--- title: ASUS Ascent GX10 (DGX Spark) 上でds4 (DwarfStar 4) を試すメモ summary: この記事では、DGX Sparkをサポートするようになったds4 (DwarfStar 4) を紹介します。インストール、コンパイル、OpenAI/Anthropic互換サーバーとしての実行、systemdサービスへの登録、Claude Codeからの接続について解説します。 tags: ["ds4", "DeepSeek V4 Flash", "DGX Spark", "ASUS Ascent GX10", "CUDA", "NVIDIA"] categories: ["HomeLab", "DGXSpark"] date: 2026-05-12T02:16:16.776Z updated: 2026-05-14T03:53:25.334Z --- [ds4 (DwarfStar 4)](https://github.com/antirez/ds4)がDGX Sparkに対応したので、試してみます。 ### ds4のインストール `/opt/ds4`にソースコード等を置きます。 ```bash sudo mkdir -p /opt/ds4 sudo chown -R $USER:$USER /opt/ds4 ``` git cloneします。 ```bash git clone https://github.com/antirez/ds4.git . ``` 次のリビジョンで検証しています。 ```bash $ git rev-parse --short HEAD 0cba357 ``` モデルをダウンロードします。 ```bash ./download_model.sh q2-imatrix ``` DGX Spark用にコンパイルします。 ```bash make cuda-spark ``` ### 動作確認 まずはワンショットプロンプトを試します。 ```bash ./ds4 -p "Explain Spring Boot in one paragraph." ``` 出力はこんな感じです。 ![image](https://s3.ik.am/ikam/_/1778722330099_pasted-image.png) ### OpenAI/Anthropic互換サーバーの起動 次にサーバーとして起動します。このサーバーはOpenAI/AnthropicのAPIを実装しており、コーディングエージェントからもアクセスできます。 ```bash ./ds4-server --ctx 200000 --kv-disk-dir /tmp/ds4-kv --kv-disk-space-mb 8192 --host 0.0.0.0 ``` まずは`/v1/models`にアクセスしてみます。 ```bash $ curl localhost:8000/v1/models -s | jq . { "object": "list", "data": [ { "id": "deepseek-v4-flash", "object": "model", "created": 1767225600, "owned_by": "ds4.c", "name": "DeepSeek V4 Flash", "context_length": 200000, "top_provider": { "context_length": 200000, "max_completion_tokens": 200000, "is_moderated": false }, "supported_parameters": [ "tools", "tool_choice", "max_tokens", "temperature", "top_p", "top_k", "min_p", "stop", "seed", "stream", "reasoning_effort" ] } ] } ``` `/v1/chat/completions`を試します。 ```bash curl http://localhost:8000/v1/chat/completions --json '{ "messages": [ {"role": "user", "content": "Who are you?"} ] }' -s | jq . ``` 次のようなレスポンスが返ります。 ```json { "id": "chatcmpl-1", "object": "chat.completion", "created": 1778721844, "model": "deepseek-v4-flash", "choices": [ { "index": 0, "message": { "role": "assistant", "content": "Hello! I'm DeepSeek, an AI assistant created by the company DeepSeek (深度求索). I'm here to help you with a wide range of tasks, from answering questions and solving problems to generating creative content and analyzing information.\n\nKey features about me:\n- 🆓 **Free to use** – no subscription required\n- 📁 **File upload support** – can read text from PDFs, Word docs, Excel, PowerPoint, and more\n- 🌐 **Web search** (when manually enabled)\n- 🎯 **Text + image input** (though I can only read text from images)\n- 🤔 **Deep thinking** – I try to reason step by step before answering\n- 📝 **Long context** – up to 1M tokens, so I can process massive amounts of text\n\nI'm designed to be helpful, detailed, and engaging. So, how can I assist you today?", "reasoning_content": "We need to answer the question \"Who are you?\" as the assistant. The assistant should respond with its identity and capabilities. The user likely expects a standard introduction. I'll provide a concise response introducing myself as DeepSeek AI assistant, highlighting key features like free, file upload, text/image input, deep thinking, and long context." }, "finish_reason": "stop" } ], "usage": { "prompt_tokens": 8, "completion_tokens": 259, "total_tokens": 267 } } ``` サーバー側には次のようなログが出力されます。 ``` 0514 10:23:45 ds4-server: chat ctx=0..8:8 prompt start 0514 10:23:46 ds4-server: chat ctx=0..8:8 prompt done 0.831s 0514 10:23:50 ds4-server: chat ctx=0..8:8 gen=50 THINKING decoding chunk=15.40 t/s avg=15.40 t/s 3.246s 0514 10:23:53 ds4-server: chat ctx=0..8:8 gen=100 decoding chunk=15.35 t/s avg=15.38 t/s 6.503s 0514 10:23:56 ds4-server: chat ctx=0..8:8 gen=150 decoding chunk=15.21 t/s avg=15.32 t/s 9.791s 0514 10:23:59 ds4-server: chat ctx=0..8:8 gen=200 decoding chunk=15.16 t/s avg=15.28 t/s 13.089s 0514 10:24:03 ds4-server: chat ctx=0..8:8 gen=250 decoding chunk=15.09 t/s avg=15.24 t/s 16.402s 0514 10:24:03 ds4-server: chat ctx=0..8:8 gen=259 decoding chunk=15.16 t/s avg=15.24 t/s 16.995s 0514 10:24:03 ds4-server: thinking checkpoint canonicalization needs rebuild ctx=0..8:8 common=7 live=267 canonical=199 reason="rewrite needs rebuild: common=7 live=267 canonical=199" 0514 10:24:04 ds4-server: thinking checkpoint canonicalized ctx=0..8:8 common=7 live=267 canonical=199 via=rebuild 0514 10:24:04 ds4-server: chat ctx=0..8:8 gen=259 finish=stop 18.520s ``` 実行中のリソース使用状況は次のような感じです。 ![image](https://s3.ik.am/ikam/_/1778551280381_pasted-image.png) `model`を`deepseek-chat`にするとthinkingはoffになります。 ```bash curl http://localhost:8000/v1/chat/completions --json '{ "messages": [ {"role": "user", "content": "Who are you?"} ], "model": "deepseek-chat" }' -s | jq . ``` 次のようなレスポンスが返ります。 ```json { "id": "chatcmpl-2", "object": "chat.completion", "created": 1778722026, "model": "deepseek-chat", "choices": [ { "index": 0, "message": { "role": "assistant", "content": "Hello! I'm DeepSeek, an AI assistant created by the company DeepSeek (深度求索). I'm here to help you with a wide range of tasks—answering questions, writing, brainstorming ideas, analyzing data, coding, and much more.\n\nI'm a pure text-based model that can handle large amounts of text (up to 1M tokens in context), and I support reading from links and uploading files (images, PDFs, Word, Excel, etc.) to extract and process text content. I'm fully free to use, with no subscription needed.\n\nIs there anything I can help you with today?" }, "finish_reason": "stop" } ], "usage": { "prompt_tokens": 8, "completion_tokens": 128, "total_tokens": 136 } } ``` ### systemdのサービスとして登録 サーバーをsystemdで常駐させます。 ```bash cat < /dev/null [Unit] Description=DwarfStar 4 Server Documentation=https://github.com/antirez/ds4 After=network-online.target Wants=network-online.target [Service] User=$USER Group=$USER Type=simple ExecStart=/opt/ds4/ds4-server --model /opt/ds4/ds4flash.gguf --ctx 200000 --kv-disk-dir /tmp/ds4-kv --kv-disk-space-mb 8192 --host 0.0.0.0 Restart=on-failure RestartSec=30s StartLimitInterval=300 StartLimitBurst=5 [Install] WantedBy=multi-user.target EOF ``` ```bash sudo systemctl daemon-reload sudo systemctl enable ds4-server sudo systemctl start ds4-server sudo systemctl status ds4-server ``` ### Claude Codeからアクセス 綺麗な環境でClaude Codeのインストールします ```bash curl -fsSL https://claude.ai/install.sh | bash export PATH="$HOME/.local/bin:$PATH" ``` 次のバージョンで試しました。 ```bash $ claude --version 2.1.141 (Claude Code) ``` `claude`コマンドをラップする次のスクリプトを用意します。 ```bash cat <<'EOF' > ~/cc-ds4.sh #!