grenade

grenade commented on issue helexa/helexa#339

2026-09-29 19:19:40 +00:00

L6 — Laya checkpoint family: language router, subfolder checkpoints, rf.internal mirroring

Done.

Detector: #345, parity with the SDK on 924 generated cases.
Routing and family loading: #348. Routing blocks are byte-identical to the reference on every recorded case.…

grenade commented on issue helexa/helexa#338

2026-09-29 19:19:38 +00:00

L5 — /v1/systemone surface on neuron, cortex and helexa-router + decision capability

Done.

Gateway half: #344 (cortex and helexa-router).
Neuron half: #348 (the handler, checkpoint routing and family loading).
Aliases: #351 (helexa/one in both alias tables).

Va…

grenade commented on issue helexa/helexa#341

2026-09-29 19:19:34 +00:00

L8 — quadbrat residency swap to Laya, fleet rollout and live validation (definition of done)

Laya is resident on quadbrat and serving the public chain (2026-09-29)

Deployed: #351, quadbrat /version 8ed8d04 (the #350 build with flash-attn; the #351 merge changed only config).…

grenade pushed to fix/bench-skip-decision-models at helexa/helexa

2026-09-29 19:19:11 +00:00

cdcd5bb092 fix(bench): chat scenarios skip decision models (#341)

grenade created branch fix/bench-skip-decision-models in helexa/helexa

2026-09-29 19:19:10 +00:00

grenade opened issue helexa/helexa#352

2026-09-29 19:07:41 +00:00

Validate Z-Image (q8_0) on quadbrat's flash-attn ampere build, and measure whether the 1024² ceiling lifts

grenade pushed to main at helexa/helexa

2026-09-29 19:06:11 +00:00

c7505a9344 Merge pull request 'deploy(laya): Laya takes quadbrat's residency; helexa/one (part of #341)' (#351) from feat/341-laya-residency into main

06cf081c47 deploy(laya): Laya takes quadbrat's residency; helexa/one (#341)

Compare 2 commits »

grenade merged pull request helexa/helexa#351

2026-09-29 19:06:10 +00:00

deploy(laya): Laya takes quadbrat's residency; helexa/one (part of #341)

grenade created pull request helexa/helexa#351

2026-09-29 19:05:55 +00:00

deploy(laya): Laya takes quadbrat's residency; helexa/one (part of #341)

grenade commented on issue helexa/helexa#341

2026-09-29 19:05:49 +00:00

L8 — quadbrat residency swap to Laya, fleet rollout and live validation (definition of done)

Image validation dropped from this rollout (operator agreed, 2026-09-29). Z-Image has never been served on quadbrat:

The catalogue entry's min_device_vram_mb = 20000 excludes the 3060 from…

grenade opened issue grenade/rustingface#6

2026-09-29 19:00:18 +00:00

First fetch of an uncached repo: 503 on the hf-hub size probe (breaks neuron cold loads) and empty bodies on resumed ranges

grenade commented on issue helexa/helexa#341

2026-09-29 18:57:21 +00:00

L8 — quadbrat residency swap to Laya, fleet rollout and live validation (definition of done)

Step 1 deployed. quadbrat /version reports 8ed8d04 with ["cuda", "cudnn", "flash-attn"].

Text path on the flash build: PASS. I ran validate-neuron.sh quadbrat Qwen/Qwen3.5-0.8B ""…

grenade pushed to main at helexa/helexa

2026-09-29 17:52:07 +00:00

8ed8d04d3a Merge pull request 'ci(neuron): flash-attn on the ampere flavour, and decision/image validation (part of #341)' (#350) from feat/341-laya-rollout into main

0a122fb26f ci(neuron): flash-attn on the ampere flavour, and decision/image validation (#341)

Compare 2 commits »

grenade merged pull request helexa/helexa#350

2026-09-29 17:52:06 +00:00

ci(neuron): flash-attn on the ampere flavour, and decision/image validation (part of #341)

grenade created pull request helexa/helexa#350

2026-09-29 17:51:45 +00:00

ci(neuron): flash-attn on the ampere flavour, and decision/image validation (part of #341)

grenade closed issue helexa/helexa#337

2026-09-29 17:51:29 +00:00

L4 — Laya performance on the 3060: batched forward, flash-attn (ampere), micro-batching — GATE for the residency swap

grenade pushed to main at helexa/helexa

2026-09-29 17:51:28 +00:00

6a0d4fbb0c Merge pull request 'perf(laya): mixed precision, fused norms, mask elision, flash-attn varlen (#337)' (#349) from feat/337-laya-perf into main

e3248cc186 perf(laya): fused norms, mask elision, and a flash-attn varlen path (#337)

5264f78b2d perf(laya): autocast-style mixed precision on the GPU (#337)

1ba160491c test(laya): device/precision-parameterised parity and a forward timing bench (#337)

Compare 4 commits »

grenade merged pull request helexa/helexa#349

2026-09-29 17:51:28 +00:00

perf(laya): mixed precision, fused norms, mask elision, flash-attn varlen (#337)

grenade created pull request helexa/helexa#349

2026-09-29 17:51:17 +00:00

perf(laya): mixed precision, fused norms, mask elision, flash-attn varlen (#337)

grenade pushed to main at helexa/helexa

2026-09-29 17:28:29 +00:00

73a376e57c ci(deploy): take benjy out of the neuron deploy matrix while it is down