[DO NOT MERGE] Kernel Interview Track: curriculum + 6 'why is this slow' challenges (134-139)

[kunal-mansukhani]

⚠️ DO NOT MERGE. This is a new paid product (Kernel Interview Track), so it needs Kunal's review first. Merging to main would auto-deploy challenges 134–139 into the public list through deploy-challenges.yml, because the challenge format has no hidden/unlisted flag yet (see "Needed from infra").

Track: tracks/kernel-interview.json (+ tracks/README.md schema)

enabled: false, status: "draft", entitlement: "interview_pass", and pricing is a placeholder only (nothing in Stripe).

#	Module	Existing challenges (id/slug)	Slow-kernel round
1	Memory coalescing	1 vector-addition, 31 matrix-copy, 3 matrix-transpose, 63 interleave-arrays	134
2	Shared memory & tiling	9 1d-convolution, 10 2d-convolution, 28 gaussian-blur, 69 2d-jacobi-stencil	135
3	Reductions & warp primitives	4 reduction, 17 dot-product, 5 softmax, 16 prefix-sum, 13 histogramming, 125 scatter-add	136, 139
4	Occupancy & latency hiding	50 rms-normalization, 113 layer-normalization, 83 fused-residual-add-and-rms-norm, 40 batch-normalization	137
5	Matrix multiplication	2 matrix-multiplication, 22 general-matrix-multiplication-gemm, 30 batched-matrix-multiplication, 57 fp16-batched-matrix-multiplication, 99 grouped-gemm	138
6	Attention & inference	6 softmax-attention, 53 causal-self-attention, 80 grouped-query-attention, 86 paged-kv-cache-attention, 128 split-kv-attention-reduction, 115 fused-qkv-projection-with-rope-and-kv-cache-update	n/a

Each module has an objective and a one-line interview tip.

6 new "why is this slow" challenges (medium, access_tier="premium" as a placeholder)

The starter starter.cu is a deliberately slow but correct kernel. This is an intentional exception to the "starter must not solve" rule. The other 5 language starters are empty, per convention.

id	Challenge	Bottleneck	Target speedup	Perf test
134	Speed Up Column Scale	uncoalesced_loads	4.0×	8192×8192
135	Speed Up Tiled Transpose	bank_conflicts	1.5×	8192×8192
136	Speed Up Row Sum	warp_divergence	3.0×	16384×4096
137	Speed Up Segment Reverse	low_occupancy	4.0×	N=67.1M, S=8192
138	Speed Up Matrix Multiply	missing_tiling	2.5×	4096³
139	Speed Up Histogram	atomic_contention	4.0×	N=50M, 32 bins

New proposed Challenge attributes (ignored by the current validator and deploy script): track, problem_type="slow_kernel", bottleneck, target_speedup, baseline_language="cuda", hints (3 progressive).

Verification

• black / isort / flake8 / clang-format are clean, and validate_challenges.py reports 0 errors.
• For every example and functional case, CPU checks show reference_impl matches a float64 ground truth within atol/rtol, AND a numpy emulation of each slow starter's thread mapping (and of the reduction tree) matches float64. So the baselines are correct.
• Not run on a GPU. No GPU or API key on the box, so the speedup targets are uncalibrated estimates for T4.

Needed from infra / frontend

1. The backend must enforce the speedup: time the user's solution and the baseline starter.cu on the same GPU for the perf test (median of N runs, or a cached baseline per GPU). Pass = correct AND baseline_ms/user_ms ≥ target_speedup. The format can't express this today.
2. Extend the update_challenges.py payload with track, problemType, bottleneck, targetSpeedup, baselineLanguage, hints, and a listed/hidden flag that the public list respects. This must land before merging this PR.
3. Track API per frontend-expectations.md: GET /api/v1/tracks/kernel-interview → {enabled, track:{…this JSON + solved/best_speedup}, progress, entitlement}. The profile endpoint's bottleneck.code uses the same enum and targetSpeedup.
4. Calibrate the targets on T4 before launch.

Full contract: /workspace/interview-track/curriculum-contract.md (mirrored in tracks/README.md).

[kunal-mansukhani]

Updated Interview Track target speedups to Infra's measured T4 values (median of 9 runs, Modal CI env; see infra #448): 134 4.7x, 135 2.0x, 136 3.8x, 137 5.6x, 138 1.3x, 139 6.0x. Targets are about 0.8x each demo's measured speedup; 139 capped at 6x (measured 18.27x) for interview fairness. Still a draft; not merging.