[Fix] Don't let a failed deep_ep import-time check kill servers that never use DeepEP - #40671

Motivation

Since #33932 (2026-08-07), sgl-deep-ep is a hard dependency of the sglang wheel, so deep_ep is always installed.
deep_ep/__init__.py runs host checks at import time:

• check_nccl_so() asserts that exactly one libnccl is mapped and that it is byte-identical to the pip NCCL (AssertionError, or FileNotFoundError when a mapping reads (deleted)).
• find_cuda_home() asserts that a CUDA home is found (AssertionError).
• init_jit() initializes the C++ JIT runtime.

The guard in token_dispatcher/deepep.py dates from when "not installed" was the only failure mode, and catches only ImportError; deepep_v2.py catches (ImportError, OSError).
Every server process imports these modules, including the launcher:

launch_server -> http_server -> engine -> data_parallel_controller -> scheduler
  -> disaggregation/decode -> managers/utils -> layers/logits_processor
  -> layers/logprob_processor -> model_executor/runner_utils/__init__
  -> runner_utils/deepep_adapter -> layers/moe/token_dispatcher -> deepep -> deep_ep

So one failed check stops the server at import time, even for dense models where DeepEP is never requested.

Reproduction (H200, Qwen3-1.7B, dense, no DeepEP flags)

# any second libnccl on the host; here a source build of NCCL 2.32.3 vs the pip 2.29.7
export LD_PRELOAD=/path/to/nccl/build/lib/libnccl.so.2
python -m sglang.launch_server --model-path Qwen/Qwen3-1.7B --port 30006

Before (main @ 00986c8): the launcher exits with code 1 during import.

File "python/sglang/srt/layers/moe/token_dispatcher/deepep.py", line 53, in <module>
File ".venv/lib/python3.12/site-packages/deep_ep/__init__.py", line 91, in <module>
AssertionError: Duplicate NCCL runtime found in the current system: .../nccl/build/lib/libnccl.so.2.32.3 and .../site-packages/nvidia/nccl/lib/libnccl.so.2

After: the server starts with the custom NCCL mapped, and /generate returns Paris. The capital of the United States is Washington, D for "The capital of France is".
The same failure also happened without LD_PRELOAD on a shared network filesystem, where the detokenizer saw the NCCL mapping as (deleted) and filecmp raised FileNotFoundError.

Modifications

• deepep.py: catch any exception from the optional import, keep it in _deepep_import_error, and chain it into the ImportError raised when a DeepEP dispatcher is created. The message now says "not available" instead of "not installed", because the package usually is installed.
• deepep_v2.py: widen (ImportError, OSError) to Exception for the deep_ep import. The v2 module already stores and reports the original error.

Users who request DeepEP still get a hard error, now with the root cause instead of a misleading "not installed".
The v1 try block also covers the sglang_per_token_group_quant_fp8 import, as before; a failure there is reported the same way.

Follow-up, not in this PR: runner_utils/__init__.py eagerly re-exports DeepEPCudaGraphRunnerAdapter, so importing runner_utils.pool from logprob_processor pulls every MoE dispatcher (and deep_ep) into the launcher, tokenizer and detokenizer. Making that re-export lazy would stop these processes from loading DeepEP at all.

Accuracy Tests

No model forward, kernel or sampling code is touched; only the import guard of the two DeepEP dispatcher modules changes.

• New CPU test test/registered/unit/layers/moe/test_deepep_import_guard.py runs a fresh interpreter with a stand-in deep_ep whose import raises AssertionError. It checks that the dispatcher package imports, that both use_deepep flags are false, and that creating a DeepEP dispatcher raises ImportError chained to the original error. It fails on main and passes with this change.
• Existing DeepEP unit tests: test_deepep_v2_buffer_lifecycle.py, test_deepep_v2_masked_slab.py, test_deepep_v2_wire_dtype.py, test_w4afp8_deepep_dtype.py, test_w4afp8_deepep_post_reorder.py, plus the new test: 39 passed.
• pre-commit run --files <changed files>: all hooks pass.
• No model code is touched; the dense end-to-end run above is the functional check.

Speed Tests and Profiling

Not applicable. The change only affects module import, not any runtime path.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations. Not needed: no user-facing option changes.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed. Not applicable: import-time change only.
• Follow the SGLang code style guidance.

CI States

Latest PR Test (Base): ❌ Run #37439252608
Latest PR Test (Extra): ❌ Run #37439252212
Latest PR Test (AMD ROCm 10): ❌ Run #37439252555