[Model][Spec Decode] Nemotron-H MTP and Mamba Speculative Decoding Support#33726

Merged

vllm-bot merged 69 commits intovllm-project:mainfrom

CentML:bchislett/mamba-nemotron-mtp

Feb 24, 2026

Commits on Dec 21, 2025

fully working mixed requests version
shaharmor98
committed
add working real checkpoint code
shaharmor98
committed
CUDA graphs work

roikoren755
authored and
shaharmor98
committed
Prefix caching works

roikoren755
authored and
shaharmor98
committed
Fix triton kernel to support varlen, and update call site

roikoren755
authored and
shaharmor98
committed
Full CUDA graphs

roikoren755
authored and
shaharmor98
committed
Prefix caching + refactored code
shaharmor98
committed
added updated load weights
shaharmor98
committed
Fix CUDA graphs compat

roikoren755
authored and
shaharmor98
committed
Working speculative code with multiple MTP layers
shaharmor98
committed
running mtp for num_speculative > 2
shaharmor98
committed
remove redundant ids handling, add eagle3 support flag
shaharmor98
committed
temporarily disable update_block_table
shaharmor98
committed

Commits on Dec 23, 2025

Commits on Dec 30, 2025

Commits on Dec 31, 2025

fix block size used in EAGLE slot mapping

benchislett
authored and
shaharmor98
committed

Commits on Jan 1, 2026

Commits on Jan 4, 2026

change mtp layer count
shaharmor98
committed

Commits on Jan 8, 2026

final cleanup
shaharmor98
committed

Commits on Jan 30, 2026

Commits on Jan 31, 2026

Commits on Feb 2, 2026

Commits on Feb 6, 2026

Commits on Feb 8, 2026

adjust other mamba-base backends to the new state_indices_tensor representation
shaharmor98
committed

Commits on Feb 11, 2026

merge main
shaharmor98
committed

Commits on Feb 15, 2026

Commits on Feb 16, 2026

Commits on Feb 17, 2026

fix cr comments
shaharmor98
committed

Commits on Feb 20, 2026

Commits on Feb 23, 2026

Commits on Feb 24, 2026