==== Podstawowe informacje ==== $ date śro, 10 cze 2026, 09:43:08 CEST [exit=0] $ uname -a Linux llama-station 6.8.0-124-generic #124~22.04.1-Ubuntu SMP PREEMPT_DYNAMIC Tue May 26 21:05:19 UTC x86_64 x86_64 x86_64 GNU/Linux [exit=0] $ cat /etc/os-release 2>/dev/null || true PRETTY_NAME="Ubuntu 22.04.5 LTS" NAME="Ubuntu" VERSION_ID="22.04" VERSION="22.04.5 LTS (Jammy Jellyfish)" VERSION_CODENAME=jammy ID=ubuntu ID_LIKE=debian HOME_URL="https://www.ubuntu.com/" SUPPORT_URL="https://help.ubuntu.com/" BUG_REPORT_URL="https://bugs.launchpad.net/ubuntu/" PRIVACY_POLICY_URL="https://www.ubuntu.com/legal/terms-and-policies/privacy-policy" UBUNTU_CODENAME=jammy [exit=0] ==== Sprzęt PCI / GPU ==== $ lspci -nnk | grep -EA4 'VGA|3D|Display' 04:00.0 VGA compatible controller [0300]: NVIDIA Corporation GK208B [GeForce GT 710] [10de:128b] (rev a1) Subsystem: Micro-Star International Co., Ltd. [MSI] GK208B [GeForce GT 710] [1462:8c93] Kernel driver in use: nouveau Kernel modules: nvidiafb, nouveau 04:00.1 Audio device [0403]: NVIDIA Corporation GK208 HDMI/DP Audio Controller [10de:0e0f] (rev a1) [exit=0] ==== Sterownik jądra AMDGPU ==== [OK] Moduł amdgpu jest załadowany. $ lsmod | grep '^amdgpu' amdgpu 17190912 0 [exit=0] $ dmesg 2>/dev/null | grep -iE "amdgpu|kfd|firmware|vram|ring|ras|gfx" | tail -n 120 || true [exit=0] ==== /dev/kfd i /dev/dri ==== [OK] /dev/kfd istnieje $ ls -l /dev/kfd crw-rw---- 1 root render 236, 0 cze 10 09:33 /dev/kfd [exit=0] [OK] /dev/dri istnieje $ ls -l /dev/dri total 0 drwxr-xr-x 2 root root 80 cze 10 08:23 by-path crw-rw----+ 1 root video 226, 1 cze 10 08:23 card1 crw-rw----+ 1 root render 226, 128 cze 10 08:23 renderD128 [exit=0] ==== Grupy i uprawnienia użytkownika ==== $ id uid=1000(adminek) gid=1000(adminek) groups=1000(adminek),4(adm),24(cdrom),27(sudo),30(dip),44(video),46(plugdev),110(render),122(lpadmin),135(lxd),136(sambashare) [exit=0] [OK] Użytkownik jest w grupie render [OK] Użytkownik jest w grupie video ==== Pakiety i narzędzia ROCm ==== [OK] Znaleziono: rocminfo -> /usr/bin/rocminfo [OK] Znaleziono: rocm-smi -> /usr/bin/rocm-smi [OK] Znaleziono: amd-smi -> /usr/bin/amd-smi [OK] Znaleziono: hipconfig -> /usr/bin/hipconfig [OK] Znaleziono: clinfo -> /usr/bin/clinfo $ dpkg -l 2>/dev/null | grep -i rocm || true ii comgr 3.0.0.60401-83~22.04 amd64 Library to provide support functions for ROCm code objects. ii hipblas 2.4.0.60401-83~22.04 amd64 ROCm BLAS marshalling library ii hipblas-dev 2.4.0.60401-83~22.04 amd64 ROCm BLAS marshalling library ii hipfft 1.0.18.60401-83~22.04 amd64 ROCm FFT marshalling library ii hipfft-dev 1.0.18.60401-83~22.04 amd64 ROCm FFT marshalling library ii hipsparse 3.2.0.60401-83~22.04 amd64 ROCm SPARSE library ii hipsparse-dev 3.2.0.60401-83~22.04 amd64 ROCm SPARSE library ii hipsparselt 0.2.3.60401-83~22.04 amd64 ROCm Structured Sparsity Matrix Multiplication marshalling library ii hipsparselt-dev 0.2.3.60401-83~22.04 amd64 ROCm Structured Sparsity Matrix Multiplication marshalling library ii hsa-rocr 1.15.0.60401-83~22.04 amd64 AMD Heterogeneous System Architecture HSA - Linux HSA Runtime for Boltzmann (ROCm) platforms ii rccl 2.22.3.60401-83~22.04 amd64 ROCm Communication Collectives Library ii rccl-dev 2.22.3.60401-83~22.04 amd64 ROCm Communication Collectives Library ii rocalution 3.2.3.60401-83~22.04 amd64 ROCm library for sparse linear systems ii rocalution-dev 3.2.3.60401-83~22.04 amd64 ROCm library for sparse linear systems ii rocblas 4.4.0.60401-83~22.04 amd64 rocBLAS is the AMD library for BLAS in ROCm. Implemented using the HIP language and optimized for AMD GPUs ii rocblas-dev 4.4.0.60401-83~22.04 amd64 rocBLAS is the AMD library for BLAS in ROCm. Implemented using the HIP language and optimized for AMD GPUs ii rocfft 1.0.32.60401-83~22.04 amd64 ROCm FFT library ii rocfft-dev 1.0.32.60401-83~22.04 amd64 ROCm FFT library ii rocm 6.4.1.60401-83~22.04 amd64 Radeon Open Compute (ROCm) software stack meta package ii rocm-cmake 0.14.0.60401-83~22.04 amd64 rocm-cmake built using CMake ii rocm-core 6.4.1.60401-83~22.04 amd64 ROCm Runtime software stack ii rocm-dbgapi 0.77.2.60401-83~22.04 amd64 Library to provide AMD GPU debugger API ii rocm-debug-agent 2.0.4.60401-83~22.04 amd64 Radeon Open Compute Debug Agent (ROCdebug-agent) ii rocm-developer-tools 6.4.1.60401-83~22.04 amd64 Radeon Open Compute (ROCm) Runtime software stack ii rocm-device-libs 1.0.0.60401-83~22.04 amd64 Radeon Open Compute - device libraries ii rocm-gdb 15.2.60401-83~22.04 amd64 ROCgdb ii rocm-hip-libraries 6.4.1.60401-83~22.04 amd64 Radeon Open Compute (ROCm) Runtime software stack ii rocm-hip-runtime 6.4.1.60401-83~22.04 amd64 Radeon Open Compute (ROCm) Runtime software stack ii rocm-hip-runtime-dev 6.4.1.60401-83~22.04 amd64 Radeon Open Compute (ROCm) Runtime software stack ii rocm-hip-sdk 6.4.1.60401-83~22.04 amd64 Radeon Open Compute (ROCm) Runtime software stack ii rocm-language-runtime 6.4.1.60401-83~22.04 amd64 Radeon Open Compute (ROCm) Runtime software stack ii rocm-llvm 19.0.0.25184.60401-83~22.04 