• Joined on 2026-02-17
grenade pushed to fix/288-bench-authenticates at helexa/helexa 2026-08-25 10:55:51 +00:00
19243d317e feat(bench): collect and chart what a decode-rate line cannot show
grenade pushed to fix/288-bench-authenticates at helexa/helexa 2026-08-25 10:48:48 +00:00
16d473227d fix(bench): authenticate, and collect what was being thrown away
grenade created branch fix/288-bench-authenticates in helexa/helexa 2026-08-25 10:48:47 +00:00
grenade opened issue helexa/helexa#288 2026-08-25 10:37:17 +00:00
helexa-bench is anonymous, so since #262 it has measured the anonymous-yield policy instead of serving capacity — every number since 2026-08-18 understates the fleet
grenade created branch feat/285-nonstreaming-through-engine in helexa/helexa 2026-08-25 10:32:21 +00:00
grenade pushed to feat/285-nonstreaming-through-engine at helexa/helexa 2026-08-25 10:32:21 +00:00
2b2563edb4 perf(inference): run non-streaming through the batch engine
grenade commented on issue helexa/helexa#285 2026-08-25 10:04:59 +00:00
Non-streaming requests bypass the batch engine and serialize completely — 1.00× at every concurrency level

Correction: the 1.12× figure was measured on the wrong path. Batching works; non-streaming is what serializes.

The issue body's headline is wrong and I'm replacing it. The #283 sampling…

grenade commented on issue helexa/helexa#283 2026-08-25 09:53:51 +00:00
[epic] Qwen3.8-27B degrades with context length on our stack — why has it never finished a task two cloud models did in ten minutes?

Protocol near-miss: H7 was about to be merged by accident

Recording this because the issue's protocol section is exactly what caught it, and the failure mode is easy to repeat.

The `feat/283-o…

grenade pushed to feat/283-operator-sampling-override at helexa/helexa 2026-08-25 09:51:47 +00:00
d7b407cca6 feat(deploy): make CI the only writer of fleet config
902f5460bf fix(deploy): default the catalogue-changed flag when the step emits nothing
821ca116aa feat(sampling): let the operator correct a model's published temperature
Compare 3 commits »
grenade pushed to fix/284-gdn-state-f32 at helexa/helexa 2026-08-25 09:49:47 +00:00
b1200cb56b fix(gdn): stop round-tripping the recurrent state through bf16
grenade opened issue helexa/helexa#287 2026-08-25 09:42:31 +00:00
Close the last hand-synced fleet configs — CI must be the only writer everywhere, not just on the model path
grenade pushed to feat/283-operator-sampling-override at helexa/helexa 2026-08-25 09:42:08 +00:00
fd19b44a47 feat(deploy): make CI the only writer of fleet config
grenade pushed to feat/283-operator-sampling-override at helexa/helexa 2026-08-25 09:32:03 +00:00
8e37ed6fd7 fix(deploy): default the catalogue-changed flag when the step emits nothing
grenade commented on issue helexa/helexa#283 2026-08-25 09:13:31 +00:00
[epic] Qwen3.8-27B degrades with context length on our stack — why has it never finished a task two cloud models did in ten minutes?

Implementation: the operator sampling override, on feat/283-operator-sampling-override

Recommended action (1) from the previous comment is built. The override surface did not exist —…

grenade commented on issue helexa/helexa#286 2026-08-25 09:12:55 +00:00
CI clippy never lints test code — cargo clippy --workspace without --all-targets misses #[cfg(test)]

Fixed on fix/286-clippy-all-targets — and the damage was worse than the issue guessed

The backlog was 19 warnings, not the years of accumulation I assumed when filing. All fixed rather…

grenade pushed to fix/286-clippy-all-targets at helexa/helexa 2026-08-25 09:12:37 +00:00
03fe47e91b fix(ci): lint test code too, and revive two tests that stopped running
grenade created branch fix/286-clippy-all-targets in helexa/helexa 2026-08-25 09:12:36 +00:00
grenade pushed to fix/284-gdn-state-f32 at helexa/helexa 2026-08-25 09:03:19 +00:00
528c39c907 fix(gdn): stop round-tripping the recurrent state through bf16
grenade created branch fix/284-gdn-state-f32 in helexa/helexa 2026-08-25 09:03:18 +00:00
grenade opened issue helexa/helexa#286 2026-08-25 08:59:39 +00:00
CI clippy never lints test code — cargo clippy --workspace without --all-targets misses #[cfg(test)]