whisper.cpp AMDGPU build

userFeb 23, 01:25:13 PM

Can you build the whisper.cpp project against my amdgpu? I just had success building ollama for the gpu with:
cmake --preset "ROCm 7.2" -B build -DAMDGPU_TARGETS="gfx1030"
cmake --build build --preset "ROCm 7.2"

 Then force a clean Go rebuild (the cache is stale):

 go clean -cache
 go run . serve

assistantFeb 23, 01:25:13 PM

z-ai/glm-5 · openrouter$0.0156

The user wants to build whisper.cpp with AMD GPU (ROCm) support. They've provided a reference for how they built ollama with ROCm 7.2 and targeting gfx1030. Let me first explore the whisper.cpp project to understand its build system and see if it has ROCm/AMD GPU support.

bashcompleted

List project root files

Output

total 124
drwxrwxr-x 1 jm jm   420 Feb 23 13:18 .
drwxr-xr-x 3 jm jm    60 Feb 23 13:23 ..
drwxrwxr-x 1 jm jm   230 Feb 23 13:16 .devops
-rw-rw-r-- 1 jm jm    25 Feb 23 13:16 .dockerignore
drwxrwxr-x 1 jm jm   138 Feb 23 13:23 .git
drwxrwxr-x 1 jm jm    18 Feb 23 13:16 .github
-rw-rw-r-- 1 jm jm   865 Feb 23 13:16 .gitignore
-rw-rw-r-- 1 jm jm 21736 Feb 23 13:16 AUTHORS
-rw-rw-r-- 1 jm jm 10024 Feb 23 13:16 CMakeLists.txt
-rw-rw-r-- 1 jm jm  1078 Feb 23 13:16 LICENSE
-rw-rw-r-- 1 jm jm  2170 Feb 23 13:16 Makefile
-rw-rw-r-- 1 jm jm 36005 Feb 23 13:17 README.md
-rw-rw-r-- 1 jm jm  6593 Feb 23 13:16 README_sycl.md
drwxrwxr-x 1 jm jm    68 Feb 23 13:16 bindings
drwxrwxr-x 1 jm jm   386 Feb 23 13:18 build
-rwxrwxr-x 1 jm jm 23200 Feb 23 13:16 build-xcframework.sh
drwxrwxr-x 1 jm jm    30 Feb 23 13:16 ci
-rw-rw-r-- 1 jm jm   939 Feb 23 13:16 close-issue.yml
drwxrwxr-x 1 jm jm   426 Feb 23 13:16 cmake
drwxrwxr-x 1 jm jm  1050 Feb 23 13:16 examples
drwxrwxr-x 1 jm jm    78 Feb 23 13:16 ggml
drwxrwxr-x 1 jm jm    70 Feb 23 13:16 grammars
drwxrwxr-x 1 jm jm    18 Feb 23 13:16 include
drwxrwxr-x 1 jm jm  1340 Feb 23 13:20 models
drwxrwxr-x 1 jm jm    66 Feb 23 13:16 samples
drwxrwxr-x 1 jm jm   398 Feb 23 13:16 scripts
drwxrwxr-x 1 jm jm   106 Feb 23 13:16 src
drwxrwxr-x 1 jm jm   314 Feb 23 13:16 tests

globcompleted

Output

/home/jm/data/code/ai/whisper.cpp/tests/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/src/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-zendnn/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-zdnn/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-webgpu/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-vulkan/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-virtgpu/backend/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-virtgpu/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-sycl/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-rpc/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-opencl/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-musa/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-metal/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-hip/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-hexagon/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-hexagon/htp/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-cuda/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-cpu/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-cann/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-blas/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/src/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/ggml/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/examples/whisper.wasm/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/examples/whisper.android.java/app/src/main/jni/whisper/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/examples/wchess/CMakeLists.txt
/home/jm/data/code/ai/whisper.cpp/examples/wchess/wchess.cmd/CMakeLists.txt
/home/jm/data/code/ai/whis...

