Support llama.cpp "Multi GPU support, CUDA refactor, CUDA scratch buffer"

[wyhanz]

Multi-GPU inference is essential for small VRAM GPU. 13B llama model cannot fit in a single 3090 unless using quantization.

llama.cpp yesterday merge multi gpu branch, which help us using small VRAM GPUS to deploy LLM.
ggml-org/llama.cpp#1703

Hope llama-cpp-python can support multi GPU inference in the future.
Many thanks!!!

[abetlen]

Pushed to v0.1.59, let me know if that works.

[devilteo911]

Can I use this with the High Level API or is it available only in the Low Level ones?

[gjmulder]

Working for me with the high level API.

$ CMAKE_ARGS="-DLLAMA_CUBLAS=on" FORCE_CMAKE=1 pip install llama-cpp-python==0.1.59 --no-cache-dir
Collecting llama-cpp-python==0.1.59
  Downloading llama_cpp_python-0.1.59.tar.gz (1.4 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.4/1.4 MB 69.9 MB/s eta 0:00:00
  Installing build dependencies ... done
  Getting requirements to build wheel ... done
  Preparing metadata (pyproject.toml) ... done
Requirement already satisfied: typing-extensions>=4.5.0 in /home/mulderg/anaconda3/envs/lcp/lib/python3.10/site-packages (from llama-cpp-python==0.1.59) (4.6.0)
Requirement already satisfied: diskcache>=5.6.1 in /home/mulderg/anaconda3/envs/lcp/lib/python3.10/site-packages/diskcache-5.6.1-py3.10.egg (from llama-cpp-python==0.1.59) (5.6.1)
Requirement already satisfied: numpy>=1.20.0 in /home/mulderg/anaconda3/envs/lcp/lib/python3.10/site-packages (from llama-cpp-python==0.1.59) (1.24.3)
Building wheels for collected packages: llama-cpp-python
  Building wheel for llama-cpp-python (pyproject.toml) ... done
  Created wheel for llama-cpp-python: filename=llama_cpp_python-0.1.59-cp310-cp310-linux_x86_64.whl size=310301 sha256=2182deb938214949a0838f7b4f649cd78636f396450a48686d2fe07dabcd1d81
  Stored in directory: /data/tmp/pip-ephem-wheel-cache-pps732gv/wheels/ae/3d/33/588d5327568faa38106293c3de8fc5ba0c3cf514d4f6eec9c5
Successfully built llama-cpp-python
Installing collected packages: llama-cpp-python
  Attempting uninstall: llama-cpp-python
    Found existing installation: llama-cpp-python 0.1.57
    Uninstalling llama-cpp-python-0.1.57:
      Successfully uninstalled llama-cpp-python-0.1.57
Successfully installed llama-cpp-python-0.1.59

$ pip list | grep llama-cpp-python
llama-cpp-python         0.1.59

$ python ./smoke_test.py -f genesis50.txt -l 5 > /dev/null
ggml_init_cublas: found 2 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3090 Ti
  Device 1: NVIDIA GeForce GTX 1080 Ti
llama.cpp: loading model from /data/llama/7B/ggml-model-f16.bin
llama_model_load_internal: format     = ggjt v1 (pre #1405)
llama_model_load_internal: n_vocab    = 32000
llama_model_load_internal: n_ctx      = 8192
llama_model_load_internal: n_embd     = 4096
llama_model_load_internal: n_mult     = 256
llama_model_load_internal: n_head     = 32
llama_model_load_internal: n_layer    = 32
llama_model_load_internal: n_rot      = 128
llama_model_load_internal: ftype      = 1 (mostly F16)
llama_model_load_internal: n_ff       = 11008
llama_model_load_internal: n_parts    = 1
llama_model_load_internal: model size = 7B
llama_model_load_internal: ggml ctx size =    0.07 MB
llama_model_load_internal: using CUDA for GPU acceleration
ggml_cuda_set_main_device: using device 0 (NVIDIA GeForce RTX 3090 Ti) as main device
llama_model_load_internal: mem required  = 2292.09 MB (+ 1026.00 MB per state)
llama_model_load_internal: allocating batch_size x 1 MB = 512 MB VRAM for the scratch buffer
llama_model_load_internal: offloading 32 layers to GPU
llama_model_load_internal: offloading output layer to GPU
llama_model_load_internal: total VRAM used: 12865 MB
...................................................................................................
llama_init_from_file: kv self size  = 4096.00 MB
AVX = 1 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | VSX = 0 | 

