背景介绍

在我参与的一个国企项目中，我们基于大语言模型开发了一些应用，但是甲方公司所有的资源环境都是纯内网。更为有趣的是，甲方公司已自主搭建并运行着一套百度机器学习平台（BML），客户要求所有的大模型部署必须依托于现有的BML平台进行，而非独立构建全新的基础设施，资源申请也相当严苛。面对这一系列限定条件，我们只能试着利用Docker容器技术进行大语言模型的部署。

前期准备

1、首先，内网环境部署docker：

这部分内容不再赘述，可参考之前写的教程。

https://zyn1994.blog.csdn.net/article/details/109516191

2、其次，使用一台具备网络环境的设备，拉取ollama的基础镜像：

docker pull ollama/ollama:latest
# 如果拉取不到，可使用下面这个
docker pull dhub.kubesre.xyz/ollama/ollama:latest

3、下载Qwen2的GGUF模型，这里为了演示方便就下载0.5B的模型了。

下载地址：https://huggingface.co/Qwen/Qwen2-0.5B-Instruct-GGUF​或者https://modelscope.cn/models/qwen/Qwen2-0.5B-Instruct-GGUF​

4、编写Modelfile文件：

# 注意GGUF模型文件的地址要与Dockerfile中保持一致
FROM /tmp/qwen2-0_5b-instruct-q4_0.gguf
TEMPLATE "{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>
"
PARAMETER stop <|im_start|>
PARAMETER stop <|im_end|>

最终得到GGUF模型文件和Modelfile文件。

-rw-r--r--. 1 root root       290 Jun 21 14:00 Modelfile
-rw-r--r--. 1 root root 352969408 Jun 21 13:44 qwen2-0_5b-instruct-q4_0.gguf

构建镜像

1、将先前拉取的基础镜像导入内网设备，然后编写Dockerfile文件：

FROM ollama:latest
EXPOSE 11434

ADD Modelfile /tmp/Modelfile
ADD qwen2-0_5b-instruct-q4_0.gguf /tmp/qwen2-0_5b-instruct-q4_0.gguf

ENTRYPOINT ["sh","-c","/bin/ollama serve"]

2、构建docker镜像，执行docker build -t ollama_qwen2-0_5b:1.0 -f Dockerfile .​：

(base) [root@localhost docker-qwen2]# docker build -t ollama_qwen2-0_5b:1.0 -f Dockerfile .
[+] Building 1.7s (8/8) FINISHED                                                                                                                                                docker:default
 => [internal] load .dockerignore                                                                                                                                                         0.4s
 => => transferring context: 2B                                                                                                                                                           0.0s
 => [internal] load build definition from Dockerfile                                                                                                                                      0.5s
 => => transferring dockerfile: 303B                                                                                                                                                      0.0s
 => [internal] load metadata for docker.io/library/ollama:latest                                                                                                                          0.0s
 => [1/3] FROM docker.io/library/ollama:latest                                                                                                                                            0.0s
 => [internal] load build context                                                                                                                                                         0.1s
 => => transferring context: 201B                                                                                                                                                         0.0s
 => CACHED [2/3] ADD Modelfile /tmp/Modelfile                                                                                                                                             0.0s
 => CACHED [3/3] ADD qwen2-0_5b-instruct-q4_0.gguf /tmp/qwen2-0_5b-instruct-q4_0.gguf                                                                                                     0.0s
 => exporting to image                                                                                                                                                                    0.1s
 => => exporting layers                                                                                                                                                                   0.0s
 => => writing image sha256:a6a949928f9bffffe1fbc5ee2c1002bd76afd9a9579dc10c6598faebb57a4885                                                                                              0.0s
 => => naming to docker.io/library/ollama_qwen2-0_5b:1.0

运行容器

1、创建并运行容器，执行docker run -itd --name ollama_qwen2 -p 11434:11434 ollama_qwen2-0_5b:1.0​

(base) [root@localhost docker-qwen2]# docker run -itd --name ollama_qwen2 -p 11434:11434 ollama_qwen2-0_5b:1.0
b034390bf79ceca1ec67bb4f9898c930c2a6efe8260bb8ba0fcbe5ffd2634f1a

2、验证docker容器是否执行成功：

(base) [root@localhost docker-qwen2]# docker ps -a
CONTAINER ID   IMAGE                     COMMAND                  CREATED         STATUS            PORTS                                             NAMES
0341f573e41f   ollama_qwen2-0_5b:1.0     "sh -c '/bin/ollama …"   6 seconds ago   Up 3 seconds      0.0.0.0:11434->11434/tcp, :::11434->11434/tcp     ollama_qwen2

