Repository navigation

[None][feat] ModelingV2: GLM-5.3-Flash block-FP8 target for sm_10x tp4 - #22082

Open

ruocheng-nv wants to merge 10 commits into

NVIDIA:mainfrom

ruocheng-nv:feat/glm53-flash-fp8-mv2-target

Open

[None][feat] ModelingV2: GLM-5.3-Flash block-FP8 target for sm_10x tp4#22082
ruocheng-nv wants to merge 10 commits into
NVIDIA:mainfrom
ruocheng-nv:feat/glm53-flash-fp8-mv2-target

Commits on Nov 11, 2026
- [None][feat] ModelingV2: GLM-5.3-Flash block-FP8 target for sm_10x tp4
  
  ruocheng-nv
  authored and
  hanjingtian
  committed
  d9616b9
- [None][chore] ModelingV2 GLM-5.3-Flash FP8: keep the target's operations in the target
  
  ruocheng-nv
  authored and
  hanjingtian
  committed
  932dd18
- [None][fix] ModelingV2 GLM-5.3-Flash FP8: name only the refused requirements; run the boundary matrix
  
  ruocheng-nv
  authored and
  hanjingtian
  committed
  ec443a2
- [None][fix] Keep model defaults from marking unset LLM args as set
  ruocheng-nv
  committed
  4860e3d
- [None][fix] Let a model keep one-token prompts out of decode graph rows
  ruocheng-nv
  committed
  030b271
- [None][fix] ModelingV2 GLM-5.3-Flash FP8: route only what the target runs; harden its kernels and tests
  ruocheng-nv
  committed
  41ceed7
Commits on Nov 12, 2026
- [None][fix] ModelingV2 GLM-5.3-Flash FP8: route on the KDA kernel sizes; cover the Lamport sentinel and shared tuning
  ruocheng-nv
  committed
  fe10409
- [None][fix] ModelingV2 GLM-5.3-Flash FP8: decline the speculation acceptance-rate gate
  ruocheng-nv
  committed
  9784275
Commits on Nov 13, 2026
- Merge remote-tracking branch 'upstream/main' into feat/glm53-flash-fp8-mv2-target
  ruocheng-nv
  committed
  41c7a98
- [None][fix] ModelingV2 GLM-5.3-Flash FP8: return KDA prefill to the auto route
  ruocheng-nv
  committed
  ec01ba7