fsdatalab/quail

Quail refuses ORDER BY, OFFSET, GROUP BY, DISTINCT, and plain column predicates in WHERE. Users apply them on the client or in the source database. Semantic top-k, the most wanted query, cannot be written: SELECT d.id, AI.SCORE(PROMPT('Is this a refund request? {0}', d.body)) AS s FROM documents d ORDER BY s DESC LIMIT 10 Plan Three PRs, grouped by the machinery they share. Everything runs on the coordinator in Arrow; the GPU executor changes only where noted. Only the Quail planners (standard and reranker) get new nodes. The stock vLLM and SGLang planners refuse the new clauses, so a query with one runs on Quail only. PR 1: ordering and the limit rule. #187. ORDER BY, OFFSET, top-k, and DISTINCT, since all four are fields on the root Project and one coordinator node. Sort keys, offset, and a distinct flag become fields on the root Project, beside LIMIT. One Sort node: select-k when there is a LIMIT, full sort otherwise. DISTINCT is an Acero aggregate with no functions in the same nod

Goal Add two devices, l40s and b200. On quail-b at scale factor 0.1, show that each one runs the queries with the right answers, at the speed its specs predict. Then run a few queries on 8×L40S to check multi-GPU scaling. This is an onboarding ticket. It touches the device specs, the planner's cost inputs, the executor's kernels, the Modal setup, and the benchmark runner. Background Architectures. L40S is Ada, CUDA capability 8.9, with 48 GB. B200 is Blackwell, capability 10.0. Tensor-core precisions. FP8 runs natively on all three GPUs. H100 (Hopper), L40S (Ada), and B200 (Blackwell) all have FP8 tensor cores, each at twice its BF16 rate. So the fp8 checkpoints Quail uses run on each one without upcasting. The fastest FP8 kernels are Hopper-only. DeepGEMM, FlashAttention 3, and CUTLASS's best block-scaled FP8 kernels use wgmma (Hopper's asynchronous warp-group matrix multiply) and TMA (its asynchronous memory copy engine). Ada has neither. On L40S, FP8 runs through synchronous mma