amd64 ROCm core compiler ii rocm-ml-libraries 6.4.1.60401-83~22.04 amd64 Radeon Open Compute (ROCm) Runtime software stack ii rocm-ml-sdk 6.4.1.60401-83~22.04 amd64 Radeon Open Compute (ROCm) Runtime software stack ii rocm-opencl 2.0.0.60401-83~22.04 amd64 clr built using CMake ii rocm-opencl-dev 2.0.0.60401-83~22.04 amd64 clr built using CMake ii rocm-opencl-runtime 6.4.1.60401-83~22.04 amd64 Radeon Open Compute (ROCm) Runtime software stack ii rocm-opencl-sdk 6.4.1.60401-83~22.04 amd64 Radeon Open Compute (ROCm) Runtime software stack ii rocm-openmp-sdk 6.4.1.60401-83~22.04 amd64 Radeon Open Compute (ROCm) OpenMP Software development Kit. ii rocm-smi-lib 7.5.0.60401-83~22.04 amd64 AMD System Management libraries ii rocm-utils 6.4.1.60401-83~22.04 amd64 Radeon Open Compute (ROCm) Runtime software stack ii rocminfo 1.0.0.60401-83~22.04 amd64 Radeon Open Compute (ROCm) Runtime rocminfo tool ii rocprofiler-compute 3.1.0.60401-83~22.04 amd64 ROCm Compute Profiler: tool for GPU performance profiling ii rocprofiler-sdk 0.6.0-83~22.04 amd64 ROCm GPU performance analysis SDK ii rocprofiler-sdk-roctx 0.6.0-83~22.04 amd64 ROCm GPU performance analysis SDK ii rocsolver 3.28.0.60401-83~22.04 amd64 AMD ROCm SOLVER library ii rocsolver-dev 3.28.0.60401-83~22.04 amd64 AMD ROCm SOLVER library ii rpp 1.9.10.60401-83~22.04 amd64 ROCm Performance Primitives library is a comprehensive high performance computer vision library for AMD CPUs and GPUs with HOST/HIP/OpenCL back-ends. ii rpp-dev 1.9.10.60401-83~22.04 amd64 ROCm Performance Primitives library is a comprehensive high performance computer vision library for AMD CPUs and GPUs with HOST/HIP/OpenCL back-ends. RPP develop package provides rpp library, header files, and license.txt [exit=0] $ apt list --installed 2>/dev/null | grep -i rocm || true rocm-cmake/jammy,now 0.14.0.60401-83~22.04 amd64 [installed,automatic] rocm-core/jammy,now 6.4.1.60401-83~22.04 amd64 [installed,automatic] rocm-dbgapi/jammy,now 0.77.2.60401-83~22.04 amd64 [installed,automatic] rocm-debug-agent/jammy,now 2.0.4.60401-83~22.04 amd64 [installed,automatic] rocm-developer-tools/jammy,now 6.4.1.60401-83~22.04 amd64 [installed,automatic] rocm-device-libs/jammy,now 1.0.0.60401-83~22.04 amd64 [installed,automatic] rocm-gdb/jammy,now 15.2.60401-83~22.04 amd64 [installed,automatic] rocm-hip-libraries/jammy,now 6.4.1.60401-83~22.04 amd64 [installed,automatic] rocm-hip-runtime-dev/jammy,now 6.4.1.60401-83~22.04 amd64 [installed,automatic] rocm-hip-runtime/jammy,now 6.4.1.60401-83~22.04 amd64 [installed,automatic] rocm-hip-sdk/jammy,now 6.4.1.60401-83~22.04 amd64 [installed,automatic] rocm-language-runtime/jammy,now 6.4.1.60401-83~22.04 amd64 [installed,automatic] rocm-llvm/jammy,now 19.0.0.25184.60401-83~22.04 amd64 [installed,automatic] rocm-ml-libraries/jammy,now 6.4.1.60401-83~22.04 amd64 [installed,automatic] rocm-ml-sdk/jammy,now 6.4.1.60401-83~22.04 amd64 [installed,automatic] rocm-opencl-dev/jammy,now 2.0.0.60401-83~22.04 amd64 [installed,automatic] rocm-opencl-runtime/jammy,now 6.4.1.60401-83~22.04 amd64 [installed,automatic] rocm-opencl-sdk/jammy,now 6.4.1.60401-83~22.04 amd64 [installed,automatic] rocm-opencl/jammy,now 2.0.0.60401-83~22.04 amd64 [installed,automatic] rocm-openmp-sdk/jammy,now 6.4.1.60401-83~22.04 amd64 [installed,automatic] rocm-smi-lib/jammy,now 7.5.0.60401-83~22.04 amd64 [installed,automatic] rocm-utils/jammy,now 6.4.1.60401-83~22.04 amd64 [installed,automatic] rocm/jammy,now 6.4.1.60401-83~22.04 amd64 [installed] rocminfo/jammy,now 1.0.0.60401-83~22.04 amd64 [installed,automatic] [exit=0] ==== Wersja ROCm / HIP ==== $ hipconfig --full || hipconfig || true HIP version: 6.4.43483-a187df25c ==hipconfig HIP_PATH :/opt/rocm-6.4.1 ROCM_PATH :/opt/rocm-6.4.1 HIP_COMPILER :clang HIP_PLATFORM :amd HIP_RUNTIME :rocclr CPP_CONFIG : -D__HIP_PLATFORM_HCC__= -D__HIP_PLATFORM_AMD__= -I/opt/rocm-6.4.1/include -I/include ==hip-clang HIP_CLANG_PATH :/opt/rocm-6.4.1/lib/llvm/bin AMD clang version 19.0.0git (https://github.com/RadeonOpenCompute/llvm-project roc-6.4.1 25184 c87081df219c42dc27c5b6d86c0525bc7d01f727) Target: x86_64-unknown-linux-gnu Thread model: posix InstalledDir: /opt/rocm-6.4.1/lib/llvm/bin Configuration file: /opt/rocm-6.4.1/lib/llvm/bin/clang++.cfg AMD LLVM version 19.0.0git Optimized build. Default target: x86_64-unknown-linux-gnu Host CPU: broadwell Registered Targets: amdgcn - AMD GCN GPUs r600 - AMD GPUs HD2XXX-HD6XXX x86 - 32-bit X86: Pentium-Pro and above x86-64 - 64-bit X86: EM64T and AMD64 hip-clang-cxxflags : -O3 hip-clang-ldflags : --driver-mode=g++ -O3 --hip-link == Environment Variables PATH =/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin:/snap/bin == Linux Kernel Hostname : llama-station Linux llama-station 6.8.0-124-generic #124~22.04.1-Ubuntu SMP PREEMPT_DYNAMIC Tue May 26 21:05:19 UTC x86_64 x86_64 x86_64 GNU/Linux Distributor ID: Ubuntu Description: Ubuntu 22.04.5 LTS Release: 22.04 Codename: jammy [exit=0] $ cat /opt/rocm/.info/version 6.4.1-83 [exit=0] ==== Wykrywanie