Port SGLang's push-based 2-buffer allreduce protocol into vLLM as a new
communicator backend for small-message reductions. The push protocol
eliminates the two explicit cross-GPU NVLink barrier round-trips used by
the existing barrier-based CustomAllreduce, replacing them with a
sentinel-based data arrival detection mechanism and double-buffered epoch
alternation.
Key advantages over the barrier-based approach:
- Zero barriers: data arrival IS the synchronization (positive-zero sentinel)
- Single NVLink round-trip instead of two barrier exchanges + remote reads
- All SMs active (SM_count CTAs vs 2 CTAs) for higher NVLink bandwidth
- No cudaMemcpy to IPC staging buffer in eager mode
- PDL (griddepcontrol) support for kernel overlap on sm_90+
The new PushAllReduce is inserted in the CudaCommunicator dispatch chain
above the existing CustomAllreduce for messages below a size threshold
(~720 KB at TP=8). Larger messages continue to use the barrier-based
path. The existing CustomAllreduce code is not modified.
Measured results on DeepSeek-V4-Pro (61 layers, TP=8, 8x NVIDIA B200,
BS=1, decode with ISL=4, OSL=33024):
- Throughput: +2.14% (84.06 vs 82.30 tokens/s)
- TPOT: -2.09% (11.90 vs 12.15 ms/token)
Correctness verified via lm_eval gsm8k 5-shot with no regression
(exact_match delta within statistical noise).
The feature can be disabled at runtime via VLLM_DISABLE_PUSH_ALLREDUCE=1
to fall back to the barrier-based path.
Signed-off-by: Alexander Matveev <amatveev@redhat.com>
