Descr 693A iption
Name and Version
./llama-cli --version
register_backend: registered backend CPU (1 devices)
register_device: registered device CPU (13th Gen Intel(R) Core(TM) i9-13900H)
load_backend: failed to find ggml_backend_init in ~/workspace/github/llama.cpp/build/bin/libggml-cpu.so
version: 4942 (fbdfefe)
built with cc (Ubuntu 12.3.0-1ubuntu1~22.04) 12.3.0 for x86_64-linux-gnu
Operating systems
Linux
GGML backends
CPU
Hardware
13th Gen Intel(R) Core(TM) i9-13900H
Models
DeepSeek-Coder-V2-Lite-Instruct-Q4_K_M.gguf
Problem description & steps to reproduce
when i used "cmake -B build -DCMAKE_BUILD_TYPE=Debug" to build debug-mode and run the deepseekv2-lite model, It doesn't work properly anymore.
After some basic debugging, I found that during the traversal of the computation graph, specifically while applying the SiLU operation to the ffn-moe-gate (executing the node ffn_moe_silu-1), a numerical overflow occurred in the operand x, causing the program to crash.
This issue appears unrelated to debug mode, as switching to release mode allowed the program to run, but the output was incorrect：”<｜begin▁of▁sentence｜>Hello my name is!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!“
First Bad Commit
According to my preliminary judgment, this problem should be caused by commit:#12181
Relevant log output
llama-simple: \~/workspace/github/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c:7670: ggml_compute_forward_silu_f32: Assertion `!isnan(x)' failed.
llama-simple: \~/workspace/github/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c:7670: ggml_compute_forward_silu_f32: Assertion `!isnan(x)' failed.
llama-simple: \~/workspace/github/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c:7671: ggml_compute_forward_silu_f32: Assertion `!isinf(x)' failed.
Aborted (core dumped)