GPU przez ROCm ==== $ rocminfo | sed -n "1,220p" ROCk module is loaded ===================== HSA System Attributes ===================== Runtime Version: 1.15 Runtime Ext Version: 1.7 System Timestamp Freq.: 1000.000000MHz Sig. Max Wait Duration: 18446744073709551615 (0xFFFFFFFFFFFFFFFF) (timestamp count) Machine Model: LARGE System Endianness: LITTLE Mwaitx: DISABLED XNACK enabled: YES DMAbuf Support: YES VMM Support: NO ========== HSA Agents ========== ******* Agent 1 ******* Name: Intel(R) Xeon(R) CPU E5-2697A v4 @ 2.60GHz Uuid: CPU-XX Marketing Name: Intel(R) Xeon(R) CPU E5-2697A v4 @ 2.60GHz Vendor Name: CPU Feature: None specified Profile: FULL_PROFILE Float Round Mode: NEAR Max Queue Number: 0(0x0) Queue Min Size: 0(0x0) Queue Max Size: 0(0x0) Queue Type: MULTI Node: 0 Device Type: CPU Cache Info: L1: 32768(0x8000) KB Chip ID: 0(0x0) ASIC Revision: 0(0x0) Cacheline Size: 64(0x40) Max Clock Freq. (MHz): 3600 BDFID: 0 Internal Node ID: 0 Compute Unit: 32 SIMDs per CU: 0 Shader Engines: 0 Shader Arrs. per Eng.: 0 WatchPts on Addr. Ranges:1 Memory Properties: Features: None Pool Info: Pool 1 Segment: GLOBAL; FLAGS: FINE GRAINED Size: 65675628(0x3ea216c) KB Allocatable: TRUE Alloc Granule: 4KB Alloc Recommended Granule:4KB Alloc Alignment: 4KB Accessible by all: TRUE Pool 2 Segment: GLOBAL; FLAGS: EXTENDED FINE GRAINED Size: 65675628(0x3ea216c) KB Allocatable: TRUE Alloc Granule: 4KB Alloc Recommended Granule:4KB Alloc Alignment: 4KB Accessible by all: TRUE Pool 3 Segment: GLOBAL; FLAGS: KERNARG, FINE GRAINED Size: 65675628(0x3ea216c) KB Allocatable: TRUE Alloc Granule: 4KB Alloc Recommended Granule:4KB Alloc Alignment: 4KB Accessible by all: TRUE Pool 4 Segment: GLOBAL; FLAGS: COARSE GRAINED Size: 65675628(0x3ea216c) KB Allocatable: TRUE Alloc Granule: 4KB Alloc Recommended Granule:4KB Alloc Alignment: 4KB Accessible by all: TRUE ISA Info: *** Done *** [exit=0] $ rocminfo 2>/dev/null | grep -iE "Name:|Marketing Name|gfx|Agent" || true HSA Agents Agent 1 Name: Intel(R) Xeon(R) CPU E5-2697A v4 @ 2.60GHz Marketing Name: Intel(R) Xeon(R) CPU E5-2697A v4 @ 2.60GHz Vendor Name: CPU [exit=0] $ rocm-smi || true ===================================== ROCm System Management Interface ===================================== =============================================== Concise Info =============================================== Device Node IDs Temp Power Partitions SCLK MCLK Fan Perf PwrCap VRAM% GPU%  (DID, GUID) (Edge) (Avg) (Mem, Compute, ID)  ============================================================================================================ ============================================================================================================ =========================================== End of ROCm SMI Log ============================================ [exit=0] $ amd-smi list || amd-smi static || true AMDSMI Tool: 25.4.2+aca1101 | AMDSMI Library version: 25.4.0 | ROCm version: 6.4.1 AMDSMI Tool: 25.4.2+aca1101 | AMDSMI Library version: 25.4.0 | ROCm version: 6.4.1 [exit=0] ==== OpenCL pomocniczo ==== $ clinfo | sed -n "1,160p" Number of platforms: 1 Platform Profile: FULL_PROFILE Platform Version: OpenCL 2.1 AMD-APP (3649.0) Platform Name: AMD Accelerated Parallel Processing Platform Vendor: Advanced Micro Devices, Inc. Platform Extensions: cl_khr_icd cl_amd_event_callback Platform Name: AMD Accelerated Parallel Processing Number of devices: 0 [exit=0] ==== Zmienne środowiskowe mogące psuć detekcję ==== $ env | grep -E "ROCM|HIP|HSA|CUDA|GGML|LLAMA" | sort || true [exit=0] ==== llama.cpp / binarki ==== $ pwd /home/adminek/llama.cpp [exit=0] $ find . -maxdepth 3 -type f \( -name "llama-cli" -o -name "main" -o -name "llama-server" -o -name "libggml*" \) 2>/dev/null | sort || true ./build/bin/libggml-base.so.0.14.0 ./build/bin/libggml-cpu.so.0.14.0 ./build/bin/libggml-hip.so.0.14.0 ./build/bin/libggml.so.0.14.0 ./build/bin/llama-cli ./build/bin/llama-server [exit=0] $ find . -maxdepth 4 -type f \( -name "CMakeCache.txt" -o -name "build.ninja" \) 2>/dev/null | sort || true ./build/CMakeCache.txt [exit=0] $ grep -RniE "GGML_HIP|GGML_CUDA|GGML_VULKAN|HIPBLAS|ROCM" . 2>/dev/null | head -n 120 || true ./docs/docker.md:22:- `ghcr.io/ggml-org/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`) ./docs/docker.md:23:- `ghcr.io/ggml-org/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`) ./docs/docker.md:24:- `ghcr.io/ggml-org/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`) ./docs/docker.md:41:The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now). ./docs/multi-gpu.md:97:There's no runtime flag for NCCL - it's selected at build time (`-DGGML_CUDA_NCCL=ON`, this is the default). Note that NCCL is **not** automatically distributed with CUDA and you may need to install it manually - when in doubt check the CMake log to see whether or not it can find the package. When llama.cpp is compiled with NCCL support it uses it automatically for cross-GPU reductions in `tensor` mode. When NCCL is missing on a multi-GPU