Qwen 3.5 with Megatron-inference

Meet error when running Qwen 3.5 RL with Megatron-inference and sglang. More details please check the attached log.

model: Qwen3.5 35B A3B VLM
params: use cuda graph with Megatron-inference
Question: which model does Megatron-inference validate?

[qwen3.5_minf_error.log](https://github.com/user-attachments/files/29499710/qwen3.5_minf_error.log)

This problem seems to be the issue of cuda graph.
But when I disable the cuda graph, it will hang with Qwen3.5 + Megatron inference FLA kernel.

Start by reading the attached qwen3.5_minf_error.log and reproducing the Qwen3.5 35B A3B VLM RL run with Megatron-inference and sglang. Compare behavior with CUDA graphs enabled and disabled, including the reported FLA-kernel hang; done means identifying the failing path and confirming which model Megatron-inference validates.

Written by the indexing model from the issue text.