PHASE3_CYSIEVE_GPU_FINDINGS.md

CY-Sieve GPU-Phase Findings (2026-06-22, COMPLETE)

Status: complete. The quality gate returned a NEGATIVE result (KILL). On real WikiText-2, trained from scratch, CY-Sieve's positional bias is +10.15% worse than the best baseline at the training context — past the >5% kill threshold. The kernel is correct (§4 PASS) and its memory claim is confirmed (§6, 8192× less bias HBM), but the positional scheme fails its quality gate and is reported as a negative result, not shipped — exactly as tests.md requires. Raw artifacts: 4_ai_hardware_attention/gpu_phase_runs/run3_20260622_*.

Hardware: NVIDIA L4 (24 GB), PyTorch 2.9.1+cu129, Triton 3.5.1, project SocrateAI. Driver: run_gpu_phase.py → tests.md §4 (parity) / §5 (quality) / §6 (perf).

---

§4 — Kernel correctness: PASS (final)

The Triton CY-Sieve kernel (cy_sieve_triton.py) reproduces the CPU reference (cy_sieve_attention.flash_sdpa_with_bias, itself ~3e-16 from dense softmax) to within FP16 tolerance (rel-err < 2⁻⁸): 4/4 tests passed across (L, D, τ) ∈ {(128,64,20), (256,64,128), (256,128,480)} + a smoke test.

Reading. The fused GPU path — online-softmax with the per-distance bias (Tier-1 exact INT64 table + Tier-3 log penalty) generated inside the kernel — computes the same attention as the dense reference. The on-the-fly bias is implemented correctly; downstream measurements test the design, not a bug. (block_n=32, num_stages=1 keep SMEM under the L4's ~100 KB.)

§6 — Performance: core memory claim holds; latency claim does not (final)

L4, D=64, fp16, τ=128, causal (run 1 measured 1K–16K before an OOM at 32768², now guarded):

L	CY-Sieve (ms)	dense SDPA (ms)	table-bias SDPA (ms)	bias-HBM reduction
1024	0.055	0.054	0.060	512×
2048	0.105	0.053	0.116	1024×
4096	0.260	0.060	0.268	2048×
8192	1.084	0.288	0.805	4096×
16384	4.156	1.016	3.938	8192×
32768	15.28	2.51	(skipped: >VRAM)	16384×

(run-3 numbers; the materialized-table baseline is skipped at 32768 because its [L,L] tensor exceeds the VRAM budget — itself a demonstration of the O(L²) wall that CY-Sieve's O(L) bias avoids.)

• ✅ Bias-path HBM — the central thesis — confirmed and measured. CY-Sieve's positional bias reads O(L) bytes (a length-L vector from the recurrence) vs O(L²) for a materialized bias table: 8192× less at L=16384 (64 KB vs 512 MB), and the gap widens with context length.
• ❌ Not yet a wall-clock win. The current (unfused) Triton kernel is ~4× slower than fused dense SDPA at 16K (4.16 vs 1.02 ms) and ~6× at 32K (15.3 vs 2.5 ms), only matching the table-bias SDPA. The HBM saving is real but not yet converted into latency — cuDNN's SDPA is far more tuned. Reported as-is per the tests.md T6.3 honesty guard; speed is never presented without this caveat.
• Moot given §5: with §5 = KILL, T6.3 forbids presenting these speed/HBM numbers as a contribution at all. They are recorded as engineering measurements only.

Net: the win is memory traffic at long context, not raw speed. The case is for bandwidth-bound / very-long-context regimes; turning the HBM saving into time requires fusing the bias deeper into the FlashAttention loop.

§5 — Quality gate: KILL (negative result) (final, real WikiText-2)

The decisive test, done the only honest way — train small GPTs from scratch, identical arch/data/compute, one per scheme — on real WikiText-2 (Salesforce/wikitext), byte-level. Validation perplexity at the training context (512) and at 2×/4× extrapolation:

scheme	@512 (train)	@1024 (2×)	@2048 (4×)
learned-absolute	4.22	12.10	20.82
ALiBi	10.74	11.73	11.35
sliding-window	4.99	5.07	5.03
CY-Sieve τ-ladder	11.33	12.31	12.05
CY-Sieve τ=20	16.02	16.81	16.49
CY-Sieve τ=128	6.80	7.12	7.00
CY-Sieve τ=512	4.65	6.08	10.62

Verdict: KILL. Best baseline (learned-absolute) = 4.22; best CY-Sieve (τ=512) = 4.65 → +10.15%, well past the >5% kill threshold. CY-Sieve as designed does not meet the quality bar.

Honest reading:

• The "great extrapolation" seen earlier was an artifact of the trivially-memorized synthetic corpus (ppl≈1.0 everywhere — a bug, since fixed). On real text the picture inverts.
• Sliding-window is the clear winner here — 4.99 essentially flat across 512→2048 (best extrapolation and near-best train ppl). A plain local window beats every CY-Sieve variant.
• CY-Sieve exposes a real tension: the geometry-fixed steep decay (log λ=3.762) is too aggressive — no single τ gives both good absolute quality and stable extrapolation. τ=512 is best at the train context but degrades 4.65→10.62 at 4×; τ=128 is stable but +61% worse than the baseline; the τ-ladder meant to resolve this lands at ~11–12 (poor on both axes).
• This is a genuine negative result for the positional scheme, reported per the project's own rules. The HBM advantage (§6) is real but irrelevant if quality fails — a fast kernel that hurts the model is a failed kernel.

What this does NOT kill: §4 correctness and the §6 O(L)-vs-O(L²) HBM property stand on their own. The negative is specifically about this bias function as a drop-in positional scheme. See "Improvement directions" below.

Improvement directions (post-KILL)

The kill is informative, not terminal — it points at specific, testable fixes:

1. Decouple the slope from the geometry. logλ=3.762 is the growth rate of S₂₀, but there's no law that the best attention slope equals it. Treat the per-head slope as a learnable/tunable parameter initialized from the geometry, à la learnable-ALiBi — the geometry sets the prior, not the value.
2. Hybrid with a local window. Sliding-window won; combine an exact local window (Tier 1, which CY-Sieve already has) with a much gentler long-range tail, instead of one steep curve everywhere.
3. Re-examine the β=2 log-curvature contribution separately from the linear slope — ablate whether the log term helps or hurts at these lengths.
4. Only then is the O(L) HBM advantage worth re-claiming — and only if a redesigned bias clears the +5% gate.

---

Reproducibility

pip install -r 4_ai_hardware_attention/requirements-gpu.txt
python 4_ai_hardware_attention/run_gpu_phase.py     # §4 + §5 + §6 → one JSON

On GCP the same is driven unattended by 4_ai_hardware_attention/gcp_startup.sh (an L4 that clones, runs, uploads to GCS, and self-terminates).