build, you'll see this one-time warning and performance will be lower: ./docs/multi-gpu.md:103:When using the "ROCm" backend (which is the ggml CUDA code translated for AMD via HIP), the AMD equivalent RCCL can be used by compiling with `-DGGML_HIP_RCCL=ON`. Note that RCCL is by default *disabled* because (unlike NCCL) it was not universally beneficial during testing. ./docs/multi-gpu.md:104:### 6. With CUDA peer-to-peer access (`GGML_CUDA_P2P`) ./docs/multi-gpu.md:106:CUDA peer-to-peer (P2P) lets GPUs transfer data directly between each other instead of going through system memory, which generally improves multi-GPU performance. It is **opt-in** at runtime - set the environment variable `GGML_CUDA_P2P` to any value to enable it: ./docs/multi-gpu.md:109:GGML_CUDA_P2P=1 llama-cli -m model.gguf -sm tensor ./docs/multi-gpu.md:127:| Crashes or corrupted outputs after setting `GGML_CUDA_P2P=1` | Some motherboards and BIOS settings (e.g. with IOMMU enabled) don't support CUDA peer-to-peer reliably. Unset `GGML_CUDA_P2P`. | ./docs/build.md:173:cmake -B build -DGGML_CUDA=ON ./docs/build.md:183:cmake -B build -DGGML_CUDA=ON -DGGML_NATIVE=OFF ./docs/build.md:213:cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="86;89" ./docs/build.md:221:cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_COMPILER=/opt/cuda-11.7/bin/nvcc -DCMAKE_INSTALL_RPATH="/opt/cuda-11.7/lib64;\$ORIGIN" -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON ./docs/build.md:273:#### GGML_CUDA_FORCE_CUBLAS_COMPUTE_32F ./docs/build.md:275:Use `GGML_CUDA_FORCE_CUBLAS_COMPUTE_32F` environment variable to use FP32 compute type on all GPUs in FP16 cuBLAS for preventing possible numerical overflows in exchange for slower prompt processing (small impact on RTX PRO/Datacenter products and significant on GeForce products). ./docs/build.md:277:#### GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F ./docs/build.md:279:Use `GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F` environment variable to force use FP16 compute type (instead of default FP32) in FP16 cuBLAS for V100, CDNA and RDNA4. ./docs/build.md:283:The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`. ./docs/build.md:287:The environment variable `GGML_CUDA_P2P` can be set to enable peer-to-peer access between multiple GPUs, allowing them to transfer data directly rather than to go through system memory. ./docs/build.md:297:| GGML_CUDA_FORCE_MMQ | Boolean | false | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. | ./docs/build.md:298:| GGML_CUDA_FORCE_CUBLAS | Boolean | false | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models. There may be issues with numerical overflows (except for V100, CDNA and RDNA4 which use FP32 compute type by default) and memory use will be higher. Prompt processing may become faster on recent datacenter GPUs (the custom kernels were tuned primarily for RTX 3000/4000). | ./docs/build.md:299:| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer | 128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. | ./docs/build.md:300:| GGML_CUDA_FA_ALL_QUANTS | Boolean | false | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer. | ./docs/build.md:350:The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. ./docs/build.md:355:Make sure to have ROCm installed. ./docs/build.md:356:You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html#rocm-install-quick). ./docs/build.md:361: cmake -S . -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \ ./docs/build.md:367: To enhance flash attention performance on RDNA3+ or CDNA architectures, you can utilize the rocWMMA library by enabling the `-DGGML_HIP_ROCWMMA_FATTN=ON` option. This requires rocWMMA headers to be installed on the build system. ./docs/build.md:369: The rocWMMA library is included by default when installing the ROCm SDK using the `rocm` meta package provided by AMD. Alternatively, if you are not using the meta package, you can install the library using the `rocwmma-dev` or `rocwmma-devel` package, depending on your system's package manager. ./docs/build.md:371: As an alternative, you can manually install the library by cloning it from the official [GitHub repository](https://github.com/ROCm/rocWMMA), checkout the corresponding version tag (e.g. `rocm-6.2.4`) and set `-DCMAKE_CXX_FLAGS="-I/library/include/"` in CMake. This also works under Windows despite not officially supported by AMD. ./docs/build.md:375: clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library ./docs/build.md:384: cmake -S . -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \ ./docs/build.md:391: cmake -S . -B build -G Ninja -DGPU_TARGETS=gfx1100 -DGGML_HIP=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release ./docs/build.md:395: Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`. ./docs/build.md:398:The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used. ./docs/build.md:399:If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3. Note that [`HSA_OVERRIDE_GFX_VERSION`] is [not supported on Windows](https://github.com/ROCm/ROCm/issues/2654) ./docs/build.md:403:On Linux it is possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1`. However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs). ./docs/build.md:431:cmake -B build -DGGML_VULKAN=ON ./docs/build.md:448:cmake -B build -DGGML_VULKAN=ON ./docs/build.md:472:cmake -B build -DGGML_VULKAN=ON ./docs/build.md:515:cmake -B build -DGGML_VULKAN=1 ./docs/build.md:525:# You should see in the output, ggml_vulkan detected your GPU. For example: ./docs/build.md:526:# ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32 ./docs/build.md:556:cmake -B build -DGGML_VULKAN=1 -DGGML_METAL=OFF ./docs/build.md:772:In most cases, it is possible to build and use multiple backends at the same time. For example, you can build llama.cpp with both CUDA and Vulkan support by using the `-DGGML_CUDA=ON -DGGML_VULKAN=ON` options with CMake. At runtime, you can specify which backend devices to use with the `--device` option. To see a list of available devices, use the `--list-devices` option. ./docs/multimodal/MobileVLM.md:197:make GGML_CUDA=1 CUDA_DOCKER_ARCH=sm_87 -j 32 ./CMakePresets.json:32: { "name": "vulkan", "hidden": true, "cacheVariables": { "GGML_VULKAN": "ON" } }, ./CMakeLists.txt:144:if (NOT DEFINED GGML_CUDA_GRAPHS) ./CMakeLists.txt:145: set(GGML_CUDA_GRAPHS_DEFAULT ON) ./CMakeLists.txt:161:llama_option_depr(FATAL_ERROR LLAMA_CUBLAS GGML_CUDA) ./CMakeLists.txt:162:llama_option_depr(WARNING LLAMA_CUDA GGML_CUDA) ./.devops/cuda.Dockerfile:30: cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \ ./.devops/nix/nixpkgs-instances.nix:38: # Ensure dependencies use ROCm consistently ./.devops/nix/nixpkgs-instances.nix:39: pkgsRocm = import inputs.nixpkgs { ./.devops/nix/nixpkgs-instances.nix:41: config.rocmSupport = true; ./.devops/nix/sif.nix:20: # ROCm is currently affected by https://github.com/NixOS/nixpkgs/issues/276846 ./.devops/nix/sif.nix:25: diskSize = 4096 + optionalInt llama-cpp.useRocm 16384; ./.devops/nix/package.nix:17: rocmPackages, ./.devops/nix/package.nix:29: useRocm ./.devops/nix/package.nix:37: useRocm ? config.rocmSupport, ./.devops/nix/package.nix:38: rocmGpuTargets ? builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets, ./.devops/nix/package.nix:67: ++ lib.optionals useRocm [ "ROCm" ] ./.devops/nix/package.nix:99: rocmBuildInputs = with rocmPackages; [ ./.devops/nix/package.nix:101: hipblas ./.devops/nix/package.nix:190: ++ optionals useRocm rocmBuildInputs ./.devops/nix/package.nix:203: (cmakeBool "GGML_CUDA" useCuda) ./.devops/nix/package.nix:204: (cmakeBool "GGML_HIP" useRocm) ./.devops/nix/package.nix:206: (cmakeBool "GGML_VULKAN" useVulkan) ./.devops/nix/package.nix:218: ++ optionals useRocm [ ./.devops/nix/package.nix:219: (cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.llvm.clang}/bin/clang") ./.devops/nix/package.nix:220: (cmakeFeature "CMAKE_HIP_ARCHITECTURES" rocmGpuTargets) ./.devops/nix/package.nix:227: # Environment variables needed for ROCm ./.devops/nix/package.nix:228: env = optionalAttrs useRocm { ./.devops/nix/package.nix:229: ROCM_PATH = "${rocmPackages.clr}"; ./.devops/nix/package.nix:230: HIP_DEVICE_LIB_PATH = "${rocmPackages.rocm-device-libs}/amdgcn/bitcode"; ./.devops/llama-cpp-cuda.srpm.spec:35:make -j GGML_CUDA=1 ./.devops/rocm.Dockerfile:4:ARG ROCM_VERSION=7.2.1 ./.devops/rocm.Dockerfile:7:# Target the ROCm build image ./.devops/rocm.Dockerfile:8:ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete ./.devops/rocm.Dockerfile:15:FROM ${BASE_ROCM_DEV_CONTAINER} AS build ./.devops/rocm.Dockerfile:19:# check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-7.2.1/reference/system-requirements.html ./.devops/rocm.Dockerfile:20:# check https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityrad/native_linux/native_linux_compatibility.html ./.devops/rocm.Dockerfile:21:# check https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityryz/native_linux/native_linux_compatibility.html ./.devops/rocm.Dockerfile:23:ARG ROCM_DOCKER_ARCH='gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1150;gfx1200;gfx1201' ./.devops/rocm.Dockerfile:25:# Set ROCm architectures ./.devops/rocm.Dockerfile:26:ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH} ./.devops/rocm.Dockerfile:43: -DGGML_HIP=ON \ ./.devops/rocm.Dockerfile:44: -DGGML_HIP_ROCWMMA_FATTN=ON \ ./.devops/rocm.Dockerfile:45: -DAMDGPU_TARGETS="$ROCM_DOCKER_ARCH" \ ./.devops/rocm.Dockerfile:63:FROM ${BASE_ROCM_DEV_CONTAINER} AS base ./.devops/vulkan.Dockerfile:20:RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \ ./tools/fit-params/README.md:12:ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no ./tools/fit-params/README.md:13:ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no ./tools/fit-params/README.md:14:ggml_cuda_init: found 1 CUDA devices: ./tools/fit-params/README.md:30:ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no ./tools/fit-params/README.md:31:ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no ./tools/fit-params/README.md:32:ggml_cuda_init: found 1 CUDA devices: ./tools/ui/package-lock.json:3226: "integrity": "sha512-dWHzHa2WqEXI/O1E9OjrocMTKJl2mSrEolh1Iomrv6U+JuNwaHXsXx9bLu5gG7BUWFIN0skIQJQ/L1rIex4X6w==", ./tools/rpc/README.md:49:cmake .. -DGGML_CUDA=ON -DGGML_RPC=ON ./tools/rpc/README.md:57:ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no ./tools/rpc/README.md:58:ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no ./tools/rpc/README.md:59:ggml_cuda_init: found 1 CUDA devices: ./.git/packed-refs:448:a710d58d8867c0b494023b799bfd41573f6a439d refs/remotes/origin/ik/try_fix_rocm_k_cache ./ggml/CMakeLists.txt:117:if (NOT GGML_CUDA_GRAPHS_DEFAULT) ./ggml/CMakeLists.txt:118: set(GGML_CUDA_GRAPHS_DEFAULT OFF) ./ggml/CMakeLists.txt:199:option(GGML_CUDA "ggml: use CUDA" OFF) ./ggml/CMakeLists.txt:201:option(GGML_CUDA_FORCE_MMQ "ggml: use mmq kernels instead of cuBLAS" OFF) ./ggml/CMakeLists.txt:202:option(GGML_CUDA_FORCE_CUBLAS "ggml: always use cuBLAS instead of mmq kernels" OFF) ./ggml/CMakeLists.txt:203:set (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING ./ggml/CMakeLists.txt:205:option(GGML_CUDA_NO_PEER_COPY "ggml: do not use peer to peer copies" OFF) ./ggml/CMakeLists.txt:206:option(GGML_CUDA_NO_VMM "ggml: do not try to use CUDA VMM" OFF) ./ggml/CMakeLists.txt:207:option(GGML_CUDA_FA "ggml: compile ggml FlashAttention CUDA kernels" ON) ./ggml/CMakeLists.txt:208:option(GGML_CUDA_FA_ALL_QUANTS "ggml: compile all quants for FlashAttention" OFF) ./ggml/CMakeLists.txt:209:option(GGML_CUDA_GRAPHS "ggml: use CUDA graphs (llama.cpp only)" ${GGML_CUDA_GRAPHS_DEFAULT}) ./ggml/CMakeLists.txt:210:option(GGML_CUDA_NCCL "ggml: use NVIDIA Collective Comm. Library" ON) ./ggml/CMakeLists.txt:211:set (GGML_CUDA_COMPRESSION_MODE "size" CACHE STRING ./ggml/CMakeLists.txt:213:set_property(CACHE GGML_CUDA_COMPRESSION_MODE PROPERTY STRINGS "none;speed;balance;size") ./ggml/CMakeLists.txt:215:option(GGML_HIP "ggml: use HIP" OFF) ./ggml/CMakeLists.txt:216:option(GGML_HIP_GRAPHS "ggml: use HIP graph" ON) [exit=0] ==== Linkowanie binarki llama.cpp (jeśli istnieje) ==== [OK] Znaleziono binarkę: ./build/bin/llama-cli $ ldd './build/bin/llama-cli' || true linux-vdso.so.1 (0x00007fff8ffda000) libllama-cli-impl.so => /home/adminek/llama.cpp/build/bin/libllama-cli-impl.so (0x00007bc7b5483000) libstdc++.so.6 => /lib/x86_64-linux-gnu/libstdc++.so.6 (0x00007bc7b5200000) libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007bc7b544b000) libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007bc7b4e00000) libllama-common.so.0 => /home/adminek/llama.cpp/build/bin/libllama-common.so.0 (0x00007bc7b4800000) libmtmd.so.0 => /home/adminek/llama.cpp/build/bin/libmtmd.so.0 (0x00007bc7b50d3000) libllama.so.0 => /home/adminek/llama.cpp/build/bin/libllama.so.0 (0x00007bc7b4400000) libggml.so.0 => /home/adminek/llama.cpp/build/bin/libggml.so.0 (0x00007bc7b543c000) libggml-base.so.0 => /home/adminek/llama.cpp/build/bin/libggml-base.so.0 (0x00007bc7b4d35000) libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007bc7b4719000) /lib64/ld-linux-x86-64.so.2 (0x00007bc7b55e9000) libggml-cpu.so.0 => /home/adminek/llama.cpp/build/bin/libggml-cpu.so.0 (0x00007bc7b427a000) libggml-hip.so.0 => /home/adminek/llama.cpp/build/bin/libggml-hip.so.0 (0x00007bc7b0e00000) libgomp.so.1 => /lib/x86_64-linux-gnu/libgomp.so.1 (0x00007bc7b5089000) libhipblas.so.2 => /opt/rocm-6.4.1/lib/libhipblas.so.2 (0x00007bc7b0cce000) librocblas.so.4 => /opt/rocm-6.4.1/lib/librocblas.so.4 (0x00007bc7adc00000) libamdhip64.so.6 => /opt/rocm-6.4.1/lib/libamdhip64.so.6 (0x00007bc7ac000000) librocsolver.so.0 => /opt/rocm-6.4.1/lib/librocsolver.so.0 (0x00007bc79a600000) libroctx64.so.4 => /opt/rocm-6.4.1/lib/libroctx64.so.4 (0x00007bc7b5433000) libhipblaslt.so.0 => /opt/rocm-6.4.1/lib/libhipblaslt.so.0 (0x00007bc79a000000) librocprofiler-register.so.0 => /opt/rocm-6.4.1/lib/librocprofiler-register.so.0 (0x00007bc7b4cb2000) libhsa-runtime64.so.1 => /opt/rocm-6.4.1/lib/libhsa-runtime64.so.1 (0x00007bc799c00000) libnuma.so.1 => /lib/x86_64-linux-gnu/libnuma.so.1 (0x00007bc7b507c000) libelf.so.1 => /lib/x86_64-linux-gnu/libelf.so.1 (0x00007bc7b505e000) libdrm.so.2 => /opt/amdgpu/lib/x86_64-linux-gnu/libdrm.so.2 (0x00007bc7b5044000) libdrm_amdgpu.so.1 => /opt/amdgpu/lib/x86_64-linux-gnu/libdrm_amdgpu.so.1 (0x00007bc7b5034000) libz.so.1 => /lib/x86_64-linux-gnu/libz.so.1 (0x00007bc7b4c96000) [exit=0] $ strings './build/bin/llama-cli' | grep -iE 'ROCm|HIP|Vulkan|CUDA|hipblas|ggml' | head -n 80 || true /home/adminek/llama.cpp/build/bin:/opt/rocm-6.4.1/lib:/opt/rocm/lib: [exit=0] $ './build/bin/llama-cli' --version || './build/bin/llama-cli' -h || true [exit=0] ==== Biblioteki ROCm w systemie ==== $ ldconfig -p 2>/dev/null | grep -iE "hip|roc|hsa|blas" | head -n 120 || true libwebrtc_audio_processing.so.1 (libc6,x86-64) => /lib/x86_64-linux-gnu/libwebrtc_audio_processing.so.1 libvxu.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvxu.so.1 libvxu.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvxu.so libvx_rpp.