grenade

grenade merged pull request helexa/helexa#110

2026-07-02 03:53:11 +00:00

perf(neuron): fused grouped-GEMM dispatch for the qwen3_next MoE block (#92)

grenade created pull request helexa/helexa#110

2026-07-02 03:53:01 +00:00

perf(neuron): fused grouped-GEMM dispatch for the qwen3_next MoE block (#92)

grenade created branch feat/92-moe-fused in helexa/helexa

2026-07-02 03:46:12 +00:00

grenade pushed to feat/92-moe-fused at helexa/helexa

2026-07-02 03:46:12 +00:00

29a7054c23 perf(neuron): fused grouped-GEMM dispatch for the qwen3_next MoE block (#92)

grenade commented on issue helexa/helexa#93

2026-07-02 00:28:44 +00:00

F2 — neuron/cortex: catalogue + load Qwen3-Next-80B-A3B family @ Q4_K_M on beast

First live load of Qwen/Qwen3-Coder-Next on beast — works end-to-end. (2026-07-02 ~00:00 UTC, build 7489a49)

Checkpoint: 162 GB bf16 (40 shards) in /archive3/llm-cache (the systemd…

grenade pushed to main at helexa/helexa

2026-07-01 22:29:51 +00:00

7489a49d24 fix(ci): render RPM release stamps in UTC regardless of committer TZ

grenade commented on issue helexa/helexa#92

2026-07-01 21:54:33 +00:00

F1 — neuron: MoE FFN into the qwen3_5 hybrid layer stack + TP expert sharding

Slice 3 merged (#109) — TP expert sharding is in: per-expert column/row slicing with a single block-end AllReduce, replicated router, TP-side fused qkvz/ba group-span de-interleave, per-rank ISQ.…

grenade deleted branch feat/92-qwen3_5-moe from helexa/helexa

2026-07-01 21:54:15 +00:00

grenade pushed to main at helexa/helexa

2026-07-01 21:54:13 +00:00

341f97f705 Merge pull request 'feat(neuron): TP expert sharding for the qwen3_next MoE block (#92)' (#109) from feat/92-qwen3_5-moe into main

a157c8fa27 fix(neuron): gate TP MoE load tests to non-cuda builds (#92)

201cc54a7b feat(neuron): TP expert sharding for the qwen3_next MoE block (#92)

Compare 3 commits »

grenade merged pull request helexa/helexa#109

2026-07-01 21:54:12 +00:00

feat(neuron): TP expert sharding for the qwen3_next MoE block (#92)

grenade created pull request helexa/helexa#109

2026-07-01 21:54:03 +00:00

feat(neuron): TP expert sharding for the qwen3_next MoE block (#92)

grenade pushed to main at helexa/helexa

2026-07-01 21:44:16 +00:00

fc56eafbf7 fix(ci): deploy LLM probe retries on 429/503 instead of hard-failing

grenade pushed to feat/92-qwen3_5-moe at helexa/helexa

2026-07-01 21:42:49 +00:00

a157c8fa27 fix(neuron): gate TP MoE load tests to non-cuda builds (#92)

grenade pushed to feat/92-qwen3_5-moe at helexa/helexa

2026-07-01 21:34:11 +00:00

201cc54a7b feat(neuron): TP expert sharding for the qwen3_next MoE block (#92)

781d92d9a3 Merge pull request 'feat(neuron): qwen3_next config + MoE FFN block, single-GPU path with exact HF parity (#92)' (#108) from feat/92-qwen3_5-moe into main

63fb8c6e76 chore(bench): enable 27B baseline capability + concurrency scenarios (#94)

a9f5ac7ab4 Merge pull request 'feat(cortex): re-expose flat max_model_len / max_input_tokens / max_output_tokens on /v1/models (#78)' (#107) from feat/78-max-model-len into main

0126144a93 ci: retrigger after runner died mid-job (#78)

Compare 6 commits »

grenade commented on issue helexa/helexa#92

2026-07-01 21:23:55 +00:00

F1 — neuron: MoE FFN into the qwen3_5 hybrid layer stack + TP expert sharding

Progress — slices 1+2 merged (#108):

Slice 1: qwen3_next config acceptance — flat-layout normalisation into the qwen3_5 shape, MoE hyperparameters on TextConfig, layer_types…

grenade merged pull request helexa/helexa#108

2026-07-01 21:23:24 +00:00

feat(neuron): qwen3_next config + MoE FFN block, single-GPU path with exact HF parity (#92)

grenade pushed to main at helexa/helexa

2026-07-01 21:23:24 +00:00

781d92d9a3 Merge pull request 'feat(neuron): qwen3_next config + MoE FFN block, single-GPU path with exact HF parity (#92)' (#108) from feat/92-qwen3_5-moe into main

9bf13f09dd feat(neuron): qwen3_next MoE FFN block, single-GPU path + HF parity (#92)

a1426f177c feat(neuron): accept qwen3_next configs into the qwen3_5 arch (#92)

Compare 3 commits »

grenade created pull request helexa/helexa#108

2026-07-01 21:23:16 +00:00

feat(neuron): qwen3_next config + MoE FFN block, single-GPU path with exact HF parity (#92)

grenade pushed to feat/92-qwen3_5-moe at helexa/helexa

2026-07-01 21:16:23 +00:00

9bf13f09dd feat(neuron): qwen3_next MoE FFN block, single-GPU path + HF parity (#92)

grenade pushed to main at helexa/helexa

2026-07-01 21:00:01 +00:00

63fb8c6e76 chore(bench): enable 27B baseline capability + concurrency scenarios (#94)