[Feature] Add --disable-normalize-embedding server arg

[fortunecookiee]

Motivation

Embedding models in SGLang hard-code whether their Pooler L2-normalizes the pooled hidden state, e.g. Pooler(pooling_type=PoolingType.LAST, normalize=True). Serving a checkpoint whose downstream consumer expects un-normalized embeddings (raw pooled vectors, e.g. for a scorer or a retrieval index that does its own normalization) currently requires editing the model file.

This adds a server-level opt-out so the same checkpoint can be served either way without touching model code.

Supersedes #23282, which was auto-closed by the stale bot. This is the same change rebased onto current main and adapted to the arg_groups / runtime-context config layout that landed in the meantime.

Modifications

• arg_groups/fields/model.py: new disable_normalize_embedding: bool = False field on the model namespace, declared next to is_embedding. The CLI flag --disable-normalize-embedding is generated from the annotation like every other boolean arg.
• layers/pooler.py: Pooler.__init__ reads the resolved value via get_model() and, when set, overrides normalize=True to False, so the L2 normalization step is skipped at pool time.

Design notes:

• No per-model change needed. Every model that constructs a Pooler with normalize=True picks up the override automatically.
• The override is deliberately one-way. A model that already sets normalize=False is unaffected, and there is no path to force-enable normalization on a model that did not opt in. This matches the flag name and avoids surprising behavior changes.
• Config is read through the model config bag (get_model().disable_normalize_embedding) rather than get_global_server_args(), matching the current resolved-config reader convention (same as the existing get_model().is_embedding reads).

Accuracy Tests

Default path is unchanged: with the flag unset, normalize keeps the value the model passed, so pooled outputs are bit-identical to main.

With --disable-normalize-embedding, the returned embedding is the pre-normalization pooled vector; normalizing it client-side reproduces the default output.

Speed Tests and Profiling

Not applicable — the flag is read once in Pooler.__init__ (construction time), not on the forward path. When set, it removes work (one L2 normalization per pooled batch).

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.

---

CI States

Latest PR Test (Base): 🚫 Run #34899602609
Latest PR Test (Extra): ❌ Run #34899602318
Latest PR Test (AMD ROCm 10): ❌ Run #34899602455

[fortunecookiee]

Updated with unit tests (test/registered/unit/layers/test_pooler_score_and_pool.py::TestDisableNormalizeEmbedding) covering: the flag disabling normalization for a normalize=True pooler, normalization staying on by default, and the override being one-way. They publish a context via override_server_args, which also pins the fact that Pooler.__init__ now reads the model config bag.

Also clarified in the help text that models whose pooler does not go through Pooler — the cross-encoder, BERT-pooler and vision-pooler paths — are unaffected by the flag.

For context on alternatives considered: --json-model-override-args cannot express this today, because ~35 of the 41 Pooler(...) call sites pass normalize=True as a literal rather than reading anything off the HF config.

[fortunecookiee]

/tag-run-ci-label

[fortunecookiee]

/rerun-failed-ci

[fortunecookiee]

Rebased onto current main (picks up the test_dsv4_nonpaged_indexer.py fix from #39101, which was the only non-infra failure in the previous run).