grenade

grenade created repository grenade/newsfeed

2026-07-08 08:52:31 +00:00

grenade pushed to main at cichlid/cichlid

2026-07-08 08:36:08 +00:00

00add390d6 fix(swarm): summarise dnf output instead of leaking raw transaction log

grenade opened issue helexa/helexa#136

2026-07-06 12:38:51 +00:00

GC — pipeline / expert parallelism (contingency: 2-slot-bridge NVL only)

grenade opened issue helexa/helexa#135

2026-07-06 11:48:23 +00:00

G7 — quantized weight load + fleet validation

grenade opened issue helexa/helexa#134

2026-07-06 11:48:17 +00:00

G6 — batched decode engine + latent-cache snapshot wiring

grenade opened issue helexa/helexa#133

2026-07-06 11:48:11 +00:00

G5 — tensor parallelism for MLA + DSA (TP=4/8 on H200)

grenade opened issue helexa/helexa#132

2026-07-06 11:48:03 +00:00

G4 — decoder assembly + glm_moe_dsa config + registration

grenade opened issue helexa/helexa#131

2026-07-06 11:47:57 +00:00

G3 — noaux_tc MoE block (sigmoid + correction bias)

grenade opened issue helexa/helexa#130

2026-07-06 11:47:48 +00:00

G2 — DSA sparse-attention indexer (masked-dense → sparse kernel)

grenade opened issue helexa/helexa#129

2026-07-06 11:47:41 +00:00

G1 — MLA attention + latent KV cache (single-GPU)

grenade opened issue helexa/helexa#128

2026-07-06 11:47:25 +00:00

[epic] GLM 5.2 support: MLA + DSA (glm_moe_dsa) bring-up on neuron

grenade commented on issue helexa/helexa#126

2026-07-05 23:48:10 +00:00

27B derived context limit reports 128k — investigate which ceiling binds and whether 256k is reachable

Resolved the "which term binds" question and, along the way, root-caused an advertised context = 0 regression observed after the beast reboot. With the #127 observability line live, a healthy…

grenade pushed to main at helexa/helexa

2026-07-05 23:27:29 +00:00

c448ad417e Merge pull request 'fix(neuron): context-limit observability + qwen3_next profile (#126)' (#127) from fix/126-limit-observability into main

98d095099a fix(neuron): context-limit observability + qwen3_next profile (#126)

Compare 2 commits »

grenade merged pull request helexa/helexa#127

2026-07-05 23:27:29 +00:00

fix(neuron): context-limit observability + qwen3_next profile (#126)

grenade created pull request helexa/helexa#127

2026-07-05 23:27:16 +00:00

fix(neuron): context-limit observability + qwen3_next profile (#126)

grenade pushed to fix/126-limit-observability at helexa/helexa

2026-07-05 23:11:48 +00:00

98d095099a fix(neuron): context-limit observability + qwen3_next profile (#126)

grenade created branch fix/126-limit-observability in helexa/helexa

2026-07-05 23:11:48 +00:00

grenade opened issue helexa/helexa#126

2026-07-05 22:47:39 +00:00

27B derived context limit reports 128k — investigate which ceiling binds and whether 256k is reachable

grenade closed issue helexa/helexa#98

2026-07-03 19:49:58 +00:00

F4d — perf(neuron): continuous batching for agentic fan-out

grenade commented on issue helexa/helexa#98

2026-07-03 19:49:50 +00:00

F4d — perf(neuron): continuous batching for agentic fan-out

Bench gate: PASS. Measured live on beast (2× RTX 5090, TP-2, max_in_flight = 8), O5 concurrency scenario, batched build vs the fc95faa serialized baseline:

Qwen/Qwen3-Coder-Next (A3B…