Merge remote-tracking branch 'upstream/concedo'

Author: YellowRoseCx

.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" ]

.gitignore

@@ -20,6 +20,7 @@ build-static/
 build-cublas/
 build-opencl/
 build-metal/
+build-mpi/
 build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
@@ -67,4 +68,6 @@ koboldcpp_failsafe.dll
 koboldcpp_openblas.dll
 koboldcpp_openblas_noavx2.dll
 koboldcpp_clblast.dll
-koboldcpp_cublas.dll
+koboldcpp_cublas.dll
+cublas64_11.dll
+cublasLt64_11.dll

CMakeLists.txt

@@ -28,6 +28,8 @@ set(LLAMA_SANITIZE_THREAD OFF)
 set(LLAMA_SANITIZE_ADDRESS OFF)
 set(LLAMA_SANITIZE_UNDEFINED OFF)
 
+option(MAKE_MISC_FILES              "MAKE_MISC_FILES"                                       OFF)
+
 # instruction set specific
 option(LLAMA_AVX                    "llama: enable AVX"                                     ON)
 option(LLAMA_AVX2                   "llama: enable AVX2"                                    ON)
@@ -73,16 +75,16 @@ if (LLAMA_CUBLAS)
 
         enable_language(CUDA)
 
-        set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h)
+        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
         set(GGML_V2_CUDA_SOURCES otherarch/ggml_v2-cuda.cu otherarch/ggml_v2-cuda.h)
         set(GGML_V2_LEGACY_CUDA_SOURCES otherarch/ggml_v2-cuda-legacy.cu otherarch/ggml_v2-cuda-legacy.h)
 
         add_compile_definitions(GGML_USE_CUBLAS)
-        add_compile_definitions(GGML_CUDA_FORCE_DMMV) #non dmmv broken for me
-        
+        #add_compile_definitions(GGML_CUDA_FORCE_DMMV) #non dmmv broken for me
+
         add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
         add_compile_definitions(GGML_CUDA_DMMV_Y=${LLAMA_CUDA_DMMV_Y})
-        add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})    
+        add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
         if (LLAMA_CUDA_DMMV_F16)
             add_compile_definitions(GGML_CUDA_DMMV_F16)
         endif()
@@ -292,7 +294,7 @@ add_library(ggml OBJECT
             ggml.h
             k_quants.h
             k_quants.c
-            ${GGML_CUDA_SOURCES})
+            ${GGML_SOURCES_CUDA})
 target_include_directories(ggml PUBLIC . ./otherarch ./otherarch/tools)
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
@@ -332,12 +334,6 @@ target_link_libraries(gpttype_adapter PRIVATE common2 ggml ${LLAMA_EXTRA_LIBS})
 set_target_properties(gpttype_adapter PROPERTIES POSITION_INDEPENDENT_CODE ON)
 
 
-if (GGML_CUDA_SOURCES)
-    message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
-    set_property(TARGET ggml PROPERTY CUDA_ARCHITECTURES OFF)
-    set_property(TARGET ggml PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
-endif()
-
 set(TARGET koboldcpp_cublas)
 add_library(${TARGET} SHARED expose.cpp expose.h)
 target_include_directories(${TARGET} PUBLIC . ./otherarch ./otherarch/tools ./examples)
@@ -347,3 +343,19 @@ set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME "koboldcpp_cublas")
 set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 target_link_libraries(${TARGET} PUBLIC ggml ggml_v1 ggml_v2 common2 gpttype_adapter ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+
+if (MAKE_MISC_FILES)
+add_library(llama
+        llama.cpp
+        llama.h
+        llama-util.h
+        )
+target_include_directories(llama PUBLIC .)
+target_compile_features(llama PUBLIC cxx_std_11) # don't bump
+target_link_libraries(llama PRIVATE
+    ggml
+    ${LLAMA_EXTRA_LIBS}
+    )
+add_subdirectory(examples)
+endif()

Makefile

@@ -144,19 +144,27 @@ ifdef LLAMA_CUBLAS
 	CUBLASLD_FLAGS = -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
 	CUBLAS_OBJS = ggml-cuda.o ggml_v2-cuda.o ggml_v2-cuda-legacy.o
 	NVCC      = nvcc
-	NVCCFLAGS = --forward-unknown-to-host-compiler -arch=native -DGGML_CUDA_FORCE_DMMV
+	NVCCFLAGS = --forward-unknown-to-host-compiler
+ifdef CUDA_DOCKER_ARCH
+	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(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
 ifdef LLAMA_CUDA_DMMV_X
 	NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 else
 	NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
 endif # LLAMA_CUDA_DMMV_X
-ifdef LLAMA_CUDA_DMMV_Y
+ifdef LLAMA_CUDA_MMV_Y
 	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
-	NVCCFLAGS += -DGGML_CUDA_DMMV_Y=$(LLAMA_CUDA_DMMV_Y)
+else ifdef LLAMA_CUDA_DMMV_Y
+	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
 else
-	NVCCFLAGS += -DGGML_CUDA_DMMV_Y=1
 	NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
-endif # LLAMA_CUDA_DMMV_Y
+endif # LLAMA_CUDA_MMV_Y
 ifdef LLAMA_CUDA_DMMV_F16
 	NVCCFLAGS += -DGGML_CUDA_DMMV_F16
 endif # LLAMA_CUDA_DMMV_F16

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));