DSpark speculator trained on live SGLang lifts decode throughput 1.89× on B200s
ying11231 · x · 2026-07-23
A team says it trained a DSpark speculator for Inkling NVFP4 using SpecForge against a live SGLang serving engine, instead of an offline dataset.
What changed
- The draft model learned directly from what the target model actually generates in production serving conditions.
- That training setup is meant to preserve accept length at real batch sizes, not just in benchmark scripts.
Reported results
- 1.89× decode throughput over a non-spec baseline at bs=64 on 8×B200, TP8.
- 7–14% faster than Inkling’s built-in MTP at the same batch size.
- 3.66 mean accept length, reaching 4.76 on GSM8K.
Under the hood
- A 5-layer Qwen3-style DFlash parallel-draft backbone
- A rank-256 Markov logit-bias head for intra-block dependencies
- A per-position confidence head to predict acceptance
- 400K online regenerations distilled from the live target engine
The team says the serving command is in the comments for anyone who wants to try it.
Related event: Open-Sourced DSpark Speculator Boosts Inkling Throughput by 1.89x(3 posts)→
More from Infra
- llama.cpp lands Flash Attention tuning for RDNA4, big prefill gains on AMD — pmttyji · 2026-09-11
- Your p99 latency benchmark may be lying: a deep dive into coordinated omission — Franc0Fernand0 · 2026-09-11
- Running MiniMax H3 on 12GB VRAM: quantization, Turbo LoRAs and attention backends compared — Possible_Mood676 · 2026-09-11
- Spomin: live KV cache compaction squeezes 500k tokens of context into 180k resident — wgaca2 · 2026-09-11
- PiPNN nearest-neighbor search wins three awards, up to 78x faster index building — khademinori · 2026-09-11
- M.2-Oculink eGPU Link Silently Downgrades to PCIe Gen1 — Here's How to Check — El_90 · 2026-09-11