apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-24 01:29:57 +08:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

[PD Disaggregation] support different tp_size for prefill and decode (#5296)

Browse Source 
* up

* up

* up

* fix
This commit is contained in:
Juncai
2025-12-01 17:50:20 +08:00
committed by GitHub
parent 54119cf07e
commit 0925d44f18
13 changed files with 584 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/cache_manager/transfer_factory/kvcache_transfer/include/kvcache_connection.h
+1
View File
@@ -142,6 +142,7 @@ struct Connection {
int wc_target_count;
// Configuration
int decode_tp_size;
int layer_number;
int block_number;
int block_byte_size;
fastdeploy/cache_manager/transfer_factory/kvcache_transfer/include/kvcache_rdma.h
+8 -2
View File
@@ -24,11 +24,15 @@ class RDMACommunicator {
std::vector<int64_t> local_key_cache,
std::vector<int64_t> local_value_cache,
int block_number,
int block_bytes);
int block_bytes,
int prefill_tp_size,
int prefill_tp_idx);
~RDMACommunicator();
// Connection management
int connect(const std::string& dst_ip, const std::string& dst_port);
int connect(const std::string& dst_ip,
const std::string& dst_port,
int dest_tp_size);
bool is_connected(const std::string& dst_ip, const std::string& dst_port);
// Core functionality
@@ -120,6 +124,8 @@ class RDMACommunicator {
int block_number; // Number of blocks
int block_size_byte; // Size of each block in bytes
int layer_number; // Number of layers
int prefill_tp_size; // tensor parallelism size for prefill
int prefill_tp_idx; // tensor parallelism index for prefill
std::vector<std::vector<void*>>
local_cache_key_ptr_per_layer; // Per-layer key pointers