• Joined on 2026-02-17
grenade pushed to feat/96-tp-multi-position at helexa/helexa 2026-09-23 10:51:20 +00:00
ec0dd319e8 feat(tp): multi-position logits across every rank (#96, S2 mirror)
grenade created branch feat/96-tp-multi-position in helexa/helexa 2026-09-23 10:51:18 +00:00
grenade commented on issue helexa/helexa#96 2026-09-23 09:22:44 +00:00
F4b — perf(neuron): wire speculative decoding into generation

S3/S4 measured on Qwen/Qwen3.5-0.8B: acceptance is workload-determined, and the headline number depends entirely on which workload you pick

neuron-mtp-probe (branch feat/96-mtp-probe)…

grenade pushed to feat/96-mtp-probe at helexa/helexa 2026-09-23 09:22:19 +00:00
744c99fc22 fix(mtp-probe): chunk the prefill, like the serving path does
grenade pushed to feat/96-mtp-probe at helexa/helexa 2026-09-23 08:59:03 +00:00
ee5f84af7b feat(neuron): MTP draft-cost and acceptance probe (#96, S3/S4)
grenade created branch feat/96-mtp-probe in helexa/helexa 2026-09-23 08:59:01 +00:00
grenade deleted branch feat/96-qwen3_5-mtp-head from helexa/helexa 2026-09-23 08:39:33 +00:00
grenade pushed to main at helexa/helexa 2026-09-23 08:39:28 +00:00
f72c1e4c05 Merge pull request 'feat(neuron): MTP draft head + multi-position logits (#96, S0–S2)' (#331) from feat/96-qwen3_5-mtp-head into main
912f7a574b test(qwen3_5): compare forward_multi to forward by tolerance, not bits
268387ae3f feat(neuron): logits at every position of a multi-token forward (#96, S2)
8fa665b428 feat(qwen3_5): load and run the MTP draft head (#96, S0/S1)
Compare 4 commits »
grenade merged pull request helexa/helexa#331 2026-09-23 08:39:27 +00:00
feat(neuron): MTP draft head + multi-position logits (#96, S0–S2)
grenade pushed to feat/96-qwen3_5-mtp-head at helexa/helexa 2026-09-23 08:06:37 +00:00
912f7a574b test(qwen3_5): compare forward_multi to forward by tolerance, not bits
grenade commented on pull request helexa/helexa#331 2026-09-23 07:33:47 +00:00
feat(neuron): MTP draft head + multi-position logits (#96, S0–S2)

S2 added to this branch: logits at every position

Qwen3_5ForCausalLM::forward_multi → ModelArch::forward_multi → Job::ForwardLogitsMulti → DeviceWorkerHandle::forward_logits_multi.…

grenade pushed to feat/96-qwen3_5-mtp-head at helexa/helexa 2026-09-23 07:33:02 +00:00
268387ae3f feat(neuron): logits at every position of a multi-token forward (#96, S2)
grenade created pull request helexa/helexa#331 2026-09-23 07:16:22 +00:00
feat(qwen3_5): load and run the MTP draft head (#96, S0/S1)
grenade pushed to feat/96-qwen3_5-mtp-head at helexa/helexa 2026-09-23 07:15:59 +00:00
8fa665b428 feat(qwen3_5): load and run the MTP draft head (#96, S0/S1)
grenade created branch feat/96-qwen3_5-mtp-head in helexa/helexa 2026-09-23 07:15:56 +00:00
grenade deleted branch fix/329-reasoning-replay-follows-the-template from helexa/helexa 2026-09-23 05:23:25 +00:00
grenade pushed to main at helexa/helexa 2026-09-23 05:23:07 +00:00
e62cf12b92 Merge pull request 'fix(responses): let the model's template decide reasoning replay (#329)' (#330) from fix/329-reasoning-replay-follows-the-template into main
016c803c48 fix(responses): let the model's template decide reasoning replay (#329)
Compare 2 commits »
grenade merged pull request helexa/helexa#330 2026-09-23 05:23:06 +00:00
fix(responses): let the model's template decide reasoning replay (#329)
grenade commented on issue helexa/helexa#79 2026-09-23 05:02:51 +00:00
Evaluate speculative decoding for neuron (EAGLE-3 first, bench-gated) — batch-1 decode latency

Follow-up: the concurrency question this issue closed on now has external data — written up on #96

z-lab/Qwen3.8-27B-DFlash2 (2B…

grenade commented on issue helexa/helexa#96 2026-09-23 05:02:34 +00:00
F4b — perf(neuron): wire speculative decoding into generation

New external evidence: a drafter now exists for our resident flagship, and it answers #79's concurrency question

Source: [z-lab/Qwen3.8-27B-DFlash2](https://huggingface.co/z-lab/Qwen3.8-27B-D…