Skip to content

Commit

Permalink
ROCm Port (#1087)
Browse files Browse the repository at this point in the history
* 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>
  • Loading branch information
8 people authored Aug 25, 2023
1 parent 3f460a2 commit 6bbc598
Show file tree
Hide file tree
Showing 12 changed files with 335 additions and 59 deletions.
44 changes: 44 additions & 0 deletions .devops/full-rocm.Dockerfile
Original file line number Diff line number Diff line change
@@ -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"]
44 changes: 44 additions & 0 deletions .devops/main-rocm.Dockerfile
Original file line number Diff line number Diff line change
@@ -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" ]
9 changes: 1 addition & 8 deletions .dockerignore
Original file line number Diff line number Diff line change
Expand Up @@ -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/*

Expand Down
15 changes: 1 addition & 14 deletions .gitignore
Original file line number Diff line number Diff line change
Expand Up @@ -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/

Expand Down
38 changes: 38 additions & 0 deletions CMakeLists.txt
Original file line number Diff line number Diff line change
Expand Up @@ -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)
Expand Down Expand Up @@ -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
Expand Down
24 changes: 24 additions & 0 deletions Makefile
Original file line number Diff line number Diff line change
Expand Up @@ -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
Expand Down
29 changes: 29 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -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.
Expand Down
4 changes: 3 additions & 1 deletion common/common.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -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");
Expand Down
4 changes: 1 addition & 3 deletions examples/llama-bench/llama-bench.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -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() {
Expand Down Expand Up @@ -504,7 +502,7 @@ struct test {

static std::string get_backend() {
if (cuda) {
return "CUDA";
return GGML_CUDA_NAME;
}
if (opencl) {
return "OpenCL";
Expand Down
Loading

0 comments on commit 6bbc598

Please sign in to comment.