Vulkan-radeon 26.1.0-1 broke the llama.cpp vulkan

[fiyaf39219]

After installing vulkan-radeon 26.1.0-1 I can no longer use llama.cpp vulkan - it just halt on the model loading stage. Downgrading to 26.0.6-2 then llama.cpp vulkan works fine.

The following command is my dowload comand:

sudo downgrade vulkan-radeon

My llama.cpp log with Mesa 26.1 is in

https://pastebin.com/5ChPqH3W

[devusbc]

Can confirm this.

It worked for me by downgrading just ‘vulkan-radeon’ to 26.0.6.

[fiyaf39219]

Yes can confirm downgrading just vulkan-radeon to 26.0.6-2 would fix the problem. Updated the first post.

[fiyaf39219]

How can I submit a bug report?

[maclinuxbravo]

[Tutorial] Reporting Bugs

Hi, We wanted to share you some information’s how to report bugs to CachyOS. Describe your Issue What is not working? Does downgrading package X fix the issue? Use the search function for equal issues Provide Logs CachyOS does provide a great tool to gather logs, called cachyos-bugreport.sh. This tool will collect logs from: dmesg journlctl When the logs are collected, it will ask the user, if this should be uploaded to our paste services. You simply run: sudo cachyos-bugreport.sh in t…

sudo cachyos-bugreport.sh

this collect :

• dmesg

• journalctl

• inxi (for hardware palteform support)

[fiyaf39219]

Thanks. The script generated a link of dmesg, journalctl but where should I report the bug (describe what happened) and provide the link? Thanks

Also attached the log file

[devusbc]

I did a full system upgrade today and now llama-cpp-vulkan works even with vulkan-radeon 26.1.0-1

not sure what the root cause was though.

[fiyaf39219]

I still have problem running llama-cpp-vulkan on AMD iGPU. Here is the llama.cpp logs with vulkan-radeon 26.1.0-1:

Loading model… |                                                                                                       /[New LWP 36355]
[New LWP 36342]
[New LWP 36341]

This GDB supports auto-downloading debuginfo from the following URLs:
https://debuginfod.cachyos.org
https://debuginfod.archlinux.org
Enable debuginfod for this session? (y or [n]) [answered N; input not from terminal]
Debuginfod has been disabled.
To make this setting permanent, add ‘set debuginfod enabled off’ to .gdbinit.
[Thread debugging using libthread_db enabled]
Using host libthread_db library “/usr/lib/libthread_db.so.1”.
0x00007fe8cdeb58b2 in ?? () from /usr/lib/libc.so.6
#0  0x00007fe8cdeb58b2 in ?? () from /usr/lib/libc.so.6
#1  0x00007fe8cdea897f in ?? () from /usr/lib/libc.so.6
#2  0x00007fe8cdf2daeb in wait4 () from /usr/lib/libc.so.6
#3  0x00007fe8cf0a01af in ggml_print_backtrace () from /usr/lib/libggml-base.so.0
#4  0x00007fe8cf0b8e70 in ?? () from /usr/lib/libggml-base.so.0
#5  0x00007fe8ce2bb7cc in ?? () from /usr/lib/libstdc++.so.6
#6  0x00007fe8ce29a650 in std::terminate() () from /usr/lib/libstdc++.so.6
#7  0x00007fe8cec9ae1f in ?? () from /usr/lib/libllama-common.so.0
#8  0x00007fe8cde471fe in __cxa_finalize () from /usr/lib/libc.so.6
#9  0x00007fe8cecfc698 in ?? () from /usr/lib/libllama-common.so.0
#10 0x00007fe8cf15d000 in ?? ()
#11 0x00007fe8cf19f0e2 in ?? () from /lib64/ld-linux-x86-64.so.2
#12 0x00007fe8cf1a3990 in ?? () from /lib64/ld-linux-x86-64.so.2
#13 0x00007fe8cde479a0 in ?? () from /usr/lib/libc.so.6
#14 0x00007fe8cde47a90 in exit () from /usr/lib/libc.so.6
#15 0x0000561f83046fd2 in ?? ()
#16 
#17 0x00007fe8cdeb58b2 in ?? () from /usr/lib/libc.so.6
#18 0x00007fe8cdea88ef in ?? () from /usr/lib/libc.so.6
#19 0x00007fe8cdea8ce7 in ?? () from /usr/lib/libc.so.6
#20 0x00007fe8cdeae5c7 in ?? () from /usr/lib/libc.so.6
#21 0x00007fe8ce30691c in std:
:join() () from /usr/lib/libstdc++.so.6
#22 0x00007fe8cdeb2384 in ?? () from /usr/lib/libc.so.6
#23 0x00007fe8cdeb24b9 in pthread_once () from /usr/lib/libc.so.6
#24 0x00007fe8c9b0097a in ?? () from /usr/lib/libggml-vulkan.so.0
#25 0x00007fe8c9bb9f45 in ?? () from /usr/lib/libggml-vulkan.so.0
#26 0x00007fe8c9bbe95c in ?? () from /usr/lib/libggml-vulkan.so.0
#27 0x00007fe8c9bc778b in ?? () from /usr/lib/libggml-vulkan.so.0
#28 0x00007fe8cf0c29ec in ggml_backend_sched_graph_compute_async () from /usr/lib/libggml-base.so.0
#29 0x00007fe8ce6c1f11 in llama_context::graph_compute(ggml_cgraph*, bool) () from /usr/lib/libllama.so.0
#30 0x00007fe8ce6b6438 in llama_context::process_ubatch(llama_ubatch const&, llm_graph_type, llama_memory_context_i*, ggml_status&) () from /usr/lib/libllama.so.0
#31 0x00007fe8ce6bafea in llama_context::decode(llama_batch const&) () from /usr/lib/libllama.so.0
#32 0x00007fe8ce6c61d3 in llama_decode () from /usr/lib/libllama.so.0
#33 0x00007fe8cedf9c22 in common_init_from_params(common_params&) () from /usr/lib/libllama-common.so.0
#34 0x0000561f830be384 in ?? ()
#35 0x0000561f83043c74 in ?? ()
#36 0x00007fe8cde27c4e in ?? () from /usr/lib/libc.so.6
#37 0x00007fe8cde27d8b in __libc_start_main () from /usr/lib/libc.so.6
#38 0x0000561f83046d55 in ?? ()                                                                                          -[Inferior 1 (process 36336) detached]
terminate called without an active exception
fish: Job 1, ‘llama-cli -m ~/model/gemma-4-E4…’ terminated by signal SIGABRT (Abort)

Also attach the mesa debug output

Downgrade back to vulkan-radeon 26.0.6-2 fix my problem.