[Docs][AMD] Add optimized GLM-5.3-Flash MI355X recipe

[andyluo7]

Motivation

Document the measured GLM-5.3-Flash FP8 serving recipe for 8x MI355X and publish the exact 1K/1K concurrency-1 and concurrency-32 results. The optimized recipe depends on the ROCm engine work in #36607 and the model-specific gfx950 BF16 GEMM tuning in ROCm/aiter#5060.

Modifications

• Select FP8 KV cache with TileLang DSA for MI355X while retaining the existing BF16/Triton/graph-disabled gfx942 recipes.
• Use TP8/EP1, AITER MoE, full decode graphs at batch sizes 1 and 32, disabled prefill graphs, and disabled radix cache for the measured MI355X recipe.
• Add median TTFT/TPOT and total-throughput-per-GPU rows for concurrency 1 and 32.
• Document exact request accounting, aggregate throughput, median interactivity, engine/AITER revisions, and the distinction between the speed run and the earlier full GSM8K accuracy run.
• Keep CUDA selections unchanged; TRT-LLM DSA remains CUDA-only.

Accuracy Tests

The existing MI355X accuracy result is retained and explicitly identified as a separate run:

• 8x MI355X, TP8, full GSM8K: 1,288/1,319 = 97.65%.
• 100% stop rate.
• Zero request errors, empty generations, or truncations.
• Accuracy source: SGLang PR [AMD] Enable GLM-5.3-Flash on gfx942 and gfx950 #36607 head 9e692c9216.

The performance run completed all 330 requests with exact 1,024-token inputs and outputs and empty error strings.

Speed Tests and Profiling

Hardware and software:

• 8x AMD Instinct MI355X (gfx950), TP8/EP1.
• SGLang PR [AMD] Enable GLM-5.3-Flash on gfx942 and gfx950 #36607 head 9d2087693988.
• AITER PR [Config] [Tune] Add GLM-5.3 BF16 GEMM configs for gfx950 ROCm/aiter#5060 head 95565e33c828.
• FP8 E4M3 KV, AITER MoE, TileLang DSA, AITER mHC dispatch, shared-expert fusion, HIP fused k-pool, full decode graphs at batch sizes 1 and 32.
• Random dataset, seed 42, ISL/OSL 1,024/1,024, ignored EOS, radix cache disabled.

Concurrency	Requests	Aggregate total throughput	Total throughput/GPU	Median TTFT	Median TPOT	Median interactivity
1	10/10	287.8493 tok/s	35.9812 tok/s/GPU	84.8435 ms	6.8675 ms	145.6126 tok/s/user
32	320/320	5,731.3649 tok/s	716.4206 tok/s/GPU	514.8845 ms	10.6696 ms	93.7246 tok/s/user

Against the recorded 8x B200 baseline, MI355X reached 90.18% / 89.11% of B200 total throughput / median interactivity at concurrency 1 and 101.07% / 98.35% at concurrency 32. The B200 raw JSON is not locally available, so the docs publish only independently preserved MI355X measurements rather than presenting the comparison as a new revalidation.

Validation

• node docs/scripts/check_cookbook_configs.mjs
• pre-commit run --files docs/cookbook/autoregressive/GLM/GLM-5.3-Flash.mdx docs/src/snippets/configs/zai-org/glm-5.3-flash.jsx docs/src/snippets/configs/zai-org/glm-5.3-flash-benchmarks.jsx
• npx -y mint@latest validate
• Static assertion of all MI355X flags and absence of TRT-LLM DSA, DeepGEMM, and --disable-cuda-graph in the MI355X cell.
• Direct extraction of TTFT, TPOT, throughput, token lengths, completion counts, and error counts from the saved benchmark JSON.

Checklist

• Format and changed-file pre-commit checks pass.
• Cookbook config validation passes.
• Mintlify build validation passes.
• Documentation updated with exact platform, topology, dependencies, and benchmark protocol.
• Accuracy and speed provenance are separated.

---

CI States

Latest PR Test (Base): ✅ Run #33117751830
Latest PR Test (Extra): ❌ Run #33117751273
Latest PR Test (AMD ROCm 7.2): ➖ No AMD PR run found for this commit.

[andyluo7]

Local validation is complete (check_cookbook_configs, changed-file pre-commit, and Mintlify build validation). The current gate failures are only the expected missing run-ci / run-ci-extra labels; could a maintainer please apply the appropriate CI label for this docs-only PR?