Check the topology identities where the layout is written, and build at the published widths

[ch-wan]

Motivation

get_parallel().override(...) validated key names and nothing else. A
self-contradicting topology -- a rank at or past its width, a tp_size that
does not factor into the attention widths, a group whose width is not the one
configured -- went in quietly and surfaced much later as a hang or a wrong
answer in a collective, with nothing pointing back at the write.

At the other end, initialize_model_parallel took seven widths as arguments,
so every caller translated the published configuration into them again: the
scheduler from a per-runner record, the weight-cache daemon from its own
fields, the media encoder from neither. Three translations of one configuration
are three places for it to come out different, and the encoder's did -- it
built a layout the context never described.

Modifications

One set of identities, checked wherever the layout is written. Every rank
inside its width; tp_size factoring into the attention triple and into the
MoE triple; the attention rank layout the attention ranks are derived from; and
each built group's width equal to the configured one. They hold
unconditionally, on the four paths that write the namespace and at the group
build. The caller owns the arithmetic: stating one leaf without the quotients
that follow from it describes no real layout, so it is refused rather than
carried. Names that cannot be read are skipped -- a process that has published
nothing can still stamp a rank.

Two callers were stating a width the rest of the namespace disagreed with. The
shared-expert scope narrows moe_ep_size to one without sharding anything, so
it now says there is no expert-parallel group rather than leaving the wider one
installed. Publish stamped the placement in two calls, and the moment between
them described a half-placed process; it stamps once.

The group build reads the widths it builds at. The seven width parameters
are gone; the function reads them from the context. What is left in the
signature is not topology: the backend is decided by the device, two flags
belong to other namespaces, and the join parameters describe this particular
call. The encoder states the layout it has always built -- tp_size ranks
wide, no pipeline, no expert or MoE-DP dimension, no decode context parallelism
-- so what it answers and what it builds are the same thing again.

An elastic scale-up and the launch topology stop sharing two names. A
scale-up admits ranks into a WORLD that was pre-allocated to --max-ep-size;
it does not rebuild the process groups, which keep the width they were
constructed with. Writing the expanded replica count over attn_dp_size and
attn_dp_rank therefore left the launch names answering for groups of a
different size, and the identities say so: the rank lands outside its width and
tp_size stops factoring into the attention triple. elastic_dp_size and
elastic_dp_rank name the expanded replica set, each answering with its launch
counterpart until a scale-up moves it, and the readers that need the expanded
one -- the WORLD gather width, DP padding, KV routing, every place a gathered
list is indexed by replica, and KV-event sharding -- ask for it by name. That
includes the gather and scatter slices themselves: the list a scale-up gathers
spans WORLD, so indexing it by this process's place among the launch replicas
selects another rank's rows. The pair is range-checked like any other; it stays out of the identities
that describe the groups, because it does not describe them.

A draft on one context shard has no context-parallel communicator, and the
sampler asks for one only when there is more than one shard to reconcile --
otherwise the scope's answer, that there is no such group, is dereferenced.

The draft scope states a group for each width it narrows. It already said its
attention-TP width is the group it installs; it now hands over that group, and
says there is no attention-CP group, the way it already said there is no
expert-parallel one. The shared-expert scope states moe_dp_size for the same
reason -- narrowing two of the three MoE factors and leaving the third is a
triple that does not multiply out.

With the build checking what it made against what was configured,
get_tensor_model_parallel_world_size and its attention sibling answer the same
question as the configured widths rather than a second one, so their eight
readers move to the context.

Accuracy Tests

Not run for this revision. The build now reads the widths the previous code
passed it, and the encoder states the layout it already built, so no
configuration changes what is constructed.

Speed Tests and Profiling

Not applicable. The identity check is integer comparison on names already in
hand; it runs where a topology is written, not in a forward.

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.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Each identity is injected in the direction that breaks it and in the direction
that keeps it, because a guard that only ever fires is as uninformative as one
that never does. Twenty-one tests that built a topology by passing widths
publish it first, and ten benchmark and example entries do the same. The
scale-up test now publishes a configuration before it asserts: without one
every term was unreadable and the check it claimed to make was skipped.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): ❌ Run #35644382491
Latest PR Test (Extra): ❌ Run #35644381975
Latest PR Test (AMD ROCm 10): ❌ Run #35644382214