Blog: present both usage modes as first-class (typed non-AD core and opt-in full stack)

[shinaoka]

Summary

The flagship blog post (blog/introducing-tenferro-rs, live in EN/JA/ZH) frames tenferro as "a differentiable tensor stack" and leads with "linear algebra, PyTorch-style eager autodiff, JAX-style traced transforms, ...". Differentiability is the headline identity. The equally-valid usage — tenferro's typed, minimal-dependency, non-AD core, usable without buying into the AD stack — is absent from the outreach narrative. This issue tracks a blog update that presents both usage modes as first-class.

This is the outreach counterpart of the tenferro-rs positioning/docs work (see the tenferro-rs repo issue for README/docs). It broadens the perceived audience: people who want a typed Rust tensor layer without the AD/graph machinery.

Evidence (from tenferro-rs)

• tenferro-tensor-core depends only on num-complex, smallvec, thiserror.
• autodiff is a non-default feature; non-AD immediate linalg exists via the explicit backend, no autodiff, no traced graph. The backend is constructed once and reused (it owns a buffer pool + context), then reused across ops:

  let mut backend = CpuBackend::new();
  let outputs = backend.svd_read(TensorView::F64(a.as_view()))?;

Direction

• Recommended: a new follow-up post, not a rewrite of the launch post (keep the launch record intact).
• Axis: "choose your altitude" — (A) typed, minimal-dependency, non-AD core; (B) opt-in AD / traced / GPU / einsum / FFT on top. Both first-class.
• Show concrete (A): the 3-dependency core, autodiff non-default, an explicit-backend non-AD immediate linalg example that constructs the backend once and reuses it (do not model per-call CpuBackend::new() — it discards the buffer pool).
• Follow the established multi-language convention (EN/JA/ZH), matching the existing introducing-tenferro-rs set.
• Fits the ongoing Schelling-point outreach cadence (SciRust monthly, Psi-k).

Mechanics

• Jekyll repo tensor4all/tensor4all.github.io; a post lives at blog/<slug>/index.md (front matter: title only); push to main deploys.

Acceptance criteria

• A new post under blog/<slug>/ presenting the two usage modes as co-equal, with a runnable (A) non-AD immediate example including linalg via the explicit backend.
• Language variants consistent with the existing convention (EN plus JA/ZH as agreed).
• Messaging consistent with the tenferro-rs README positioning update.

Related

• tenferro-rs positioning/docs: Positioning & docs: make both usage modes first-class (typed non-AD core and opt-in full stack) tenferro-rs#1318

[shinaoka]

Editorial audit note: a follow-up post is the right shape here. I would not rewrite the launch post unless there is a concrete factual error.

Suggested dependency: let tensor4all/tenferro-rs#1318 settle the repository wording first, then reuse that framing here so the site and README do not diverge.

Suggested framing for the post:

• tenferro-rs has two first-class routes: typed/eager tensor compute without AD, and the opt-in traced/AD/full-stack route;
• backend reuse is part of the normal usage story, not an optimization footnote;
• backend-explicit linalg belongs in the typed/non-AD route;
• avoid promising CUDA integer compute parity until Integer dtype compute parity on the CUDA/CubeCL GPU backend tenferro-rs#1314 and Unify backend capability representation across the typed and erased tensor faces tenferro-rs#1315 have landed or are clearly caveated.

For the EN/JA/ZH variants, I would keep the structure identical and localize examples/phrasing only where it improves readability. That makes future corrections much cheaper.