SGLang adds configurable alt-stream support for Qwen3.5 inference on AMD GPUs, but disables it for shared expert fusion and certain sub-layers to avoid performance regressions, with the net effect being a small throughput improvement when alt-stream is enabled alone.
Enables a controllable alt-stream execution path for Qwen3.5 on AMD GPUs, providing a ~1.8% TPOT uplift when used without shared expert fusion, while automatically disabling the stream in configurations where it causes regressions.
Developers deploying Qwen3.5 on AMD MI355X hardware can now use a new flag to slightly improve inference throughput by better utilizing the GPU's compute streams, but they must carefully manage the new flags to avoid performance regressions in specific model layers. The optimization is conditional: enabling shared expert fusion (a larger performance win) automatically disables the alt-stream, forcing a trade-off. Operators should monitor whether the modest throughput gain justifies the added configuration complexity for their specific workload, and watch for future work that might reconcile the two optimizations.