apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 00:17:25 +08:00
Code Issues Actions 19 Packages Projects Releases Wiki Activity

[Feature][MTP] Support cacheKV transfer in per_chunk mode (#2890)

Browse Source 
* support chunk_prefill both normal and speculative_decoding(mtp)

* optimize pd-disaggregation config

* fix bug
This commit is contained in:
freeliuzc
2025-07-17 17:58:08 +08:00
committed by GitHub
parent 67180c1ff9
commit d49f8fb30a
10 changed files with 110 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files
custom_ops/gpu_ops/get_output_ep.cc
+3 -3
View File
@@ -36,9 +36,9 @@ void GetOutputKVSignal(const paddle::Tensor& x,
int* out_data = const_cast<int*>(x.data<int>());
int ret = -1;
if (!wait_flag) {
ret = msgrcv(msgid, &msg_rcv, (MAX_BSZ * 2 + 2) * 4, 0, IPC_NOWAIT);
ret = msgrcv(msgid, &msg_rcv, (MAX_BSZ * 3 + 2) * 4, 0, IPC_NOWAIT);
} else {
ret = msgrcv(msgid, &msg_rcv, (MAX_BSZ * 2 + 2) * 4, 0, 0);
ret = msgrcv(msgid, &msg_rcv, (MAX_BSZ * 3 + 2) * 4, 0, 0);
}
if (ret == -1) {
out_data[0] = -1;
@@ -47,7 +47,7 @@ void GetOutputKVSignal(const paddle::Tensor& x,
}
int encoder_count = msg_rcv.mtext[0];
for (int i = 0; i < encoder_count * 2 + 2; i++) {
for (int i = 0; i < encoder_count * 3 + 2; i++) {
out_data[i] = msg_rcv.mtext[i];
}
return;