llama_print_timings:        load time =   594.61 ms
llama_print_timings:      sample time =  1273.07 ms /  2048 runs   (    0.62 ms per token)
llama_print_timings: prompt eval time =   594.55 ms /   109 tokens (    5.45 ms per token)
llama_print_timings:        eval time = 215376.56 ms /  2047 runs   (  105.22 ms per token)
llama_print_timings:       total time = 269666.63 ms

[wyhanz]

Can I use this with the High Level API or is it available only in the Low Level ones?

Check class Llama, the parameter in __init__() (n_parts: Number of parts to split the model into. If -1, the number of parts is automatically determined.) can realize the feature. And I think high-level api is just a wrapper for low-level api to help us use more easily

[wyhanz]

Here is a brief experimental result.

I used ziya, a large-scale pre-trained model based on LLaMA with 13 billion parameters and quantized to q8_0.

The model was successfully divided into two GPUs. However, I feel that the same prompts are slower than using single 3090, perhaps need more experiments.

ggml_init_cublas: found 2 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3090
  Device 1: NVIDIA GeForce RTX 3090
llama.cpp: loading model from /data0/zhangyuhan/llama_cpp/ggml-model-q8_0.bin
llama_model_load_internal: format     = ggjt v3 (latest)
llama_model_load_internal: n_vocab    = 39424
llama_model_load_internal: n_ctx      = 512
llama_model_load_internal: n_embd     = 5120
llama_model_load_internal: n_mult     = 256
llama_model_load_internal: n_head     = 40
llama_model_load_internal: n_layer    = 40
llama_model_load_internal: n_rot      = 128
llama_model_load_internal: ftype      = 7 (mostly Q8_0)
llama_model_load_internal: n_ff       = 13824
llama_model_load_internal: n_parts    = 1
llama_model_load_internal: model size = 13B
llama_model_load_internal: ggml ctx size =    0.09 MB
llama_model_load_internal: using CUDA for GPU acceleration
ggml_cuda_set_main_device: using device 0 (NVIDIA GeForce RTX 3090) as main device
llama_model_load_internal: mem required  = 2457.17 MB (+ 1608.00 MB per state)
llama_model_load_internal: allocating batch_size x 1 MB = 512 MB VRAM for the scratch buffer
llama_model_load_internal: offloading 40 layers to GPU
llama_model_load_internal: total VRAM used: 13370 MB

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.60.11    Driver Version: 525.60.11    CUDA Version: 12.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  NVIDIA GeForce ...  Off  | 00000000:1A:00.0 Off |                  N/A |
| 30%   44C    P2   117W / 370W |   7189MiB / 24576MiB |      7%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   1  NVIDIA GeForce ...  Off  | 00000000:68:00.0 Off |                  N/A |
| 52%   66C    P2   157W / 370W |  23693MiB / 24576MiB |     15%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A   2653794      C   python                           7186MiB |
|    1   N/A  N/A   2260259      C   python                          16496MiB |
|    1   N/A  N/A   2653794      C   python                           7186MiB |
+-----------------------------------------------------------------------------+

The inference time was slower than before probably because I was running the evaluation code on another card at the same time 😄

[gjmulder]

The model was successfully divided into two GPUs. However, I feel that the same prompts are slower than using single 3090, perhaps need more experiments.

Yeah, I got worse performance using two GPUs than using one GPU as well. However, my second GPU is a 1080Ti, so I assumed it was my environment.

[AlphaAtlas]

@wyzhangyuhan I think that is correct behavior? If its like the CPU/GPU spit, the gpus will run sequentially, not in parallel. It will run some on one 3090, and then on the other.

But I think you can fit Q4 65B on 2x 3090s, with very minimal quality loss 👍

[wyhanz]

@AlphaAtlas Thanks for sharing your thoughts, I had mistakenly thought that multiple GPUs would enable faster inference. Also thanks for the suggestion, multi-GPU really solved my problem with LLM deployment! 👍