Answers are grounded in this paper and cite their sources.
Language models can draft radiology impressions, but a 4-bit copy of a model does not always reproduce the output of its full-precision version. TRIASE-SLO routes each request between FP16 and AWQ INT4 copies of one LoRA-adapted Qwen2.5-3B-Instruct model. A frozen rule-based hazard score computed from the FINDINGS text sends every request scoring at least 0.5 to FP16, and a secondary latency-aware score places the remaining requests. We evaluated the design in a locked confirmation in which every arm reserved the same two A100 MIG slices: four arms, seven loads and 12 paired deployment blocks of 450 previously inspected test reports (151,200 requests, none failed). The gate improved agreement with FP16. With full TRIASE-SLO routing, agreement with finding-bearing FP16 references was 99.11–99.49%, compared with 66.15–82.31% under hazard-blind routing, while 34–48% of requests still used INT4. A consistent serving-efficiency advantage was not demonstrated: full routing did not improve tail latency or throughput over two FP16 workers (median p95 change +0.33%; slower at C = 2 after Holm correction), and a consistent speed advantage of INT4 appeared only under hazard-blind routing. The demonstrated benefits of the INT4 copy are a smaller checkpoint and a larger KV-cache capacity. The incremental value of the score remains uncertain, because its prespecified C = 8 throughput benefit over gate-only routing (+1.86%, 95% CI −3.27% to +7.27%) was not confirmed. No INT4-using arm met the zero-drift engineering floor. All quality measures are rule-based FP16-consistency proxies from one public dataset, not evidence of clinical correctness or safety.
No takes yet. Share an insight, caveat, or question.
Zeniarja et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: