[Copilot]

Qwen3-VL fails to launch on XPU with BackendCompilerFailed because CUDA-specific fusion passes are imported conditionally but instantiated unconditionally in PostGradPassManager.configure().

Changes

• Guard CUDA-alike passes (is_cuda_alike()): SequenceParallelismPass, RMSNormQuantFusionPass, ActivationQuantFusionPass, AttnFusionPass, SplitCoalescingPass, QKNormRoPEFusionPass
• Guard CUDA-only passes (is_cuda()): AsyncTPPass, AllReduceFusionPass

Before:

if self.pass_config.fuse_norm_quant:
    self.passes += [RMSNormQuantFusionPass(config)]  # NameError on XPU

After:

if current_platform.is_cuda_alike():
    if self.pass_config.fuse_norm_quant:
        self.passes += [RMSNormQuantFusionPass(config)]  # Only when imported

Platform behavior:

• CUDA: All passes available
• ROCm: CUDA-alike passes only
• XPU/other: Platform-agnostic passes only

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[hongbolv]

@copilot

1. rebase the PR to latest version
2. What's the behavior for XPU and other?