Add support for topk metadata transferring for PD (#10616)
Signed-off-by: Shangming Cai <csmthu@gmail.com>
This commit is contained in:
@@ -421,6 +421,8 @@ class SchedulerDisaggregationPrefillMixin:
|
||||
last_hidden_index = (
|
||||
hidden_state_offset + extend_input_len_per_req[i] - 1
|
||||
)
|
||||
req.output_topk_p = batch.spec_info.topk_p[i]
|
||||
req.output_topk_index = batch.spec_info.topk_index[i]
|
||||
if self.spec_algorithm.is_eagle3():
|
||||
req.hidden_states_tensor = (
|
||||
batch.spec_info.hidden_states[i].cpu().clone()
|
||||
|
||||
Reference in New Issue
Block a user