feat(cuda-oxide): RoPE (adjacent-pair) pure-Rust #[kernel] port + gx10 sm_121 A/B vs hand-PTX (PMAT-921)

[noahgift]

PMAT-921 — cuda-oxide RoPE (adjacent-pair) port: GB10 sm_121 A/B vs hand-PTX

Advances the cuda-oxide marquee track (pure-Rust #[kernel] → PTX on Blackwell). Ports the standard adjacent-pair RoPE apply kernel (hand-PTX RopeKernel, entry rope, in crates/aprender-gpu/src/kernels/elementwise/rope/standard.rs) to a pure-Rust cuda-oxide #[kernel] and runs a true on-device A/B vs the hand-PTX baseline on the GB10 (sm_121 / compute_cap 12.1). Mirrors the established PMAT-882/893/894 port pattern (attention / RMSNorm / SwiGLU — all GO).

Why RoPE (per the rule)

RoPE is decode-hot (applied to Q and K every layer, every token) and is pure f32 FMA + sin/cos/ex2 transcendentals (ZERO DP4A) — squarely the established GO class. The rule from the prior verdicts: port FMA/softmax/elementwise/transcendental kernels; keep DP4A/INT8-bound Q4K GEMV/FFN (PMAT-881, the lone NO-GO) on hand-PTX. RoPE was the next-best un-ported candidate.

VERDICT: GO (statistical tie) — F-OXIDE-ROPE-PARITY-001

PARITY vs an f64 CPU reference — PASS at every shape (incl. Qwen2.5 high-freq theta=1M):

heads	head_dim	theta	cos	maxdiff (A / B)
32	128	10000	1.0000000	1.07e-6 / 6.86e-7
14	128	1000000	1.0000000	2.83e-6 / 2.83e-6
128	128	10000	1.0000000	1.65e-5 / 7.47e-6

PERF — FAIR matched single-launch A/B vs the real hand-PTX rope (RopeKernel::emit_ptx_for_target("sm_121"), committed in baseline-ptx/), same data + same grid (num_heads × head_dim/2) + GPU-event median 5×100. Gate oxide_us/handptx_us ≤ 1.2 = GO:

heads	head_dim	theta	oxA µs	oxB µs	hand-PTX µs	best ratio	verdict
32	128	10000	2.070	2.073	1.863	A 1.111×	GO
14	128	1000000	2.032	2.052	2.030	A 1.001×	GO
128	128	10000	2.215	2.159	2.041	B 1.058×	GO

Best ratio swings 0.98×–1.12× over 4 runs, centered ~1.00–1.05× — a clean statistical tie at the DRAM-bandwidth roofline (~2 µs both sides), exactly as predicted for an elementwise trig kernel. PASS at every shape.

Honest framing

This is a tie, not a speedup, and is reported as one. Adjacent-pair RoPE is memory-bandwidth-bound (1 load + 1 store per element), so neither hand-PTX nor oxide can beat the wall. The GO value: the pure-Rust #[kernel] matches the hand-PTX with no perf loss, so this decode-hot kernel can be migrated off hand-PTX (and off the GH-480 Blackwell-JIT workaround) for free — another north-star datapoint where pure-Rust→PTX replaces hand-PTX. Production migration of the live trueno RopeKernel is a follow-up, not this PR.

What's in the PR (no CI impact)

The spike is an isolated crate with its own [workspace] and the cuda-oxide toolchain (nightly-2026-04-03 + LLVM-21 + cargo-oxide 0.2.1) — it builds ONLY on gx10 and never affects the normal cargo build / CI.

• experiments/cuda-oxide/rope/ — spike crate (src/main.rs, 2 variants) + 2 committed sm_121 hand-PTX baselines + RESULTS.md
• contracts/cuda-oxide-rope-parity-v1.yaml — F-OXIDE-ROPE-PARITY-001..005, pv validate clean, full-dir pv lint PASS (0 errors)
• evidence/pmat-921-cuda-oxide-rope-2026-06-24/ — findings.json + ab_run.txt (GPU-evidence archive, per convention)
• experiments/cuda-oxide/README.md — adds the rope/ row (+ backfills the missing swiglu/ PMAT-894 row)

Reproduce on gx10

ssh gx10
export PATH="$HOME/.cargo/bin:/usr/lib/llvm-21/bin:$PATH"
export LLVM_SYS_211_PREFIX=/usr/lib/llvm-21
cd experiments/cuda-oxide/rope
cargo oxide run   # parity (2 variants × 3 shapes) + fair matched-launch A/B

🤖 Generated with Claude Code