[AMD] Add GLM-5-FP8 nightly performance benchmarks for MI30x and MI35x

[michaelzhang-ai]

Summary

Add GLM-5-FP8 nightly perf benchmarks (bench_one_batch) for MI30x and MI35x. Both accuracy and perf use zai-org/GLM-5-FP8 with NSA tilelang backend, TP=8, FP8 KV cache, and --reasoning-parser=glm45 --tool-call-parser=glm47 (matching NV/InferenceX configs).

Changes

• New: test/registered/amd/perf/mi30x/test_glm5_perf_amd.py (suite: nightly-perf-8-gpu-glm5)
• New: test/registered/amd/perf/mi35x/test_glm5_perf_mi35x.py (suite: nightly-perf-8-gpu-mi35x-glm5)
• Modified: accuracy tests in mi30x/ and mi35x/ — switch model to zai-org/GLM-5-FP8, add parser flags
• Modified: nightly-test-amd.yml and nightly-test-amd-rocm720.yml — add perf step after accuracy in each GLM-5 job

Workflow behavior

• Accuracy step has no continue-on-error — if it fails, perf is skipped and the job fails
• Perf step has continue-on-error: true — perf failures don't block CI

Dependencies

• [AMD] Fix GLM-5 fp8 KV quant path dispatch on MI300 #22314: Fix MI300 FP8 KV quant path dispatch (enables --kv-cache-dtype fp8_e4m3 on MI30x)
• Reduce unnecessary kernels and copies in the NSA indexer #22232: Fix NSA indexer GPU memory fault at batch_size=64 (fixes MI35x perf crash)

Server config

	MI30x	MI35x
--kv-cache-dtype	fp8_e4m3	fp8_e4m3
--mem-fraction-static	0.85	0.85
--model-loader-extra-config	{"enable_multithread_load": true}	{"enable_multithread_load": true, "num_threads": 8}
Env	SGLANG_USE_AITER=1	SGLANG_ROCM_FUSED_DECODE_MLA=0, ROCM_QUICK_REDUCE_QUANTIZATION=INT4, SAFETENSORS_FAST_GPU=1

CI validation

• Nightly Test (AMD): https://github.com/sgl-project/sglang/actions/runs/24118470025
• Nightly Test (AMD ROCm 7.2): https://github.com/sgl-project/sglang/actions/runs/24118470693

MI30x perf results (earlier run, without FP8 KV cache)

From AMD run / ROCm 7.2 run:

batch	ISL	latency (s)	input tok/s	output tok/s	ITL (ms)
1	4096	25.50	1735	22.1	45.2
8	4096	33.76	5172	149.4	53.6
16	4096	40.23	6313	274.5	58.3
64	4096	85.64	6351	738.7	86.6

[gemini-code-assist]

Code Review

This pull request introduces nightly performance benchmarks for the GLM-5 model on AMD MI30x and MI35x platforms. The reviewer identified several areas for improvement, including refactoring duplicated report generation logic into a shared utility, fixing potential division-by-zero errors in throughput calculations, and enhancing test portability by avoiding hardcoded local paths. Additionally, it was noted that certain environment variables should be consistently applied across different GPU configurations to ensure optimal performance.

[gemini-code-assist]

SGLANG_USE_AITER is missing from the environment variables for MI35x, whereas it is enabled for MI30x. Since both use the same model and attention backend (tilelang), this might be an oversight that could lead to suboptimal performance results on MI35x.

            "env_vars": {
                "SGLANG_USE_AITER": "1",
            },

[1am9trash]

Maybe we can add --reasoning-parser=glm45 --tool-call-parser=glm47 to the GLM-5-FP8 AMD test configs for consistency.
These parsers are already used in NV unit tests and InferenceX tests, so aligning AMD test settings would reduce cross-platform behavior drift.

NV cmd
InferenceX cmd