14,917in / 129out / 70reasoning$0.0156

assistantFeb 23, 01:25:46 PM

z-ai/glm-5 · openrouter$0.0175

I see there's a `ggml-hip` directory which is the ROCm/AMD GPU backend. Let me check the main CMakeLists.txt and the ggml-hip CMakeLists.txt to understand how to build with ROCm support.

readcompleted

CMakeLists.txt

Output

<path>/home/jm/data/code/ai/whisper.cpp/CMakeLists.txt</path>
<type>file</type>
<content>1: cmake_minimum_required(VERSION 3.5) # for add_link_options and implicit target directories.
2: project("whisper.cpp" C CXX)
3: project("whisper.cpp" VERSION 1.8.3)
4: include(CheckIncludeFileCXX)
5: 
6: set(SOVERSION 1)
7: 
8: #set(CMAKE_WARN_DEPRECATED YES)
9: set(CMAKE_WARN_UNUSED_CLI YES)
10: 
11: set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
12: 
13: if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
14:     set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
15:     set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
16: endif()
17: 
18: # Add path to modules
19: list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
20: 
21: set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
22: 
23: if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
24:     set(WHISPER_STANDALONE ON)
25: 
26:     include(git-vars)
27: 
28:     # configure project version
29:     configure_file(${CMAKE_SOURCE_DIR}/bindings/javascript/package-tmpl.json ${CMAKE_SOURCE_DIR}/bindings/javascript/package.json @ONLY)
30: else()
31:     set(WHISPER_STANDALONE OFF)
32: endif()
33: 
34: if (EMSCRIPTEN)
35:     set(BUILD_SHARED_LIBS_DEFAULT OFF)
36: 
37:     set(CMAKE_CXX_STANDARD 17)
38:     set(CMAKE_CXX_STANDARD_REQUIRED ON)
39: 
40:     option(WHISPER_WASM_SINGLE_FILE "whisper: embed WASM inside the generated whisper.js" ON)
41: 
42:     # TODO: without these, we get the following error:
43:     #       wasm-ld: error: --shared-memory is disallowed by whisper.cpp.o because it was not compiled with 'atomics' or 'bulk-memory' features.
44:     set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread")
45:     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread")
46: 
47:     set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -s TOTAL_STACK=5242880")
48:     set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -s TOTAL_STACK=5242880")...

readcompleted

ggml/src/ggml-hip/CMakeLists.txt

Output

<path>/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-hip/CMakeLists.txt</path>
<type>file</type>
<content>1: if (NOT EXISTS $ENV{ROCM_PATH})
2:     if (NOT EXISTS /opt/rocm)
3:         set(ROCM_PATH /usr)
4:     else()
5:         set(ROCM_PATH /opt/rocm)
6:     endif()
7: else()
8:     set(ROCM_PATH $ENV{ROCM_PATH})
9: endif()
10: 
11: list(APPEND CMAKE_PREFIX_PATH  ${ROCM_PATH})
12: list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
13: 
14: # CMake on Windows doesn't support the HIP language yet
15: if (WIN32)
16:     set(CXX_IS_HIPCC TRUE)
17: else()
18:     string(REGEX MATCH "hipcc(\.bat)?$" CXX_IS_HIPCC "${CMAKE_CXX_COMPILER}")
19: endif()
20: 
21: if (CXX_IS_HIPCC)
22:     if (LINUX)
23:         if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
24:             message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
25:         endif()
26: 
27:         message(WARNING "Setting hipcc as the C++ compiler is legacy behavior."
28:                 " Prefer setting the HIP compiler directly. See README for details.")
29:     endif()
30: else()
31:     # Forward (AMD)GPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
32:     if(AMDGPU_TARGETS AND NOT GPU_TARGETS)
33:         set(GPU_TARGETS ${AMDGPU_TARGETS})
34:     endif()
35:     if(GPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
36:         set(CMAKE_HIP_ARCHITECTURES ${GPU_TARGETS})
37:     endif()
38:     cmake_minimum_required(VERSION 3.21)
39:     enable_language(HIP)
40: endif()
41: 
42: find_package(hip     REQUIRED)
43: find_package(hipblas REQUIRED)
44: find_package(rocblas REQUIRED)
45: 
46: if (${hip_VERSION} VERSION_LESS 6.1)
47:     message(FATAL_ERROR "At least ROCM/HIP V6.1 is required")
48: endif()
49: 
50: message(STATUS "HIP and hipBLAS found")
51: 
52: # Workaround old compilers
53: set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} --gpu-max-threads-per-block=1024")
54: 
55: file(GLOB   GGML_HEADERS_ROCM "../ggml-cuda/*.cuh")
56: list(APPEND GGML_HEADERS_ROCM "../../include/ggml-...

