Dockerで最新ROCmを環境を汚さずに試す ～仮想コンテナで重要ソフトを安全に試す llama-cppでLLMテスト～

rcat@rcat-EVO-X2:~$ docker image ls
IMAGE                                                           ID             DISK USAGE   CONTENT SIZE   EXTRA
rocm/pytorch:rocm7.1_ubuntu24.04_py3.12_pytorch_release_2.9.1   c61158cfb7f3       31.3GB             0B

rcat@rcat-EVO-X2:~$ docker run --rm -it rocm/pytorch:rocm7.1_ubuntu24.04_py3.12_pytorch_release_2.9.1

docker run -it --network=host --device=/dev/kfd --device=/dev/dri --group-add=video --ipc=host --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --shm-size 16G -v ~/dockerdir:/usr/local/bin/host rocm/pyt
orch:rocm7.1_ubuntu24.04_py3.12_pytorch_release_2.9.1

(venv) root@rcat-EVO-X2:/usr/local/bin# python
Python 3.12.3 (main, Aug 14 2025, 17:47:21) [GCC 13.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from llama_cpp import Llama
>>> llm = Llama(model_path='/usr/local/bin/host/gemma-3-4b-it-Q4_K_M.gguf', n_ctx=2048,n_gpu_layers=-1)


ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
  Device 0: AMD Radeon 8060S, gfx1151 (0x1151), VMM: no, Wave Size: 32
llama_model_load_from_file_impl: using device ROCm0 (AMD Radeon 8060S) - 41666 MiB free
llama_model_loader: loaded meta data with 40 key-value pairs and 444 tensors from /usr/local/bin/host/gemma-3-4b-it-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.

省略

>>> stop_tokens: list = ["<|eot_id|>", "<|im_end|>","<end_of_turn>"]
>>> messages = [
... {"role": "system", "content": "あなたは親切なアシスタントです。問いかけに対して丁寧に回答します"},
... {"role": "user", "content": "こんにちわ!ねこってかわいいよね?"}
... ]
>>> output = llm.create_chat_completion(
...     messages=messages,
...     max_tokens=1024,
...     stream=False,
...     stop=stop_tokens,
... )
llama_perf_context_print:        load time =     101.74 ms
llama_perf_context_print: prompt eval time =     101.60 ms /    36 tokens (    2.82 ms per token,   354.34 tokens per second)
llama_perf_context_print:        eval time =    1656.74 ms /   103 runs   (   16.08 ms per token,    62.17 tokens per second)
llama_perf_context_print:       total time =    1936.83 ms /   139 tokens
llama_perf_context_print:    graphs reused =         99
>>> print(output["choices"][0]["message"]["content"])
こんにちは！そうですね、猫は本当に可愛いです！ふわふわの毛並みや、つぶらな瞳、そして気まぐれな仕草…見ているだけで癒されますよね。

猫について何か話したいことや、知りたいことなどありますか？例えば、

*   猫の品種について
*   猫の飼い方について
*   猫の好きな食べ物について
*   猫の面白い行動について

など、どんなことでも構いませんよ！😊

rcat@rcat-EVO-X2:~$ docker container ls -a
CONTAINER ID   IMAGE                                                           COMMAND       CREATED          STATUS                      PORTS     NAMES
60ca8d47a5ec   rocm/pytorch:rocm7.1_ubuntu24.04_py3.12_pytorch_release_2.9.1   "/bin/bash"   27 minutes ago   Exited (0) 17 seconds ago             goofy_knuth


rcat@rcat-EVO-X2:~$ docker start -ai  60ca8d47a5ec
root@rcat-EVO-X2:/# ls
apex-1.9.0a0+rocm7.1.0.git07c3ee53-cp312-cp312-linux_x86_64.whl  lib                opt   tmp
bin                                                              lib.usr-is-merged  proc  torch-2.9.1+rocm7.1.0.lw.git351ff442-cp312-cp312-linux_x86_64.whl
boot                                                             lib32              root  torchaudio-2.9.0+rocm7.1.0.gite3c6ee2b-cp312-cp312-linux_x86_64.whl
dev                                                              lib64              run   torchvision-0.24.0+rocm7.1.0.gitb919bd0c-cp312-cp312-linux_x86_64.whl
etc                                                              libx32             sbin  triton-3.5.1+rocm7.1.0.gita272dfa8-cp312-cp312-linux_x86_64.whl
home                                                             media              srv   usr
install                                                          mnt                sys   var
root@rcat-EVO-X2:/# cd usr/local/bin/
root@rcat-EVO-X2:/usr/local/bin# ls
host  venv