Block a user
F4b — perf(neuron): wire speculative decoding into generation
S3/S4 measured on
S3/S4 measured on Qwen/Qwen3.5-0.8B: acceptance is workload-determined, and the headline number depends entirely on which workload you pick
neuron-mtp-probe (branch feat/96-mtp-probe)…
feat(neuron): MTP draft head + multi-position logits (#96, S0–S2)
feat(neuron): MTP draft head + multi-position logits (#96, S0–S2)
S2 added to this branch: logits at every position
Qwen3_5ForCausalLM::forward_multi → ModelArch::forward_multi → Job::ForwardLogitsMulti → DeviceWorkerHandle::forward_logits_multi.…
feat(qwen3_5): load and run the MTP draft head (#96, S0/S1)
grenade
deleted branch fix/329-reasoning-replay-follows-the-template from helexa/helexa
2026-09-23 05:23:25 +00:00
fix(responses): let the model's template decide reasoning replay (#329)
Evaluate speculative decoding for neuron (EAGLE-3 first, bench-gated) — batch-1 decode latency
Follow-up: the concurrency question this issue closed on now has external data — written up on #96
F4b — perf(neuron): wire speculative decoding into generation
New external evidence: a drafter now exists for our resident flagship, and it answers #79's concurrency question
Source: [z-lab/Qwen3.8-27B-DFlash2](https://huggingface.co/z-lab/Qwen3.8-27B-D…