16,947in / 107out / 47reasoning$0.0175

assistantFeb 23, 01:25:54 PM

z-ai/glm-5 · openrouter$0.0229

Good, I can see that whisper.cpp has HIP/ROCm support via the `ggml-hip` backend. Looking at the ggml/src/ggml-hip/CMakeLists.txt, I can see: 1. It looks for ROCM_PATH (defaulting to /opt/rocm or /usr) 2. It supports AMDGPU_TARGETS/GPU_TARGETS which gets forwarded to CMAKE_HIP_ARCHITECTURES 3. It requires ROCm/HIP 6.1 or higher Let me check the main ggml CMakeLists.txt to see what options enable HIP:

grepcompleted

GGML_HIP

Output

Found 15 matches
/home/jm/data/code/ai/whisper.cpp/ggml/src/ggml-hip/CMakeLists.txt:
  Line 105: if (GGML_HIP_GRAPHS)
  Line 106:     add_compile_definitions(GGML_HIP_GRAPHS)
  Line 109: if (GGML_HIP_NO_VMM)
  Line 110:     add_compile_definitions(GGML_HIP_NO_VMM)
  Line 113: if (GGML_HIP_ROCWMMA_FATTN)
  Line 114:     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN)
  Line 117: if (NOT GGML_HIP_MMQ_MFMA)
  Line 118:     add_compile_definitions(GGML_HIP_NO_MMQ_MFMA)
  Line 121: if (GGML_HIP_EXPORT_METRICS)

/home/jm/data/code/ai/whisper.cpp/ggml/CMakeLists.txt:
  Line 210: option(GGML_HIP                             "ggml: use HIP"                                   OFF)
  Line 211: option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
  Line 212: option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
  Line 213: option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
  Line 214: option(GGML_HIP_MMQ_MFMA                    "ggml: enable MFMA MMA for CDNA in MMQ"           ON)
  Line 215: option(GGML_HIP_EXPORT_METRICS              "ggml: enable kernel perf metrics output"         OFF)

readcompleted

ggml/CMakeLists.txt

Output

<path>/home/jm/data/code/ai/whisper.cpp/ggml/CMakeLists.txt</path>
<type>file</type>
<content>1: cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit target directories.
2: project("ggml" C CXX ASM)
3: 
4: ### GGML Version
5: set(GGML_VERSION_MAJOR 0)
6: set(GGML_VERSION_MINOR 9)
7: set(GGML_VERSION_PATCH 6)
8: set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
9: 
10: find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
11: if(GIT_EXE)
12:     # Get current git commit hash
13:     execute_process(COMMAND ${GIT_EXE} rev-parse --short HEAD
14:         WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
15:         OUTPUT_VARIABLE GGML_BUILD_COMMIT
16:         OUTPUT_STRIP_TRAILING_WHITESPACE
17:         ERROR_QUIET
18:     )
19: 
20:     # Check if the working directory is dirty (i.e., has uncommitted changes)
21:     execute_process(COMMAND ${GIT_EXE} diff-index --quiet HEAD -- .
22:         WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
23:         RESULT_VARIABLE GGML_GIT_DIRTY
24:         ERROR_QUIET
25:     )
26: endif()
27: 
28: set(GGML_VERSION "${GGML_VERSION_BASE}")
29: 
30: if(NOT GGML_BUILD_COMMIT)
31:     set(GGML_BUILD_COMMIT "unknown")
32: endif()
33: 
34: # Build the commit string with optional dirty flag
35: if(DEFINED GGML_GIT_DIRTY AND GGML_GIT_DIRTY EQUAL 1)
36:     set(GGML_BUILD_COMMIT "${GGML_BUILD_COMMIT}-dirty")
37: endif()
38: 
39: include(CheckIncludeFileCXX)
40: 
41: set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
42: 
43: if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
44:     set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
45:     set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
46: endif()
47: 
48: if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
49:     set(GGML_STANDALONE ON)
50: 
51:     set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
52: 
53:     # configure project version
54:     # ...

