Add memra as a model library

[avifenesh]

Adds memra to model-libraries.ts with a snippet, per Adding a library to the Hub.

What it is. A from-scratch inference engine in Rust + CUDA for NVIDIA Blackwell workstation and consumer cards (sm_120a, with a compile-gated Hopper lane), serving GGUF and safetensors over an OpenAI-compatible HTTP API. MIT. I maintain it.

Prerequisite. A model already carries the tag: ?filter=memra returns Avifenesh/Qwen3.8-27B-NVFP4-MTP-GGUF (~3k downloads).

filter: false — one tagged model today, far under the 100-model bar for the models-page filter.

countDownloads: path_extension:"gguf" OR path:"config.json". Two artifact shapes, one counted file each:

• GGUF files are self-contained, so one file is one load — the same reasoning behind the Hub's own GGUF special case, and the pattern rkllm and litert-lm already use.
• safetensors checkpoints are counted by config.json, which is one per repo. Counting path_extension:"safetensors" would overestimate a sharded model by its shard count, which the docs explicitly warn against.

Snippet. Three short blocks:

1. Install, with the hardware requirement stated first — Blackwell-class card, Linux x86_64, driver 580+, no CUDA toolkit — so nobody installs it expecting an unsupported card to work.
2. One chat-templated generation via the engine's hf: resolver, noting that a repo containing several GGUF files needs :<substring> to pick one (which is exactly what the resolver's error asks for).
3. The OpenAI-compatible server, one line.

Checks. In packages/tasks: pnpm run check (tsc) clean, pnpm test 22 passed, pnpm run format:check clean.

---

Note

Low Risk
Additive registry and static snippet strings in packages/tasks; no auth, inference, or download-count logic changes in this diff.

Overview
Registers memra as a Hub model library so repos tagged with library_name: memra get the standard model-page treatment (label, repo link, usage snippets). The library is not added to the models-page filter (filter: false), consistent with a small catalog.

Model pages show three snippets: install via the upstream install.sh (with Blackwell / driver notes), one-shot chat generation with run-gen and hf:${model.id} (including multi-GGUF :<substring> selection), and an OpenAI-compatible memra-server one-liner bound to the current model.

^{Reviewed by Cursor Bugbot for commit 88f4041. Bugbot is set up for automated code reviews on this repo. Configure here.}

[Wauplin]

Looks good thank you!

Expect 2-3 days before getting it live on the Hub 🤗