Block a user
[epic] Decode efficiency: we are at ~16% of roofline on beast and do not know where the rest goes
Decode decay measured out to 70k: 29.5 → 11.0 tok/s (2.7×)
From the #283 probe ladder. Extends this issue's 30 → 31k window by more than a factor of two in context, on a single stream…
6-way batched decode yields ~1.12× single-stream throughput, not ~6× — batching is buying almost nothing
GatedDeltaNet recurrent state round-trips through bf16 every decode token — ~1% relative error across 48 of 64 layers
grenade
created branch feat/283-operator-sampling-override in helexa/helexa
2026-08-25 08:58:00 +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?
H5 settled:
H5 settled: temperature=1.0 is the defect source (p = 0.0031) — plus one new numerical defect, three hypotheses excluded, and a confound in the curve
All below taken 2026-08-25 against…
[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?
The stack under evaluation, and one change deliberately held back
Results here are only attributable if the build they were taken on is recorded. Everything in the issue body was measured on…
[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?
grenade
created branch fix/replay-only-the-latest-reasoning in helexa/helexa
2026-08-25 05:05:00 +00:00
grenade
created branch fix/developer-role-maps-to-system in helexa/helexa
2026-08-24 19:11:06 +00:00
/no_think is honoured on chat/completions but ignored on /v1/responses
Items (1) and (2) implemented and merged —
Items (1) and (2) implemented and merged — d94b780e
Four commits on fix/223-reasoning-effort-off, all four validation jobs
green including the CUDA type-check:
f5fd6782— `reasoning:…
/no_think is honoured on chat/completions but ignored on /v1/responses
Design for item (1): a reasoning budget, not an on/off switch
effort: none landed today and is the wrong tool for the common case.
The evidence says so: the one run that produced files spent…