Android / SM8750: same-frame HTP + GPU causal prefill with Xiaomi 15 evidence #17

Could Jetlink support an explicitly experimental SM8750 processor that overlaps independent HTP and GPU work within the same model frame?

I opened a companion implementation for review: #16

On Cinque Terre V3, historical attention queries cannot read current image tokens. The prototype computes HTP vision and GPU processing of 256 historical tokens in parallel, then decodes 32 current queries on HTP against all 288 original keys/values and assembles the original outputs/recurrent queues on CPU. It keeps the current-frame feedback contract and original weights.

Measured on a Xiaomi 15 (SM8750 / Snapdragon 8 Elite, HTP v79, 12 GB RAM, Android 17), ORT 1.29.0 and matched QAIRT 2.47:

The sustained run processed 12,001 frames with zero above 50 ms. The unchanged original CPU accuracy gates passed all 128 identical synthetic frames before and after the run; both sides maintained their own recurrent state.

Reproduction and build instructions · Full accuracy logs, timing windows and conditions

Important limits: hardware results are from the prototype. The separate port onto current main builds as both default and opt-in Android APKs, passes 59 Android tests per variant, 33 targeted Swift tests and eight adapter tests, and reproduces all four graph hashes plus manifest exactly. The final port APK has not been rerun on the phone. Physical comma USB, car power and road behavior are unvalidated. This is not a general claim for all Snapdragon devices.

Tests used PC USB power with airplane mode on and Wi-Fi/Bluetooth off for stability. Battery temperature rose 36.6 → 45.4 °C and charge fell 62% → 55%, despite all sampled Android thermal statuses being 0. Sequential controls were not randomized or temperature/clock locked. Approximately 22.1 ms of provider-call overlap does not establish hardware occupancy or 100% NPU utilization.

The PR keeps existing defaults, gates the new option to SM8750 plus a build opt-in, verifies the exact model bundle, and includes no SDK binaries/headers, model weights or signing material. Feedback on upstream integration and the hardware/accuracy validation matrix would be appreciated.