编译运行 llama.cpp (vulkan, Intel GPU SYCL)

llama.cpp 是一个运行 AI (神经网络) 语言大模型的推理程序, 支持多种 后端 (backend), 也就是不同的具体的运行方式, 比如 CPU 运行, GPU 运行等.

但是编译运行 llama.cpp 并不是那么容易的, 特别是对于 SYCL 后端 (用于 Intel GPU), 坑那是一大堆. 只有特定版本的 llama.cpp, 特定版本的 Linux 系统和 GPU 驱动程序, 才可能成功运行, 否则都是失败. 能够运行的版本还不是最新版本, 经过了大量尝试和失败, 才获得了本文的结果. 本文适用于 Intel GPU (A770) 和 Linux 操作系统.

这里是 穷人小水滴, 专注于 穷人友好型 低成本技术.

---

相关文章:

• 《QEMU/KVM 虚拟机显卡透传 (vfio-pci)》 https://blog.csdn.net/secext2022/article/details/141473097
• 《自制神之嘴: fish-speech 容器镜像 (文本转语音 AI 大模型)》 https://blog.csdn.net/secext2022/article/details/141224704

目录

• 1 下载 llama.cpp 源代码
• 2 编译 llama.cpp
  • 2.1 编译 vulkan 后端
  • 2.2 编译 SYCL (Intel oneAPI) 后端
• 3 运行测试
  • 3.1 vulkan 运行测试
  • 3.2 SYCL 运行测试
• 4 总结与展望

1 下载 llama.cpp 源代码

可以从网页下载: https://github.com/ggerganov/llama.cpp

也可以使用 git 命令 (下载 b3600 版本):

git clone https://github.com/ggerganov/llama.cpp --branch b3600 --single-branch --depth=1

下载 b3038 版本:

git clone https://github.com/ggerganov/llama.cpp --branch b3038 --single-branch --depth=1

---

vulkan 后端参考文档:

• https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md#vulkan
• Run LLMs on Any GPU: GPT4All Universal GPU Support: https://www.nomic.ai/blog/posts/gpt4all-gpu-inference-with-vulkan

SYCL 后端参考文档:

• https://github.com/ggerganov/llama.cpp/blob/master/docs/backend/SYCL.md

下载 GGUF 模型文件 (llama-2-7b.Q4_K_M.gguf): https://hf-mirror.com/TheBloke/Llama-2-7B-GGUF

2 编译 llama.cpp

为了方便, 窝们使用容器 (podman) 来进行编译. 首先构建基础镜像, Dockerfile 如下:

# ubuntu-intel-base (ubuntu 22.04)
FROM quay.io/jitesoft/ubuntu:22.04

#RUN apt update && apt install -y ca-certificates
# 修复网络错误
RUN touch /etc/apt/apt.conf.d/99-verify-peer.conf && echo >>/etc/apt/apt.conf.d/99-verify-peer.conf "Acquire { https::Verify-Peer false }"

# 设置 apt 镜像 (加速软件包下载)
RUN sed -i 's/http:\/\/archive.ubuntu.com/https:\/\/mirror.sjtu.edu.cn/g' /etc/apt/sources.list
# 更新系统, 安装软件包, 清理
RUN apt update && apt upgrade -y && apt install -y ca-certificates curl gpg cmake git && apt clean

# 添加 intel apt 仓库
RUN curl https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor > /usr/share/keyrings/oneapi-archive-keyring.gpg
RUN echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" > /etc/apt/sources.list.d/oneAPI.list
# https://dgpu-docs.intel.com/driver/client/overview.html
RUN curl https://repositories.intel.com/gpu/intel-graphics.key | gpg --dearmor > /usr/share/keyrings/intel-graphics.gpg
RUN echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/gpu/ubuntu jammy client" > /etc/apt/sources.list.d/intel-gpu-jammy.list
RUN apt update

# 安装 GPU 驱动
RUN apt install -y clinfo hwinfo intel-opencl-icd intel-level-zero-gpu level-zero intel-level-zero-gpu-raytracing mesa-vulkan-drivers intel-igc-cm level-zero-dev && apt clean
# 安装 oneAPI
RUN apt install -y intel-oneapi-dpcpp-cpp-2024.2=2024.2.1-1079 intel-oneapi-mkl-devel=2024.2.1-103 intel-oneapi-ccl-devel=2021.13.1-31 && apt clean

CMD /bin/bash

执行命令:

podman build -t ubuntu-intel-base .