到这里，我们已经部署好了docker版的ollama。这时ollama里并没有运行任何的模型，还需要我们进入容器创建加载一下。

加载Qwen2模型

1、首先进入我们刚刚运行的容器：

docker exec -it 0341f573e41f /bin/bash

2、执行ollama create​命令，创建及加载Qwen2模型：

root@0341f573e41f:/# ollama create qwen:0.5b -f /tmp/Modelfile
transferring model data
using existing layer sha256:aca679832ded61145239ce7f5c5ebddb1c57ada786c9c23733899c3888e0596f
creating new layer sha256:62fbfd9ed093d6e5ac83190c86eec5369317919f4b149598d2dbb38900e9faef
creating new layer sha256:f02dd72bb2423204352eabc5637b44d79d17f109fdb510a7c51455892aa2d216
creating new layer sha256:a82ce90cbc26a4c59e0985d2bceffa1d2616a1579218ff3cb656a3252cafdcc0
writing manifest
success

3、以上内容显示Qwen2模型已经成功在ollama中运行，然后输入exit​退出容器即可。

问答测试

1、基于我自己写的open-ai-java的框架，访问Ollama服务的代码：

# 访问代码
public static void test0(){
    OpenAIChat openAIChat = OpenAIChat.builder()
            .endpointUrl("http://10.8.xxx.xxx:11434/v1")
            .model("qwen:0.5b")
            .build().init();
    String stringFlux = openAIChat.chat("0dbe1580-60ae-4440-9462-df0a8f629f2c","你好");
    System.out.println(stringFlux);
}

# Idea中的响应日志
17:05:32.370 [main] INFO com.xxx.openai.llms.OpenAIChat - OpenAI 请求参数: {top_p=0.78, max_tokens=20000, temperature=0.9, messages=[Message(role=user, content=你好)], model=qwen:0.5b}
17:05:34.547 [main] INFO com.xxx.openai.llms.OpenAIChat - OpenAI 处理成功 响应结果为:
您好！有什么我可以帮助您的吗？
{"id":"chatcmpl-675","object":"chat.completion","created":1718960749,"model":"qwen:0.5b","system_fingerprint":"fp_ollama","choices":[{"index":0,"message":{"role":"assistant","content":"您好！有什么我可以帮助您的吗？"},"finish_reason":"stop"}],"usage":{"prompt_tokens":9,"completion_tokens":9,"total_tokens":18}}

2、查询docker容器中的日志，可以看到服务运行良好：

(base) [root@localhost docker-qwen2]# docker logs -f 0341f573e41f
Couldn't find '/root/.ollama/id_ed25519'. Generating new private key.
Your new public key is:

ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAID82w1ycq7Ej68YHdCJHVPh4xOy09uzzzCk2c9hvJcvg