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvx_rpp.so.1 libvx_rpp.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvx_rpp.so libvx_opencv.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvx_opencv.so.1 libvx_opencv.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvx_opencv.so libvx_nn.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvx_nn.so.1 libvx_nn.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvx_nn.so libvx_amd_migraphx.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvx_amd_migraphx.so.1 libvx_amd_migraphx.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvx_amd_migraphx.so libvx_amd_media.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvx_amd_media.so.1 libvx_amd_media.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvx_amd_media.so libvx_amd_custom.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvx_amd_custom.so.1 libvx_amd_custom.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libvx_amd_custom.so librpp.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librpp.so.1 librpp.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librpp.so libroctx64.so.4 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libroctx64.so.4 libroctx64.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libroctx64.so libroctracer64.so.4 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libroctracer64.so.4 libroctracer64.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libroctracer64.so librocsparse.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocsparse.so.1 librocsparse.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocsparse.so librocsolver.so.0 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocsolver.so.0 librocsolver.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocsolver.so librocrand.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocrand.so.1 librocrand.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocrand.so librocprofiler64.so.2 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprofiler64.so.2 librocprofiler64.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprofiler64.so.1 librocprofiler64.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprofiler64.so librocprofiler-sdk.so.0 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprofiler-sdk.so.0 librocprofiler-sdk.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprofiler-sdk.so librocprofiler-sdk-roctx.so.0 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprofiler-sdk-roctx.so.0 librocprofiler-sdk-roctx.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprofiler-sdk-roctx.so librocprofiler-register.so.0 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprofiler-register.so.0 librocprofiler-register.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprofiler-register.so librocprof-sys.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprof-sys.so.1 librocprof-sys.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprof-sys.so librocprof-sys-user.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprof-sys-user.so.1 librocprof-sys-user.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprof-sys-user.so librocprof-sys-rt.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprof-sys-rt.so.1 librocprof-sys-rt.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprof-sys-rt.so librocprof-sys-dl.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprof-sys-dl.so.1 librocprof-sys-dl.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocprof-sys-dl.so librocm_smi64.so.7 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocm_smi64.so.7 librocm_smi64.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocm_smi64.so librocm-debug-agent.so.2 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocm-debug-agent.so.2 librocm-dbgapi.so.0 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocm-dbgapi.so.0 librocm-dbgapi.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocm-dbgapi.so librocm-core.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocm-core.so.1 librocm-core.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocm-core.so librocfft.so.0 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocfft.so.0 librocfft.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocfft.so librocblas.so.4 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocblas.so.4 librocblas.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocblas.so librocalution_hip.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocalution_hip.so.1 librocalution_hip.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocalution_hip.so librocalution.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocalution.so.1 librocalution.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librocalution.so librccl.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/librccl.so.1 librccl.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/librccl.so libprocps.so.8 (libc6,x86-64) => /lib/x86_64-linux-gnu/libprocps.so.8 libpostproc.so.55 (libc6,x86-64) => /lib/x86_64-linux-gnu/libpostproc.so.55 libopenvx.