ROCm unavailable for gfx103X, gfx110X and gfx120X

[lordlessknight]

Platform

Linux/Arch

Lemonade Version

10.8.0

GPU / APU Model

No response

Component

llama.cpp

Bug Description

This particular change broke the rocm backend for any AMD GPU from the gfx103X, gfx110X and gfx120X families.
I believe that this happens because the new code returns the specific name, for example, gfx1201, instead of returning gfx120X.

Steps to Reproduce

1. Try to run the command lemonade backends install llamacpp:rocm with an AMD GPU from the gfx103X, gfx110X or gfx120X families.
2. See that it returns the error: "GPU Unsupported: <gpu_family>"
3. You can also run the command lemonade backends and see that it returns the same error wherever the rocm backend is available.

Expected vs Actual Behavior

No response

Log Output

Additional Context

No response

[ianbmacdonald]

I can reproduce this on a Radeon RX 7900 XT (gfx1100, Linux). lemonade backends reports:

llamacpp    rocm    unsupported    Unsupported GPU: gfx1100

and /api/v1/system-info shows "family": "gfx1100" (KFD hands the name through as "110000").

Root cause matches your diagnosis: identify_rocm_arch_from_name() in system_info.cpp returns the specific arch (e.g. gfx1100, gfx1201) instead of the family download target (gfx110X, gfx120X) that the backend support set expects. Commit 2a7aa18 (#2093) removed ROCM_ARCH_MAPPING, so the regex and KFD-ISA detection paths stopped collapsing the specific arch to its family — the marketing-name heuristics still return families (gfx110X for a "7900" name), which is why only the gfx-string / KFD-numeric path regressed.

Taking this — I'll put up a fix that restores the specific→family normalization for those two paths (keeping the gfx115x iGPUs as exact targets), with a unit test covering gfx103X/110X/120X.

[ckuethe]

same as #2302 ?