Circle STARK for LLMs and classifiers¶
Hallmark proves with Circle STARKs: SASH for the quantized LLM and direct STWO for classifiers. The two paths share the proof family but prove different computations with different provers.
The current division¶
| LLM path | Classifier path | |
|---|---|---|
| Backend | SASH Circle STARK | Direct STWO 2.2.0 Circle STARK |
| Workload | Complete fixed quantized TinyLlama response | Rule, trained MLP and contextual byte CNN; 96-byte demo semantics |
| Proof structure | 23 linked segments | One fixed-gate proof per classifier/request |
| Verification | SASH; Linux and adapted Mac variants | Fresh native STWO processes; shared engine/AIR implementation |
| Setup/enrollment | Expected model, shape and calibration roots | Public preprocessing, program/model mapping and concrete parameters |
| Recorded proof files | 738 MB across 23 proofs | 1.91–2.16 MB canonical JSON per proof |
LLM/classifier artifact sizes cover different computations. A proof-size ratio between them is not a meaningful comparison.
Circle STARKs: a small-field route¶
Circle STARKs organize polynomial work over points on the circle \(x^2 + y^2 = 1\), enabling efficient domains for suitable fields, including the Mersenne prime \(2^{31}-1\). The construction changes how polynomial evaluation and interpolation domains are organized.
SASH turns its specialized quantized inference relation into constraints checked by its Circle STARK stack. Hallmark uses hash-pinned upstream source and command-line adapters, with changes for preparation, native Mac verification and CUDA proving. The constraints describe inference directly, without a general-purpose VM.
Read the original Circle STARKs paper by Haböck, Levit and Papini, then Hallmark's SASH integration. The construction's mathematical properties do not replace a security review of this implementation, parameters and composition.
A fair comparison needs the same claim¶
Before comparing prover backends, freeze:
- the same model, weights, input bytes, coverage and arithmetic;
- the same public inputs and caller binding;
- honest verification and mutation/rejection gates;
- hardware, threads, preprocessing and timing boundaries;
- proof, key, enrollment and transfer sizes.
The same-host GPU comparison¶
The H200 experiment compares SIMD and CUDA inside the same SASH relation with the same inputs, parameters and four Rayon threads. Generated constraint kernels eliminate CPU constraint fallback. All 69 files match byte for byte.
CUDA took 53.94 s versus SIMD's 547.75 s for summed proving, a 10.15× ratio. Complete segment commands took 155.89 s versus 679.49 s, a 4.36× ratio. Witness/interaction generation stays on CPU. Only one matched pair was measured, so these timings do not describe run-to-run variation. The benchmark guide preserves the raw evidence and the earlier slower fallback configuration.
Continue with verify one request.