Eval bug: 'ggml-impl.h:319: fatal error' starting certain models

[munchmunch77]

Name and Version

On main branch:

version: 0.4.6-dev (build 11794, commit 78af832)
built with MSVC 19.51.36256.0 for x64
ggml op count: 105

Operating systems

Windows

GGML backends

CUDA

Hardware

RTX 5070Ti, Ryzen 9900x

Models

No response

Problem description & steps to reproduce

Certain models don't start, same command starts on llama.cpp:

llama-server.exe' --ctx-size 4096
--batch-size 512 --ubatch-size 512
--flash-attn on --load-mode mlock
--gpu-layers all --fit-target 320
--model 'C:\AI Models\unsloth\Llama-3.2-3B-Instruct-GGUF\Llama-3.2-3B-Instruct-UD-Q4_K_XL.gguf' --temp 0.75
--top-k 40 --top-p 0.90
--min-p 0.05 --repeat-penalty 1.12
--presence-penalty 0.1 --parallel 1
--host 0.0.0.0 `
--jinja

First Bad Commit

No response

Relevant log output

Logs

0.00.031.700 I srv    load_model: loading model 'C:\AI Models\unsloth\Llama-3.2-3B-Instruct-GGUF\Llama-3.2-3B-Instruct-UD-Q4_K_XL.gguf'
0.01.388.633 I cmn          init: llama threadpool init, n_threads = 12
C:\source\beellama.cpp\ggml\src\ggml-impl.h:319: fatal error
Process exited (code=3221226505, signal=null)