---
title: "Backends and installation"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Backends and installation}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
resource_files:
  - figures/logo.svg
  - figures/logo-dark.svg
---

<!-- cttir-vignette-logo: start -->
<style>
.cttir-vignette-logo { float: right; width: 120px; max-width: 30%; margin: 0 0 1rem 1.5rem; }
.cttir-vignette-logo img { width: 100%; height: auto; border: 0; background: transparent; box-shadow: none; }
.cttir-vignette-logo .cttir-vignette-logo-light { display: block; }
.cttir-vignette-logo .cttir-vignette-logo-dark { display: none; }
@media (prefers-color-scheme: dark) {
  .cttir-vignette-logo .cttir-vignette-logo-light { display: none; }
  .cttir-vignette-logo .cttir-vignette-logo-dark { display: block; }
}
[data-bs-theme="light"] .cttir-vignette-logo .cttir-vignette-logo-light { display: block; }
[data-bs-theme="light"] .cttir-vignette-logo .cttir-vignette-logo-dark { display: none; }
[data-bs-theme="dark"] .cttir-vignette-logo .cttir-vignette-logo-light { display: none; }
[data-bs-theme="dark"] .cttir-vignette-logo .cttir-vignette-logo-dark { display: block; }
</style>
<div class="cttir-vignette-logo">
  <img class="cttir-vignette-logo-light" src="figures/logo.svg" alt="hypermodR logo" width="120" height="139" />
  <img class="cttir-vignette-logo-dark" src="figures/logo-dark.svg" alt="hypermodR logo" width="120" height="139" />
</div>
<!-- cttir-vignette-logo: end -->

```{r setup, include=FALSE}
library(hypermodR)
knitr::opts_chunk$set(collapse = TRUE, comment = "#>", error = FALSE)
```
The core R workflow needs no Python runtime. Capability inspection is offline
and does not install anything. A capability probe is not model qualification.

```{r capabilities}
hm_backend_capabilities()
```

The automatic metadata resolver currently admits Linux x86_64 with an explicit
existing Python 3.11 or 3.12 and exactly uv 0.11.26. `reference_cpu` resolves
NumPy 2.2.6, Torch 2.13.0 and setuptools 84.1.0. Resolution is an explicit network operation and is
therefore omitted from offline builds. This setup route is not a blanket
platform/model support claim.

The earlier Torch 2.6.0 generated profile is retired following the publisher's
[checkpoint advisory](https://github.com/pytorch/pytorch/security/advisories/GHSA-63cw-57p8-fm3p).
Automatic `vllm_linux_cuda` resolution is blocked pending a reviewed replacement
for its vulnerable 0.11.0 profile; see the publisher's
[video-processing advisory](https://github.com/vllm-project/vllm/security/advisories/GHSA-4r2x-xpjr-7cvv).
Existing files are preserved, but those generated profiles cannot be installed
or started again. Resolve a new CPU profile and review its artifacts/terms, or
use a separately reviewed current custom manifest. Exact upstream/model/Linux
qualification remains separate from these version and admission repairs.

```{r resolve, eval=identical(Sys.getenv("HYPERMODR_RESOLVE_EXAMPLE"), "true")}
# Opt in only with explicit executables and a new persistent external directory.
resolution <- hm_resolve_backend("reference_cpu",
  model_id = "reference:cnn1d",
  python = Sys.getenv("HYPERMODR_PYTHON"), uv = Sys.getenv("HYPERMODR_UV"),
  directory = Sys.getenv("HYPERMODR_RESOLUTION_DIR")
)
resolution$hard_gates
```

Inspect every artifact's full license/notice text and exact identity before
setting its reviewed flag. `hm_admit_backend(resolution, reviewed_terms)`
creates an installation plan only after that review. Pip configuration files
and inherited pip options do not redirect installation outside the owned environment. `hm_backend_plan()` also
accepts an explicitly pinned custom profile. `hm_install_backend(plan,
accept_terms = TRUE)` performs the separately requested installation into an
external cache; `hm_start_backend()`, `hm_backend_status()` and
`hm_stop_backend()` own the process lifecycle. `hm_worker_request()` uses bounded
numeric interchange, never arbitrary Python deserialization.

A separately reviewed custom vLLM setup requires exact NVIDIA driver/device ABI admission and an exact
local checkpoint manifest. `hm_resolve_checkpoint()` resolves a public immutable
40-character revision without downloading weights and rejects custom remote
code; review checkpoint terms separately. Use `hm_vllm_chat()` only for a
compatible served language/vision model, with explicit consent and disclosure
for remote transmission. Native HSI architectures do not run through stock
vLLM. `hm_render_hsi()` records the information lost in three-channel views;
`hm_prompt()` binds prompt origins. Generated text remains auxiliary.

Install receipts, artifact hashes, exact versions and `hm_check_backend()`
smoke/parity evidence must travel with a support claim. Offline misses, corrupt
artifacts and unsupported platforms fail explicitly. Repeat installs verify
integrity; interrupted staging and ownership locks are not silently stolen.
No Windows, macOS, GPU or named upstream model qualification follows from this
article. No installation is triggered by loading the package or opening the app.