[BUG] CuTe DSL: cute.compile stalls for minutes when no GPU is visible #3681

Which component has the problem?

CuTe DSL

Bug Report

Describe the bug

With no GPU visible and CUTE_DSL_ARCH unset, cute.compile stalls for minutes with no output. It then compiles for the default arch. On 4.8.0 a trivial kernel took 773 s; on 4.7.0 it took 325 s.

Arch detection recurses through its own error message:

1. env_manager.arch calls detect_gpu_arch, which calls cuInit.
2. cuInit fails with CUDA_ERROR_NO_DEVICE, so checkCudaErrors raises DSLCudaRuntimeError.
3. That error's constructor calls _get_friendly_cuda_error_message, which reads env_manager.arch again. _arch is still unset, so this is step 1 again.

Each round pays a failing cuInit, about 2 s here, until Python's recursion limit. Then the except in detect_gpu_arch unwinds it. A faulthandler dump shows these frames repeating:

base_dsl/runtime/cuda.py:163  get_compute_capability_major_minor
base_dsl/env_manager.py:422   detect_gpu_arch
base_dsl/env_manager.py:999   arch
base_dsl/common.py:302        _get_friendly_cuda_error_message
base_dsl/common.py:459        __init__
base_dsl/runtime/cuda.py:359  checkCudaErrors

Steps/Code to reproduce bug

import time
import cuda.bindings.driver as cuda
import cutlass
import cutlass.cute as cute
from cutlass.cute.runtime import make_fake_stream


@cute.kernel
def noop_kernel(x: cutlass.Int32):
    pass


@cute.jit
def noop(x: cutlass.Int32, stream: cuda.CUstream):
    noop_kernel(x).launch(grid=(1, 1, 1), block=(32, 1, 1), stream=stream)


start = time.perf_counter()
cute.compile(noop, cutlass.Int32(0), make_fake_stream(), options="--enable-tvm-ffi")
print(f"cutlass-dsl {cutlass.__version__}: compiled in {time.perf_counter() - start:.0f} s", flush=True)

$ CUDA_VISIBLE_DEVICES= python repro.py
cutlass-dsl 4.8.0: compiled in 773 s
$ CUDA_VISIBLE_DEVICES= CUTE_DSL_ARCH=sm_100a python repro.py
cutlass-dsl 4.8.0: compiled in 0 s

Expected behavior

With no device, compile quickly for the default arch, or raise an error that points at CUTE_DSL_ARCH. Building an error message shouldn't re-enter device detection.

Environment details (please complete the following information):

• Environment location: Bare-metal, run with CUDA_VISIBLE_DEVICES= to hide the GPU
• nvidia-cutlass-dsl 4.8.0 and 4.7.0 (cu13), cuda-python 13.4.1
• B200 host, driver 580.159.03, Ubuntu 22.04, Python 3.12

Additional context

This hits CPU-only build machines that compile kernels ahead of time. Workaround: set CUTE_DSL_ARCH.