[Fix] cuda_graph: Disable full prefill CUDA graph for hybrid linear attention models (#48593) - #48680

Motivation

Fixes #48593

When attempting to capture full prefill CUDA graphs (--cuda-graph-config '{"prefill": {"backend": "full"}}') for hybrid linear attention models (e.g., Mamba2, GDN, KDA, NemotronH, GraniteMoe), the engine crashes during capture initialization with:

IndexError: list index out of range
  in HybridLinearAttnBackend._replay_metadata:
    indices = self.state_indices_list[bs - 1]

Root Cause

1. In cuda_graph_setup.py, PrefillCudaGraphRunner runs capture() before DecodeCudaGraphRunner is constructed.
2. self.state_indices_list is only allocated and populated inside init_cuda_graph_state(), which is called exclusively by DecodeCudaGraphRunner. PrefillCudaGraphRunner never calls init_cuda_graph_state().
3. In python/sglang/srt/arg_groups/cuda_graph_hook.py, disable_full_prefill_cudagraph_if_incompatible() had an empty rule list (rules = []), so hybrid linear attention models were never filtered out during configuration compatibility resolution.
4. If an operator explicitly passed --cuda-graph-config '{"prefill": {"backend": "full"}}', validate_cuda_graph_config() did not check architecture compatibility.

Modifications

1. python/sglang/srt/arg_groups/cuda_graph_hook.py:
   • Populated disable_full_prefill_cudagraph_if_incompatible() to check mambaish_config(model_config_of(server_args)). When detected during compatibility resolution, it logs a warning and disables the prefill CUDA graph.
   • Added validation in validate_cuda_graph_config(): if an operator explicitly locks full prefill CUDA graph on a hybrid linear attention model, it raises a descriptive ValueError before engine capture begins.
2. python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py:
   • Added a defensive guard in _replay_metadata() to verify state_indices_list bounds, raising an explicit RuntimeError if an uninitialized replay is ever invoked.
3. test/registered/unit/server_args/test_server_args.py:
   • Added TestHybridLinearPrefillCudaGraphPolicy covering:
     • Automatic disabling of full prefill CUDA graph for hybrid linear architectures.
     • Preservation of full prefill CUDA graph for standard non-hybrid architectures (e.g. Llama).
     • Validation rejection when full prefill CUDA graph is explicitly locked on hybrid models.

Accuracy Tests

N/A: Configuration validation and runtime defensive bounds checks; no alterations to model tensor outputs or numeric kernels.

Speed Tests and Profiling

N/A: Prevents initialization crash during full prefill CUDA graph capture.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Follow the SGLang code style guidance.

CI States

Latest PR Test (Base): ❌ Run #42066902800
Latest PR Test (Extra): ❌ Run #42066902358
Latest PR Test (AMD ROCm 10): ❌ Run #42066902499

Find the registered tests this PR affects and run them with /rerun-test, following https://github.com/sgl-project/sglang/blob/main/.agents/skills/rerun-test/SKILL.md