Add Super VL and Mamba DisAgg compatibility to vLLM 0.28 - #44

Merged

TomerBN-Nvidia merged 7 commits into

TomerBN-Nvidia:super_vl_evals_v0.28.0from

natsala13:version/v028-eval-pr55178

Sep 8, 2026

Commits on Aug 27, 2026

Add NemotronH_Omni_Reasoning_V3 as a supported Nemotron architecture (#52929)

Naveassaf
authored and
TomerBN-Nvidia
committed
07ebd2a
Add MTP support for Nemotron VL models (#53121)

danisereb
authored and
TomerBN-Nvidia
committed
92278a5
[Bugfix][NIXL] Fix Mamba prefill truncation ordering (#53523)

authored and
TomerBN-Nvidia
committed
1c54364
fix: port remaining Super VL checkpoint compatibility
TomerBN-Nvidia
committed
0c4dd4e

Commits on Sep 8, 2026

Fix Mamba state for padded prompt tails

natsala13
authored and
nsalaNvidia
committed
ce316fc
Refine Mamba prompt-tail classification

natsala13
authored and
nsalaNvidia
committed
471ff44
Update vllm/v1/attention/backends/mamba2_attn.py

authored and
nsalaNvidia
committed
e6ee5b8