21,898in / 171out / 106reasoning$0.0229

assistantFeb 23, 01:26:19 PM

z-ai/glm-5 · openrouter$0.0305

I found the relevant option: `GGML_HIP` is the option to enable HIP/ROCm support. Looking at the ggml-hip CMakeLists.txt, I can see that: 1. AMDGPU_TARGETS gets forwarded to GPU_TARGETS, which then gets forwarded to CMAKE_HIP_ARCHITECTURES 2. ROCm/HIP 6.1 is required So to build whisper.cpp with AMD GPU support using ROCm, the command should be: ```bash cmake -B build -DGGML_HIP=ON -DAMDGPU_TARGETS="gfx1030" cmake --build build ``` Let me clean the existing build directory first and then run the build.

I found it - the option is GGML_HIP=ON to enable ROCm/HIP support. Let me clean and rebuild:

bashcompleted

Clean and recreate build directory

29,333in / 202out / 138reasoning$0.0305

assistantFeb 23, 01:26:27 PM

z-ai/glm-5 · openrouter$0.0065

Good, now let me configure the build with HIP support and the user's GPU target.

bashcompleted

Configure with HIP/ROCm support

Output

CMake Deprecation Warning at CMakeLists.txt:1 (cmake_minimum_required):
  Compatibility with CMake < 3.10 will be removed from a future version of
  CMake.

  Update the VERSION argument <min> value.  Or, use the <min>...<max> syntax
  to tell CMake that the project requires at least <min> but has been updated
  to work with policies introduced by <max> or earlier.


-- The C compiler identification is GNU 15.2.0
-- The CXX compiler identification is GNU 15.2.0
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found Git: /usr/bin/git (found version "2.51.0")
-- The ASM compiler identification is GNU
-- Found assembler: /usr/bin/cc
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Success
-- Found Threads: TRUE
-- ccache found, compilation results will be cached. Disable with GGML_CCACHE=OFF.
-- CMAKE_SYSTEM_PROCESSOR: x86_64
-- GGML_SYSTEM_ARCH: x86
-- Including CPU backend
-- Found OpenMP_C: -fopenmp (found version "4.5")
-- Found OpenMP_CXX: -fopenmp (found version "4.5")
-- Found OpenMP: TRUE (found version "4.5")
-- x86 detected
-- Adding CPU backend variant ggml-cpu: -march=native 
-- The HIP compiler identification is unknown
CMake Error at /usr/share/cmake-4.2/Modules/CMakeDetermineHIPCompiler.cmake:197 (message):
  Failed to find ROCm root directory.
Call Stack (most recent call first):
  ggml/src/ggml-hip/CMakeLists.txt:39 (enable_language)


-- Configuring incomplete, errors occurred!

228in / 74out / 20reasoning$0.0065

assistantFeb 23, 01:26:34 PM

z-ai/glm-5 · openrouter$0.0131

It failed to find the ROCm root directory. Let me check if ROCm is installed and where:

bashcompleted

Find ROCm installation

8,224in / 100out / 22reasoning$0.0131

assistantFeb 23, 01:26:40 PM

z-ai/glm-5 · openrouter$0.0072

It seems ROCm might not be installed, or installed in a non-standard location. Let me check more locations:

bashcompleted

Check for HIP/ROCm tools

Output

/usr/bin/bash: line 1: hipcc: command not found
/usr/bin/bash: line 1: rocminfo: command not found

905in / 74out / 27reasoning$0.0072

assistantFeb 23, 01:26:45 PM

z-ai/glm-5 · openrouter

userFeb 23, 01:27:15 PM

Sorry, you are in a sandbox. Do you want to give me commands and I can paste back responses?

assistantFeb 23, 01:27:15 PM

z-ai/glm-5 · openrouter$0.0184

The user is saying I'm in a sandbox and can't access the ROCm installation directly. They're offering to run commands for me and paste back the results. I should give them the commands to check their ROCm installation so I can figure out the right cmake configuration.

