v0.7.1 — --cuda-graphs surfaced with measured A100+A10G perf

v0.7.1 — surfaces --cuda-graphs with measured perf

The reflex serve --cuda-graphs flag has been silently in main since v0.5.0 (commits 49110d2 → 7dfcab6, April 24-25). v0.7.1 surfaces it with the Day 8-9 A/B numbers from Modal so users actually know to enable it.

Measured perf

A100-80GB (pi0.5 num_steps=10, n=200 paired iterations):

session	OFF mean	ON mean	speedup	OFF p99	ON p99	jitter Δ
vlm_prefix	93.45 ms	87.34 ms	1.07×	94.33 ms	87.84 ms	1.4× tighter
expert_denoise (10× / chunk)	17.74 ms	12.04 ms	1.47×	20.91 ms	12.44 ms (-40%)	4.1× tighter

Combined per-chunk: 270.85 → 207.74 ms = 1.30× speedup, p99 304 → 212 ms (-30%).

A10G (n=100 paired): expert_denoise 1.18× mean, p99 -35%, jitter 14× tighter. vlm_prefix capacity-bounded — gracefully degrades to eager per ADR tier-aware semantics.

Win shape: predictability > raw mean

Robotics control loops care about p99 + jitter more than mean. CUDA graph replay eliminates per-op kernel-launch jitter, which is why the jitter improvements are larger than the mean improvements. expert_denoise runs 10× per chunk in flow-matching denoise, so the inner-loop jitter reduction compounds.

Hardware tier guidance

• A100 / H100: capture both sessions
• A10G: capture expert_denoise only; vlm_prefix gracefully degrades
• Orin Nano / smaller: not yet measured. Run reflex doctor after reflex serve --cuda-graphs to check capture status.

Validation

• Day-0 spike: 2026-04-25 (atol=1e-6 parity verified)
• Day 8-9 A/B Modal A10G: $2 — see reflex_context/03_experiments/2026-04-29-cuda-graphs-ab-modal-a10g.md
• Day 8-9 A/B Modal A100: $1.50 — see reflex_context/03_experiments/2026-04-29-cuda-graphs-ab-modal-a100.md

5/6 ship gates clear with margin (CI no-overlap, mean improvement, p99/p99.9 caps, jitter cap). Parity gate carried forward from Day-0 spike.

Added (this release)

• tests/test_cuda_graphs_integration.py — Day 7 integration tests (7 mock + 1 CUDA-gated)
• scripts/modal_cuda_graphs_ab.py — Day 8-9 A/B benchmark with 30s progress logging
• CHANGELOG entry surfacing the flag with measured numbers

Notes

• The --cuda-graphs flag was always live in reflex serve --help since v0.5.0. v0.7.1 simply documents it with measured numbers.
• Total Modal cost for the validation: $3.50.