Skip to content

[None][feat] KDA verify: optional per-token states in the V2 hybrid cache manager #19817

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
vsabavat wants to merge 8 commits into from
Closed

[None][feat] KDA verify: optional per-token states in the V2 hybrid cache manager #19817

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
f6cb23b
[None][feat] KDA verify: optional per-token states in the V2 hybrid c…
vsabavat Oct 1, 2026
f3cf23f
[None][fix] KV cache manager V2: keep the SSM pool at its live floor …
vsabavat Oct 2, 2026
e30477f
[None][fix] KV cache manager V2: take the per-token KDA states out of…
vsabavat Oct 3, 2026
a9938c2
[None][fix] KV cache manager V2: count the live-floor KDA token-state…
vsabavat Oct 4, 2026
54b87a2
[None][fix] KV cache manager V2: refuse per-token KDA states off the …
vsabavat Oct 4, 2026
dc903ae
[None][fix] KDA verify: size kda_state_tok for the drafts' records
vsabavat Oct 7, 2026
c05a0c3
[None][fix] KDA verify: drop the replay caches the drafts' records re…
vsabavat Oct 7, 2026
de2f7e1
[None][fix] KV cache manager V2: count every live-floor SSM slot in t…
vsabavat Oct 7, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 8 additions & 0 deletions tensorrt_llm/_torch/pyexecutor/_util.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3048,6 +3048,14 @@ def _create_kv_cache_manager(
if is_kda_mtp_verify_available():
kda_extra_kwargs["kda_replay_num_spec"] = (
spec_config.tokens_per_gen_step - 1)
# A model whose KDA verify replays the accepted drafts from
# its own per-draft records (instead of the replay caches)
# asks for a buffer of them with `kda_token_states`.
backbone = getattr(getattr(model_engine, "model", None),
"model", None)
if (issubclass(kv_cache_manager_cls, MambaHybridCacheManagerV2)
and getattr(backbone, "kda_token_states", False)):
kda_extra_kwargs["kda_token_states"] = True
if is_glm5_next:
# The manager places an indexer buffer on every attention layer.
from ..attention.backends.sparse.glm_kpool import \
Expand Down
Loading