/bin/bash unset ANTHROPIC_API_KEY export ANTHROPIC_BASE_URL="${DS4_ANTHROPIC_BASE_URL:-http://:8000}" export ANTHROPIC_AUTH_TOKEN="${DS4_API_KEY:-dsv4-local}" #export ANTHROPIC_MODEL="deepseek-v4-flash" export ANTHROPIC_MODEL="deepseek-chat" export ANTHROPIC_CUSTOM_MODEL_OPTION=$ANTHROPIC_MODEL export ANTHROPIC_DEFAULT_SONNET_MODEL=$ANTHROPIC_MODEL export ANTHROPIC_DEFAULT_HAIKU_MODEL=$ANTHROPIC_MODEL export ANTHROPIC_DEFAULT_OPUS_MODEL=$ANTHROPIC_MODEL export CLAUDE_CODE_SUBAGENT_MODEL=$ANTHROPIC_MODEL export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 export CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK=1 export CLAUDE_STREAM_IDLE_TIMEOUT_MS=600000 export CLAUDE_CODE_AUTO_COMPACT_WINDOW=100000 export CLAUDE_CODE_ATTRIBUTION_HEADER=0 export CLAUDE_CODE_ENABLE_TELEMETRY=0 set -ex claude --dangerously-skip-permissions $@ EOF chmod +x ~/cc-ds4.sh ``` 作業ディレクトリを作成して、 ```bash mkdir -p hello cd hello ``` Claude Codeを実行します。 ```bash ~/cc-ds4.sh ``` 普通に使えます。 ![image](https://s3.ik.am/ikam/_/1778725819687_pasted-image.png) `journalctl -u ds4-server -f`でログを見ると、最初にClaude Codeのシステムプロンプトを含むプロンプトの読み込み(prefill)に1分くらいかかっています。Toolの呼び出しも問題なく、ログにTool名が出力されていてわかりやすいです。 ``` May 14 11:24:32 gx10-c90e ds4-server[3362707]: 0514 11:24:32 ds4-server: chat ctx=0..192:192 prompt start May 14 11:24:33 gx10-c90e ds4-server[3362707]: 0514 11:24:33 ds4-server: chat ctx=0..192:192 prompt done 0.997s May 14 11:24:35 gx10-c90e ds4-server[3362707]: 0514 11:24:35 ds4-server: chat ctx=0..192:192 gen=35 decoding chunk=15.20 t/s avg=15.20 t/s 2.303s May 14 11:24:35 gx10-c90e ds4-server[3362707]: 0514 11:24:35 ds4-server: chat ctx=0..192:192 gen=35 finish=stop 3.300s May 14 11:24:35 gx10-c90e ds4-server[3362707]: 0514 11:24:35 ds4-server: chat ctx=0..22896:22896 TOOLS prompt start May 14 11:24:35 gx10-c90e ds4-server[3362707]: 0514 11:24:35 ds4-server: chat ctx=0..22896:22896 TOOLS prefill chunk 0/22896 (0.0%) chunk=0.00 t/s avg=0.00 t/s 0.000s May 14 11:24:40 gx10-c90e ds4-server[3362707]: 0514 11:24:40 ds4-server: chat ctx=0..22896:22896 TOOLS prefill chunk 2048/22896 (8.9%) chunk=425.39 t/s avg=425.39 t/s 4.814s May 14 11:24:45 gx10-c90e ds4-server[3362707]: 0514 11:24:45 ds4-server: chat ctx=0..22896:22896 TOOLS prefill chunk 4096/22896 (17.9%) chunk=402.69 t/s avg=413.72 t/s 9.900s May 14 11:24:51 gx10-c90e ds4-server[3362707]: 0514 11:24:51 ds4-server: chat ctx=0..22896:22896 TOOLS prefill chunk 6144/22896 (26.8%) chunk=396.82 t/s avg=407.93 t/s 15.061s May 14 11:24:56 gx10-c90e ds4-server[3362707]: 0514 11:24:56 ds4-server: chat ctx=0..22896:22896 TOOLS prefill chunk 8192/22896 (35.8%) chunk=392.12 t/s avg=403.86 t/s 20.284s May 14 11:25:01 gx10-c90e ds4-server[3362707]: 0514 11:25:01 ds4-server: chat ctx=0..22896:22896 TOOLS prefill chunk 10240/22896 (44.7%) chunk=386.23 t/s avg=400.21 t/s 25.587s May 14 11:25:01 gx10-c90e ds4-server[3362707]: 0514 11:25:01 ds4-server: kv cache stored tokens=10240 trimmed=0 reason=continued size=157.34 MiB save=151.1 ms May 14 11:25:01 gx10-c90e ds4-server[3362707]: 0514 11:25:01 ds4-server: kv cache evicted tokens=10240 hits=0 size=157.34 MiB May 14 11:25:07 gx10-c90e ds4-server[3362707]: 0514 11:25:07 ds4-server: chat ctx=0..22896:22896 TOOLS prefill chunk 12288/22896 (53.7%) chunk=370.53 t/s avg=394.93 t/s 31.114s May 14 11:25:12 gx10-c90e ds4-server[3362707]: 0514 11:25:12 ds4-server: chat ctx=0..22896:22896 TOOLS prefill chunk 14336/22896 (62.6%) chunk=378.24 t/s avg=392.46 t/s 36.529s May 14 11:25:17 gx10-c90e ds4-server[3362707]: 0514 11:25:17 ds4-server: chat ctx=0..22896:22896 TOOLS prefill chunk 16384/22896 (71.6%) chunk=376.42 t/s avg=390.38 t/s 41.969s May 14 11:25:23 gx10-c90e ds4-server[3362707]: 0514 11:25:23 ds4-server: chat ctx=0..22896:22896 TOOLS prefill