[Spec] Windowed draft-decode attention for built-in EAGLE / MTP drafts (#32673)

This commit is contained in:
avalliappan-nvidia
2026-09-22 08:17:57 +08:00
committed by GitHub
parent 9fdb71732a
commit 61d0cf2074
8 changed files with 560 additions and 44 deletions
@@ -232,6 +232,7 @@ class TestServerArgsHook(_RegistryIsolated):
decrypted_draft_config_file=None,
trust_remote_code=False,
speculative_draft_window_size=None,
speculative_draft_sink_size=None,
speculative_skip_dp_mlp_sync=False,
speculative_adaptive=False,
)