Problem

Since 1.1.95, Windows users on CUDA 13 hosts get "Failed to download GPU backend". The "Better Configuration Available → CUDA 13" prompt offers a backend that 404s on download.

Root cause: ggml-org/llama.cpp renamed the Windows CUDA-13 release asset from llama-{tag}-bin-win-cuda-13.1-x64.zip to ...-cuda-13.3-x64.zip (visible on release b9495). The llamacpp-upstream provider hardcodes the 13.1 minor as the canonical backend id, so:

• detectIdealBackendType() recommends the literal win-cuda-13.1-x64 to every CUDA-13 host (the prompt), and
• getBackendDownloadUrl() turns it into …/b9495/llama-b9495-bin-win-cuda-13.1-x64.zip → HTTP 404.

Observed download attempt from a reporter (1.1.95):

Downloading backend b9495/win-cuda-13.1-x64 from
https://github.com/ggml-org/llama.cpp/releases/download/b9495/llama-b9495-bin-win-cuda-13.1-x64.zip
Error: Failed to get file size: HTTP status 404 Not Found

The release only ships llama-b9495-bin-win-cuda-13.3-x64.zip.

Fix (hotfix)

Bump the canonical Windows CUDA-13 minor 13.1 → 13.3 everywhere it's treated as an identity string, across the TS extension and the Rust plugin:

• recommendation tier (detectIdealBackendType), static variant list, remote-asset whitelist
• cudart sidecar filename + toolkit-version maps + backend regex
• category matcher + priority list + is_cuda_installed cudart lib lookup
• map_old_backend_to_new canonical target

Back-compat: legacy cuda-13.1 is still recognized by the category matcher and is migrated forward to 13.3 by map_old_backend_to_new, so an already-installed win-cuda-13.1-x64 folder is not orphaned. Updated the affected Rust unit tests accordingly.

Out of scope / follow-ups

• Driver gate unchanged. The CUDA-13 gate (>= 581.15) is the documented minimum for Toolkit 13.1; CUDA 13.3's minimum Windows driver should be verified separately. Low risk here because the runtime device probe already degrades to 12.4 / Vulkan / CPU when a recommended tier can't enumerate a GPU.
• Root-cause hardening (so this never recurs on the next upstream minor bump) is tracked separately — see the companion issue: resolve the CUDA-13.x asset name dynamically from the release instead of hardcoding the minor.

Testing

• CI (TS typecheck/vitest + cargo test for tauri-plugin-llamacpp-upstream)
• Manual: on a Windows CUDA-13 host, accept the "CUDA 13" upgrade prompt and confirm the backend downloads (no 404).

🤖 Generated with Claude Code