Deployment blocked: FlashInfer "Unsupported architecture" on DGX Spark (SM121)
#4
by JayceHuang - opened
Feedback on DGX Spark deployment:
Just sharing our debugging results here. The deployment is currently blocked at the underlying FlashInfer kernel in the official dedicated image: GLM-5.3's 0-RoPE sparse MLA returns an Unsupported architecture error on DGX Spark (SM121).
Model weights and multi-node connections are both fine. The issue is strictly related to kernel/architecture support rather than downloads, NVFP4 quantization, or memory limits. Hopefully, this helps the team look into FlashInfer support for SM121/SM120 architectures!
Hey! So I updated the model card to give pointers for a dual DGX spark cluster using sglang. Also fixed some bugs in config for it.