Simulation distance 3 magic state cultivation

[bachase]

First off, thank you for building this tool and open sourcing it. I'm building a similar near-Clifford simulator using different techniques. I was interested in comparing performance on the magic state cultivation distance 3 circuit. Before running on a GPU, I was doing a trial on my CPU but the compilation stage seems to hang. I let it run for 5 minutes with no progress. Is there a different configuration I should be using? Sidenote that the version number inside the package differs from the PyPI package version.

Reproducer

#!/usr/bin/env python3
"""Minimal reproducer: tsim hangs compiling d=3 cultivation circuit.

Requirements:
    pip install bloqade-tsim==0.1.1

Usage:
    python tsim_hang_repro_fetch.py

On first run, downloads the circuit locally from CIRCUIT_URL.
The script appears to hang at compile_detector_sampler().
"""

from __future__ import annotations

import time
import urllib.request
from pathlib import Path

CIRCUIT_URL = (
    "https://raw.githubusercontent.com/haoliri0/SOFT/refs/heads/main/"
    "magic_state_cultivation/circuits/circuit_d3_p0.001.stim"
)
LOCAL_CIRCUIT = Path(__file__).with_name("circuit_d3_p0.001.stim")


def load_circuit_text() -> str:
    if not LOCAL_CIRCUIT.exists():
        print(f"Downloading circuit to {LOCAL_CIRCUIT} ...", flush=True)
        with urllib.request.urlopen(CIRCUIT_URL) as response:
            LOCAL_CIRCUIT.write_text(response.read().decode("utf-8"))
        print("  done.")
    return LOCAL_CIRCUIT.read_text()


def main() -> None:
    import tsim

    print(f"tsim version: {getattr(tsim, '__version__', 'unknown')}")
    print("Circuit: d=3 magic state cultivation (15 qubits)")
    print(f"Source:  {CIRCUIT_URL}")
    print()

    circuit_stim = load_circuit_text()

    print("Creating tsim.Circuit ...", flush=True)
    tc = tsim.Circuit(circuit_stim)
    print("  done.")

    print("Calling compile_detector_sampler() ...", flush=True)
    t0 = time.time()
    sampler = tc.compile_detector_sampler()
    elapsed = time.time() - t0
    print(f"  done in {elapsed:.1f}s.")

    print("Sampling 1000 shots ...", flush=True)
    t0 = time.time()
    result = sampler.sample(1000, separate_observables=True)
    elapsed = time.time() - t0
    print(f"  done in {elapsed:.1f}s.")
    print(f"  detector samples shape: {result[0].shape}")
    print(f"  observable samples shape: {result[1].shape}")


if __name__ == "__main__":
    main()

Expected

compile_detector_sampler() finishes or raises an explicit error.

Actual

The script prints:

tsim version: 0.1.0
Circuit: d=3 magic state cultivation (15 qubits)
Source:  https://raw.githubusercontent.com/haoliri0/SOFT/refs/heads/main/magic_state_cultivation/circuits/circuit_d3_p0.001.stim

Creating tsim.Circuit ...
  done.
Calling compile_detector_sampler() ...

Then it produces no further output for at least 180 seconds.

Environment

• macOS 15.7.4 (24G517)
• arm64 Apple Silicon
• Python 3.13.10
• bloqade-tsim==0.1.1
• runtime tsim.__version__ == 0.1.0
• jax==0.9.2
• jaxlib==0.9.2
• numpy==2.4.4
• stim==1.15.0

[rafaelha]

Hi, the problem is that the circuit has 29 T gates (after doubling to compute marginal probabilities, 58 T gates need to be decomposed). Here, the default decomposition strategy fails.

However, the cutting decomposition works better (~10s for me):

compile_detector_sampler(strategy="cutting")

I did, however, encounter an underflow issue, which is now fixed in v0.1.2. Could you try that one?

On a GH200 GPU I timed 23us per shoot (with around batch_size=50_000). For the d=3 cultivation results in the paper we used the circuits from https://github.com/kh428/accel-cutting-magic-state with only 16 T gates.

Also, thank you for pointing out the wrong version number.

[bachase]

Thanks @rafaelha I can confirm the cutting decomposition does work for me now. Is there a rule of thumb to know when to use the different decomposition strategies and how they impact performance?