[Spec Decode] Support hybrid attention models in extract_hidden_states (#39949)
Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 4.6
parent
5794c65f8c
commit
2f821faeae
@@ -2347,7 +2347,13 @@ class GPUModelRunner(
|
||||
|
||||
if self.speculative_config and spec_decode_common_attn_metadata is None:
|
||||
if isinstance(
|
||||
self.drafter, (EagleProposer, DFlashProposer, Gemma4Proposer)
|
||||
self.drafter,
|
||||
(
|
||||
EagleProposer,
|
||||
DFlashProposer,
|
||||
Gemma4Proposer,
|
||||
ExtractHiddenStatesProposer,
|
||||
),
|
||||
):
|
||||
if self.drafter.kv_cache_gid == kv_cache_gid:
|
||||
spec_decode_common_attn_metadata = cm
|
||||
|
||||
Reference in New Issue
Block a user