• Joined on 2026-02-17
grenade created branch fix/223-reasoning-effort-off in helexa/helexa 2026-08-24 16:11:13 +00:00
grenade commented on issue helexa/helexa#271 2026-08-24 15:52:04 +00:00
[epic] Sampling correctness: honour the caller's parameters and the model's own generation_config

The repetition hypothesis is refuted — on the long-form task this epic said it needed

The opening of this epic attributed a 19-minute, 24,576-token run to untruncated sampling causing the…

grenade commented on issue helexa/helexa#223 2026-08-24 15:25:30 +00:00
/no_think is honoured on chat/completions but ignored on /v1/responses

Item (3) has a client-visible symptom now, in a client's own words

The pi harness (pi-ai openai-responses), against beast, 2026-08-24. Ten minutes of generation ended with this on the user's…

grenade commented on issue helexa/helexa#282 2026-08-24 15:13:02 +00:00
[epic] Decode efficiency: we are at ~16% of roofline on beast and do not know where the rest goes

Baseline: idle box, single stream

The missing row. Co-tenant CCTV workload stopped (lairball-vision.service on quadbrat), beast confirmed idle — in_flight: 0, both GPUs at 0% for a…

grenade commented on issue helexa/helexa#257 2026-08-24 14:11:18 +00:00
VRAM backstop rejects before admission, killing long agentic sessions — admission counts requests, capacity is KV bytes

The device fault under saturation, with a full trace

Reproduced on beast 2026-08-24 16:54:39, build acef01b, during a live agentic session. This is the failure this issue anticipated and had…

grenade commented on issue helexa/helexa#270 2026-08-24 13:35:46 +00:00
Prefix-cache reuse looks low on long agentic contexts — 2068 tokens reused against a 42K prompt

Measured after #277, on a real agentic turn

Live DSH session against beast, 2026-08-24, build acef01b. Step 2 of a tower-defence build — the step that replays step 1's 27k-token think block…

grenade opened issue helexa/helexa#282 2026-08-24 13:25:02 +00:00
[epic] Decode efficiency: we are at ~16% of roofline on beast and do not know where the rest goes
grenade closed issue helexa/helexa#280 2026-08-24 13:07:11 +00:00
Every model neuron downloads itself is served with our prompt format, not its own
grenade commented on issue helexa/helexa#280 2026-08-24 13:07:07 +00:00
Every model neuron downloads itself is served with our prompt format, not its own

Verified live — closing

Merged acef01bc, deployed fleet-wide (acef01b on beast, benjy and quadbrat).

benjy's snapshot now carries what its readers look for:

config.json  generation_c…
grenade opened issue helexa/helexa#281 2026-08-24 13:04:00 +00:00
The advertised context window is still a gauge, not a promise — clients cannot cache it
grenade commented on issue helexa/helexa#223 2026-08-24 12:26:11 +00:00
/no_think is honoured on chat/completions but ignored on /v1/responses

Status after #277: one of three items done, and the repro no longer

reproduces — which is not the same as fixed

Re-ran the issue's exact repro against beast on the deployed build (74df8fd): …

grenade closed issue helexa/helexa#254 2026-08-24 12:25:57 +00:00
neuron ignores max_completion_tokens, so the output cap is unenforced for current OpenAI clients
grenade closed issue helexa/helexa#278 2026-08-24 12:25:56 +00:00
limit.output is a KV planning reserve advertised as an output ceiling — nothing reconciles the two
grenade closed issue helexa/helexa#277 2026-08-24 12:25:55 +00:00
The Responses translator silently discards conversation state — a truncated agentic turn can never be continued
grenade commented on issue helexa/helexa#278 2026-08-24 12:25:51 +00:00
limit.output is a KV planning reserve advertised as an output ceiling — nothing reconciles the two

Landed and verified — closing, with one item handed to #223

Merged 74df8fd0, deployed, checked against beast and through cortex.

The advertisement is now the ceiling:

neuron …
grenade commented on issue helexa/helexa#254 2026-08-24 12:25:34 +00:00
neuron ignores max_completion_tokens, so the output cap is unenforced for current OpenAI clients

Fixed and verified — closing

Landed as part of #278 (74df8fd0) rather than the branch described above: the four generation loops now take the cap from effective_max_tokens() (current…

grenade commented on issue helexa/helexa#277 2026-08-24 12:25:28 +00:00
The Responses translator silently discards conversation state — a truncated agentic turn can never be continued

Verified live — closing

Merged 6945cf98, deployed to the fleet, probed against beast's Qwen3.8-27B on 74df8fd.

The test plants a fact that exists only inside a replayed reasoning item…

grenade pushed to main at helexa/helexa 2026-08-24 12:24:46 +00:00
acef01bc4c Merge branch 'fix/fetch-the-models-own-prompt-format'
441cddea37 fix(load): fetch the model's own prompt format, not just its weights
Compare 2 commits »
grenade opened issue helexa/helexa#280 2026-08-24 11:56:55 +00:00
Every model neuron downloads itself is served with our prompt format, not its own
grenade pushed to fix/fetch-the-models-own-prompt-format at helexa/helexa 2026-08-24 11:56:17 +00:00
441cddea37 fix(load): fetch the model's own prompt format, not just its weights