Merge branch 'mpi' into refactor-mpi

.devops/full-cuda.Dockerfile

@@ -0,0 +1,33 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=11.7.1
+
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
+
+RUN apt-get update && \
+    apt-get install -y build-essential python3 python3-pip
+
+COPY requirements.txt requirements.txt
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable cuBLAS
+ENV LLAMA_CUBLAS=1
+
+RUN make
+
+ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/main-cuda.Dockerfile

@@ -0,0 +1,32 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=11.7.1
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the CUDA runtime image
+ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
+
+RUN apt-get update && \
+    apt-get install -y build-essential
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable cuBLAS
+ENV LLAMA_CUBLAS=1
+
+RUN make
+
+FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
+
+COPY --from=build /app/main /main
+
+ENTRYPOINT [ "/main" ]

.github/workflows/build.yml

@@ -16,7 +16,10 @@ on:
     paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
 
 env:
- BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+  GGML_NLOOP: 3
+  GGML_NITER: 1
+  GGML_N_THREADS: 1
 
 jobs:
   ubuntu-focal-make:
@@ -64,7 +67,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose
+          ctest --verbose --timeout 900
 
   ubuntu-latest-cmake-sanitizer:
     runs-on: ubuntu-latest
@@ -99,11 +102,17 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose
+          ctest --verbose --timeout 900
 
   ubuntu-latest-cmake-mpi:
     runs-on: ubuntu-latest
 
+    continue-on-error: true
+
+    strategy:
+      matrix:
+        mpi_library: [mpich, libopenmpi-dev]
+
     steps:
       - name: Clone
         id: checkout
@@ -113,7 +122,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential mpich
+          sudo apt-get install build-essential ${{ matrix.mpi_library }}
 
       - name: Build
         id: cmake_build
@@ -175,10 +184,11 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose
+          ctest --verbose --timeout 900
 
   windows-latest-cmake:
     runs-on: windows-latest
+
     env:
       OPENBLAS_VERSION: 0.3.23
       OPENCL_VERSION: 2023.04.17
@@ -277,7 +287,7 @@ jobs:
         if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # Test AVX-512 only when possible
         run: |
           cd build
-          ctest -C Release --verbose
+          ctest -C Release --verbose --timeout 900
 
       - name: Get commit hash
         id: commit

CMakeLists.txt

@@ -218,6 +218,9 @@ if (LLAMA_BLAS)
         message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
         add_compile_options(${BLAS_LINKER_FLAGS})
         add_compile_definitions(GGML_USE_OPENBLAS)
+        if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${LLAMA_BLAS_VENDOR} MATCHES "Generic" OR ${LLAMA_BLAS_VENDOR} MATCHES "Intel"))
+            add_compile_definitions(GGML_BLAS_USE_MKL)
+        endif()
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES})
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
 

Makefile

@@ -172,7 +172,12 @@ ifdef LLAMA_CUBLAS
 	LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
 	OBJS      += ggml-cuda.o
 	NVCC      = nvcc
-	NVCCFLAGS = --forward-unknown-to-host-compiler -arch=native
+	NVCCFLAGS = --forward-unknown-to-host-compiler
+ifdef CUDA_DOCKER_ARCH
+	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
+else
+	NVCCFLAGS += -arch=native
+endif # CUDA_DOCKER_ARCH
 ifdef LLAMA_CUDA_FORCE_DMMV
 	NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
@@ -196,6 +201,7 @@ ifdef LLAMA_CUDA_KQUANTS_ITER
 else
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
+
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif # LLAMA_CUBLAS

README.md

@@ -734,7 +734,7 @@ export LD_LIBRARY_PATH=/vendor/lib64:$LD_LIBRARY_PATH
 
 For easy and swift re-execution, consider documenting this final part in a .sh script file. This will enable you to rerun the process with minimal hassle.
 