结果:

> podman images
REPOSITORY                     TAG       IMAGE ID      CREATED        SIZE
localhost/ubuntu-intel-base    latest    f0991290cd93  5 minutes ago  7.43 GB

2.1 编译 vulkan 后端

Dockerfile 如下:

# llama.cpp vulkan
FROM ubuntu-intel-base as build

# 安装 vulkan-sdk
RUN curl https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
RUN curl -o /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
RUN apt update && apt-get install -y vulkan-sdk && apt clean

WORKDIR /app
COPY llama.cpp-b3600 .

RUN cmake -B build -DGGML_VULKAN=1 -DBUILD_SHARED_LIBS=OFF && \
    cmake --build build --config Release --target llama-cli

# 阶段 2
FROM ubuntu-intel-base as runtime

COPY --from=build /app/build/bin/llama-cli /usr/bin/llama-cli

CMD /bin/bash

执行命令:

podman build -t ubuntu-llamacpp-vulkan .

结果:

> podman images
REPOSITORY                          TAG       IMAGE ID      CREATED        SIZE
localhost/ubuntu-llamacpp-vulkan    latest    ee12a2a1e6f0  2 minutes ago  7.44 GB

2.2 编译 SYCL (Intel oneAPI) 后端

Dockerfile 如下:

# llama.cpp-b3038/.devops/main-intel.Dockerfile
FROM ubuntu-intel-base as build

WORKDIR /app
COPY llama.cpp-b3038 .

RUN . /opt/intel/oneapi/setvars.sh && \
    cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON && \
    cmake --build build --config Release --target main

# 阶段 2
FROM ubuntu-intel-base as runtime

COPY --from=build /app/build/bin/main /usr/bin/main

CMD /bin/bash

执行命令:

podman build -t ubuntu-llamacpp-sycl .

结果:

> podman images
REPOSITORY                        TAG       IMAGE ID      CREATED         SIZE
localhost/ubuntu-llamacpp-sycl    latest    2baacc3bb758  35 seconds ago  7.44 GB

3 运行测试

由于 Intel SYCL 只能在特定系统的特定驱动版本才能正常运行, 所以使用了虚拟机 GPU 透传 (详见文章 《QEMU/KVM 虚拟机显卡透传 (vfio-pci)》).

在这篇文章的虚拟机的基础上, 需要额外安装软件包:

sudo apt -y install gawk dkms linux-headers-$(uname -r) libc6-dev
sudo apt install -y intel-i915-dkms

重启虚拟机. 虚拟机内的相关信息如下:

a2@a2s:~$ uname -a
Linux a2s 6.8.0-40-generic #40~22.04.3-Ubuntu SMP PREEMPT_DYNAMIC Tue Jul 30 17:30:19 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
a2@a2s:~$ clinfo -l
Platform #0: Intel(R) OpenCL Graphics
 `-- Device #0: Intel(R) Arc(TM) A770 Graphics
a2@a2s:~$ source /opt/intel/oneapi/setvars.sh
 
:: initializing oneAPI environment ...
   -bash: BASH_VERSION = 5.1.16(1)-release
   args: Using "$@" for setvars.sh arguments: 
:: ccl -- latest
:: compiler -- latest
:: debugger -- latest
:: dev-utilities -- latest
:: mkl -- latest
:: mpi -- latest
:: tbb -- latest
:: oneAPI environment initialized ::
 
a2@a2s:~$ sycl-ls
[opencl:cpu][opencl:0] Intel(R) OpenCL, AMD Ryzen 5 5600G with Radeon Graphics          OpenCL 3.0 (Build 0) [2024.18.7.0.11_160000]
[opencl:gpu][opencl:1] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [24.22.29735.27]
[level_zero:gpu][level_zero:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.29735]

把上面编译的 llama.cpp 程序 (以及 gguf 模型文件) 复制到虚拟机:

a2@a2s:~$ ls -l
total 11906072
-r--r--r-- 1 a2 a2 4081004224 Aug 21 05:13 llama-2-7b.Q4_K_M.gguf
-rwxr-xr-x 1 a2 a2    5898128 Aug 23 03:39 llama-cli
-rwxr-xr-x 1 a2 a2    6378544 Aug 23 03:39 main
-r--r--r-- 1 a2 a2 8098522912 Aug 21 06:25 qwen2-7b-instruct-q8_0.gguf
a2@a2s:~$ ./llama-cli --version
version: 1 (2fb9267)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
a2@a2s:~$ ./main --version
version: 1 (fb76ec3)
built with Intel(R) oneAPI DPC++/C++ Compiler 2024.2.1 (2024.2.1.20240711) for x86_64-unknown-linux-gnu