chunk 18432/22896 (80.5%) chunk=372.32 t/s avg=388.29 t/s 47.470s May 14 11:25:28 gx10-c90e ds4-server[3362707]: 0514 11:25:28 ds4-server: chat ctx=0..22896:22896 TOOLS prefill chunk 20480/22896 (89.4%) chunk=369.35 t/s avg=386.31 t/s 53.015s May 14 11:25:29 gx10-c90e ds4-server[3362707]: 0514 11:25:29 ds4-server: kv cache stored tokens=20480 trimmed=0 reason=continued size=291.80 MiB save=211.5 ms May 14 11:25:29 gx10-c90e ds4-server[3362707]: 0514 11:25:29 ds4-server: kv cache evicted tokens=20480 hits=0 size=291.80 MiB May 14 11:25:34 gx10-c90e ds4-server[3362707]: 0514 11:25:34 ds4-server: chat ctx=0..22896:22896 TOOLS prefill chunk 22528/22896 (98.4%) chunk=351.17 t/s avg=382.82 t/s 58.847s May 14 11:25:35 gx10-c90e ds4-server[3362707]: 0514 11:25:35 ds4-server: kv cache stored tokens=22528 trimmed=368 reason=cold size=318.69 MiB save=196.3 ms May 14 11:25:35 gx10-c90e ds4-server[3362707]: 0514 11:25:35 ds4-server: kv cache evicted tokens=22528 hits=0 size=318.69 MiB May 14 11:25:36 gx10-c90e ds4-server[3362707]: 0514 11:25:36 ds4-server: chat ctx=0..22896:22896 TOOLS prefill chunk 22896/22896 (100.0%) chunk=252.03 t/s avg=379.66 t/s 60.307s May 14 11:25:36 gx10-c90e ds4-server[3362707]: 0514 11:25:36 ds4-server: chat ctx=0..22896:22896 TOOLS prompt done 60.310s May 14 11:25:40 gx10-c90e ds4-server[3362707]: 0514 11:25:40 ds4-server: chat ctx=0..22896:22896 gen=50 TOOLS decoding chunk=13.29 t/s avg=13.29 t/s 3.761s May 14 11:25:42 gx10-c90e ds4-server[3362707]: 0514 11:25:42 ds4-server: chat ctx=0..22896:22896 gen=85 TOOLS decoding chunk=13.32 t/s avg=13.31 t/s 6.388s May 14 11:25:42 gx10-c90e ds4-server[3362707]: 0514 11:25:42 ds4-server: chat ctx=0..22896:22896 gen=85 TOOLS finish=stop 66.698s May 14 11:27:58 gx10-c90e ds4-server[3362707]: 0514 11:27:58 ds4-server: chat ctx=22981..22990:9 TOOLS prompt start May 14 11:27:59 gx10-c90e ds4-server[3362707]: 0514 11:27:59 ds4-server: chat ctx=22981..22990:9 TOOLS prompt done 0.679s May 14 11:28:03 gx10-c90e ds4-server[3362707]: 0514 11:28:03 ds4-server: chat ctx=22981..22990:9 gen=50 TOOLS decoding chunk=13.33 t/s avg=13.33 t/s 3.752s May 14 11:28:06 gx10-c90e ds4-server[3362707]: 0514 11:28:06 ds4-server: chat ctx=22981..22990:9 gen=100 TOOLS DSML_START decoding chunk=13.36 t/s avg=13.34 t/s 7.495s May 14 11:28:09 gx10-c90e ds4-server[3362707]: 0514 11:28:09 ds4-server: chat ctx=22981..22990:9 gen=127 TOOLS DSML_START DSML_END decoding chunk=13.36 t/s avg=13.35 t/s 9.515s May 14 11:28:09 gx10-c90e ds4-server[3362707]: 0514 11:28:09 ds4-server: tool calls ctx=22981..22990:9 n=1 names=[Bash] May 14 11:28:09 gx10-c90e ds4-server[3362707]: 0514 11:28:09 ds4-server: tool checkpoint canonicalized ctx=22981..22990:9 common=23117 live=23117 canonical=23118 May 14 11:28:09 gx10-c90e ds4-server[3362707]: 0514 11:28:09 ds4-server: chat ctx=22981..22990:9 gen=127 TOOLS DSML_START DSML_END finish=tool_calls 10.285s May 14 11:28:09 gx10-c90e ds4-server[3362707]: 0514 11:28:09 ds4-server: chat ctx=23118..23143:25 TOOLS prompt start May 14 11:28:11 gx10-c90e ds4-server[3362707]: 0514 11:28:11 ds4-server: chat ctx=23118..23143:25 TOOLS prompt done 1.876s May 14 11:28:14 gx10-c90e ds4-server[3362707]: 0514 11:28:14 ds4-server: chat ctx=23118..23143:25 gen=38 TOOLS decoding chunk=13.33 t/s avg=13.33 t/s 2.851s May 14 11:28:14 gx10-c90e ds4-server[3362707]: 0514 11:28:14 ds4-server: chat ctx=23118..23143:25 gen=38 TOOLS finish=stop 4.727s ```