Add Nemotron-H speculative decoding support (#2145)
* Add Nemotron-H speculative decoding support * Forward projected_context through DFlash2 and Gemma4 DSpark decoder layers DFlashDecoderLayer.__call__ gained a projected_context argument, but the DFlash2 and Gemma4 DSpark layers override __call__ and still used the four-argument signature, so the shared _hidden() call raised TypeError. Gemma4 DSpark has no test covering that path; only DFlash2 surfaced in CI.
A
Alazer Manakelew committed
2a73dc2dbb0e19edd321bced2be852fe77927aa7
Parent: 8fd4dd3
Committed by GitHub <noreply@github.com>
on 9/3/2026, 5:45:32 PM