3.1 vulkan 运行测试

使用模型 llama-2-7b.Q4_K_M.gguf, 生成长度 200:

a2@a2s:~$ ./llama-cli -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 200 -ngl 33
Log start
main: build = 1 (2fb9267)
main: built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: seed  = 1724384928
llama_model_loader: loaded meta data with 19 key-value pairs and 291 tensors from llama-2-7b.Q4_K_M.gguf (version GGUF V2)
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = llama
llama_model_loader: - kv   1:                               general.name str              = LLaMA v2
llama_model_loader: - kv   2:                       llama.context_length u32              = 4096
llama_model_loader: - kv   3:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   4:                          llama.block_count u32              = 32
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 11008
llama_model_loader: - kv   6:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv   7:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   8:              llama.attention.head_count_kv u32              = 32
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                          general.file_type u32              = 15
llama_model_loader: - kv  11:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  12:                      tokenizer.ggml.tokens arr[str,32000]   = ["", "", "", "<0x00>", "<...
llama_model_loader: - kv  13:                      tokenizer.ggml.scores arr[f32,32000]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  14:                  tokenizer.ggml.token_type arr[i32,32000]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  15:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  16:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  17:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  18:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q4_K:  193 tensors
llama_model_loader: - type q6_K:   33 tensors
llm_load_vocab: special tokens cache size = 3
llm_load_vocab: token to piece cache size = 0.1684 MB
llm_load_print_meta: format           = GGUF V2
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = SPM
llm_load_print_meta: n_vocab          = 32000
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 4096
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 32
llm_load_print_meta: n_rot            = 128
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 128
llm_load_print_meta: n_embd_head_v    = 128
llm_load_print_meta: n_gqa            = 1
llm_load_print_meta: n_embd_k_gqa     = 4096
llm_load_print_meta: n_embd_v_gqa     = 4096
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-05
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             = 11008
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        = 0
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 4096
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       = 7B
llm_load_print_meta: model ftype      = Q4_K - Medium
llm_load_print_meta: model params     = 6.74 B
llm_load_print_meta: model size       = 3.80 GiB (4.84 BPW) 
llm_load_print_meta: general.name     = LLaMA v2
llm_load_print_meta: BOS token        = 1 ''
llm_load_print_meta: EOS token        = 2 ''
llm_load_print_meta: UNK token        = 0 ''
llm_load_print_meta: LF token         = 13 '<0x0A>'
llm_load_print_meta: max token length = 48
ggml_vulkan: Found 1 Vulkan devices:
Vulkan0: Intel(R) Arc(tm) A770 Graphics (DG2) (Intel open-source Mesa driver) | uma: 0 | fp16: 1 | warp size: 32
llm_load_tensors: ggml ctx size =    0.27 MiB
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 33/33 layers to GPU
llm_load_tensors:        CPU buffer size =    70.31 MiB
llm_load_tensors: Intel(R) Arc(tm) A770 Graphics (DG2) buffer size =  3820.93 MiB
..................................................................................................
llama_new_context_with_model: n_ctx      = 4096
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init: Intel(R) Arc(tm) A770 Graphics (DG2) KV buffer size =  2048.00 MiB
llama_new_context_with_model: KV self size  = 2048.00 MiB, K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
llama_new_context_with_model: Vulkan_Host  output buffer size =     0.12 MiB
llama_new_context_with_model: Intel(R) Arc(tm) A770 Graphics (DG2) compute buffer size =   296.00 MiB
llama_new_context_with_model: Vulkan_Host compute buffer size =    16.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 2

system_info: n_threads = 4 / 4 | 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 = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | 
sampling: 
	repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
	top_k = 40, tfs_z = 1.000, top_p = 0.950, min_p = 0.050, typical_p = 1.000, temp = 0.800
	mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampling order: 
CFG -> Penalties -> top_k -> tfs_z -> typical_p -> top_p -> min_p -> temperature 
generate: n_ctx = 4096, n_batch = 2048, n_predict = 200, n_keep = 1


 hello, this is a very very long story. I'm in 10th grade and I'm a girl.

(此处省略部分输出结果)

llama_print_timings:        load time =    2264.36 ms
llama_print_timings:      sample time =       6.68 ms /   200 runs   (    0.03 ms per token, 29958.06 tokens per second)
llama_print_timings: prompt eval time =     440.55 ms /    10 tokens (   44.05 ms per token,    22.70 tokens per second)
llama_print_timings:        eval time =    7684.85 ms /   199 runs   (   38.62 ms per token,    25.90 tokens per second)
llama_print_timings:       total time =    8149.27 ms /   209 tokens
Log end

生成速度约为 25.90 tokens per second, 也就是每秒输出 25.9 个字符.

