DeepInfer native sm80 support (beyond the kernel port); DSpark + suffix-decoding hybrid for DS V4.1 on the GNmac V1-runner path (reasoning acceptance stuck at 2.9 of 5); DFlash2 drafter f...

  • Swimlane: Default swimlane
  • Column: Idea
  • Position: 12
  • Assignee: not assigned
  • Creator: Ivan
  • Started:
  • Created: 10/30/2026 09:17
  • Modified: 10/30/2026 09:17
  • Moved: 10/30/2026 09:17
  • 8 GB
  • 10 GB
Description

Full title: DeepInfer native sm80 support (beyond the kernel port); DSpark + suffix-decoding hybrid for DS V4.1 on the GNmac V1-runner path (reasoning acceptance stuck at 2.9 of 5); DFlash2 drafter for MiMo Flash; EAGLE3 head path at PP>1 for the hybrid

Reported by / on it: @c133, @reala7096, @nathang6866, @darkmaniac7

Source: https://discord.com/channels/1518634379604136178/1528826330454429706/1562917910165323808

Wiki page: Engines (Speculative decoding)

Cards: both

Card D-096, seeded from the cmpunlocker Discord with credit (reference discord:1562917910165323808).

Sub-Tasks
Internal links
Comments