Add DFlash speculative decoding support to NeMo SpeechLM vLLM plugin - #16129

Closed

Slyne wants to merge 20 commits into

NVIDIA-NeMo:mainfrom

Slyne:slyne/dflash-inference

Commits on Aug 24, 2026

fix(speechlm2): harden SALM MTP vLLM support
Slyne
committed
8a4df2d
fix(speechlm2): derive audio token index at runtime
Slyne
committed
b723aa6
fix(speechlm2): address final SALM review findings
Slyne
committed
ad2514c

Commits on Aug 25, 2026

fix(speechlm2): make MTP override spawn-pickleable
Slyne
committed
4ad950b
fix(speechlm2): align export trust and test imports
Slyne
committed
5bbfd2f
fix(tests): normalize SpeechLM config import
Slyne
committed
33957a2
fix(speechlm2): normalize speculative config imports
Slyne
committed
ecf30a9
style(tests): satisfy black formatting
Slyne
committed
c5ff502

Commits on Aug 26, 2026

fix(speechlm2): use bundled vLLM backbone config
Slyne
committed
67639ba
Add DFlash support to SpeechLM vLLM
Slyne
committed
2e4a4a9
Add DFlash2 bootstrap support for SpeechLM vLLM
Slyne
committed
e809666
Keep DFlash2 support inference-only
Slyne
committed
06bb29b
Pin vLLM DFlash2 runtime commit
Slyne
committed
61b7176
fix(speechlm2): load embedded PE inference checkpoints
Slyne
committed
71c2ef1
fix(speechlm2): accept Automodel DFlash drafts
Slyne
committed
c942cd7

Commits on Sep 4, 2026

feat(speechlm2): route Automodel DFlash2 through vLLM
Slyne
committed
e095c3c