🚀 The feature, motivation and pitch
The P2pNcclConnector supports sending the request to the P and D instances simultaneously. However, the vllm-router currently awaits the response from P before sending the request to D (like the toy proxy), which is a waste of TTFT as the P2pNcclConnectort already supports async transfer. We should allow also running the router in concurrent dispatch mode to save a full HTTP roundtrip and scheduling overhead on the D side.
Alternatives
No response
Additional context
No response
Before submitting a new issue...
🚀 The feature, motivation and pitch
The P2pNcclConnector supports sending the request to the P and D instances simultaneously. However, the vllm-router currently awaits the response from P before sending the request to D (like the toy proxy), which is a waste of TTFT as the P2pNcclConnectort already supports async transfer. We should allow also running the router in concurrent dispatch mode to save a full HTTP roundtrip and scheduling overhead on the D side.
Alternatives
No response
Additional context
No response
Before submitting a new issue...