[Bugfix][Model Runner V2] Order uniform decodes first so spec decodes aren't misclassified as prefills (#47381)
Signed-off-by: Woosuk Kwon <woosuk@inferact.ai> Signed-off-by: Nick Hill <nickhill123@gmail.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
co-authored by
Claude Fable 5
Nick Hill
parent
b83be00cdd
commit
85b3a7264b
@@ -1156,7 +1156,12 @@ class DeepseekV2MLAAttention(nn.Module):
|
||||
cache_config,
|
||||
quant_config,
|
||||
prefix,
|
||||
skip_topk=_skip_topk,
|
||||
# MTP layers must never start with skip_topk=True: their indexer
|
||||
# computes indices at draft step 0, and the runtime toggle
|
||||
# (set_skip_topk, index_share_for_mtp_iteration) only exists in
|
||||
# the V1 proposer. A frozen True would leave the draft reading a
|
||||
# never-written topk buffer.
|
||||
skip_topk=_skip_topk and not is_mtp_layer,
|
||||
)
|
||||
|
||||
def forward(
|
||||
|
||||
Reference in New Issue
Block a user