-Place your desired model into the `/llama.cpp/models/` directory and execute the `./main (...)` script.
+Place your desired model into the `~/llama.cpp/models/` directory and execute the `./main (...)` script.
 
 ### Docker
 
@@ -770,6 +770,38 @@ or with a light image:
 docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512
 ```
 
+### Docker With CUDA
+
+Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
+
+#### Building Locally
+
+```bash
+docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
+docker build -t local/llama.cpp:light-cuda -f .devops/main-cuda.Dockerfile .
+```
+
+You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
+
+The defaults are:
+
+- `CUDA_VERSION` set to `11.7.1`
+- `CUDA_DOCKER_ARCH` set to `all`
+
+The resulting images, are essentially the same as the non-CUDA images:
+
+1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
+2. `local/llama.cpp:light-cuda`: This image only includes the main executable file.
+
+#### Usage
+
+After building locally, Usage is similar to the non-CUDA examples, but you'll need to add the `--gpus` flag. You will also want to use the `--n-gpu-layers` flag.
+
+```bash
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+```
+
 ### Contributing
 
 - Contributors can open PRs
@@ -790,5 +822,10 @@ docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /mode
 
 ### Docs
 
-- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
+- [main](./examples/main/README.md)
+- [server](./examples/server/README.md)
+- [embd-input](./examples/embd-input/README.md)
+- [jeopardy](./examples/jeopardy/README.md)
+- [BLIS](./docs/BLIS.md)
 - [Performance troubleshooting](./docs/token_generation_performance_tips.md)
+- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)

convert.py

@@ -828,6 +828,7 @@ def lazy_load_torch_file(outer_fp: IO[bytes], path: Path) -> ModelPlus:
 
 
 SAFETENSORS_DATA_TYPES: Dict[str, DataType] = {
+    'BF16': DT_BF16,
     'F16': DT_F16,
     'F32': DT_F32,
     'I32': DT_I32,

examples/baby-llama/baby-llama.cpp

@@ -31,6 +31,17 @@ float frand_normal(struct random_normal_distribution * rnd) {
     return ((r < rnd->min) ? (rnd->min) : (r > rnd->max) ? (rnd->max) : r);
 }
 
+void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
+
+    if (plan.work_size > 0) {
+        buf.resize(plan.work_size);
+        plan.work_data = buf.data();
+    }
+
+    ggml_graph_compute(graph, &plan);
+}
+
 struct ggml_tensor * randomize_tensor(
         struct ggml_tensor * tensor,
         int ndims,
@@ -1569,6 +1580,8 @@ int main(int argc, char ** argv) {
     int n_tokens = model.hparams.n_ctx;
     int n_vocab  = model.hparams.n_vocab;
 
+    std::vector<uint8_t> work_buffer;
+
     for (int ex=0; ex<n_examples; ++ex) {
         struct ggml_init_params params = {
             /*.mem_size   =*/ compute_size,
@@ -1586,7 +1599,6 @@ int main(int argc, char ** argv) {
         int n_past = 0;
 
         ggml_cgraph gf = {};
-        gf.n_threads = 1;
 
         get_example_targets_batch(ctx0, 64*ex+0,  tokens_input, targets);
 
@@ -1595,7 +1607,7 @@ int main(int argc, char ** argv) {
         struct ggml_tensor * e = square_error_loss(ctx0, targets, logits);
 
         ggml_build_forward_expand(&gf, e);
-        ggml_graph_compute(ctx0, &gf);
+        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
 
         float error_before_opt = ggml_get_f32_1d(e, 0);
 
@@ -1611,7 +1623,7 @@ int main(int argc, char ** argv) {
         ggml_opt(ctx0, opt_params_lbfgs, e);
         //
         ggml_build_forward_expand(&gf, e);
-        ggml_graph_compute(ctx0, &gf);
+        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
 
         float error_after_opt = ggml_get_f32_1d(e, 0);
 
@@ -1659,13 +1671,12 @@ int main(int argc, char ** argv) {
