Add XPU and CPU branches to check_env so it stops raising NameError

[kpjeeja]

Motivation

python3 -m sglang.check_env can crash when the environment does not match one of the accelerator backends handled by the existing dispatch logic. The dispatch chain in __main__ had no XPU branch and no fallback else branch, so env could remain undefined:

$ python3 -m sglang.check_env
Traceback (most recent call last):
  File "/usr/lib/python3.12/runpy.py", line 198, in _run_module_as_main
    return _run_code(code, main_globals, None,
  ...
  File ".../python/sglang/check_env.py", line 591, in <module>
    env.check_env()
    ^^^
NameError: name 'env' is not defined

This is particularly problematic when collecting environment information for a bug report. For example, an Intel GPU/XPU environment, a CPU-engine environment, or a host where the accelerator stack is installed but currently unusable can produce a traceback instead of an environment report.

This change makes environment selection explicit and ensures that check_env always produces a report, including for XPU, CPU-engine, and otherwise unsupported environments.

Modifications

python/sglang/check_env.py

• select_env() — moves environment selection out of if __name__ == "__main__" into a module-level function that always returns a BaseEnv. This makes the selection logic importable and unit-testable. __main__ is reduced to select_env().check_env(). The selection order is CUDA → HIP → NPU → MUSA → MPS → XPU → CPU → unknown; existing backends retain their current ordering and behavior.
• XPUEnv — adds XPU-specific environment reporting, including XPU availability, devices grouped by name and memory, Level-Zero driver versions, the XPU build version reported by torch.version.xpu, and xpu-smi topology -m. It also reports relevant packages: intel-extension-for-pytorch, pytorch-triton-xpu, intel-sycl-rt, and oneccl. The xpu-smi probe uses a 15-second timeout so that a hung Level-Zero driver does not block the environment report indefinitely.
• is_xpu_v2() — detects an XPU-enabled PyTorch build using torch.version.xpu is not None, rather than relying on is_xpu(), which also depends on a usable/visible XPU device. This allows check_env to report an XPU environment even when the XPU device is currently unavailable, instead of incorrectly falling through to the generic fallback.
• CPUEnv — reports SGLANG_USE_CPU_ENGINE, platform.machine(), the CPU model, and the availability of avx512f, avx512_bf16, amx_bf16, and amx_int8 as reported by lscpu. If lscpu is unavailable, only the CPU ISA information is omitted.
• UnknownEnv — provides a fallback for environments where no supported accelerator or CPU-engine configuration is detected. It reports Accelerator: none detected together with the standard Python, PyTorch, package, and ulimit information.
• BaseEnv.__init__ — copies PACKAGE_LIST into self.package_list instead of aliasing the module-level list. Previously, a subclass extending self.package_list could modify the global PACKAGE_LIST, causing packages from one backend to leak into subsequently created environment objects in the same process.
• PACKAGE_LIST — adds torch_memory_saver.
• Moves the platform import from MPSEnv.get_info() to the module-level imports because CPUEnv also uses it.

test/registered/unit/hardware_backend/test_check_env.py (new, register_cpu_ci(est_time=5, suite="base-a-test-cpu"))

• Verifies that select_env() returns UnknownEnv instead of raising when no detector matches, and that the fallback produces a complete report.
• Verifies dispatch priority: an XPU-enabled PyTorch build is selected before the CPU fallback; a CUDA build takes precedence over later probes; and a CPU-engine environment is not reported as an unsupported/unknown environment.
• Verifies that XPUEnv reports XPU available: False when no usable XPU device is available, without attempting device-property queries.
• Verifies device grouping: identical devices are collapsed into a single entry, while different device configurations are reported separately, with driver information included.
• Verifies that missing, timed-out, or non-zero xpu-smi execution degrades gracefully to an empty topology rather than aborting the environment report.
• Verifies CPUEnv using captured lscpu output for both a Xeon configuration with AVX512/AMX support and an EPYC configuration without AVX512/AMX, as well as the missing-lscpu case.

No runtime or model-execution code is modified; this change is limited to the check_env diagnostic entry point and its unit tests.

Accuracy Tests

Not applicable — this change does not modify kernels, model execution, or inference behavior.

Speed Tests and Profiling

Not applicable — this change only affects the diagnostic CLI.

Checklist

• Format your code according to the [Format code with pre-commit](https://docs.sglang.io/developer_guide/contribution_guide.html#format-code-with-pre-commit).
• Add unit tests according to the [Run and add unit tests]
• Follow the SGLang code style [guidance](https://docs.sglang.io/developer_guide/contribution_guide.html#code-style-guidance).

---

CI States

Latest PR Test (Base): ❌ Run #35207328925
Latest PR Test (Extra): ❌ Run #35207328657
Latest PR Test (AMD ROCm 10): ❌ Run #35207328975

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[airMeng]

ipex is in maintain mode

[kpjeeja]

fixed

[airMeng]

pytorch-triton-xpu has been replaced by triton-xpu

[kpjeeja]

fixed

[airMeng]

@kpjeeja what is the difference between this PR and #38873? can you merge as one PR?

[kpjeeja]

@kpjeeja what is the difference between this PR and #38873? can you merge as one PR?

yes, this is Duplicate PR