Dual RTX 3090 Qwen3.8-27B SGLang Speculative Decoding Setup

Setup: Dual RTX 3090 Qwen3.8-27B SGLang 0.5.19 AutoRound INT4 FP8 KV cache DFlash2 with the z-lab BF16 block drafter 262K context Benchmark results (thinking off, 3 warm-up runs + 5 measured runs): Code decoding: 219.8 tok/s Narrative: 124.6 tok/s TTFT: 64 ms Code CV: 2.9% Narrative CV: 3.3% Prefill: ~10K tokens: 1,543 tok/s ~90K tokens: 1,195 tok/s Memory: GPU 1: ~22.5 GiB GPU 2: ~22.1 GiB Memory leak: ~20 MiB Comparison: Built-in MTP path (same quant): ~112 code tok/s Dead drafter path: ~39 tok/s DFlash2 block drafter: 219.8 code tok/s vLLM dual-3090 comparison: Dual-superfast: ~141 code tok/s Ultrafast: ~231 code tok/s (but switches to BF16 KV cache, gives up full 262K context) Caveats: SGLang doesn't emit vLLM SpecDecoding line, so acceptance length couldn't be scraped directly. With thinking on, visible chat output appears slower (reasoning tokens included in generation).