Block a user
qwen4_exp I: host-resident experts — and a placement decision measured per machine, not chosen once
Re-measured on
Re-measured on 76c93a0: 2.48 tok/s, 2.5× — and the attribution holds
Same load config as the baseline (q4k, devices:[0], expert_residency:host, max_in_flight:1), same prompt,…
Epic: serve Qwen3.8-Flash-Next (qwen4_exp) — a 180B sparse MoE whose architecture is built for beast's shape
Milestone: flash-next generates coherent text on beast
Qwen/Qwen3.8-Flash-Next loads all 48 layers and generates. The chain that got there, all merged:
qwen4_exp I: host-resident experts — and a placement decision measured per machine, not chosen once
Slice 1 result: host-resident experts work, and they are not the bottleneck
First generation with expert_residency: host on beast: 64 tokens, prefill 10,253 ms, decode 64,421 ms — **0.99…
grenade
created branch feat/318-per-layer-activation-hop in helexa/helexa
2026-09-08 09:44:30 +00:00
Decided: malloc_trim, not a candle fork — why host-resident experts need the heap coaxed
Decided: the trim is the answer, and the slab is not
Operator call, and it corrects the framing this issue was filed with.
The "fix that is not a trim" above proposes putting a slab…
Host-resident experts are 73,728 separate allocations, which is why the heap has to be coaxed into giving them back