Yes, please run these commands:

# Find ROCm installation
ls -la /opt/rocm 2>/dev/null || echo "no /opt/rocm"

# Check for HIP
hipcc --version 2>&1

# Check environment
echo "ROCM_PATH=$ROCM_PATH"
echo "HIP_PATH=$HIP_PATH"

# Check what cmake presets exist
cmake --list-presets

14,678in / 146out / 67reasoning$0.0184

userFeb 23, 01:28:14 PM

y[whisper.cp ❯ # Find ROCm installation master
ls -la /opt/rocm 2>/dev/null || echo "no /opt/rocm"
# Check for HIP
hipcc --version 2>&1
# Check environment
echo "ROCM_PATH=$ROCM_PATH"
echo "HIP_PATH=$HIP_PATH"
# Check what cmake presets exist
cmake --list-preset
lrwxrwxrwx 1 root root 22 Feb 11 21:40 /opt/rocm -> /etc/alternatives/rocm/
HIP version: 7.2.26015-fc0010cf6a
AMD clang version 22.0.0git (https://github.com/RadeonOpenCompute/llvm-project roc-7.2.0 26014 7b800a19466229b8479a78de19143dc33c3ab9b5)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /opt/rocm-7.2.0/lib/llvm/bin
Configuration file: /opt/rocm-7.2.0/lib/llvm/bin/clang++.cfg
ROCM_PATH=
HIP_PATH=
CMake Error: Unknown argument --list-preset
CMake Error: Run 'cmake --help' for all supported options.
[whisper.cp ❯ cmake --list-preset master
CMake Error: Unknown argument --list-preset
CMake Error: Run 'cmake --help' for all supported options.
[whisper.cp ❯ cmake --list-presets master
CMake Error: Could not read presets from /home/jm/data/code/ai/whisper.cpp:
File not found: /home/jm/data/code/ai/whisper.cpp/CMakePresets.json

assistantFeb 23, 01:28:14 PM

z-ai/glm-5 · openrouter$0.0071

Good, ROCm is installed at /opt/rocm-7.2.0. I need to set ROCM_PATH and use the HIP compiler. Let me provide the user with the correct cmake configure command.

ROCm 7.2 is installed. Run this:

# Clean and configure
rm -rf build && mkdir build
ROCM_PATH=/opt/rocm-7.2.0 cmake -B build -DGGML_HIP=ON -DAMDGPU_TARGETS="gfx1030"

Then if that succeeds:

cmake --build build -j$(nproc)

