Repository navigation
[Bug] core dump in CPU backend if using FP8 (e4m3) tensors #2139
Description
Git commit
Operating System & Version
Ubuntu 24.04
GGML backends
CPU
Command-line arguments used
./sd-cli -m ./RealVisXL_V5_Lightning_Native_FP8_00001_.safetensors -p "a cute cat" --steps 5 --cfg-scale 1
Steps to reproduce
Use a CPU version of sd-cli version and run above command line.
The model comes from here: https://huggingface.co/sharp-y/realvisxl_v5.0_turbo_fp8
Also you could use any other sd-cli (with CUDA or so) if adding option --backend cpu , or run sd-cli with any other T2I model using fp8_e4m3 tensors, e.g.: https://huggingface.co/DanielMarius/t5xxl_fp8_e4m3fn.safetensors (add --clip-on-cpu here).
What you expected to happen
Create a PNG
What actually happened
Core dump, see below
Logs / error messages / stack trace
Program terminated with signal SIGSEGV, Segmentation fault.
GDB backtrace:
#0 0x0000000000000000 in ?? ()
#1 0x00005c79db5a7de4 in ggml_compute_forward_mul_mat_one_chunk (params=0x7c7d249317c0, dst=0x7c7d344fb3b0, type=GGML_TYPE_F8_E4M3, num_rows_per_vec_dot=1, ir0_start=48, ir0_end=64,
ir1_start=0, ir1_end=16) at /home/xxx/stable-diffusion.cpp/ggml/src/ggml-cpu/ggml-cpu.c:1394
#2 0x00005c79db5a8cec in ggml_compute_forward_mul_mat (params=0x7c7d249317c0, dst=0x7c7d344fb3b0) at /home/xxx/stable-diffusion.cpp/ggml/src/ggml-cpu/ggml-cpu.c:1615
#3 0x00005c79db5aa8fc in ggml_compute_forward (params=0x7c7d249317c0, tensor=0x7c7d344fb3b0) at /home/xxx/stable-diffusion.cpp/ggml/src/ggml-cpu/ggml-cpu.c:2026
#4 0x00005c79db5acf03 in ggml_graph_compute_thread (data=0x5c79f3d8ddb0) at /home/xxx/stable-diffusion.cpp/ggml/src/ggml-cpu/ggml-cpu.c:3326
#5 0x00005c79db5ae328 in ggml_graph_compute._omp_fn.0 () at /home/xxx/stable-diffusion.cpp/ggml/src/ggml-cpu/ggml-cpu.c:3616
#6 0x00007c7d8f0f679f in ?? () from /lib/x86_64-linux-gnu/libgomp.so.1
#7 0x00007c7d60e9cb84 in start_thread (arg=<optimized out>) at ./nptl/pthread_create.c:447
#8 0x00007c7d60f29ecc in clone3 () at ../sysdeps/unix/sysv/linux/x86_64/clone3.S:78
Additional context / environment details
The problem exists since 2026-Aug-30 / commit 2540a4f .
Activity