Add registration API for external linear attention backend

[charlotte12l]

Motivation

SGLang currently hardcodes hybrid model support (GDN, KDA, Mamba2, Lightning) across 5 core files via isinstance checks and architecture name lists. Adding a new linear attention hybrid model requires modifying all 5 files:

• model_runner.py — mambaish_config property
• attention_registry.py — backend dispatch in attn_backend_wrapper
• scheduler.py — is_hybrid_ssm flag for MambaRadixCache
• server_args.py — _handle_model_specific_adjustments
• triton_backend.py — v_head_dim check

This makes it difficult for external or experimental linear attention models (e.g., custom KDA variants) to integrate with SGLang without forking the core. A registration API lets external models self-register and plug into all 5 integration points with zero modifications to SGLang source.

Modifications

New file: python/sglang/srt/configs/linear_attn_model_registry.py

• LinearAttnModelSpec dataclass — captures config class, backend class name, arch names, and cache behavior flags
• register_linear_attn_model() — appends a spec to the global registry (called at import time)
• get_linear_attn_config() — isinstance-based lookup, returns (spec, resolved_config) or None
• get_linear_attn_spec_by_arch() — arch name lookup for server_args dispatch
• import_backend_class() — lazy import of the backend class from a fully-qualified dotted name

Integration (all purely additive, existing behavior unchanged):

• model_runner.py — new linear_attn_model_spec property; mambaish_config falls back to registry after existing hardcoded checks
• attention_registry.py — attn_backend_wrapper falls back to registry when no hardcoded backend matches
• scheduler.py — is_hybrid_ssm checks registry for uses_mamba_radix_cache
• server_args.py — _handle_model_specific_adjustments calls _handle_mamba_radix_cache for registered models
• triton_backend.py — v_head_dim check includes linear_attn_model_spec

Usage:

from sglang.srt.configs.linear_attn_model_registry import (
    register_linear_attn_model, LinearAttnModelSpec,
)

register_linear_attn_model(LinearAttnModelSpec(
    config_class=MyLinearAttnConfig,
    backend_class_name="sglang.srt.layers.attention.linear.kda_backend.KDAAttnBackend",
    arch_names=["MyModelForCausalLM"],
    uses_mamba_radix_cache=True,
))

Accuracy Tests

Unit tests covering all registry functions (11 tests, all passing):

test_register_and_lookup_by_config ... ok
test_lookup_no_match ... ok
test_lookup_empty_registry ... ok
test_unwrap_text_config ... ok
test_unwrap_text_config_no_match ... ok
test_lookup_by_arch ... ok
test_lookup_by_arch_empty_registry ... ok
test_multiple_registrations ... ok
test_first_match_wins ... ok
test_import_backend_class ... ok
test_spec_defaults ... ok

No accuracy regression is possible — the registry is purely additive. All existing hardcoded checks execute first; the registry is only consulted as a fallback when no existing model matches. When the registry is empty (default), all code paths are identical to before this change.

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.

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.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[merrymercy]

when is this used?

[charlotte12l]

This check is inspired from https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/layers/attention/attention_registry.py#L194-L196

[merrymercy]

/tag-and-rerun-ci

[charlotte12l]

@merrymercy I've addressed the comments, those timeout errors are not related...

[merrymercy]

/tag-and-rerun-ci