             struct ggml_context * ctx0 = ggml_init(params);
 
             ggml_cgraph gf = {};
-            gf.n_threads = 1;
 
             int n_past = 0;
             struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, &gf, tokens_input, sample_ctx, n_past);
 
             ggml_build_forward_expand(&gf, logits);
-            ggml_graph_compute(ctx0, &gf);
+            ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
 
             struct ggml_tensor * best_samples = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sample_ctx);
             struct ggml_tensor * probs        = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_vocab, sample_ctx);
@@ -1687,10 +1698,11 @@ int main(int argc, char ** argv) {
     }
 
     print_matrix(model.tok_embeddings);
-
     printf("done\n");
+
     // ggml_free(kv_self.ctx);
     // ggml_free(model_lora.ctx);
     ggml_free(model.ctx);
+
     return 0;
 }

examples/benchmark/benchmark-matmult.cpp

@@ -20,6 +20,17 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
+
+    if (plan.work_size > 0) {
+        buf.resize(plan.work_size);
+        plan.work_data = buf.data();
+    }
+
+    ggml_graph_compute(graph, &plan);
+}
+
 float tensor_sum_elements(const ggml_tensor * tensor) {
     float sum = 0;
     if (tensor->type==GGML_TYPE_F32) {
@@ -159,13 +170,14 @@ int main(int argc, char ** argv)  {
     // printf("Creating compute graph\n");
     struct ggml_cgraph gf = ggml_build_forward(m11xm2);
 
-    gf.n_threads=benchmark_params.n_threads;
-    printf("cgraph->n_threads=%i\n",gf.n_threads);
+    printf("n_threads=%i\n", benchmark_params.n_threads);
 
     TENSOR_DUMP(m11);
     TENSOR_DUMP(m2);
 
-    ggml_graph_compute(ctx, &gf);
+    std::vector<uint8_t> work_buffer;
+
+    ggml_graph_compute_helper(work_buffer, &gf, benchmark_params.n_threads);
 
     TENSOR_DUMP(gf.nodes[0]);
 
@@ -187,7 +199,6 @@ int main(int argc, char ** argv)  {
 
     // printf("Creating compute graph\n");
     struct ggml_cgraph gf31 = ggml_build_forward(q31);
-    gf31.n_threads=benchmark_params.n_threads;
 
     // Set up a second graph computation to make sure we override the CPU cache lines
     // printf("Creating new tensor q12 & Running quantize\n");
@@ -199,8 +210,7 @@ int main(int argc, char ** argv)  {
 
     //printf("Creating compute graph\n");
     struct ggml_cgraph gf32 = ggml_build_forward(q32);
-    gf32.n_threads=benchmark_params.n_threads;
-    printf("cgraph->n_threads=%i\n",gf31.n_threads);
+    printf("n_threads=%i\n", benchmark_params.n_threads);
 
     const int dimx = sizex;
     const int dimy = sizey;
@@ -221,14 +231,15 @@ int main(int argc, char ** argv)  {
 
         long long int start = ggml_time_us();
         //printf("Running ggml_graph_compute\n");
-        ggml_graph_compute(ctx, &gf31);
+        ggml_graph_compute_helper(work_buffer, &gf31, benchmark_params.n_threads);
+
         long long int stop = ggml_time_us();
         long long int usec = stop-start;
         double gflops = (double)(flops_per_matrix)/usec/1000.0;
         gflops_sum += gflops;
         printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%10.2f\n",
             i,
-            gf31.n_threads,
+            benchmark_params.n_threads,
             sizex, sizey, sizez, flops_per_matrix,
             usec,gflops);
 
@@ -253,7 +264,7 @@ int main(int argc, char ** argv)  {
         }
 
         // Running a different graph computation to make sure we override the CPU cache lines
-        ggml_graph_compute(ctx, &gf32);
+        ggml_graph_compute_helper(work_buffer, &gf32, benchmark_params.n_threads);
     }
     printf("\n");
     printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations));