Skip to main content

Module gpu_inference_providers

Module gpu_inference_providers 

Source
Expand description

Stage S6 (VS03-12) GPU compute (EtalonGpuComputeProvider, re-exported from etalon-vibe) and governed inference (GovernedInferenceProvider) extension providers with canonical contract digests and direct-call reference oracles.

Structs§

EtalonGpuComputeProvider
Zero-heap Etalon GPU/Tensor compute provider (https://qualiadb.org/ext/etalon/gpu).
GovernedInferenceProvider
Governed inference extension provider (VS03-12).

Functions§

oracle_governed_token_step
Direct-call oracle for infer::decode::governed_token_step.
oracle_gpu_affine_f32
Direct-call oracle for gpu::compute::affine_f32.
oracle_gpu_gemv4_f32
Direct-call oracle for gpu::compute::gemv4_f32.
oracle_gpu_rmsnorm_f32
Reference oracle for (0, 3) rmsnorm_f32: y[i] = x[i] / sqrt(mean(x^2) + eps) * weight[i].
oracle_gpu_rope_f32
Reference oracle for (0, 4) rope_f32: 2D rotary position embedding on pairs (0,1) and (2,3).
oracle_gpu_topk2_f32
Direct-call oracle for gpu::compute::topk2_f32.
oracle_rms_norm_f64
Direct-call oracle for infer::activations::rms_norm_f64.
oracle_softmax_f64
Direct-call oracle for infer::activations::softmax_f64.