517in / 132out / 40reasoning$0.0071

userFeb 23, 01:29:05 PM

...
[ 84%] Building HIP object ggml/src/ggml-hip/CMakeFiles/ggml-hip.dir//ggml-cuda/template-instances/mmf-instance-ncols_8.cu.o
[ 84%] Building HIP object ggml/src/ggml-hip/CMakeFiles/ggml-hip.dir//ggml-cuda/template-instances/mmf-instance-ncols_9.cu.o
[ 85%] Building HIP object ggml/src/ggml-hip/CMakeFiles/ggml-hip.dir//ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_0.cu.o
[ 85%] Building HIP object ggml/src/ggml-hip/CMakeFiles/ggml-hip.dir//ggml-cuda/template-instances/fattn-vec-instance-q8_0-q8_0.cu.o
[ 86%] Building HIP object ggml/src/ggml-hip/CMakeFiles/ggml-hip.dir/__/ggml-cuda/template-instances/fattn-vec-instance-f16-f16.cu.o
[ 87%] Linking HIP shared library libggml-hip.so
[ 87%] Built target ggml-hip
[ 88%] Building CXX object ggml/src/CMakeFiles/ggml.dir/ggml-backend-reg.cpp.o
[ 88%] Building CXX object ggml/src/CMakeFiles/ggml.dir/ggml-backend-dl.cpp.o
[ 89%] Linking CXX shared library libggml.so
[ 89%] Built target ggml
[ 89%] Building CXX object src/CMakeFiles/whisper.dir/whisper.cpp.o
[ 90%] Linking CXX shared library libwhisper.so
[ 90%] Built target whisper
[ 90%] Building CXX object examples/CMakeFiles/common.dir/common.cpp.o
[ 90%] Building CXX object examples/CMakeFiles/common.dir/common-whisper.cpp.o
[ 92%] Building CXX object examples/CMakeFiles/common.dir/common-ggml.cpp.o
[ 92%] Building CXX object examples/CMakeFiles/common.dir/grammar-parser.cpp.o
[ 93%] Building CXX object examples/bench/CMakeFiles/whisper-bench.dir/bench.cpp.o
[ 93%] Linking CXX executable ../../bin/whisper-bench
[ 93%] Linking CXX static library libcommon.a
[ 93%] Built target whisper-bench
[ 93%] Built target common
[ 94%] Building CXX object tests/CMakeFiles/test-vad-full.dir/test-vad-full.cpp.o
[ 94%] Building CXX object tests/CMakeFiles/test-vad.dir/test-vad.cpp.o
[ 94%] Building CXX object examples/cli/CMakeFiles/whisper-cli.dir/cli.cpp.o
[ 96%] Building CXX object examples/server/CMakeFiles/whisper-server.dir/server.cpp.o
[ 96%] Building CXX object examples/quantize/CMakeFiles/whisper-quantize.dir/quantize.cpp.o
[ 97%] Building CXX object examples/vad-speech-segments/CMakeFiles/whisper-vad-speech-segments.dir/speech.cpp.o
[ 99%] Linking CXX executable ../bin/test-vad-full
[ 99%] Linking CXX executable ../bin/test-vad
[100%] Linking CXX executable ../../bin/whisper-vad-speech-segments
[100%] Built target test-vad-full
[100%] Built target test-vad
[100%] Linking CXX executable ../../bin/whisper-cli
[100%] Linking CXX executable ../../bin/whisper-quantize
[100%] Linking CXX executable ../../bin/whisper-server
[100%] Built target whisper-vad-speech-segments
[100%] Built target whisper-cli
[100%] Built target whisper-quantize
[100%] Built target whisper-server
[whisper.cp ❯ master

assistantFeb 23, 01:29:05 PM

z-ai/glm-5 · openrouter$0.0075

The build completed successfully. Now the user can test it. Let me give them a command to verify the HIP support is working.

Build succeeded. Test it:

# Check that HIP is detected
./build/bin/whisper-cli --help | grep -i gpu

# Quick test with your model
./build/bin/whisper-cli -m models/ggml-small.en.bin -f your_audio.wav

995in / 88out / 31reasoning$0.0075

userFeb 23, 01:35:46 PM

Can you help me run personaplex 7b model of nvidias with the build I have?

assistantFeb 23, 01:35:46 PM

z-ai/glm-5 · openrouter$0.0326

The user wants to run a "personaplex 7b" model with the whisper.cpp build they just made. But there's a fundamental issue here: 1. whisper.cpp is for speech-to-text (transcription) using OpenAI's Whisper models 2. "personaplex 7b" sounds like a large language model (LLM), probably a 7 billion parameter model These are two completely different things: - whisper.cpp runs Whisper models for audio transcription - LLMs like "personaplex 7b" would need llama.cpp or similar I should clarify this with the user - they may be confused about what whisper.cpp does, or they may be asking about something else entirely. Let me ask for clarification.

whisper.cpp is for speech-to-text transcription. "Personaplex 7b" sounds like a text LLM (7 billion parameter language model).

Those are different tools:

whisper.cpp → audio transcription (uses Whisper models)
llama.cpp → text generation (uses LLMs like Personaplex 7b)

Did you mean:

Run audio transcription with a Whisper model?
Or run the Personaplex 7b LLM with a different tool (llama.cpp)?

