[PD][Spec] Fix hidden state transfer for spec decode (#7516)
Signed-off-by: Shangming Cai <caishangming@linux.alibaba.com>
This commit is contained in:
@@ -579,11 +579,11 @@ class DecodeTransferQueue:
|
||||
idx = decode_req.metadata_buffer_index
|
||||
(
|
||||
output_id,
|
||||
output_hidden_states,
|
||||
output_token_logprobs_val,
|
||||
output_token_logprobs_idx,
|
||||
output_top_logprobs_val,
|
||||
output_top_logprobs_idx,
|
||||
output_hidden_states,
|
||||
) = self.metadata_buffers.get_buf(idx)
|
||||
|
||||
decode_req.req.output_ids.append(output_id[0].item())
|
||||
|
||||
Reference in New Issue
Block a user