sglang-runtime-context

Read resolved values from namespace config bags such as get_exec(), get_schedule(), or get_model() rather than get_server_args().field, which is a read-ratchet failure. Bag leaves are plain attributes, safe inside torch.compile-traced code, and reflect post-publish overrides.

Use get_context().override(source, **fields), the only post-publish mutation entry point. It writes bag leaves in place and records provenance; the ServerArgs instance stays pristine with no write-through.

A read ratchet pins direct ServerArgs field reads in business code at zero because the record holds raw operator input, not resolved values. Read the corresponding namespace bag leaf or a runtime_context derived accessor instead.

Declare it in sglang/srt/arg_groups/overrides.py: use MODEL_OVERRIDES for constants, @register_model_override for derived values, or a post-process pass for normalization. Never assign server_args fields from model code; only resolvable=True fields are declarable.

Tests should publish config via get_context().override_server_args(field=...) and install it, rather than injecting SimpleNamespace stand-ins. Production code reads bags, so injected stubs stop working once a reader migrates to bag accessors.

Logic consulting extensible registries, like out-of-tree platforms registering attention backends at model_runner import, must stay at load time in ModelRunner init and write via get_context().override(). __post_init__ runs before any model or platform import.