In a previous announcement, I mentioned creating a tool that performs QUBO annealing on a GPU using PyTorch, and the benchmarks were very promising. This time, I&#39;ll introduce GPU computing using MPS through PyTorch for Macs that do not have NVIDIA GPUs. For reference, the previous article discussed...<a href="https://blueqat.com/yuichiro_minato2/e09f3c90-f004-49bf-b50a-2e8e5e73f39b" rel="noopener noreferrer" target="_blank">https://blueqat.com/yuichiro_minato2/e09f3c90-f004-49bf-b50a-2e8e5e73f39b</a> 以前PyTorchを使ってGPUでのQUBOアニーリングを実行するツールを作ったとアナウンスしました。ベンチマークも大変いい感じです。今回はMac向けにNVIDIA GPUを積んでいない場合でも、PyTorch経由でMPSを使ったGPU計算を紹介します。ちなみに以前の記事は、<a href="https://blueqat.com/yuichiro_minato2/ffdbc078-5258-47ce-a45c-e84bb03982f5" rel="noopener noreferrer" target="_blank">https://blueqat.com/yuichiro_minato2/ffdbc078-5258-47ce-a45c-e84bb03982f5</a> In the current code, MPS did not work as is (though it&#39;s probable that this article will be used as a basis for corrections in the future). Essentially, Torch tensors need to be explicitly set to float32. This was achieved by specifying float32 in the following section of the ArminSampler. 現在のコードではMPSはそのままでは動きませんでした。（おそらく今後この記事をベースに修正されると思いますが。）基本的には、Torchテンソルをfloat32にする必要があります。該当箇所はArminSamplerの下記部分でfloat32を明示的に指示すればOKでした。 <pre class="ql-syntax" spellcheck="false">qmatrix = torch.tensor(qmatrix, dtype=torch.float32, device=self.device).float()
</pre> <pre class="ql-syntax" spellcheck="false">dtype=torch.float32
</pre> So finally we get,最終的には、 <pre class="ql-syntax" spellcheck="false">from tytan.tytan import *
import random
import time

N = 5000

# qubits
q = symbols_list(N, &#39;q{}&#39;)

# hamiltonian
H = 0

# biases
for i in range(N):
  H += random.randint(-10, 10) * q[i]

# Jij but set 1000 as it takes a long time to finish if we set all of connections
for i in range(1000):
  H += random.choice([-1, 1]) * q[random.randint(0,N-1)] * q[random.randint(0,N-1)]

# compile
qubo, offset = Compile(H).get_qubo()

# sampler
solver = sampler.ArminSampler(seed=None, mode=&#39;GPU&#39;, device=&#39;cuda:0&#39;, verbose=1)

start = time.time()

#sampling
result = solver.run(qubo, shots=1)

print(time.time() - start)
</pre> It felt like a difference only began to emerge around 5000 qubits, similar to what you&#39;d expect with NVIDIA GPUs.NVIDIAのGPUと同様に5000量子ビットでようやく差が出るような感じでした。 <pre class="ql-syntax" spellcheck="false">MODE: GPU
DEVICE: mps:0
54.36705923080444
</pre> <pre class="ql-syntax" spellcheck="false">solver = sampler.ArminSampler(seed=None, mode=&#39;CPU&#39;, verbose=1)

start = time.time()
result = solver.run(qubo, shots=1)
print(time.time() - start)
</pre> <pre class="ql-syntax" spellcheck="false">MODE: CPU
DEVICE: cpu
105.90340304374695
</pre> It nicely achieved acceleration with MPS as well.いい感じでMPSでも高速化ができましたね。

QUBO annealing with MPS on Mac + Torch Tytan / macのMPS + Torch TytanでQUBOアニーリング

Yuichiro Minato