光合开发者社区-论坛

数学库

提问帖

有回复

Qwen3.6+vllm0.18报错，

2026-04-30 10:11:55

359

报错：No available shared memory broadcast block found in 60 seconds. This typically happens when some processes are hanging or doing some time-consuming work (e.g., compilation, weight/kv cache quantization). 镜像：harbor.sourcefind.cn:5443/dcu/admin/base/custom:vllm018-ubuntu22.04-dtk26.04-nemotron-20260422 设置了1、2卡并行什么原因

HIP_VISIBLE_DEVICES="0,1" \

OMP_NUM_THREADS=32 \

ALLREDUCE_STREAM_WITH_COMPUTE=1 \

NCCL_MIN_NCHANNELS=16 \

NCCL_MAX_NCHANNELS=16 \

VLLM_RPC_TIMEOUT=1800000 \

HSA_FORCE_FINE_GRAIN_PCIE=1 \

HIP_PERF=1 \

SAFETENSORS_FAST_GPU=1 \

vllm serve /models/cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit \

--served-model-name Qwen3.6-35B-A3B \

--trust-remote-code \

--tensor-parallel-size 2 \

--dtype float16 \

--max-model-len 204800 \

--kv-cache-dtype fp8 \

--max-num-batched-tokens 32768 \

--max-num-seqs 2 \

--gpu-memory-utilization 0.6 \

--enable-auto-tool-choice \

--tool-call-parser qwen3_coder \

--reasoning-parser qwen3 \

--limit-mm-per-prompt '{"video": 1, "image": 2}' \

--mm-encoder-tp-mode data \

--speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":3}' \

--enable-chunked-prefill \

--port 8001

我也遇到这个问题，然后是通过上面的启动命令解决的，生效的估计是ALLREDUCE_STREAM_WITH_COMPUTE=1参数，豆包说开启该参数通过统一 CUDA/HIP 流执行 AllReduce 与计算任务，消除了多流异步导致的流同步等待与资源竞争，从而避免共享内存广播块长时间无法获取的超时错误。