2024/06/21 09:00:37 routes.go:1060: INFO server config env="map[CUDA_VISIBLE_DEVICES: GPU_DEVICE_ORDINAL: HIP_VISIBLE_DEVICES: HSA_OVERRIDE_GFX_VERSION: OLLAMA_DEBUG:false OLLAMA_FLASH_ATTENTION:false OLLAMA_HOST:http://0.0.0.0:11434 OLLAMA_INTEL_GPU:false OLLAMA_KEEP_ALIVE: OLLAMA_LLM_LIBRARY: OLLAMA_MAX_LOADED_MODELS:1 OLLAMA_MAX_QUEUE:512 OLLAMA_MAX_VRAM:0 OLLAMA_MODELS:/root/.ollama/models OLLAMA_NOHISTORY:false OLLAMA_NOPRUNE:false OLLAMA_NUM_PARALLEL:1 OLLAMA_ORIGINS:[http://localhost https://localhost http://localhost:* https://localhost:* http://127.0.0.1 https://127.0.0.1 http://127.0.0.1:* https://127.0.0.1:* http://0.0.0.0 https://0.0.0.0 http://0.0.0.0:* https://0.0.0.0:* app://* file://* tauri://*] OLLAMA_RUNNERS_DIR: OLLAMA_SCHED_SPREAD:false OLLAMA_TMPDIR: ROCR_VISIBLE_DEVICES:]"
time=2024-06-21T09:00:37.756Z level=INFO source=images.go:725 msg="total blobs: 0"
time=2024-06-21T09:00:37.756Z level=INFO source=images.go:732 msg="total unused blobs removed: 0"
time=2024-06-21T09:00:37.756Z level=INFO source=routes.go:1106 msg="Listening on [::]:11434 (version 0.1.45)"
time=2024-06-21T09:00:37.757Z level=INFO source=payload.go:30 msg="extracting embedded files" dir=/tmp/ollama1087015210/runners
time=2024-06-21T09:00:40.123Z level=INFO source=payload.go:44 msg="Dynamic LLM libraries [cpu cpu_avx cpu_avx2 cuda_v11 rocm_v60101]"
time=2024-06-21T09:00:40.125Z level=INFO source=types.go:98 msg="inference compute" id=0 library=cpu compute="" driver=0.0 name="" total="62.6 GiB" available="41.1 GiB"
[GIN] 2024/06/21 - 09:02:27 | 404 |    2.961658ms |     10.8.10.196 | POST     "/v1/chat/completions"
[GIN] 2024/06/21 - 09:04:31 | 200 |      20.105µs |       127.0.0.1 | HEAD     "/"
[GIN] 2024/06/21 - 09:05:05 | 200 |      13.876µs |       127.0.0.1 | HEAD     "/"
[GIN] 2024/06/21 - 09:05:06 | 201 |  854.761714ms |       127.0.0.1 | POST     "/api/blobs/sha256:aca679832ded61145239ce7f5c5ebddb1c57ada786c9c23733899c3888e0596f"
[GIN] 2024/06/21 - 09:05:08 | 200 |   1.59888621s |       127.0.0.1 | POST     "/api/create"
time=2024-06-21T09:05:48.013Z level=INFO source=memory.go:309 msg="offload to cpu" layers.requested=-1 layers.model=25 layers.offload=0 layers.split="" memory.available="[41.1 GiB]" memory.required.full="662.7 MiB" memory.required.partial="0 B" memory.required.kv="24.0 MiB" memory.required.allocations="[662.7 MiB]" memory.weights.total="217.0 MiB" memory.weights.repeating="79.1 MiB" memory.weights.nonrepeating="137.9 MiB" memory.graph.full="298.5 MiB" memory.graph.partial="405.0 MiB"
time=2024-06-21T09:05:48.013Z level=INFO source=server.go:359 msg="starting llama server" cmd="/tmp/ollama1087015210/runners/cpu_avx2/ollama_llama_server --model /root/.ollama/models/blobs/sha256-aca679832ded61145239ce7f5c5ebddb1c57ada786c9c23733899c3888e0596f --ctx-size 2048 --batch-size 512 --embedding --log-disable --parallel 1 --port 36844"
time=2024-06-21T09:05:48.047Z level=INFO source=sched.go:382 msg="loaded runners" count=1
time=2024-06-21T09:05:48.047Z level=INFO source=server.go:547 msg="waiting for llama runner to start responding"
time=2024-06-21T09:05:48.048Z level=INFO source=server.go:585 msg="waiting for server to become available" status="llm server error"
INFO [main] build info | build=1 commit="7c26775" tid="139630217590656" timestamp=1718960748
INFO [main] system info | n_threads=8 n_threads_batch=-1 system_info="AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | " tid="139630217590656" timestamp=1718960748 total_threads=8
INFO [main] HTTP server listening | hostname="127.0.0.1" n_threads_http="7" port="36844" tid="139630217590656" timestamp=1718960748
llama_model_loader: loaded meta data with 26 key-value pairs and 290 tensors from /root/.ollama/models/blobs/sha256-aca679832ded61145239ce7f5c5ebddb1c57ada786c9c23733899c3888e0596f (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2
llama_model_loader: - kv   1:                               general.name str              = qwen2-0_5b-instruct
llama_model_loader: - kv   2:                          qwen2.block_count u32              = 24
llama_model_loader: - kv   3:                       qwen2.context_length u32              = 32768
llama_model_loader: - kv   4:                     qwen2.embedding_length u32              = 896
llama_model_loader: - kv   5:                  qwen2.feed_forward_length u32              = 4864
llama_model_loader: - kv   6:                 qwen2.attention.head_count u32              = 14
llama_model_loader: - kv   7:              qwen2.attention.head_count_kv u32              = 2
llama_model_loader: - kv   8:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv   9:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  10:                          general.file_type u32              = 2
