[Bug] DeepSeek-V4 compressed attention backend: no SM120 fallback for Lightning Indexer

[rs-ipps]

Checklist

• I searched related issues but found no solution.
• The bug persists in the latest version.
• Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback.
• If this is not a bug report but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
• Please use English. Otherwise, it will be closed.

Describe the bug

Summary

SGLang's compressed attention backend for DeepSeek-V4 calls get_paged_mqa_logits_metadata directly from DeepGEMM without a fallback path. On SM120 (RTX Pro 6000 Blackwell, RTX 5090), this kernel has no implementation in DeepGEMM, so V4 cannot complete a forward pass.

In contrast, SGLang's mHC implementation (layers/mhc.py) routes through TileLang and works correctly on SM120. If a similar TileLang or Triton path were available for the paged MQA logits metadata, V4 would be fully functional on SM120.

Behavior

• Server starts successfully, loads model (88.05 GiB per GPU), accepts HTTP requests
• On the first forward pass: RuntimeError: Assertion error (csrc/apis/attention.hpp:211): Unsupported architecture

Full stack trace:

sglang/srt/layers/attention/deepseek_v4_backend_radix.py:491 in init_forward_metadata_prefill
→ sglang/srt/layers/attention/deepseek_v4_backend_radix.py:410 in init_forward_metadata_indexer
→ sglang/srt/layers/attention/compressed/metadata.py:142 in __post_init__
→ self.deep_gemm_metadata = get_paged_mqa_logits_metadata(...)
RuntimeError: Assertion error (csrc/apis/attention.hpp:211): Unsupported architecture

Request

Would the team consider adding an SM120 fallback path (TileLang or Triton) for get_paged_mqa_logits_metadata / paged_mqa_logits in the compressed attention backend, similar to how mHC is handled via TileLang? This would unblock V4 on all consumer/workstation Blackwell GPUs.

Upstream DeepGEMM issue: [https://github.com/deepseek-ai/DeepGEMM/issues/317]

Reproduction

SGLANG_DISABLE_DEEP_GEMM=1 \
python -m sglang.launch_server \
  --model-path /models/DeepSeek-V4-Flash \
  --tp 2 --trust-remote-code \
  --fp8-gemm-backend triton --moe-runner-backend triton \
  --kv-cache-dtype fp8_e4m3 \
  --mem-fraction-static 0.96 --context-length 65536 \
  --disable-cuda-graph --max-running-requests 8 \
  --reasoning-parser deepseek-v4 --tool-call-parser deepseekv4

Setting --nsa-prefill-backend tilelang and --nsa-decode-backend tilelang has no effect because V4 auto-selects the compressed attention backend, not nsa.

Environment

• GPU: 2× NVIDIA RTX Pro 6000 Blackwell (96 GB, SM120)
• Image: lmsysorg/sglang:deepseek-v4-blackwell (container CUDA 12.9.1)
• Model: deepseek-ai/DeepSeek-V4-Flash

[wuwenthink]

i have the same question