[Bugfix] Make xgrammar an import-time optional dependency - #56565

Purpose

Fixes #56559. On platforms without an xgrammar wheel (e.g. s390x), from vllm import LLM and vllm serve crash at import time even for workloads that never use structured output.

On current main the first failure is not backend_xgrammar.py (already lazy-loaded) but vllm/parser/harmony.py, pulled in via vllm/v1/structured_output/__init__.py -> vllm.parser; vllm/tool_parsers/structural_tag_registry.py has the same unconditional imports and the OpenAI server imports both. Once those are fixed, the next failure is the dataclass field annotations in backend_xgrammar.py (xgr.GrammarMatcher), which evaluate the LazyLoader at class-definition time.

Changes:

• harmony.py, structural_tag_registry.py: wrap the xgrammar imports in try/except ImportError and bind the names to PlaceholderModule("xgrammar"), so structural tag builders raise a clear ImportError on first use instead of breaking import. Module-level xgrammar objects (_JSON_CONTENT, _ANY_CONTENT, the runtime-evaluated StructuralTagBuilder alias) are made lazy so nothing touches the placeholders at import time.
• backend_xgrammar.py: from __future__ import annotations to defer the field annotations (same line as [Bugfix] Defer xgrammar annotations to fix startup crash when xgrammar is unavailable #56561, which is necessary but not sufficient on its own).
• tests/standalone_tests/lazy_imports.py: add xgrammar to the modules that must not be imported by import vllm (already run in CI).
• tests/tool_parsers/test_structural_tag_registry.py: subprocess regression test that blocks xgrammar, imports LLM, and asserts a builder raises ImportError mentioning xgrammar.

Why this is not a duplicate

Searched open PRs for 56559, "xgrammar import", "xgrammar s390x". #56561 only adds the __future__ line to backend_xgrammar.py; with xgrammar blocked, from vllm import LLM on main still fails earlier in harmony.py, so that PR alone does not fix the issue. This PR includes that line and fixes the remaining import chain; see my comment on the issue.

Test Plan

.venv/bin/python tests/standalone_tests/lazy_imports.py
.venv/bin/python -m pytest tests/parser/test_harmony.py tests/tool_parsers/test_structural_tag_registry.py
.venv/bin/python -m pytest tests/v1/structured_output/test_backend_xgrammar_stop_tokens.py tests/v1/structured_output/test_utils.py tests/v1/structured_output/test_validation.py
pre-commit run --files # incl. mypy-3.10 / mypy-3.12
Manual: with sys.modules["xgrammar"] = None, from vllm import LLM, import vllm.entrypoints.openai.api_server and import vllm.entrypoints.cli.main all succeed on this branch and fail on main.

Test Result

• lazy_imports.py: passes (fails on main after adding xgrammar to the list)
• harmony + structural_tag_registry: 147 passed; new test: 1 passed
• structured_output: 25 passed
• pre-commit (ruff, typos, mypy 3.10/3.12, SPDX, lazy-import checks): all passed

Model output

No change to model execution. Sanity check on an RTX 5090 with Qwen/Qwen2.5-1.5B-Instruct and structured_outputs_config={"backend": "xgrammar"}: JSON-schema constrained generation returns valid JSON, and a hermes structural tag from get_model_structural_tag yields <tool_call>{"name": "get_weather", "arguments": {"city": "Tokyo"}}</tool_call>.