vllm - 💡(How to fix) Fix [Feature]: NIXL P/D Disaggregation: GDN support (Qwen3.5) [1 participants]

🚀 The feature, motivation and pitch

Add GDN (Gated Delta Net) model support for NIXL-based Prefill/Decode disaggregation. GDN models (e.g. Qwen3.5) have a different SSM layout than Mamba2 — conv state is [K, K, V] instead of [x, B, C], and temporal state shape is (num_v_heads, v_dim, k_dim) instead of (num_heads, head_dim).

PR: #41869

Verified Configs

Model: Qwen/Qwen3.5-0.8B | GSM8K 5-shot (baseline ~0.323)

• Homogeneous TP (1p1d, 2p2d, 4p4d): all pass across all backends (FLASHINFER, FLASH_ATTN, TRITON_ATTN) and block sizes (16, 32, 64, 128).
• Heterogeneous TP (1p2d, 2p1d): all pass across all backends and block sizes.
• Heterogeneous TP with TP=4 (FA replication) (1p4d, 4p1d, 2p4d, 4p2d): pass on FLASHINFER/FLASH_ATTN with block sizes 32, 64, 128.

Known Issues / TODOs

• Heterogeneous TP not supported when P and D have mismatched kernel block sizes — e.g. TP=4 configs with block_size=16 on FLASHINFER/FLASH_ATTN, or all TP=4 configs on TRITON_ATTN.
• Async scheduling not supported (#37285)
• Prefix caching not supported

Alternatives

N/A - extends the existing NIXL Mamba2 transfer path.

Additional context

Tracking issue: #40017

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.