Request Stable Rng Kernels#
- void trt_edgellm::kernel::launchRequestStableSpecUniforms(
- uint64_t const *requestSeeds,
- uint64_t const *nextAbsolutePositions,
- float *proposalUniforms,
- float *acceptUniforms,
- int32_t batchSize,
- int32_t proposalLen,
- cudaStream_t stream
Fill all per-round random streams in one launch. Seeds and absolute positions identify logical requests, so compaction cannot perturb them.