• Joined on 2026-02-17
grenade pushed to feat/neuron-phase-timing at helexa/helexa 2026-09-08 13:46:37 +00:00
e129765918 feat(neuron): split the decode loop's wall-clock into phases
grenade created branch feat/neuron-phase-timing in helexa/helexa 2026-09-08 13:46:36 +00:00
grenade pushed to main at blackbeard/observer 2026-09-08 13:30:50 +00:00
53dac3717b feat: block explorer routes, addressed by hash
grenade commented on issue helexa/helexa#318 2026-09-08 13:16:34 +00:00
qwen4_exp I: host-resident experts — and a placement decision measured per machine, not chosen once

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,…

grenade commented on issue helexa/helexa#307 2026-09-08 13:07:17 +00:00
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:

grenade pushed to main at blackbeard/observer 2026-09-08 12:55:14 +00:00
0ce58094e4 feat(data): record what a reorg displaced
grenade commented on issue helexa/helexa#318 2026-09-08 12:21:39 +00:00
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 pushed to main at helexa/helexa 2026-09-08 12:19:15 +00:00
76c93a0304 Merge branch 'fix/qwen4-ple-conv-launch-bound'
bf1b4e6d36 perf(qwen4_exp): write PLE's depthwise conv as taps, not a grouped conv1d
Compare 2 commits »
grenade pushed to main at blackbeard/observer 2026-09-08 12:01:38 +00:00
337d1ef516 ci: format the frontend with prettier and gate it
grenade pushed to main at blackbeard/observer 2026-09-08 11:59:36 +00:00
9cbc2cc274 ui: put the miner count before the height tile
grenade pushed to main at blackbeard/observer 2026-09-08 11:39:34 +00:00
7e5c2de936 feat: sparkline history under the headline tiles
grenade created branch fix/qwen4-ple-conv-launch-bound in helexa/helexa 2026-09-08 11:29:49 +00:00
grenade pushed to fix/qwen4-ple-conv-launch-bound at helexa/helexa 2026-09-08 11:29:49 +00:00
bf1b4e6d36 perf(qwen4_exp): write PLE's depthwise conv as taps, not a grouped conv1d
grenade pushed to main at helexa/helexa 2026-09-08 10:05:42 +00:00
bcf6d94b36 Merge branch 'feat/318-per-layer-activation-hop'
da6284f2c9 feat(qwen4_exp): cross to the experts once per layer, and measure what dominates
Compare 2 commits »
grenade created branch feat/318-per-layer-activation-hop in helexa/helexa 2026-09-08 09:44:30 +00:00
grenade pushed to feat/318-per-layer-activation-hop at helexa/helexa 2026-09-08 09:44:30 +00:00
da6284f2c9 feat(qwen4_exp): cross to the experts once per layer, and measure what dominates
grenade pushed to main at helexa/helexa 2026-09-08 09:03:27 +00:00
a2719a60fb Merge branch 'fix/318-release-host-memory'
d2ade80c44 fix(neuron): trim the host heap on the load path too (#318)
f3298b8a85 fix(neuron): give host memory back when a model is unloaded (#318)
Compare 3 commits »
grenade commented on issue helexa/helexa#325 2026-09-08 08:34:41 +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…

grenade pushed to fix/318-release-host-memory at helexa/helexa 2026-09-08 08:34:29 +00:00
d2ade80c44 fix(neuron): trim the host heap on the load path too (#318)
grenade opened issue helexa/helexa#325 2026-09-08 08:31:03 +00:00
Host-resident experts are 73,728 separate allocations, which is why the heap has to be coaxed into giving them back