fix: force to use fgpu-rdc to compile

[zhenhuang12]

From JIRA AIMA-212, it shows that DeepEP intranode-combine timeout on rocm7.2 docker image, but works well on rocm7.1.

I found a workaround to solve the timeout which forcely add -fgpu-rdc to compile and disable -amdgpu-function-calls=false

[Copilot]

Pull request overview

This PR addresses a timeout issue in DeepEP intranode-combine on ROCm 7.2 by forcing the use of the -fgpu-rdc (Relocatable Device Code) compiler flag across all compilation stages.

Changes:

• Added -fgpu-rdc flag to CXX compilation flags
• Added -fgpu-rdc and --hip-link flags to linker arguments
• Added -fgpu-rdc flag to HIP compilation flags

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The extra_link_args list is formatted on a single line, making it harder to read and maintain compared to the original multi-line format. Consider restoring the multi-line format for consistency and readability.

Suggested change

	extra_link_args = ["-Wl,-rpath,/opt/rocm/lib", f"-L/usr/lib/{arch}-linux-gnu", "-fgpu-rdc", "--hip-link"]
	extra_link_args = [
	"-Wl,-rpath,/opt/rocm/lib",
	f"-L/usr/lib/{arch}-linux-gnu",
	"-fgpu-rdc",
	"--hip-link",
	]

[Copilot]

The -fgpu-rdc flag is duplicated across cxx_flags and hip_flags. Consider extracting this to a shared constant or variable to ensure consistency and make future updates easier to manage.

[Copilot]

The -fgpu-rdc flag is duplicated across cxx_flags and hip_flags. Consider extracting this to a shared constant or variable to ensure consistency and make future updates easier to manage.

[xiaobochen-amd]

I remember this compile option caused issues before.