From d8791fb5e33e30e8c469010b5b5569aa292cfac5 Mon Sep 17 00:00:00 2001 From: Cheng Wan <54331508+ch-wan@users.noreply.github.com> Date: Thu, 9 Jul 2026 02:11:36 -0700 Subject: [PATCH] [docs] Add the sglang-runtime-context skill (#30494) --- .../skills/sglang-runtime-context/SKILL.md | 195 ++++++++++++++++++ 1 file changed, 195 insertions(+) create mode 100644 .claude/skills/sglang-runtime-context/SKILL.md diff --git a/.claude/skills/sglang-runtime-context/SKILL.md b/.claude/skills/sglang-runtime-context/SKILL.md new file mode 100644 index 000000000..cb9a66a17 --- /dev/null +++ b/.claude/skills/sglang-runtime-context/SKILL.md @@ -0,0 +1,195 @@ +--- +name: sglang-runtime-context +description: How SGLang's runtime configuration and process-global state are organized (RuntimeContext tiers, resolve-at-end ServerArgs, override entry point, resource/stream/buffer leases, per-forward flags), the CI guardrails that enforce the design, and the idioms for developing and testing against it. Load this before touching server_args, model overrides, module-level state, or per-forward state in sglang. +--- + +# SGLang runtime-context architecture + +One container owns process-static runtime state: `sglang.srt.runtime_context.RuntimeContext` +(a process singleton reached via `get_context()`). Everything below is a tier on it. + +| Tier | Accessor | Holds | Lifecycle | +|------|----------|-------|-----------| +| config | `get_server_args()` | the process-wide **resolved** `ServerArgs` | resolved once in `__post_init__`; audited mutations only | +| runtime flags | `get_flags()` | state that is *not* a pure function of config: `capture` (cuda-graph lifecycle), `moe` (ACTIVE backends, swappable), `dp` (DP-attention runtime flags) | materialized at subsystem init; groups offer `override()` for tests | +| resources | `get_resources()`, `get_stream(name)`, `get_buffer(name, factory)` | process-level handles: graph pools, EPLB state, EP dispatcher state, named side streams, workspace buffers | lazy; cleared by `reset_context()` | +| per-forward | `get_forward()` | forward-scoped flags (multi-stream switch, MoE output buffer, attn-TP inputs, extend-in-batch) | contextvar-backed; `scoped(**kw)` restores on exit; new threads see defaults | +| parallel | `get_parallel()` | read-through wrapper over topology getters (tp/pp/moe/attn sizes, ranks, groups) | stateless; `override()` for tests | + +`reset_context()` (unit-test teardown) drops the published server_args and installs fresh +flags/resources/forward tiers. + +## Config: the resolve-at-end contract + +**After `ServerArgs.__post_init__` returns, the fields ARE the resolved configuration.** +Model overrides and normalization passes *declare* values during resolution (into a +provenance stash); `materialize_declarations()` applies them once at the very end of +`__post_init__` (gate order, last writer wins). Consequences: + +- **Reading config**: read fields directly, in any process, at any time after construction. + For global access use `runtime_context.get_server_args()` (the blessed accessor — + `get_global_server_args()` is a legacy shim over the same slot and its call-site count is + ratcheted; do not add new ones). +- **Mutating config after resolution**: the ONLY entry point is + `ServerArgs.override(source, **fields)`. It records provenance (`_runtime_mutations`), + keeps whitelisted resolvable fields consistent with the declaration stash (so a republish + resolves the same values), and bypasses the strict guard. Bare `server_args.x = ...` + after resolution **raises** under `SGLANG_STRICT_CONFIG_MUTATION=1`, which the test + harness (`sglang.test.test_utils`) turns on by default. +- **Mid-resolution code** (inside `server_args.py` / `arg_groups/` only): fields are + read-only during resolution; handlers and hooks read the in-flight state through + `resolved_view(server_args)` / `self._resolved()`. This is pipeline-internal — never use + `resolved_view` outside the pipeline. (One sanctioned exception: helpers that the pipeline + itself invokes mid-resolution, e.g. `adaptive_spec_params`.) + +### Adding a model-specific config adjustment + +Never assign `server_args` fields from model code. Declare instead +(`sglang/srt/arg_groups/overrides.py`): + +- Constant per-arch values → `MODEL_OVERRIDES["MyArchForCausalLM"] = {...}`. +- Derived values → `@register_model_override("MyArchForCausalLM")` returning a dict; the + callable receives *pristine* `server_args` + `hf_config` and must not write. +- Normalization that must see earlier declarations → a post-process pass invoked via + `run_post_process_pass` at its slot (reads a view, returns a declaration dict). +- Values only knowable at weight-load time → `declare_load_time_override(source, {...