llama_model_loader: - kv  11:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  12:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  14:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
time=2024-06-21T09:05:48.299Z level=INFO source=server.go:585 msg="waiting for server to become available" status="llm server loading model"
llama_model_loader: - kv  15:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  16:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  17:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  19:                    tokenizer.chat_template str              = {% for message in messages %}{% if lo...
llama_model_loader: - kv  20:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  21:               general.quantization_version u32              = 2
llama_model_loader: - kv  22:                      quantize.imatrix.file str              = ../Qwen2/gguf/qwen2-0_5b-imatrix/imat...
llama_model_loader: - kv  23:                   quantize.imatrix.dataset str              = ../sft_2406.txt
llama_model_loader: - kv  24:             quantize.imatrix.entries_count i32              = 168
llama_model_loader: - kv  25:              quantize.imatrix.chunks_count i32              = 1937
llama_model_loader: - type  f32:  121 tensors
llama_model_loader: - type q4_0:  165 tensors
llama_model_loader: - type q4_1:    3 tensors
llama_model_loader: - type q8_0:    1 tensors
llm_load_vocab: special tokens cache size = 293
llm_load_vocab: token to piece cache size = 0.9338 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = qwen2
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 151936
llm_load_print_meta: n_merges         = 151387
llm_load_print_meta: n_ctx_train      = 32768
llm_load_print_meta: n_embd           = 896
llm_load_print_meta: n_head           = 14
llm_load_print_meta: n_head_kv        = 2
llm_load_print_meta: n_layer          = 24
llm_load_print_meta: n_rot            = 64
llm_load_print_meta: n_embd_head_k    = 64
llm_load_print_meta: n_embd_head_v    = 64
llm_load_print_meta: n_gqa            = 7
llm_load_print_meta: n_embd_k_gqa     = 128
llm_load_print_meta: n_embd_v_gqa     = 128
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-06
llm_load_print_meta: f_clamp_kqv      = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: f_logit_scale    = 0.0e+00
llm_load_print_meta: n_ff             = 4864
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: causal attn      = 1
llm_load_print_meta: pooling type     = 0
llm_load_print_meta: rope type        = 2
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 1000000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 32768
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: ssm_d_conv       = 0
llm_load_print_meta: ssm_d_inner      = 0
llm_load_print_meta: ssm_d_state      = 0
llm_load_print_meta: ssm_dt_rank      = 0
llm_load_print_meta: model type       = 1B
llm_load_print_meta: model ftype      = Q4_0
llm_load_print_meta: model params     = 494.03 M
llm_load_print_meta: model size       = 330.95 MiB (5.62 BPW)
llm_load_print_meta: general.name     = qwen2-0_5b-instruct
llm_load_print_meta: BOS token        = 151643 '<|endoftext|>'
llm_load_print_meta: EOS token        = 151645 '<|im_end|>'
llm_load_print_meta: PAD token        = 151643 '<|endoftext|>'
llm_load_print_meta: LF token         = 148848 'ÄĬ'
llm_load_print_meta: EOT token        = 151645 '<|im_end|>'
llm_load_tensors: ggml ctx size =    0.14 MiB
llm_load_tensors:        CPU buffer size =   330.95 MiB
llama_new_context_with_model: n_ctx      = 2048
llama_new_context_with_model: n_batch    = 512
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 1000000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:        CPU KV buffer size =    24.00 MiB
llama_new_context_with_model: KV self size  =   24.00 MiB, K (f16):   12.00 MiB, V (f16):   12.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.58 MiB
llama_new_context_with_model:        CPU compute buffer size =   298.50 MiB
llama_new_context_with_model: graph nodes  = 846
llama_new_context_with_model: graph splits = 1
INFO [main] model loaded | tid="139630217590656" timestamp=1718960748
time=2024-06-21T09:05:48.801Z level=INFO source=server.go:590 msg="llama runner started in 0.75 seconds"
[GIN] 2024/06/21 - 09:05:49 | 200 |  1.930408217s |     10.8.10.196 | POST     "/v1/chat/completions"

‍

# docker # 容器
Logo
欢迎加入ModelScope魔搭中文开源社区

ModelScope旨在打造下一代开源的模型即服务共享平台，为泛AI开发者提供灵活、易用、低成本的一站式模型服务产品，让模型应用更简单！

更多推荐

cover

可控高清视频生成： CogVideoX+DiffSynth-Studio = “配置拉满”

avatar ModelScope魔搭社区
cover

小钢炮进化，MiniCPM 3.0 开源！4B参数超GPT3.5性能，无限长文本，超强RAG三件套！模型推理、微调实战来啦！

avatar ModelScope魔搭社区
cover

一键服务化：从魔搭开源模型到OpenAI API服务

avatar ModelScope魔搭社区