Block a user
F4e — cortex: vision cold-swap policy (27B-VL ⇄ frontier)
F4d — perf(neuron): continuous batching for agentic fan-out
F4c — perf(neuron): KV-cache quantization (measurement-gated)
F4b — perf(neuron): wire speculative decoding into generation
F4a — perf(neuron): enable flash-attention on full-attn layers
F3 — bench: frontier A/B decision gate, pick resident model
F2 — neuron/cortex: catalogue + load Qwen3-Next-80B-A3B family @ Q4_K_M on beast
F1 — neuron: MoE FFN into the qwen3_5 hybrid layer stack + TP expert sharding
O7 — helexa-bench: capability probe scenario (quality, not just speed)
O6 — helexa-bench: cold-load + model-swap cost scenario
O5 — helexa-bench: concurrency / agentic-load scenario
O4 — helexa-bench: context-length scaling scenario (128 → 128k)
O3 — helexa-bench: per-run VRAM high-water + GPU util/temp
O2 — helexa-bench: raw per-sample retention + p50/p95/p99 (schema migration)
O1 — neuron: emit prefill/decode timing breakdown in wire Finish event
[epic] Reasoning frontier: 80B-A3B MoE on beast (qwen3_5 hybrid + high-sparsity MoE)
[epic] Performance observability: per-phase profiling baseline before engine work
grenade
deleted branch fix/responses-usage-reasoning-tokens from helexa/helexa
2026-06-26 18:30:10 +00:00
fix(neuron): surface reasoning_tokens in non-streaming /v1/responses usage