ROCm Port (#1087)

* use hipblas based on cublas * Update Makefile for the Cuda kernels * Expand arch list and make it overrideable * Fix multi GPU on multiple amd architectures with rocblas_initialize() (#5) * add hipBLAS to README * new build arg LLAMA_CUDA_MMQ_Y * fix half2 decomposition * Add intrinsics polyfills for AMD * AMD assembly optimized __dp4a * Allow overriding CC_TURING * use "ROCm" instead of "CUDA" * ignore all build dirs * Add Dockerfiles * fix llama-bench * fix -nommq help for non CUDA/HIP --------- Co-authored-by: YellowRoseCx <80486540+YellowRoseCx@users.noreply.github.com> Co-authored-by: ardfork <134447697+ardfork@users.noreply.github.com> Co-authored-by: funnbot <22226942+funnbot@users.noreply.github.com> Co-authored-by: Engininja2 <139037756+Engininja2@users.noreply.github.com> Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> Co-authored-by: jammm <2500920+jammm@users.noreply.github.com> Co-authored-by: jdecourval <7315817+jdecourval@users.noreply.github.com>
2024-12-24 02:14:35 +00:00 · 2023-08-25 12:09:42 +03:00 · 2023-08-25 12:09:42 +03:00 · 6bbc598a63
commit 6bbc598a63
parent 3f460a2b72
12 changed files with 335 additions and 59 deletions
--- a/.devops/full-rocm.Dockerfile
+++ b/.devops/full-rocm.Dockerfile
@ -0,0 +1,44 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt requirements.txt
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV LLAMA_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+RUN make
+
+ENTRYPOINT ["/app/.devops/tools.sh"]
--- a/.devops/main-rocm.Dockerfile
+++ b/.devops/main-rocm.Dockerfile
@ -0,0 +1,44 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt requirements.txt
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV LLAMA_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+RUN make
+
+ENTRYPOINT [ "/app/main" ]
--- a/.dockerignore
+++ b/.dockerignore
@ -5,14 +5,7 @@
 .vscode/
 .DS_Store

-build/
-build-em/
-build-debug/
-build-release/
-build-static/
-build-no-accel/
-build-sanitize-addr/
-build-sanitize-thread/
+build*/

 models/*

--- a/.gitignore
+++ b/.gitignore
@ -16,20 +16,7 @@
 .vs/
 .vscode/

-build/
-build-em/
-build-debug/
-build-release/
-build-ci-debug/
-build-ci-release/
-build-static/
-build-cublas/
-build-opencl/
-build-metal/
-build-mpi/
-build-no-accel/
-build-sanitize-addr/
-build-sanitize-thread/
+build*/
 out/
 tmp/

--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -74,6 +74,7 @@ set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kern
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
 option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some calculations"   OFF)
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
+option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
@ -352,6 +353,43 @@ if (LLAMA_CLBLAST)
    endif()
 endif()

+if (LLAMA_HIPBLAS)
+    list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
+
+    if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
+    endif()
+    if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+    endif()
+
+    find_package(hip)
+    find_package(hipblas)
+    find_package(rocblas)
+
+    if (${hipblas_FOUND} AND ${hip_FOUND})
+        message(STATUS "HIP and hipBLAS found")
+        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
+        if (LLAMA_CUDA_FORCE_DMMV)
+            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
+        endif()
+        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
+        target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+        target_compile_definitions(ggml-rocm PRIVATE CC_TURING=1000000000)
+        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
+
+        if (LLAMA_STATIC)
+            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+        endif()
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-rocm)
+    else()
+        message(WARNING "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
+    endif()
+endif()
+
 if (LLAMA_ALL_WARNINGS)
    if (NOT MSVC)
        set(c_flags
--- a/24
+++ b/24
@ -280,6 +280,30 @@ ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif # LLAMA_CLBLAST

