fix: fallback to triton for attention-sink models (flashinfer unsupported)

[alphabetc1]

Motivation

The FlashInfer attention backend does not accept a sinks kwarg, but hybrid-SWA models like GptOssForCausalLM and MiMoV2FlashForCausalLM (when add_{swa,full}_attention_sink_bias=True) pass per-head sink scalars to the attention call. Two auto-selection paths used to fall back to flashinfer without considering this, which could leave sink models on an unsupported backend.

Modifications

• Add ModelConfig.has_attention_sinks, derived in _derive_hybrid_model via a new _detect_attention_sinks() helper:
  • GptOssForCausalLM always has sinks
  • MiMoV2FlashForCausalLM / MiMoV2MTP have sinks only when add_swa_attention_sink_bias or add_full_attention_sink_bias is set on hf_text_config
• In ServerArgs._get_default_attn_backend, the non-MLA fallback branch now picks triton instead of flashinfer when model_config.has_attention_sinks is true
• In ServerArgs._resolve_io_decode_attention_compatibility, the hicache kernel-io + FA3 avoidance path applies the same check before setting decode_attention_backend

Scope of Behavior Change

Only the implicit auto-selection paths are affected:

Scenario	Before	After
Sink model, no --attention-backend, non-Hopper / non-Blackwell GPU	flashinfer (unsupported)	triton
Sink model, hicache io_backend=kernel with FA3 effective decode	flashinfer (unsupported)	triton
Non-sink models, all configurations	unchanged	unchanged
Explicit --attention-backend ...	unchanged	unchanged

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]

Code Review

This pull request introduces a mechanism to detect attention sinks in models and ensures the attention backend defaults to Triton when sinks are present, as FlashInfer currently lacks support for them. The review feedback highlights that the detection logic may need to include additional hybrid SWA architectures and suggests implementing validation to prevent users from manually selecting FlashInfer when attention sinks are active, which would lead to incorrect results.

[hnyls2002]

/tag-and-rerun-ci