).
Fri	Take-home sport. Official month target < 3,000.
Sat	Month 2 artifact (below).
Sun	Off.

Month 2 exit criteria

CSAPP Ch 2, 3, 5, 6, 7, 9 worked (problems). Ch 4/8 at least read.
LA book: through SVD statement and one computed thin SVD.
PMPP 1–6 + one profiled tiled GEMM + Triton twin.
FlashAttention 2022 algorithm + IO argument reproduced on paper.
Online softmax + attention backward still cold (month 1 must not decay).
Take-home < 3,000, tests green, changelog.
Tiny GPT still trains (do not abandon the repo).

Artifact

Public commit: GEMM worklog (TFLOPS table vs a library baseline on your GPU), notes/fa1-algorithm.md, take-home cycle number, 800–1500 word writeup titled as the Greenhouse paragraph so far.

Do not start FA1-in-Triton until the algorithm page is something you can rewrite from a blank sheet. That is month 3.