+ifdef LLAMA_HIPBLAS
+	ROCM_PATH	?= /opt/rocm
+	HIPCC	    ?= $(ROCM_PATH)/bin/hipcc
+	GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+	LLAMA_CUDA_DMMV_X       ?= 32
+	LLAMA_CUDA_MMV_Y        ?= 1
+	LLAMA_CUDA_KQUANTS_ITER ?= 2
+	CFLAGS      += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	CXXFLAGS    += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	LDFLAGS     += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
+	LDFLAGS		+= -lhipblas -lamdhip64 -lrocblas
+	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
+	HIPFLAGS    += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
+	HIPFLAGS    += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
+	HIPFLAGS    += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+	HIPFLAGS    += -DCC_TURING=1000000000
+ifdef LLAMA_CUDA_FORCE_DMMV
+	HIPFLAGS 	+= -DGGML_CUDA_FORCE_DMMV
+endif # LLAMA_CUDA_FORCE_DMMV
+	OBJS        += ggml-cuda.o
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+endif # LLAMA_HIPBLAS
+
 ifdef LLAMA_METAL
 	CFLAGS   += -DGGML_USE_METAL -DGGML_METAL_NDEBUG
 	CXXFLAGS += -DGGML_USE_METAL
--- a/README.md
+++ b/README.md
@ -422,6 +422,35 @@ Building the program with BLAS support may lead to some performance improvements
  | LLAMA_CUDA_F16          | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |

+- #### hipBLAS
+
+  This provide BLAS acceleation on HIP supported GPU like AMD GPU.
+  Make sure to have ROCm installed.
+  You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html).
+  Windows support is coming soon...
+
+  - Using `make`:
+    ```bash
+    make LLAMA_HIPBLAS=1
+    ```
+  - Using `CMake`:
+    ```bash
+    mkdir build
+    cd build
+    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DLLAMA_HIPBLAS=ON
+    cmake --build .
+    ```
+
+  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.
+  If your GPU is not officialy supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 or 11.0.0 on RDNA3.
+  The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
+
+  | Option                  | Legal values           | Default | Description |
+  |-------------------------|------------------------|---------|-------------|
+  | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_MMV_Y        | Positive integer       |       1 | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
+
 - #### CLBlast

  OpenCL acceleration is provided by the matrix multiplication kernels from the [CLBlast](https://github.com/CNugteren/CLBlast) project and custom kernels for ggml that can generate tokens on the GPU.
--- a/common/common.cpp
+++ b/common/common.cpp
@ -613,9 +613,11 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
+#ifdef GGML_USE_CUBLAS
    fprintf(stdout, "  -nommq, --no-mul-mat-q\n");
-    fprintf(stdout, "                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
+    fprintf(stdout, "                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
    fprintf(stdout, "                        Not recommended since this is both slower and uses more VRAM.\n");
+#endif // GGML_USE_CUBLAS
 #endif
    fprintf(stdout, "  --mtest               compute maximum memory usage\n");
    fprintf(stdout, "  --export              export the computation graph to 'llama.ggml'\n");
--- a/examples/llama-bench/llama-bench.cpp
+++ b/examples/llama-bench/llama-bench.cpp
@ -18,9 +18,7 @@
 #include "llama.h"
 #include "common.h"
 #include "build-info.h"
-#ifdef GGML_USE_CUBLAS
 #include "ggml-cuda.h"
-#endif

 // utils
 static uint64_t get_time_ns() {
@ -504,7 +502,7 @@ struct test {

    static std::string get_backend() {
        if (cuda) {
-            return "CUDA";
+            return GGML_CUDA_NAME;
        }
        if (opencl) {
            return "OpenCL";
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@ -6,15 +6,116 @@
 #include <atomic>
 #include <assert.h>

