grenade

grenade opened issue helexa/helexa#322

2026-09-07 05:08:01 +00:00

In-situ quantisation is redone on every load and discarded on unload — ~26 minutes per load for qwen4_exp

grenade commented on issue helexa/helexa#307

2026-09-06 19:05:31 +00:00

Epic: serve Qwen3.8-Flash-Next (qwen4_exp) — a 180B sparse MoE whose architecture is built for beast's shape

The clean number: 16 of 48 layers on one card

Third window, guards in place — CI idle, activation.state == "ready", 27B resident and evicted first, restored from the trap afterwards (`loaded…

grenade commented on issue helexa/helexa#307

2026-09-06 18:55:22 +00:00

Epic: serve Qwen3.8-Flash-Next (qwen4_exp) — a 180B sparse MoE whose architecture is built for beast's shape

Correction: the experts exhaust the card long before PLE is reached

My previous comment said the load would die at layer 1's PLE table. That is wrong about ordering, and the difference…

grenade pushed to main at helexa/helexa

2026-09-06 18:21:14 +00:00

7593ab3d5d Merge branch 'fix/qwen4-exp-expert-row-width'

bf1e80ed55 fix(qwen4_exp): 640 is not a whole number of k-quant blocks

Compare 2 commits »

grenade pushed to fix/qwen4-exp-expert-row-width at helexa/helexa

2026-09-06 17:59:45 +00:00

bf1e80ed55 fix(qwen4_exp): 640 is not a whole number of k-quant blocks

grenade commented on issue helexa/helexa#307

2026-09-06 17:32:50 +00:00

Epic: serve Qwen3.8-Flash-Next (qwen4_exp) — a 180B sparse MoE whose architecture is built for beast's shape

First cold load on our own engine (2026-09-06) — and a re-ordering of C/I/J

Ran the guarded window on beast against a14d1af4. Evict, load, restore, all from a trap; the 27B came back…

grenade created branch fix/qwen4-exp-expert-row-width in helexa/helexa

2026-09-06 17:32:18 +00:00

grenade pushed to fix/qwen4-exp-expert-row-width at helexa/helexa

2026-09-06 17:32:18 +00:00

154fc6d3e0 fix(qwen4_exp): 640 is not a whole number of k-quant blocks

grenade pushed to main at helexa/helexa

2026-09-06 16:58:03 +00:00

a14d1af447 Merge branch 'fix/qwen4-exp-load-reporting-and-pool-trim'

f1af926c30 fix(qwen4_exp): route the arch on the path that actually loads it

430b6be62b fix(neuron): let the loaded model answer for itself, and give the pool back

Compare 3 commits »

grenade pushed to fix/qwen4-exp-load-reporting-and-pool-trim at helexa/helexa

2026-09-06 16:36:55 +00:00

f1af926c30 fix(qwen4_exp): route the arch on the path that actually loads it

grenade opened issue helexa/helexa#321

2026-09-06 16:11:52 +00:00

A neuron flavour can build and then not be packaged, so a host silently keeps running the previous binary

grenade created branch fix/qwen4-exp-load-reporting-and-pool-trim in helexa/helexa

2026-09-06 16:10:19 +00:00

grenade pushed to fix/qwen4-exp-load-reporting-and-pool-trim at helexa/helexa

2026-09-06 16:10:19 +00:00

430b6be62b fix(neuron): let the loaded model answer for itself, and give the pool back

grenade deleted branch fix/320-verbosity-preserve-candle-trace from helexa/helexa

2026-09-06 15:06:02 +00:00

grenade created branch fix/320-verbosity-preserve-candle-trace in helexa/helexa

2026-09-06 15:03:59 +00:00

grenade pushed to fix/320-verbosity-preserve-candle-trace at helexa/helexa

2026-09-06 15:03:59 +00:00

a33602424e fix(deploy): manage the RUST_LOG the fleet is actually running

grenade pushed to main at helexa/helexa

2026-09-06 15:03:18 +00:00

af40bac9b8 Merge branch 'fix/320-isq-reachable-and-log-levels'

d5738bfb89 fix(neuron): read quant as ISQ when the repo has no GGUF (#320)

Compare 2 commits »

grenade pushed to main at blackbeard/observer

2026-09-04 10:56:47 +00:00

b552b08b10 docs: record the mesh-bind, hairpin and runner-label gotchas

grenade pushed to main at blackbeard/observer

2026-09-04 10:48:19 +00:00

3db86cfc66 fix(ci): health-probe the address the API actually binds, not loopback

grenade pushed to main at blackbeard/observer

2026-09-04 10:37:37 +00:00

151a30ae1a ci: run the build on the rust runner, drop corepack, pin the site check to loopback