Repository navigation

Prism b10683 Ganglion Helios port multi slot concurrency speedup from kernel reuse - #243

Open

synchronic1 wants to merge 6 commits into

PrismML-Eng:prismfrom

synchronic1:prism-b10683-helios-port

Open

Prism b10683 Ganglion Helios port multi slot concurrency speedup from kernel reuse#243
synchronic1 wants to merge 6 commits into
PrismML-Eng:prismfrom
synchronic1:prism-b10683-helios-port

Commits on Sep 22, 2026
- llama-helios-port onto PrismML prism-b10683-d8f26ee
  synchronic1
  committed
  2d843b2
- cuda: multi-column PQ2_0 MMVQ weight reuse on Ampere
  synchronic1
  committed
  2c9103d
- cuda: tune Ampere MMVQ rows per block for PQ2_0 at 2 columns
  synchronic1
  committed
  356b782
- bench: add Prism helios Ampere decode perf chart
  synchronic1
  committed
  1366363
Commits on Sep 23, 2026
- ggml-cuda : skip f16 K/V reservation for the fused q8_0 flash-attn path
  synchronic1
  committed
  b1960db
- cuda: skip the f16 K/V reservation when the fused q8_0 FA path runs
  synchronic1
  committed
  1d9ef0c