feat(blog): MI355X Qwen3.5 SGLang v0.5.12 up-to-17x in 3 months

[functionstackx]

Summary

• New blog post: 13 weeks after Alibaba's Qwen3.5-397B-A17B release, AMD MI355X SGLang FP8 throughput per GPU on the 8k/1k workload moved up to 17.7x at iso-interactivity at 40 tok/s/user (210 → 3,709 tok/s/GPU), peak per-GPU throughput 1.3k → 6.4k tok/s/GPU
• Three-date comparison: v0.5.8.post1 (Feb 20) → v0.5.10rc0 (Apr 16) → v0.5.12 (May 19) on the same MI355X silicon
• Walks through the three AMD-authored upstream PRs that drove the April jump:
  • sgl-project/sglang#20736 — fuses shared expert with routed experts in Qwen2/Qwen3.5 MoE as topk+1 in a single AITER dispatch
  • sgl-project/sglang#21188 — adds forward_hip path to GemmaRMSNorm (AITER fused kernels instead of native fallback)
  • sgl-project/sglang#21421 — integrates AITER fused_topk into SGLang's softmax MoE top-K
• Covers Qwen3.5 architecture (397B / 17B active, 512 experts, hybrid Gated DeltaNet + Gated Attention, 256K context) and the TP=8 → TP=2/TP=4 retune that bundled with the April image bump
• All numbers sourced from the InferenceX 2026-05-19 run; chart preset linked from both DashboardCTA blocks

Test plan

• Replace benchmark-dark.png — I copied the light chart you shared into both light and dark slots so the post renders; a real dark-theme export should land before merging
• Verify chart preset URL resolves to the GLM-5… err, Qwen3.5 FP8 view across all 3 dates on MI355X SGLang
• Spot-check the iso-interactivity ratios at one or two band points (interpolation is linear on the Pareto frontier of each date)
• Confirm the canonical dashboard URL is inferencex.semianalysis.com (your link was a Vercel preview — I normalized to canonical)

🤖 Generated with Claude Code

---

Note

Low Risk
Content-only addition (MDX); no application logic, auth, or data pipeline changes.

Overview
Adds a new InferenceX blog post (mi355x-qwen3-5-sglang-v0-5-12-up-to-17x.mdx) documenting ~17.7x MI355X SGLang FP8 throughput-per-GPU gains on Qwen3.5-397B-A17B (8k/1k) from Feb → May 2026 on unchanged silicon.

The article walks three SGLang image milestones (v0.5.8.post1, v0.5.10rc0, v0.5.12), ties the April jump to three AITER-gated upstream SGLang PRs (#20736 shared-expert MoE fusion, #21188 HIP GemmaRMSNorm, #21421 fused_topk), and the bundled TP=8 → TP=2/TP=4 InferenceX recipe retune. It includes concurrency tables, iso-interactivity Pareto ratios, a Figure pointing at /images/mi355x-qwen3-5-sglang-v0-5-12-up-to-17x/, twin DashboardCTA links to the canonical InferenceX preset, and JsonLd FAQ schema.

Note: chart assets are referenced but not in this diff; the PR description calls out replacing benchmark-dark.png before merge.

^{Reviewed by Cursor Bugbot for commit b13f595. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview, Comment	May 25, 2026 11:29pm