feat(blog): MI355X DeepSeek-V4-Pro SGLang — 110.5x throughput per GPU in 26 days

[functionstackx]

Summary

• New blog post: MI355X DeepSeek-V4-Pro on SGLang — 110.5x throughput per GPU in 26 days (/blog/mi355x-deepseek-v4-pro-sglang-110x-in-26-days). Time-series story of the sgl-project/sglang amd/deepseek_v4 side branch: first-light 20.4 tok/s/GPU at 2.4 tok/s/user on 2026-04-25 (FP8, recipe needed SGLANG_HACK_FLASHMLA_BACKEND=torch to compile) → 2,256 tok/s/GPU at 9.4 tok/s/user on 2026-05-21 (FP4, DP attention, lmsysorg/sglang:v0.5.12-rocm720-mi35x). Both axes climb together: 110.5x throughput-per-GPU × 3.85x interactivity.
• Decomposes the 31 numbered side-branch PRs into 5 mechanism buckets — DSA attention (TileLang indexer + Triton sparse MLA), mHC fusion, RoPE+Hadamard fusion, MoE (FlyDSL + FP4 + fused hash topk), AITER + misc fusions — and traces each container-image bump in the InferenceX recipe loop (rocm/sgl-dev → lmsysorg/sglang).
• Adds 9 files: the MDX post + 4 light/dark image pairs (benchmark chart, DSv4-Pro-Max vs Claude Opus 4.6 / GPT-5.4 / Gemini 3.1 quality bars, MI355X-vs-B200 SXM /gpu-specs radar, B200-vs-MI355X SGLang FP4 performance curve).
• On-Paper Specs section in What's Next anchors the residual ~5x gap to B200 SGLang as software, not silicon: MI355X has 1.60x HBM capacity, the same 8 TB/s HBM BW, and 1.12x dense FP4 / FP8 / BF16 vs B200 SXM; only scale-up BW (576 vs 900 GB/s uni-di, 0.64x) favors B200. Pulled from packages/app/src/lib/gpu-specs.ts.
• Iso-interactivity table uses the bundled iso_interactivity.py helper (monotone cubic Hermite on the upper-left Pareto frontier — matches d3.curveMonotoneX exactly so the table values track the rendered chart, not a linear approximation). Unreachable cells render as _unreachable_.
• What's Next explicitly flags the upstream-to-main migration story via PR #24933 (kk, merged 2026-05-18, +3,678 / -70) — first chunk landing DSv4 on ROCm in main in eager mode, with the perf-critical fusions / TileLang indexer / FlyDSL MoE / SGLANG_OPT_* toggles still side-branch-only. Hard reproducibility note: SGLang main will under-perform this post by an order of magnitude until those upstream.
• FAQ JSON-LD covers the 5 questions readers actually ask: peak number, what shipped, why 04-25 was so slow, NVIDIA comparison, what's still uncovered.

Test plan

• pnpm dev and visit /blog/mi355x-deepseek-v4-pro-sglang-110x-in-26-days — verify all 4 figures render in light + dark modes
• Post appears in /blog listing with correct title, subtitle, publish date (2026-05-26)
• OG image renders correctly for /blog/mi355x-deepseek-v4-pro-sglang-110x-in-26-days
• DashboardCTA at top + bottom links land on the preset MI355X DSv4-Pro 5-date view on inferencex.semianalysis.com
• Live chart link inside body opens the same preset view
• Sitemap / RSS feed / llms.txt include the new post
• All 31 SGLang PR links + PR #24933 + DeepSeek announcement + SemiAnalysis tweet resolve

🤖 Generated with Claude Code

---

Note

Low Risk
Content-only addition (MDX + referenced static images); no application logic, auth, or data-path changes.

Overview
Adds a new long-form benchmark article at /blog/mi355x-deepseek-v4-pro-sglang-110x-in-26-days documenting 26 days of MI355X SGLang tuning for DeepSeek-V4-Pro on the amd/deepseek_v4 fork—from first-light ~20 tok/s/GPU (2026-04-25 FP8) to ~2,256 tok/s/GPU (2026-05-21 FP4)—with throughput and interactivity rising together.

The post ties gains to 31 side-branch PRs (DSA/TileLang + Triton sparse MLA, mHC/RoPE/Hadamard fusions, FlyDSL MoE, FP4 path, InferenceX container bumps), publishes date-stamped concurrency tables, an iso-interactivity comparison, and MI355X vs B200 software-gap framing. It uses existing blog MDX patterns: Figure (light/dark assets under /images/mi355x-deepseek-v4-pro-sglang-110x-in-26-days/), DashboardCTA + live InferenceX links, and JsonLd FAQ schema for SEO.

^{Reviewed by Cursor Bugbot for commit 036d473. 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 26, 2026 4:21am

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 036d473. Configure here.}

[cursor]

Iso-interactivity table starts with unreachable ratio row

Medium Severity

The iso-interactivity table's first row (8 tok/s/user) shows _unreachable_ for 05-21 and _∞_ in the ratio column. This happens because the 05-21 Pareto frontier ends at 9.37 tok/s/user (conc 256); conc 512 at 5.59 tok/s/user is dominated and excluded from the frontier. The project's SKILL.md row-pruning heuristic explicitly states "The first row of the table must have two real numbers — never start with an _unreachable_ row" and warns that "a table that opens with _∞_ reads like the data is missing." Starting at 10 tok/s/user — where both 05-02 and 05-21 have real values — avoids this.

 

^{Reviewed by Cursor Bugbot for commit 036d473. Configure here.}

[cursor]

Dashboard URL dates mismatch blog's discussed five dates

Medium Severity

The i_dates URL parameter decodes to six dates (04-25, 05-03, 05-04, 05-08, 05-19, 05-21), but the blog tables and figure captions reference five dates (04-25, 05-02, 05-04, 05-10, 05-21). Two dates are offset — blog says 05-02 but URL has 05-03, blog says 05-10 but URL has 05-08 — and 05-19 appears in the URL but is never discussed. The text at line 158 says "across the 5 measured dates" while the link shows 6 curves. Readers clicking the CTA will see different date labels and an extra undiscussed curve.

Additional Locations (2)

• packages/app/content/blog/mi355x-deepseek-v4-pro-sglang-110x-in-26-days.mdx#L157-L158
• packages/app/content/blog/mi355x-deepseek-v4-pro-sglang-110x-in-26-days.mdx#L204-L207

 

^{Reviewed by Cursor Bugbot for commit 036d473. Configure here.}