+#if defined(GGML_USE_HIPBLAS)
+#include <hip/hip_runtime.h>
+#include <hipblas/hipblas.h>
+#include <hip/hip_fp16.h>
+#ifdef __HIP_PLATFORM_AMD__
+// for rocblas_initialize()
+#include "rocblas/rocblas.h"
+#endif
+#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
+#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N HIPBLAS_OP_N
+#define CUBLAS_OP_T HIPBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH 0
+#define CUDA_R_16F  HIPBLAS_R_16F
+#define CUDA_R_32F  HIPBLAS_R_32F
+#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
+#define cublasCreate hipblasCreate
+#define cublasGemmEx hipblasGemmEx
+#define cublasHandle_t hipblasHandle_t
+#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
+#define cublasSetStream hipblasSetStream
+#define cublasSgemm hipblasSgemm
+#define cublasStatus_t hipblasStatus_t
+#define cudaDeviceProp hipDeviceProp_t
+#define cudaDeviceSynchronize hipDeviceSynchronize
+#define cudaError_t hipError_t
+#define cudaEventCreateWithFlags hipEventCreateWithFlags
+#define cudaEventDisableTiming hipEventDisableTiming
+#define cudaEventRecord hipEventRecord
+#define cudaEvent_t hipEvent_t
+#define cudaEventDestroy hipEventDestroy
+#define cudaFree hipFree
+#define cudaFreeHost hipHostFree
+#define cudaGetDevice hipGetDevice
+#define cudaGetDeviceCount hipGetDeviceCount
+#define cudaGetDeviceProperties hipGetDeviceProperties
+#define cudaGetErrorString hipGetErrorString
+#define cudaGetLastError hipGetLastError
+#define cudaMalloc hipMalloc
+#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#define cudaMemcpy hipMemcpy
+#define cudaMemcpy2DAsync hipMemcpy2DAsync
+#define cudaMemcpyAsync hipMemcpyAsync
+#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
+#define cudaMemcpyKind hipMemcpyKind
+#define cudaMemset hipMemset
+#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
+#define cudaSetDevice hipSetDevice
+#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
+#define cudaStreamNonBlocking hipStreamNonBlocking
+#define cudaStreamSynchronize hipStreamSynchronize
+#define cudaStreamWaitEvent(stream, event) hipStreamWaitEvent(stream, event, 0)
+#define cudaStream_t hipStream_t
+#define cudaSuccess hipSuccess
+#else
 #include <cuda_runtime.h>
 #include <cublas_v2.h>
 #include <cuda_fp16.h>
+#endif

 #include "ggml-cuda.h"
 #include "ggml.h"

 #define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
+#ifndef CC_TURING
 #define CC_TURING   700
+#endif
+
+#if defined(GGML_USE_HIPBLAS)
+#define __CUDA_ARCH__ 1300
+
+typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
+static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
+    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+    const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
+    return reinterpret_cast<const int&>(c);
+}
+
+static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
+#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__)
+    c = __builtin_amdgcn_sdot4(a, b, c, false);
+#elif defined(__gfx1100__)
+    c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
+#elif defined(__gfx1010__) || defined(__gfx900__)
+    int tmp1;
+    int tmp2;
+    asm("\n \
+        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \
+        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \
+        v_add3_u32 %0, %1, %2, %0 \n \
+        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \
+        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \
+        v_add3_u32 %0, %1, %2, %0 \n \
+        "
+        : "+v"(c), "=&v"(tmp1), "=&v"(tmp2)
+        : "v"(a), "v"(b)
+    );
+#else
+    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+    c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3];
+#endif
+    return c;
+}
+#endif

 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@ -424,8 +525,8 @@ static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const in
 static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
    const block_q4_1 * x = (const block_q4_1 *) vx;

-    const dfloat d = x[ib].dm.x;
-    const dfloat m = x[ib].dm.y;
+    const dfloat d = __low2half(x[ib].dm);
+    const dfloat m = __high2half(x[ib].dm);

    const int vui = x[ib].qs[iqs];

@ -467,8 +568,8 @@ static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const in
 static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
    const block_q5_1 * x = (const block_q5_1 *) vx;

