Eval bug: llama-bench seems to be broken

[electroglyph]

Name and Version

version: 5215 (5f5e39e)
built with MSVC 19.43.34808.0 for x64

i've tested CPU, Vulkan, and SYCL, llama-bench either crashes and burns, or outputs the following 2 lines and then exits:

| model                          |       size |     params | backend    | ngl |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |

Operating systems

Windows

GGML backends

CPU

Hardware

Ryzen 7900X + Intel A770

Models

i tried several which are currently working with llama-server

Problem description & steps to reproduce

.\llama-bench.exe -m model

First Bad Commit

No response

Relevant log output

PS C:\Users\ANON\repos\AI_Grotto\llama.cpp\Windows\CPU\AVX512> .\llama-bench.exe -m C:\LLM\google-gemma-3-12b-it-qat-q4_0-gguf-small\gemma-3-12b-it-q4_0_s.gguf
| model                          |       size |     params | backend    | ngl |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
PS C:\Users\ANON\repos\AI_Grotto\llama.cpp\Windows\CPU\AVX512>

[gnusupport]

llama-bench -m Qwen3-0.6B-Q8_0.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes

model	size	params	backend	ngl	test	t/s
terminate called after throwing an instance of 'std::bad_alloc'
what(): std::bad_alloc
Aborted

[pt13762104]

./bin/llama-bench -ot ".ffn_.*_exps.=CPU" -m ~/Qwen_Qwen3-30B-A3B-Q4_K_L.gguf -ngl 99                                                                                                           
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce GTX 1660 Ti, compute capability 7.5, VMM: yes
| model                          |       size |     params | backend    | ngl | ot                    |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------------- | --------------: | -------------------: |
[1]    28437 segmentation fault (core dumped)  ./bin/llama-bench -ot ".ffn_.*_exps.=CPU" -m ~/Qwen_Qwen3-30B-A3B-Q4_K_L.gguf

[Alcpz]

I've had problems as well since the merge of #13096. I found an inconsistency between the values and the fields which caused llama-bench to crash when printing the output of the tests. Opened a PR, not sure if that is the correct fix, but I hope it at least gets help to bring visibility to the issue

[pt13762104]

It worked:

| model                          |       size |     params | backend    | ngl | ot                    |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------------- | --------------: | -------------------: |
| qwen3moe 30B.A3B Q4_K - Medium |  16.49 GiB |    30.53 B | CUDA       |  99 | .ffn_.*_exps.=CPU     |           pp512 |        107.80 ± 6.18 |
| qwen3moe 30B.A3B Q4_K - Medium |  16.49 GiB |    30.53 B | CUDA       |  99 | .ffn_.*_exps.=CPU     |           tg128 |         15.05 ± 0.41 |