Improve BLIS blog: technical rigor, structure, and diagrams

[chcost]

Summary

Editorial and structural improvements to the BLIS blog post (builds on #344):

• Tone: Tightened prose to match the technical register of published llm-d blogs (KV-cache, Intelligent Scheduling)
• Structure: Added :::tip key takeaways, :::info fidelity validation scope callout, numbered figure captions (FIGURE 1–5), BLIS-vs-cluster comparison table, admission control results table, and standard "Get Involved with llm-d" closing section
• Diagrams: Replaced original PNG diagrams with purpose-built SVGs verified against the inference-sim codebase:
  • BLIS architecture (inputs → simulator internals → outputs)
  • AI-native inner/outer loop pipeline
  • Admission control latency overlay chart
• Content: Added "What's next" roadmap, "Further reading" links, and "Limitations" section refinements

Builds on top of #344 — intended as a complementary PR with editorial improvements.

[netlify]

✅ Deploy Preview for elaborate-kangaroo-25e1ee ready!

Name	Link
🔨 Latest commit	e0f1da5
🔍 Latest deploy log	https://app.netlify.com/projects/elaborate-kangaroo-25e1ee/deploys/6a4537b38f03a100081b4e87
😎 Deploy Preview	https://deploy-preview-363--elaborate-kangaroo-25e1ee.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

---

To edit notification comments on pull requests, go to your Netlify project configuration.

[github-actions]

Unsigned commits detected! Please sign your commits.

For instructions on how to set up GPG/SSH signing and verify your commits, please see GitHub Documentation.

[chcost]

@ahg-g @robertgshaw2-redhat PTAL

[mtoslalibu]

Two suggestions:

---

1. Clarify the code-generation paragraph (line ~147)

The last sentence is hard to parse — "stabilizing the disaggregation fraction near 50%" doesn't clearly convey what's happening or why that matters. Suggested rewrite:

Code generation: 30K-token prefix, approximately 1,500 uncached tokens per turn, 15 turns per session. At N=16, the threshold fires for 100% of requests, routing all of them to the prefill pool. This can saturate prefill and inflate TTFT by up to 20× relative to always-local. EDPP avoids this: when disaggregated requests start missing the TTFT SLO, its feedback loop reduces the disaggregation rate — in this workload, settling near 50% — so the prefill pool stays below saturation and TTFT remains bounded.

The key change: make explicit what the feedback loop does (reduces disaggregation rate), why (TTFT SLO misses), and what the outcome is (prefill stays below saturation), rather than stating the steady-state number without context.

---

2. Fix "The physics of high-fidelity distributed inference platform simulation" link

The current link points to Medium. It should point to the BLIS project blog:

- [The physics of high-fidelity distributed inference platform simulation](https://medium.com/modeling-distributed-inference/the-physics-of-high-fidelity-distributed-inference-platform-simulation-28fe27b59da2)
+ [The physics of high-fidelity distributed inference platform simulation](https://inference-sim.github.io/inference-sim/latest/blog/2026/04/09/the-physics-of-high-fidelity-distributed-inference-platform-simulation/)

[mtoslalibu]

Link should point to the BLIS project blog, not Medium:

Suggested change

	- [From simulation to production: the admission controller case study](https://ai-native-systems-research.github.io/ai-native-systems-research/blog/2026/05/13/from-simulation-to-production-how-an-ai-native-pipeline-discovered-a-better-admission-controller-for-llm-d/)
	- [The physics of high-fidelity distributed inference platform simulation](https://inference-sim.github.io/inference-sim/latest/blog/2026/04/09/the-physics-of-high-fidelity-distributed-inference-platform-simulation/)

[ahg-g]

Non blocking questions:

1. I know that simulation makes it faster to iterate over a wide range of configurations. But for the configurations that turned out to be good candidates, did we run actual experiments to cross check and validate?
2. How does this compare with Dynamo's AI Configurator?

[davidgs]

LGTM