[yctseng0211]

Motivation

#34484 blocked the (bf16_as_f32 * scale) -> fp16 reassociation by naming v_cvt_pk_f16_f32 in inline asm, but that instruction is not part of the CDNA ISA, so every gfx942 build now fails to assemble (error: instruction not supported on this GPU) and takes the AMD CI down with it.
What stopped the reassociation was making the scaled values opaque to the optimizer rather than the choice of instruction, so this swaps the named instruction for an empty asm barrier and leaves the narrowing to __float22half2_rn — portable to every target, and the same round-to-nearest conversion the quantized path in this file already uses.
Verified on gfx942, gfx950 and gfx90a: the generated gfx942 asm keeps v_mul_f32 ahead of v_cvt_f16_f32, and removing the barrier makes the multiply vanish entirely, which is exactly the reassociation #34484 set out to prevent.

Modifications

Accuracy Tests

Speed Tests and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): 🚫 Run #33351561701
Latest PR Test (Extra): ❌ Run #33351561593
Latest PR Test (AMD ROCm 7.2): 🚫 Run #33351561744

[yctseng0211]

unblocked amd install dependency :