})` + (validates the whitelist, routes through `override()`). + +Declarable fields form a whitelist: `Arg(..., resolvable=True)` in the `ServerArgs` +dataclass. A declaration against a non-whitelisted field fails at its slot. + +### Load-time vs resolution-time (critical) + +`__post_init__` runs in the launcher process before any model/platform import. Logic that +consults an **extensible registry** (e.g. out-of-tree platforms registering attention +backends in `init_backend()`, which runs at `model_runner` import) must stay at load time +(ModelRunner init), writing through `override()`. Before moving any load-time logic into +resolution, verify everything it reads is already complete at construction time. + +## Runtime flags (`get_flags()`) + +For state that init-time code *derives* and runtime code reads — parsed enums, platform +probes, swappable ACTIVE values. Not for config mirrors (those died with resolve-at-end: +read the field). + +- Groups are typed dataclasses on `Flags` (`capture` / `moe` / `dp`): typo-safe writes, + transactional test-only `override(**kw)` context manager. +- `flags.moe` is materialized by `initialize_moe_config(server_args)` at scheduler init; + accessors (`get_moe_a2a_backend` etc.) are thin shims with lazy defaults. The speculative + contexts (`speculative_moe_backend_context`) swap the ACTIVE leaves around draft forwards. +- `flags.dp` is materialized by `initialize_dp_attention`; `is_dp_attention_enabled()` is a + shim over `flags.dp.enabled`. +- Adding a leaf: declare the dataclass field with a default equal to the pre-init behavior, + materialize it at the owning subsystem's init, keep any public accessor as a shim. + +## Resources (`get_resources()`) + +Named slots + two keyed-lazy registries: + +- `get_stream(name)` — get-or-create a named CUDA side stream; `set_stream(name, stream)` + installs explicitly. **Name leases by subsystem ROLE**: all model alternate streams share + `"alt"`; the offloader's copy stream is `"offload"`; DP-TBO comm is `"dp_tbo_comm"`; LoRA + side stream is `"lora_side"`. Two call sites may share a name only if their work belongs + on one stream — sharing across roles serializes intended overlap. +- `get_buffer(name, factory)` — get-or-create a named persistent buffer. Grow-only or + per-device semantics manage their `resources.buffers` entries directly (see tokenspeed / + SM120 split / Marlin workspace). Buffer names are per-backend today; do not silently + share. +- Singletons with manager semantics (EP dispatcher buffers, EPLB recorder/metadata, graph + memory pool) keep their owning accessors/classes as facades; only the *state* lives in a + resources entry. Preserve exact semantics in the shim: lazy defaults (the EPLB recorder + defaults to a Noop instance, not None), publish-once asserts, event-reuse contracts. +- Stream/buffer creation is a driver call — it must happen outside cuda-graph capture; + keep lease points at init/warmup time. + +## Per-forward flags (`get_forward()`) + +Contextvar-backed; a new thread sees the defaults; `scoped(**kw)` is the regular write path +(transactional, restores on exit and on exception); `set(name, value)` exists for legacy +sticky setters (`is_extend_in_batch` is intentionally sticky within a thread). Use this +tier for anything set-per-forward and read-within-forward. Before adding cross-thread +state here, prove the readers' thread affinity: contextvars do NOT propagate to already- +running or newly spawned threads. Note TBO ("two-batch overlap") interleaves ubatches on +ONE thread — do not design for TBO threads that don't exist. + +## Testing idioms + +- **Force a code path by overriding causes, not effects**: compose + `get_context().override_server_args(**fields)` (config tier: publishes a fresh + dummy-boundary `ServerArgs` carrying the overrides — `with`-scoped, or + `install()`/`restore()` for fixture-lifetime use) + `get_parallel().override(...)` + + `get_flags()..override(...)` + `get_forward().scoped(...)` + + `get_resources().override(...)`. All are scoped and transactional. Tests control + execution through the context — do not hand-build and publish config objects. + Note `override_server_args` is itself transitional (to be deprecated): it exists + while production still branches on raw `server_args` fields at runtime; prefer the + finer-grained tier overrides wherever they already cover the path you need. +- **Never monkeypatch import bindings** (`module.get_x = lambda: ...