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libopenvx.so.1 libopenvx.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libopenvx.so libopencv_ximgproc.so.4.5d (libc6,x86-64) => /lib/x86_64-linux-gnu/libopencv_ximgproc.so.4.5d libopencv_ximgproc.so (libc6,x86-64) => /lib/x86_64-linux-gnu/libopencv_ximgproc.so libopencv_imgproc.so.4.5d (libc6,x86-64) => /lib/x86_64-linux-gnu/libopencv_imgproc.so.4.5d libopencv_imgproc.so (libc6,x86-64) => /lib/x86_64-linux-gnu/libopencv_imgproc.so liboam.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/liboam.so.1 liboam.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/liboam.so libmigraphx_py_3.10.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libmigraphx_py_3.10.so libmigraphx_py_3.8.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libmigraphx_py_3.8.so libmigraphx_py.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libmigraphx_py.so libmigraphx_c.so.3 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libmigraphx_c.so.3 libmigraphx_c.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libmigraphx_c.so libhsa-runtime64.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhsa-runtime64.so.1 libhsa-runtime64.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhsa-runtime64.so libhsa-amd-aqlprofile64.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhsa-amd-aqlprofile64.so.1 libhsa-amd-aqlprofile64.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhsa-amd-aqlprofile64.so libhiptensor.so.0 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhiptensor.so.0 libhiptensor.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhiptensor.so libhipsparselt.so.0 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhipsparselt.so.0 libhipsparselt.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhipsparselt.so libhipsparse.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhipsparse.so.1 libhipsparse.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhipsparse.so libhipsolver.so.0 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhipsolver.so.0 libhipsolver.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhipsolver.so libhiprtc.so.6 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhiprtc.so.6 libhiprtc.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhiprtc.so libhiprtc-builtins.so.6 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhiprtc-builtins.so.6 libhiprtc-builtins.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhiprtc-builtins.so libhiprand.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhiprand.so.1 libhiprand.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhiprand.so libhipfft.so.0 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhipfft.so.0 libhipfft.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhipfft.so libhipblaslt.so.0 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhipblaslt.so.0 libhipblaslt.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhipblaslt.so libhipblas.so.2 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhipblas.so.2 libhipblas.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libhipblas.so libblas.so.3 (libc6,x86-64) => /lib/x86_64-linux-gnu/libblas.so.3 libamdocl64.so.2 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libamdocl64.so.2 libamdocl64.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libamdocl64.so libamdhip64.so.6 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libamdhip64.so.6 libamdhip64.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libamdhip64.so libamd_smi.so.25 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libamd_smi.so.25 libamd_smi.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libamd_smi.so libamd_comgr.so.3 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libamd_comgr.so.3 libamd_comgr.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libamd_comgr.so libMIOpen.so.1 (libc6,x86-64) => /opt/rocm-6.4.1/lib/libMIOpen.so.1 libMIOpen.so (libc6,x86-64) => /opt/rocm-6.4.1/lib/libMIOpen.so [exit=0] $ find /opt/rocm -maxdepth 3 -type f 2>/dev/null | grep -iE "libamdhip64|libhipblas|libhsa-runtime|librocblas" | sort || true [exit=0] ==== Test dostępu do urządzeń ==== $ test -r /dev/kfd && echo readable:/dev/kfd || echo not-readable:/dev/kfd readable:/dev/kfd [exit=0] $ for x in /dev/dri/renderD*; do [ -e "$x" ] && { test -r "$x" && echo readable:$x || echo not-readable:$x; }; done readable:/dev/dri/renderD128 [exit=0] ==== Wnioski wstępne ==== Interpretacja: - Jeśli brak /dev/kfd albo amdgpu: problem jest na poziomie sterownika/kernela. - Jeśli rocminfo nie pokazuje agenta gfx*: ROCm nie widzi GPU. - Jeśli rocminfo działa, ale llama.cpp nie: problem jest zwykle w buildzie/linkowaniu backendu. - Jeśli użytkownik nie jest w grupie render/video: może być problem z uprawnieniami do /dev/kfd lub /dev/dri. - Jeśli karta to Radeon VII / Pro VII (gfx906): nowsze ROCm może już nie wspierać jej poprawnie i trzeba sprawdzić wersję stacku. Gotowe. Wklej cały output albo plik loga z uruchomienia tego skryptu.