[zijiexia]

What

Builds on the GB300 NVFP4 scaffolding already in main (#29380 + #29466) with the newly-measured single-node 4×GB300 (TP4) recipes and benchmark numbers for nvidia/GLM-5.2-NVFP4, plus a small --trust-remote-code cleanup across the NVFP4 cells.

Recipes (glm-5.2.jsx) — GB300 NVFP4

Strategy	Before (main)	After
Low-Latency	mem-frac 0.8	mem-frac 0.85 (MTP 5-1-6 unchanged)
Balanced	bare TP4	+ DP-Attention (dp4) + MTP 2-1-3 + max-running 256 + mem-frac 0.92
High-Throughput	(absent)	new — DP-Attention (dp4), greedy, max-running 512, mem-frac 0.92

Also drop --trust-remote-code from every NVFP4 cell (B300 + GB300) — GLM-5.2 (glm_moe_dsa) is built into sglang and the FP8 / BF16 cells never used it, so the NVFP4 cells now match. --quantization modelopt_fp4 stays right after --tp.

Benchmarks (glm-5.2-benchmarks.jsx) — GB300 NVFP4

Fills the previously-pending stubs for all three strategies (ISL/OSL 8192/1024; the card's per-GPU column = total server output tok/s ÷ 4 GPUs):

Strategy	concurrency	output tok/s (total)	TPOT	TTFT
Low-Latency	1	337	2.23 ms	238 ms
Low-Latency	16	1248	11.9 ms	315 ms
Balanced	64	1162	58 ms	1169 ms
Balanced	256	1695	167 ms	6389 ms
High-Throughput	1024	1730	321 ms	156 s

• AIME25 = 89.58% (per-cell override of the variant default; on par with FP8's 87.7% within AIME25 run-to-run variance).

Notes

• Measured on the lmsysorg/sglang:dev-glm52-nvfp4 preview image — already wired as the per-quant Docker image for b300|nvfp4 / gb300|nvfp4 in main.
• B300 NVFP4 is touched only by the --trust-remote-code removal; its recipes/benchmarks are otherwise unchanged. FP8 / BF16 and the MDX page are untouched (the NVFP4 prose landed in [Docs] Add NVFP4 quantization to GLM-5.2 cookbook #29380).

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ✅ Run #28283086079
Latest PR Test (Extra): ❌ Run #28283086032

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
lmsysorg	🟢 Ready	View Preview	Jun 27, 2026, 5:35 AM

💡 Tip: Enable Workflows to automatically generate PRs for you.

[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!