If it's #2, do you have llama.cpp built with ROCm?

31,058in / 281out / 170reasoning$0.0326

userFeb 23, 01:36:26 PM

Run audio transcription, speech to text ( to voice commands)

assistantFeb 23, 01:36:26 PM

z-ai/glm-5 · openrouter$0.0322

The user wants to run audio transcription with whisper.cpp for voice commands. They have successfully built whisper.cpp with HIP/ROCm support. Now they need to: 1. Download a Whisper model 2. Have an audio file to transcribe 3. Run the transcription Let me help them get started.

Run these to get set up:

# Check available models
ls -la models/

# If empty, download a model (base.en is fast for voice commands)
bash ./models/download-ggml-model.sh base.en

# Test with a sample audio file
./build/bin/whisper-cli -m models/ggml-base.en.bin -f /path/to/your/audio.wav

For real-time voice commands, there's also a streaming example:

./build/bin/whisper-stream -m models/ggml-base.en.bin --step 500 --length 5000

Do you have an audio file to test with?

31,267in / 202out / 75reasoning$0.0322

userFeb 23, 01:36:53 PM

samples/jfk.wav

assistantFeb 23, 01:36:53 PM

z-ai/glm-5 · openrouter$0.0069

The user wants to test with the sample audio file samples/jfk.wav. Let me give them the command to run.

./build/bin/whisper-cli -m models/ggml-base.en.bin -f samples/jfk.wav

If you don't have a model yet:

bash ./models/download-ggml-model.sh base.en
./build/bin/whisper-cli -m models/ggml-base.en.bin -f samples/jfk.wav

180in / 102out / 26reasoning$0.0069

userFeb 23, 01:37:52 PM

[whisper.cp ❯ ./build/bin/whisper-cli -m models/ggml-base.en.bin -f samples/jfk.wav master ?14
whisper_init_from_file_with_params_no_state: loading model from 'models/ggml-base.en.bin'
whisper_init_with_params_no_state: use gpu = 1
whisper_init_with_params_no_state: flash attn = 1
whisper_init_with_params_no_state: gpu_device = 0
whisper_init_with_params_no_state: dtw = 0
ggml_cuda_init: found 2 ROCm devices:
Device 0: AMD Radeon RX 6600 XT, gfx1030 (0x1030), VMM: no, Wave Size: 32
Device 1: AMD Radeon Graphics, gfx1030 (0x1030), VMM: no, Wave Size: 32
whisper_init_with_params_no_state: devices = 3
whisper_init_with_params_no_state: backends = 2
whisper_model_load: loading model
whisper_model_load: n_vocab = 51864
whisper_model_load: n_audio_ctx = 1500
whisper_model_load: n_audio_state = 512
whisper_model_load: n_audio_head = 8
whisper_model_load: n_audio_layer = 6
whisper_model_load: n_text_ctx = 448
whisper_model_load: n_text_state = 512
whisper_model_load: n_text_head = 8
whisper_model_load: n_text_layer = 6
whisper_model_load: n_mels = 80
whisper_model_load: ftype = 1
whisper_model_load: qntvr = 0
whisper_model_load: type = 2 (base)
whisper_model_load: adding 1607 extra tokens
whisper_model_load: n_langs = 99
whisper_model_load: ROCm0 total size = 147.37 MB
whisper_model_load: model size = 147.37 MB
whisper_backend_init_gpu: device 0: ROCm0 (type: 1)
whisper_backend_init_gpu: found GPU device 0: ROCm0 (type: 1, cnt: 0)
whisper_backend_init_gpu: using ROCm0 backend
whisper_init_state: kv self size = 6.29 MB
whisper_init_state: kv cross size = 18.87 MB
whisper_init_state: kv pad size = 3.15 MB
whisper_init_state: compute buffer (conv) = 17.24 MB
whisper_init_state: compute buffer (encode) = 23.09 MB
whisper_init_state: compute buffer (cross) = 4.66 MB
whisper_init_state: compute buffer (decode) = 97.29 MB

main: processing 'samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 processors, 5 beams + best of 5, lang = en, task = transcribe, timestamps = 1 ...