3.2 SYCL 运行测试

使用模型 llama-2-7b.Q4_K_M.gguf, 生成长度 200:

a2@a2s:~$ ./main -m llama-2-7b.Q4_K_M.gguf -p "hello, this is a very very long story" -n 200 -ngl 33
Log start
main: build = 1 (fb76ec3)
main: built with Intel(R) oneAPI DPC++/C++ Compiler 2024.2.1 (2024.2.1.20240711) for x86_64-unknown-linux-gnu
main: seed  = 1724384798

(此处省略部分输出结果)

[SYCL] call ggml_init_sycl
ggml_init_sycl: GGML_SYCL_DEBUG: 0
ggml_init_sycl: GGML_SYCL_F16: yes
found 3 SYCL devices:
|  |                   |                                       |       |Max    |        |Max  |Global |                     |
|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
| 0| [level_zero:gpu:0]|                Intel Arc A770 Graphics|    1.3|    512|    1024|   32| 16225M|            1.3.29735|
| 1|     [opencl:gpu:0]|                Intel Arc A770 Graphics|    3.0|    512|    1024|   32| 16225M|       24.22.29735.27|
| 2|     [opencl:cpu:0]|AMD Ryzen 5 5600G with Radeon Graphics         |    3.0|      4|    8192|   64|  8327M|2024.18.7.0.11_160000|
ggml_backend_sycl_set_mul_device_mode: true
detect 1 SYCL GPUs: [0] with top Max compute units:512
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
llm_load_tensors: ggml ctx size =    0.30 MiB
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 33/33 layers to GPU
llm_load_tensors:      SYCL0 buffer size =  3820.94 MiB
llm_load_tensors:        CPU buffer size =    70.31 MiB
..................................................................................................
llama_new_context_with_model: n_ctx      = 512
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  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:      SYCL0 KV buffer size =   256.00 MiB
llama_new_context_with_model: KV self size  =  256.00 MiB, K (f16):  128.00 MiB, V (f16):  128.00 MiB
llama_new_context_with_model:  SYCL_Host  output buffer size =     0.12 MiB
llama_new_context_with_model:      SYCL0 compute buffer size =    70.50 MiB
llama_new_context_with_model:  SYCL_Host compute buffer size =     9.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 2

system_info: n_threads = 4 / 4 | 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 = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | 
sampling: 
	repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
	top_k = 40, tfs_z = 1.000, top_p = 0.950, min_p = 0.050, typical_p = 1.000, temp = 0.800
	mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampling order: 
CFG -> Penalties -> top_k -> tfs_z -> typical_p -> top_p -> min_p -> temperature 
generate: n_ctx = 512, n_batch = 2048, n_predict = 200, n_keep = 1


 hello, this is a very very long story, so i am going to break it up in two parts.
i started out by saying that i am a very good girl. and that i am.

(此处省略部分输出结果)

llama_print_timings:        load time =    1727.57 ms
llama_print_timings:      sample time =       5.67 ms /   200 runs   (    0.03 ms per token, 35248.50 tokens per second)
llama_print_timings: prompt eval time =     377.79 ms /    10 tokens (   37.78 ms per token,    26.47 tokens per second)
llama_print_timings:        eval time =    6517.30 ms /   199 runs   (   32.75 ms per token,    30.53 tokens per second)
llama_print_timings:       total time =    6917.86 ms /   209 tokens
Log end

生成速度约为 30.53 tokens per second, 也就是每秒输出 30.5 个字符.

4 总结与展望

本文使用容器 (podman) 编译了 llama.cpp 的 vulkan 后端和 SYCL 后端, 并成功在 Intel GPU (A770) 运行, 获得了较快的语言模型推理速度.

SYCL 后端比 vulkan 后端稍微快一点, 但不多. 使用的模型 (gguf), 生成长度, 软件驱动版本, 运行参数设置等很多因素, 都可能影响模型推理的速度, 所以本文中的运行速度仅供参考.

SYCL 比 vulkan 快不了多少, 但是使用 SYCL (Intel oneAPI) 却非常麻烦 !! 所以, 至少目前为止, 对于 A770 (16GB) 显卡来说, 使用 vulkan 即可, 强行使用 SYCL 的意义不大.

Intel 和 llama.cpp 对于 SYCL 还需要继续努力, 希望能够更方便, 更快速的运行大模型.

---

本文使用 CC-BY-SA 4.0 许可发布.