尝试在openi启智社区的dcu环境安装ollama最新版0.15.2（失败）

先上结论，ollama的rocm版本应该是支持amd的，而不是国产DCU的。

下载

先下载

复制代码

wget -c ollama-linux-amd64-rocm.tar.zst

ollama的rocm版本，后缀是zst，请问怎么解开和安装？ollama-linux-amd64-rocm.tar.zst

大约跟普通gz后缀一样？

直接tar xvf失败。

安装zstd

复制代码

apt update
apt install zstd

解压

复制代码

tar -I 'zstd' -xf ollama-linux-amd64-rocm.tar.zst -C /

解压到

标准步骤

先安装普通linux cpu版本的ollama

下载解压

复制代码

wget -c https://github.com/ollama/ollama/releases/download/v0.15.2/ollama-linux-amd64.tar.zst

解压

复制代码

tar -I 'zstd' -xf ollama-linux-amd64.tar.zst -C /usr

安装dcu rocm ollama插件

下载

解压

复制代码

tar -I 'zstd' -xf ollama-linux-amd64-rocm.tar.zst -C /

下载模型

复制代码

ollama pull lfm2.5-thinking:1.2b

启动模型

复制代码

ollama run lfm2.5-thinking:1.2b

速度很慢

推理速度很慢

复制代码

root@notebook-2015948306967752706-denglf-89748:~# ollama run lfm2.5-thinking:1.2b
>>> 天为什么是蓝色的？
Thinking...
Okay, let's tackle this question: "Why is the sun (or Earth?) blue?" Hmm, the user is asking why something appears

看看ollama serve的跟踪信息

这是纯cpu推理啊

复制代码

time=2026-01-27T02:26:23.516Z level=INFO source=server.go:245 msg="enabling flash attention"
time=2026-01-27T02:26:23.517Z level=INFO source=server.go:429 msg="starting runner" cmd="/usr/bin/ollama runner --ollama-engine --model /root/.ollama/models/blobs/sha256-a7b19185f31650af480c3fa28dd240c75862182e0f30b118dbfadb192e4beb0a --port 33665"
time=2026-01-27T02:26:23.517Z level=INFO source=sched.go:452 msg="system memory" total="1007.4 GiB" free="896.3 GiB" free_swap="0 B"
time=2026-01-27T02:26:23.517Z level=INFO source=server.go:755 msg="loading model" "model layers"=17 requested=-1
time=2026-01-27T02:26:23.549Z level=INFO source=runner.go:1405 msg="starting ollama engine"
time=2026-01-27T02:26:23.550Z level=INFO source=runner.go:1440 msg="Server listening on 127.0.0.1:33665"
time=2026-01-27T02:26:23.561Z level=INFO source=runner.go:1278 msg=load request="{Operation:fit LoraPath:[] Parallel:1 BatchSize:512 FlashAttention:Enabled KvSize:4096 KvCacheType: NumThreads:127 GPULayers:[] MultiUserCache:false ProjectorPath: MainGPU:0 UseMmap:false}"
time=2026-01-27T02:26:23.584Z level=INFO source=ggml.go:136 msg="" architecture=lfm2 file_type=Q4_K_M name="" description="" num_tensors=148 num_key_values=28
load_backend: loaded CPU backend from /usr/lib/ollama/libggml-cpu-haswell.so
time=2026-01-27T02:26:23.592Z level=INFO source=ggml.go:104 msg=system CPU.0.SSE3=1 CPU.0.SSSE3=1 CPU.0.AVX=1 CPU.0.AVX2=1 CPU.0.F16C=1 CPU.0.FMA=1 CPU.0.BMI2=1 CPU.0.LLAMAFILE=1 CPU.1.LLAMAFILE=1 compiler=cgo(gcc)
time=2026-01-27T02:26:23.614Z level=INFO source=runner.go:1278 msg=load request="{Operation:alloc LoraPath:[] Parallel:1 BatchSize:512 FlashAttention:Enabled KvSize:4096 KvCacheType: NumThreads:127 GPULayers:[] MultiUserCache:false ProjectorPath: MainGPU:0 UseMmap:false}"
time=2026-01-27T02:26:23.684Z level=INFO source=runner.go:1278 msg=load request="{Operation:commit LoraPath:[] Parallel:1 BatchSize:512 FlashAttention:Enabled KvSize:4096 KvCacheType: NumThreads:127 GPULayers:[] MultiUserCache:false ProjectorPath: MainGPU:0 UseMmap:false}"
time=2026-01-27T02:26:23.685Z level=INFO source=device.go:245 msg="model weights" device=CPU size="799.8 MiB"
time=2026-01-27T02:26:23.685Z level=INFO source=device.go:256 msg="kv cache" device=CPU size="48.1 MiB"
time=2026-01-27T02:26:23.685Z level=INFO source=device.go:267 msg="compute graph" device=CPU size="64.0 MiB"
time=2026-01-27T02:26:23.684Z level=INFO source=ggml.go:482 msg="offloading 0 repeating layers to GPU"
time=2026-01-27T02:26:23.685Z level=INFO source=ggml.go:486 msg="offloading output layer to CPU"
time=2026-01-27T02:26:23.685Z level=INFO source=ggml.go:494 msg="offloaded 0/17 layers to GPU"
time=2026-01-27T02:26:23.685Z level=INFO source=device.go:272 msg="total memory" size="911.9 MiB"
time=2026-01-27T02:26:23.685Z level=INFO source=sched.go:526 msg="loaded runners" count=1
time=2026-01-27T02:26:23.685Z level=INFO source=server.go:1347 msg="waiting for llama runner to start responding"
time=2026-01-27T02:26:23.690Z level=INFO source=server.go:1381 msg="waiting for server to become available" status="llm server loading model"
time=2026-01-27T02:26:24.192Z level=INFO source=server.go:1385 msg="llama runner started in 0.67 seconds"
[GIN] 2026/01/27 - 02:26:24 | 200 |  864.235109ms |       127.0.0.1 | POST     "/api/generate"
[GIN] 2026/01/27 - 02:28:52 | 200 | 55.662835879s |       127.0.0.1 | POST     "/api/chat"

用ps 看一下，果然在cpu推理的

复制代码

ollama ps
NAME                    ID              SIZE      PROCESSOR    CONTEXT    UNTIL              
lfm2.5-thinking:1.2b    95bd9d45385f    956 MB    100% CPU     4096       3 minutes from now

咨询一下：dcu计算卡可以用ollama的rocm进行推理吗？

根据现有信息，目前‌Ollama 并未提供专门针对 DCU（国产异构加速卡）的 ROCm 版本‌。虽然 Ollama 支持通过 ROCm 后端在 AMD GPU 上运行模型，但 DCU 属于国产芯片，其软件栈（如 DTK）与 NVIDIA CUDA 或 AMD ROCm 不同。

所以，应该是无法使用了...