[00:00:00.000 --> 00:00:11.000] And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.

whisper_print_timings: load time = 67.63 ms
whisper_print_timings: fallbacks = 0 p / 0 h
whisper_print_timings: mel time = 7.93 ms
whisper_print_timings: sample time = 23.12 ms / 133 runs ( 0.17 ms per run)
whisper_print_timings: encode time = 65.00 ms / 1 runs ( 65.00 ms per run)
whisper_print_timings: decode time = 19.37 ms / 3 runs ( 6.46 ms per run)
whisper_print_timings: batchd time = 75.58 ms / 126 runs ( 0.60 ms per run)
whisper_print_timings: prompt time = 0.00 ms / 1 runs ( 0.00 ms per run)
whisper_print_timings: total time = 266.87 ms

assistantFeb 23, 01:37:52 PM

z-ai/glm-5 · openrouter$0.0089

The build is working perfectly with ROCm. The user has successfully: 1. Built whisper.cpp with HIP/ROCm support for their AMD GPU (RX 6600 XT, gfx1030) 2. Run transcription on the sample jfk.wav file 3. The GPU was detected and used for acceleration The transcription completed successfully - 11 seconds of audio transcribed in 267ms total, with the encode happening on the GPU in 65ms. This is great performance. The user originally asked about voice commands. They might want to: 1. Use the streaming example for real-time voice input 2. Integrate this into an application I should let them know it's working and briefly mention options for real-time use.

Works perfectly. Your RX 6600 XT is being used (65ms encode time for 11s audio).

For real-time voice commands, try the stream example with a microphone:

./build/bin/whisper-stream -m models/ggml-base.en.bin --step 500 --length 3000 -t 4

1,243in / 225out / 174reasoning$0.0089

userFeb 23, 01:38:28 PM

How can I run it against the audio in from my audio device (scarlett 2i2)?

assistantFeb 23, 01:38:28 PM

z-ai/glm-5 · openrouter

The user wants to run whisper.cpp with real-time audio from their Scarlett 2i2 audio interface. The whisper-stream example should be able to capture from the microphone. Let me check what options are available for the stream example.

userFeb 23, 01:40:06 PM

	PersonaPlex	whisper.cpp
Type	Speech-to-speech conversation	Speech-to-text transcription
I/O	Audio in → Audio out	Audio in → Text out
Architecture	Moshi/Mimi	Whisper
Format	Safetensors (PyTorch)	ggml (C/C++)

Metric	whisper.cpp	voxtral.cpp
Total time (JFK sample)	708.74 ms	15,001 ms (21× slower)
Encode time	484.97 ms	5,503.2 ms
Decode time	7.06 ms	5,110.9 ms
Model size	148MB (base.en)	2.4GB (Q4_0)
GPU acceleration	ROCm HIP (65ms encode)	CPU only (no GPU)
Parameters	~74M	~4B (54× larger)

Use Case	Recommendation	Reason
Real-time voice commands	whisper.cpp	Faster (~65ms GPU encode), lower latency
High-accuracy transcription	voxtral.cpp	Larger model, better accuracy
Live streaming	Both (test both)	Voxtral designed for real-time latency
CPU-only environment	whisper.cpp	Smaller model, faster on CPU

whisper.cpp AMDGPU build

Quick Start: Create the Plugin

Summary

✅ Completed

🛠️ Key Files Created

🎯 Ready for Production

📋 Next Steps (Outside Sandbox)

Why Not Needed

Current GPU Options

Optimal Build Command

When Separate Build Might Help

🚀 Performance Comparison

🔍 Why the Difference?

1. Model Size & Architecture

2. GPU Acceleration

3. Backend Optimization

⚡ Fair Comparison Needed

Test voxtral with GPU:

Rebuild with optimizations:

🎯 Use Case Considerations

📊 Expected Performance After Optimization

🎚️ Trade-off Summary