Misc. bug: HIP compilation together with -DGGML_CPU_ALL_VARIANTS=ON does not load the model or detects the GPU

[luiznpi]

Name and Version

current version b4820, but somewhere from 29/01/2025

Operating systems

Windows

Which llama.cpp modules do you know to be affected?

libllama (core library)

Command line

Problem description & steps to reproduce

No longer possible to use the compilation with all the CPUs support and HIP together.

cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1030 -DGGML_HIP=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
-DGGML_CPU_ALL_VARIANTS=ON -DGGML_BACKEND_DL=ON

The compilation completes, but when loading the llama.dll / ggml.dll the GPU is no longer detected. The model is said to be correctly initialized, but when trying to use it complains is not loaded.

First Bad Commit

No response

Relevant log output

flutter: [2025-03-04 11:32:16.766962] ¤ New available Input Devices: []
flutter: [2025-03-04 11:32:16.801966] ¤ Initializing Llama...
flutter: [2025-03-04 11:32:16.803963] ¤ llama_isolate_manager.dart ¤ Spawning Llama isolate...
llama_model_load_from_file_impl: invalid value for main_gpu: 0 (available devices: 0)
flutter: [2025-03-04 11:32:17.141965] ¤ transcription_service.dart ¤ Llama model initialized: {requestId: 1, message: Llama model initialized successfully at Llama3-8b-Q4_K_M.gguf}
flutter: [2025-03-04 11:32:17.141965] ¤ transcription_service.dart ¤ Llama model initialization is complete!

flutter: [2025-03-04 11:44:19.024409] ¤ transcription_service.dart ¤ Llama error: Error creating Llama context: Bad state: Model must be loaded before creating a context.#0

[slaren]

Try to reproduce the issue with a debug build of llama-cli and include a full log.