0
1
0
0
k100 vllm qwen3
2026-02-04 17:06:45
408
我想部署qwen3系列模型，使用社区vllm镜像，显卡为K100
harbor.sourcefind.cn:5443/dcu/admin/base/vllm:0.9.2-ubuntu22.04-dtk25.04.2-rocblas1121-das1.7-py3.10-20251203-bw
以qwen3-14b为例
启动命令为
python -m vllm.entrypoints.openai.api\_server --model /Qwen3-14B --trust-remote-code --gpu-memory-utilization 0.9 --tensor-parallel-size 4 --max-num-seqs 32 --max-model-len 40000
报错信息如下，我想知道具体该用哪个镜像可以部署，十分感谢您的回复
DCU K100AI支持Roce吗
--
楼主
0 点赞
0 帖子
楼主热帖
暂无数据常见问题新帖
暂无数据
rocBLAS error: No Tensile solution found for { a: 0x7fcf09a00000, a_type: "bf16_r", b: 0x7fcb59a00000, b_type: "bf16_r", c: 0x7fcb4ec00000, c_type: "bf16_r", d: 0x7fcb4ec00000, d_type: "bf16_r", compute_type: "f32_r", transA: 'N', transB: 'N', M: 3584, N: 10240, K: 5120, alpha: 1, row_stride_a: 1, col_stride_a: 3584, row_stride_b: 1, col_stride_b: 5120, row_stride_c: 1, col_stride_c: 3584, row_stride_d: 1, col_stride_d: 3584, beta: 0, batch_count: 1, strided_batch: true, stride_a: 1, stride_b: 1, stride_c: 1, stride_d: 1, offset_a: 0, offset_b: 0, offset_c: 0, offset_d: 0, atomics_mode: atomics_allowed }
This message will be only be displayed once, unless the ROCBLAS_VERBOSE_TENSILE_ERROR environment variable is set.
ERROR 02-04 10:27:40 [multiproc_executor.py:135] Worker proc VllmWorker-1 died unexpectedly, shutting down executor.
Process EngineCore_0:
Traceback (most recent call last):
File "/usr/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
self.run()
File "/usr/lib/python3.10/multiprocessing/process.py", line 108, in run
self._target(*self._args, **self._kwargs)
File "/usr/local/lib/python3.10/dist-packages/vllm/v1/engine/core.py", line 607, in run_engine_core
raise e
File "/usr/local/lib/python3.10/dist-packages/vllm/v1/engine/core.py", line 594, in run_engine_core
engine_core = EngineCoreProc(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/vllm/v1/engine/core.py", line 421, in __init__
super().__init__(vllm_config, executor_class, log_stats,
File "/usr/local/lib/python3.10/dist-packages/vllm/v1/engine/core.py", line 84, in __init__
self._initialize_kv_caches(vllm_config)
File "/usr/local/lib/python3.10/dist-packages/vllm/v1/engine/core.py", line 144, in _initialize_kv_caches
available_gpu_memory = self.model_executor.determine_available_memory()
File "/usr/local/lib/python3.10/dist-packages/vllm/v1/executor/abstract.py", line 76, in determine_available_memory
output = self.collective_rpc("determine_available_memory")
File "/usr/local/lib/python3.10/dist-packages/vllm/v1/executor/multiproc_executor.py", line 215, in collective_rpc
result = get_response(w, dequeue_timeout)
File "/usr/local/lib/python3.10/dist-packages/vllm/v1/executor/multiproc_executor.py", line 202, in get_response
raise RuntimeError(
RuntimeError: Worker failed with error 'CUDA error: HIPBLAS_STATUS_NOT_SUPPORTED when calling `hipblasGemmEx( handle, opa, opb, m, n, k, &falpha, a, HIP_R_16BF, lda, b, HIP_R_16BF, ldb, &fbeta, c, HIP_R_16BF, ldc, compute_type, HIPBLAS_GEMM_DEFAULT)`', please check the stack trace above for the root cause
Traceback (most recent call last):