[AMD] Add GLM-5.3-Flash recipes for MI300X, MI325X, and MI355X

[andyluo7]

Motivation

Add copy-pasteable AMD ROCm deployment recipes for zai-org/GLM-5.3-Flash alongside the existing NVIDIA recipes.

The AMD commands require the engine changes in #36607, which is stacked on the model-support PR #36507.

Modifications

• Add single-node TP8 ROCm recipes for MI300X, MI325X, and MI355X.
• Use the validated BF16 KV + TileLang DSA configuration with SGLANG_USE_AITER=1, Triton MoE, and CUDA graphs disabled.
• Mark MI300X and MI355X as verified from full GSM8K runs.
• Keep MI325X explicitly unverified: it shares gfx942 with MI300X, but it was not measured directly.
• Disable the unvalidated AMD MTP/low-latency lane and unvalidated backend combinations.
• Add accuracy-only benchmark records with exact model revision, engine PR/source manifest, image, and request accounting. No throughput or latency result is presented.

This PR is based on current main (02dfd3782601277bcf7e74c7a836a70011183165). It does not duplicate the separate NVIDIA HiCache/fusion cleanup in #36544.

Accuracy Tests

GPU	Recipe status	Full GSM8K	Request accounting
MI300X / gfx942	Verified	1,284 / 1,319 = 97.35%	1,319 unique records; 100% stop; 0 errors, empty generations, or truncations
MI325X / gfx942	Not verified; inferred from MI300X	—	No MI325X hardware run
MI355X / gfx950	Verified	1,288 / 1,319 = 97.65%	1,319 unique records; 100% stop; 0 errors, empty generations, or truncations

Validated model revision: 3f1971b7b5f7a528c9c4ef6212c8785298a8c24a.

Validated engine source:

• PR [AMD] Enable GLM-5.3-Flash on gfx942 and gfx950 #36607 head: 9e692c9216c3b5e5c443fecf6b995700eb68d2e4
• Source manifest SHA256: 2c240e0e01d5fdf04acc485ebfa25f8a1793ba45fb07f165eecedfba7ec1db80
• MI300X image: lmsysorg/sglang:v0.5.18-rocm720-mi30x
• MI355X image: lmsysorg/sglang:v0.5.18-rocm720-mi35x

Speed Tests and Profiling

Not measured. These are accuracy/correctness results only.

Documentation validation

• node docs/scripts/check_cookbook_configs.mjs
• focused pre-commit run --files ...
• npx --yes mintlify broken-links
• npx --yes mintlify validate

All passed. The final rebase changed only the parent with an unrelated test-file commit; the validated docs/ tree hash remained fc5d92e264b3271bd347f5acd91e06dde38fba2f.

Checklist

• Document platform, image, model revision, TP/DP topology, and required environment settings.
• Mark measured and inferred hardware separately.
• Run the cookbook config checker and docs validation.
• Avoid performance claims without a matched throughput benchmark.

---

CI States

Latest PR Test (Base): ✅ Run #33037066265
Latest PR Test (Extra): ❌ Run #33037402765
Latest PR Test (AMD ROCm 7.2): ➖ No AMD PR run found for this commit.