Backend manifest 404s: whispercpp:rocm and vllm:rocm request gfx1201 assets published as gfx120X (RDNA4 / R9700)

[The-Monk]

Summary

On RDNA4 (gfx1201, Radeon AI PRO R9700), two backend installs fail with HTTP 404 because the backend manifest requests assets named gfx1201, but the corresponding release repos publish those assets under the family target name gfx120X. The two ROCm backends that do install correctly (llamacpp:rocm, sd-cpp:rocm) already use gfx120X, so this looks like a stale target-name pin in two specific backend entries.

Affected backends:

• whispercpp:rocm
• vllm:rocm

Environment

• Lemonade: lemonade-server 10.8.1~24.04 (installed from the lemonade-team/stable PPA)
• OS: Zorin OS 18.1 (Ubuntu 24.04 / noble base)
• GPU: 2× AMD Radeon AI PRO R9700 (Navi 48, gfx1201, RDNA4)
• Working ROCm backends on the same machine: llamacpp:rocm (b9752), sd-cpp:rocm (master-92a3b73) — both pulled therock-dist-linux-**gfx120X**-all-7.13.0.tar.gz

Reproduction

lemonade backends install whispercpp:rocm
lemonade backends install vllm:rocm

Actual behavior

whispercpp:rocm:

[1/1] whisper-v1.8.4-linux-rocm-gfx1201.tar.gz (0.0 MB)
Last error: HTTP error 404 for URL:
https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-linux-rocm-gfx1201.tar.gz

vllm:rocm:

[1/1] vllm0.20.1-rocm7.12.0-gfx1201-x64.tar.gz (0.0 MB)
Last error: HTTP error 404 for URL:
https://github.com/lemonade-sdk/vllm-rocm/releases/download/vllm0.20.1-rocm7.12.0-gfx1201/vllm0.20.1-rocm7.12.0-gfx1201-x64.tar.gz

Root cause / what's actually published

lemonade-sdk/whisper.cpp-rocm @ v1.8.4 — version is correct, only the target name is wrong:

• requested: whisper-v1.8.4-linux-rocm-**gfx1201**.tar.gz → 404
• published: whisper-v1.8.4-linux-rocm-**gfx120X**.tar.gz ✅ (also gfx110X, gfx1150, gfx1151)

lemonade-sdk/vllm-rocm — both the version and the target name are stale:

• requested: vllm**0.20.1**-rocm**7.12.0**-**gfx1201**-x64.tar.gz → 404 (release tag doesn't exist)
• published: vllm**0.22.1**-rocm**7.13.0**-**gfx120X**-x64 (split .part01/.part02), plus 0.22.0 (same naming)

Expected behavior

Both backends should install on gfx1201. The manifest entries should point at the gfx120X family asset names (and for vllm:rocm, bump the pin to the current 0.22.x / ROCm 7.13 split-archive release).

Notes

• whispercpp:cpu and whispercpp:vulkan install fine, so whisper STT still works via Vulkan on RDNA4 in the meantime.
• Happy to test a fix on real gfx1201 hardware (2× R9700) if useful.