-    const dfloat d = x[ib].dm.x;
-    const dfloat m = x[ib].dm.y;
+    const dfloat d = __low2half(x[ib].dm);
+    const dfloat m = __high2half(x[ib].dm);

    uint32_t qh;
    memcpy(&qh, x[ib].qh, sizeof(qh));
@ -520,8 +621,8 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, float
    const uint8_t q = x[i].qs[32*n + l];
    float * y = yy + i*QK_K + 128*n;

-    float dall = x[i].dm.x;
-    float dmin = x[i].dm.y;
+    float dall = __low2half(x[i].dm);
+    float dmin = __high2half(x[i].dm);
    y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
    y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
    y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
@ -531,8 +632,8 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, float
    const int il = tid%16;  // 0...15
    const uint8_t q = x[i].qs[il] >> (2*is);
    float * y = yy + i*QK_K + 16*is + il;
-    float dall = x[i].dm.x;
-    float dmin = x[i].dm.y;
+    float dall = __low2half(x[i].dm);
+    float dmin = __high2half(x[i].dm);
    y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
    y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
 #endif
@ -618,8 +719,8 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, float

    float * y = yy + i*QK_K + 64*il + n*ir;

-    const float dall = x[i].dm.x;
-    const float dmin = x[i].dm.y;
+    const float dall = __low2half(x[i].dm);
+    const float dmin = __high2half(x[i].dm);

    const uint8_t * q = x[i].qs + 32*il + n*ir;

@ -657,8 +758,8 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, float

    float * y = yy + i*QK_K + 64*il + 2*ir;

-    const float dall = x[i].dm.x;
-    const float dmin = x[i].dm.y;
+    const float dall = __low2half(x[i].dm);
+    const float dmin = __high2half(x[i].dm);

    const uint8_t * ql = x[i].qs + 32*il + 2*ir;
    const uint8_t * qh = x[i].qh + 2*ir;
@ -770,8 +871,8 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
        const float   * y = yy + i * QK_K + y_offset;
        const uint8_t * q = x[i].qs + q_offset;

-        const float dall = x[i].dm.x;
-        const float dmin = x[i].dm.y;
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);

        const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
        aux[0] = a[0] & 0x0f0f0f0f;
@ -991,8 +1092,8 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
        const float   * y1 = yy + i*QK_K + y_offset;
        const float   * y2 = y1 + 128;

-        const float dall = x[i].dm.x;
-        const float dmin = x[i].dm.y;
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);

        const uint16_t * a = (const uint16_t *)x[i].scales;
        aux[0] = a[im+0] & kmask1;
@ -1124,8 +1225,8 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
        const float   * y1  = yy + i*QK_K + y_offset;
        const float   * y2  = y1 + 128;

-        const float dall = x[i].dm.x;
-        const float dmin = x[i].dm.y;
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);

        const uint16_t * a = (const uint16_t *)x[i].scales;
        aux[0] = a[im+0] & kmask1;
@ -1348,8 +1449,8 @@ static __global__ void quantize_q8_1(const float * __restrict__ x, void * __rest
        return;
    }

-    y[ib].ds.x = d;
-    y[ib].ds.y = sum;
+    reinterpret_cast<half&>(y[ib].ds.x) = d;
+    reinterpret_cast<half&>(y[ib].ds.y) = sum;
 }

 template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
@ -2346,7 +2447,7 @@ static __device__ __forceinline__ float vec_dot_q8_0_q8_1(
        u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
    }

-    return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMVQ>(v, u, bq8_0->d, bq8_1->ds.x);
+    return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMVQ>(v, u, bq8_0->d, __low2half(bq8_1->ds));
 }

 template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q8_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
@ -2432,7 +2533,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1(
 #pragma unroll
    for (int i = 0; i < QR2_K; ++ i) {
        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
-        d8[i] = bq8_1[bq8_offset + i].ds.x;
+        d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
    }

    return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8);
@ -2551,7 +2652,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
 #pragma unroll
    for (int i = 0; i < QR3_K; ++i) {
        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
-        d8[i] = bq8_1[bq8_offset + i].ds.x;
+        d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
    }

    return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
@ -2720,7 +2821,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(

    for (int i = 0; i < QR4_K; ++i) {
        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        d8[i] = bq8i->ds.x;
+        d8[i] = __low2half(bq8i->ds);

        const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
        u[2*i+0] = q8[0];
@ -2747,8 +2848,8 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
    const float dall = bq4_K->d[0];
    const float dmin = bq4_K->d[1];