`): production code may + read a different accessor over the same slot and your patch silently stops intercepting. + Publish/inject for real: `get_context().override_server_args(...)` for config; + `monkeypatch.setattr(get_resources(), "slot", fake)` for resources. +- Fixtures standing in for `ServerArgs` need an `override` method if the code under test + mutates config (`_fake_server_args`-style: SimpleNamespace + write-through override). + MagicMock swallows `override()` calls silently — prefer SimpleNamespace so misses raise. +- `reset_context()` in teardown; `_IsolatedServerArgs`-style save/restore when a test + publishes. +- `ServerArgs(model_path="dummy")` early-returns `__post_init__` (no materialization, no + strict guard) — fine for lightweight fixtures. + +## Guardrails (these fail CI; what to do when they fire) + +1. **Strict mutation guard** (`SGLANG_STRICT_CONFIG_MUTATION=1`, default-on in tests): + bare `server_args.x = ...` after resolution raises. Fix: route through + `override(source, ...)`, or move genuinely config-decidable logic into resolution. +2. **Mutation ratchet** (`test_server_args_mutation_ratchet.py`, exact pin 0 over the whole + package minus the pipeline / multimodal_gen / the mock-fixture factory): textual scan + for assignment forms. Never raise the baseline. +3. **Legacy-accessor ratchet** (`test_legacy_global_ratchet.py`): `get_global_server_args` + call sites must not grow — new code uses `runtime_context.get_server_args()`. +4. **Module-state ratchet** (`test_module_state_ratchet.py`): `global` statements in the + flag-owning layers are pinned by name. A new module-level runtime global belongs on a + flags group / resources slot instead; migrating a pinned survivor must shrink the pin. + +## Hard-won pitfalls (check these before/while refactoring) + +- **Moving code drops first-line guards**: early returns (`if self.is_draft_worker: return`) + are the easiest thing to lose when relocating a method body. Draft workers share the + target's `server_args` object — a draft-side write poisons the target. +- **Registry-completeness timing**: a gate that consults an extensible list is only correct + after the registrars ran (platform `init_backend()` at module import). See "load-time vs + resolution-time". +- **Late function-scope imports shadow module names** for the WHOLE function + (UnboundLocalError at earlier lines). Audit moves with AST, not grep. +- **Lease names are per-role**, not per-API-shape (the offloader-vs-"alt" lesson). +- **Storage matrix for state read inside torch.compile-traced model code** + (piecewise cuda graph compiles the whole model forward): contextvars are + untraceable (hard error); dict-slot values are guarded per value — for a + per-forward int that is one recompile per distinct size, straight into the + recompile limit; **class/instance attributes are the only compile-friendly + form** (attribute-source ints get automatic-dynamic after the first size + change). Bools (≤2 values) are tolerable in any form — see + `ForwardFlags._GRAPH_VISIBLE`. Before moving such state, prove its readers + sit outside compile coverage; a piecewise-prefill boot of a small model is + the fast check (recompile storms show as `torch._dynamo hit + config.recompile_limit` during the compile pass). +- **Engine-booting e2e tests are the only coverage for launcher-path code**; a child crash + kills the process tree and pytest dies silently — run with `PYTHONUNBUFFERED=1` and read + child logs. +- CI arms `SGLANG_ENABLE_ASYNC_ASSERT=1` (device-side `torch._assert_async` probes, e.g. + KV-cache OOB): a fired device assert kills the tree with no Python traceback, and the + same bug is *silent corruption* locally with the flag off. Arm it when reproducing CI + crashes. +- CI startup logs print the full `server_args=ServerArgs(...)`; diffing that dump between + runs is the fastest config-divergence check. + +## Where to read the code + +Key source files: `python/sglang/srt/runtime_context.py` (the container and every tier), +`python/sglang/srt/arg_groups/overrides.py` (override registry, passes, +`declare_load_time_override`, `resolved_view`), `python/sglang/srt/server_args.py` +(`override`, `__setattr__`, `materialize_declarations` call, +`_handle_model_capability_adjustments`), and the guardrail tests under +`test/registered/unit/` (`test_server_args_mutation_ratchet.py`, +`test_legacy_global_ratchet.py`, `test_module_state_ratchet.py`, +`test_runtime_context.py` — the last one doubles as executable documentation of +every tier's semantics).