[Kernel][MoE] Add VMM-backed SharedEP for native FP4 decode on SM100 - #50392

Draft

foraxe wants to merge 19 commits into

vllm-project:mainfrom

foraxe:agent/shared-ep-vmm-native-gb200

Commits on Jul 28, 2026

Fix NVIDIA GLM-5.2 DCP attention path
foraxe
committed
1e452b1
Support NVIDIA DCP all-to-all output merge
foraxe
committed
42b5b4e
Add CUDA peer-memory primitives
foraxe
committed
82abbb5
Fix CUDA VMM pre-commit checks
foraxe
committed
39aa02b
Add consumer-side VMM DCP output merge
foraxe
committed
ee01e54
Add consumer-side VMM DCP top-k merge
foraxe
committed
50822e2
Integrate bounded DCP top-k VMM dispatch
foraxe
committed
24f240d
Avoid dynamic Top-K route logging in serving hot path
foraxe
committed
006475f
Add CUDA VMM producer fanout for DCP queries

foraxe
and
taoyuanyuan
committed
efc1bc1
Merge current main into Shared-DCP
foraxe
committed
a9a0f8f
Fix VMM close synchronization test mock
foraxe
committed
f01c10b

Commits on Jul 29, 2026

Optimize Shared-DCP output workspace reuse
foraxe
committed
d5465d4
Publish DCP output directly into owner VMM storage
foraxe
committed
5a3cad4
Size direct VMM output slots for bounded decode
foraxe
committed
c44bca4
Prefer bounded VMM output over collective backend
foraxe
committed
cecb936
Remove superseded staged VMM kernels
foraxe
committed
b8e2b2b
Cover the production VMM output row bound
foraxe
committed
f5c8cc4

Commits on Jul 30, 2026

Merge remote-tracking branch 'origin/main' into shared-ep-publish-current
foraxe
committed
a3400f3
[Kernel][MoE] Add VMM-backed SharedEP for native FP4 decode

foraxe
and
codex
committed
35395b1