-    const float d8_1 = bq8_1[0].ds.x;
-    const float d8_2 = bq8_1[1].ds.x;
+    const float d8_1 = __low2float(bq8_1[0].ds);
+    const float d8_2 = __low2float(bq8_1[1].ds);

    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
@ -2901,7 +3002,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
 #pragma unroll
    for (int i = 0; i < QR5_K; ++i) {
        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        d8[i] = bq8i->ds.x;
+        d8[i] = __low2float(bq8i->ds);

        const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
        u[2*i+0] = q8[0];
@ -2919,8 +3020,8 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(

    const float d = bq5_K->d;

-    const float d8_1 = bq8_1[0].ds.x;
-    const float d8_2 = bq8_1[1].ds.x;
+    const float d8_1 = __low2half(bq8_1[0].ds);
+    const float d8_2 = __low2half(bq8_1[1].ds);

    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
@ -3075,7 +3176,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
 #pragma unroll
    for (int i = 0; i < QR6_K; ++i) {
        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
-        d8[i] = bq8_1[bq8_offset + 2*i].ds.x;
+        d8[i] = __low2half(bq8_1[bq8_offset + 2*i].ds);
    }

    return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
@ -3243,7 +3344,7 @@ static __device__ __forceinline__ void mul_mat_q(
                    *dsi_dst = *dsi_src;
                } else {
                    float * dfi_dst = (float *) dsi_dst;
-                    *dfi_dst = (*dsi_src).x;
+                    *dfi_dst = __low2half(*dsi_src);
                }
            }

@ -4944,10 +5045,18 @@ void ggml_init_cublas() {
    static bool initialized = false;

    if (!initialized) {
+
+#ifdef __HIP_PLATFORM_AMD__
+        // Workaround for a rocBLAS bug when using multiple graphics cards:
+        // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
+        rocblas_initialize();
+        CUDA_CHECK(cudaDeviceSynchronize());
+#endif
+
        CUDA_CHECK(cudaGetDeviceCount(&g_device_count));
        GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES);
        int64_t total_vram = 0;
-        fprintf(stderr, "%s: found %d CUDA devices:\n", __func__, g_device_count);
+        fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, g_device_count);
        for (int id = 0; id < g_device_count; ++id) {
            cudaDeviceProp prop;
            CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
--- a/ggml-cuda.h
+++ b/ggml-cuda.h
@ -2,6 +2,14 @@

 #include "ggml.h"

+#ifdef GGML_USE_HIPBLAS
+#define GGML_CUDA_NAME "ROCm"
+#define GGML_CUBLAS_NAME "hipBLAS"
+#else
+#define GGML_CUDA_NAME "CUDA"
+#define GGML_CUBLAS_NAME "cuBLAS"
+#endif
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
--- a/llama.cpp
+++ b/llama.cpp
@ -1836,7 +1836,7 @@ static void llm_load_tensors(
    (void) main_gpu;
    (void) mul_mat_q;
 #if defined(GGML_USE_CUBLAS)
-    LLAMA_LOG_INFO("%s: using CUDA for GPU acceleration\n", __func__);
+    LLAMA_LOG_INFO("%s: using " GGML_CUDA_NAME " for GPU acceleration\n", __func__);
    ggml_cuda_set_main_device(main_gpu);
    ggml_cuda_set_mul_mat_q(mul_mat_q);
 #define LLAMA_BACKEND_OFFLOAD       GGML_BACKEND_GPU