[AMD] Fix GLM-5 fp8 KV quant path dispatch on MI300

[1am9trash]

Motivation

On MI300, running GLM-5-fp8 with FP8 KV cache can fail (see CI log).
The root cause is that the quant path does not dispatch the correct kernel (set_mla_kv_buffer_triton_fp8_quant).

Modifications

The flag self.nsa_kv_cache_store_fp8 is true only when KV cache is stored in fp8 with scaling. Our attention path uses fp8 KV cache without scaling, so it should not be gated by this flag.
This change moves the HIP + fp8 quant path out of the scaling-specific branch, ensuring MI300 dispatches the correct fused kernel (set_mla_kv_buffer_triton_fp8_quant).

This change only affects the MI300 code path.

Accuracy Tests

GLM-5-fp8 with fp8 kvcache Accuracy: 0.945
Also validated with the new CI script test_glm5_perf_amd.py prepared in PR #21710.

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.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[kkHuang-amd]

You can import "from sglang.srt.layers.quantization.fp8_kernel import fp8_dtype"
and use "self.dtype == fp8_dtype" to do condition check.

fp8_dtype is torch.float8_e4m3fnuz on mi300x and torch.float8_e4m3fn on mi35x

[1am9trash]

Fixed and reran well. Really appreciate the reminder.

[kkHuang-amd]

Remove Line 1585, when you use from sglang.srt.layers.quantization.fp8_kernel import fp8_dtype