mpi : fix after master merge

Not tested yet

.github/workflows/build.yml

@@ -104,6 +104,40 @@ jobs:
           cd build
           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
+        uses: actions/checkout@v1
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential ${{ matrix.mpi_library }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake -DLLAMA_MPI=ON ..
+          cmake --build . --config Release
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest --verbose
+
   macOS-latest-make:
     runs-on: macos-latest
 

CMakeLists.txt

@@ -75,6 +75,7 @@ option(LLAMA_CUDA_DMMV_F16                   "llama: use 16 bit floats for dmmv
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 OFF)
+option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 
@@ -308,6 +309,23 @@ if (LLAMA_METAL)
         )
 endif()
 
+if (LLAMA_MPI)
+    cmake_minimum_required(VERSION 3.10)
+    find_package(MPI)
+    if (MPI_C_FOUND)
+        message(STATUS "MPI found")
+        set(GGML_SOURCES_MPI ggml-mpi.c ggml-mpi.h)
+        add_compile_definitions(GGML_USE_MPI)
+        add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
+        set(cxx_flags ${cxx_flags} -Wno-cast-qual)
+        set(c_flags   ${c_flags}   -Wno-cast-qual)
+        set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${MPI_C_LIBRARIES})
+        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
+    else()
+        message(WARNING "MPI not found")
+    endif()
+endif()
+
 if (LLAMA_CLBLAST)
     find_package(CLBlast)
     if (CLBlast_FOUND)
@@ -476,6 +494,7 @@ add_library(ggml OBJECT
             ${GGML_SOURCES_CUDA}
             ${GGML_SOURCES_OPENCL}
             ${GGML_SOURCES_METAL}
+            ${GGML_SOURCES_MPI}
             ${GGML_SOURCES_EXTRA}
             )
 

Makefile

@@ -147,6 +147,15 @@ ifndef LLAMA_NO_ACCELERATE
 	endif
 endif # LLAMA_NO_ACCELERATE
 
+ifdef LLAMA_MPI
+	CFLAGS += -DGGML_USE_MPI -Wno-cast-qual
+	CXXFLAGS += -DGGML_USE_MPI -Wno-cast-qual
+	OBJS     += ggml-mpi.o
+
+ggml-mpi.o: ggml-mpi.c ggml-mpi.h
+	$(CC) $(CFLAGS) -c $< -o $@
+endif # LLAMA_MPI
+
 ifdef LLAMA_OPENBLAS
 	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
 	LDFLAGS += -lopenblas

README.md

@@ -268,6 +268,45 @@ Any value larger than 0 will offload the computation to the GPU. For example:
 ./main -m ./models/7B/ggml-model-q4_0.bin -n 128 -ngl 1
 ```
 
+### MPI Build
+
+MPI lets you distribute the computation over a cluster of machines. Because of the serial nature of LLM prediction, this won't yield any end-to-end speed-ups, but it will let you run larger models than would otherwise fit into RAM on a single machine.
+
+First you will need MPI libraries installed on your system. The two most popular (only?) options are [MPICH](https://www.mpich.org) and [OpenMPI](https://www.open-mpi.org). Either can be installed with a package manager (`apt`, Homebrew, MacPorts, etc).
+
+Next you will need to build the project with `LLAMA_MPI` set to true on all machines; if you're building with `make`, you will also need to specify an MPI-capable compiler (when building with CMake, this is configured automatically):
+
+- Using `make`:
+
+  ```bash
+  make CC=mpicc CXX=mpicxx LLAMA_MPI=1
+  ```
+
+- Using `CMake`:
+
+  ```bash
+  cmake -S . -B build -DLLAMA_MPI=ON
+  ```
+
+Once the programs are built, download/convert the weights on all of the machines in your cluster. The paths to the weights and programs should be identical on all machines.
+
+Next, ensure password-less SSH access to each machine from the primary host, and create a `hostfile` with a list of the hostnames and their relative "weights" (slots). If you want to use localhost for computation, use its local subnet IP address rather than the loopback address or "localhost".
+
+Here is an example hostfile:
+
+```
+192.168.0.1:2
+malvolio.local:1
+```
+
+The above will distribute the computation across 2 processes on the first host and 1 process on the second host. Each process will use roughly an equal amount of RAM. Try to keep these numbers small, as inter-process (intra-host) communication is expensive.
+
+Finally, you're ready to run a computation using `mpirun`:
+
+```bash
+mpirun -hostfile hostfile -n 3 ./main -m ./models/7B/ggml-model-q4_0.bin -n 128
+```
+
 ### BLAS Build
 
 Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). BLAS doesn't affect the normal generation performance. There are currently three different implementations of it:

examples/embd-input/embd-input-lib.cpp

@@ -34,7 +34,7 @@ struct MyModel* create_mymodel(int argc, char ** argv) {
     }
     fprintf(stderr, "%s: seed  = %d\n", __func__, params.seed);
 
-    llama_init_backend(params.numa);
+    llama_backend_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;

examples/embedding/embedding.cpp

@@ -35,7 +35,7 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_init_backend(params.numa);
+    llama_backend_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -93,5 +93,7 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
     llama_free_model(model);
 
+    llama_backend_free();
+
     return 0;
 }

examples/main/main.cpp

@@ -105,7 +105,7 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_init_backend(params.numa);
+    llama_backend_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -671,5 +671,7 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
     llama_free_model(model);
 
+    llama_backend_free();
+
     return 0;
 }

examples/perplexity/perplexity.cpp

@@ -147,7 +147,7 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_init_backend(params.numa);
+    llama_backend_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -172,5 +172,7 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
     llama_free_model(model);
 
+    llama_backend_free();
+
     return 0;
 }

examples/quantize/quantize.cpp

@@ -180,7 +180,7 @@ int main(int argc, char ** argv) {
         usage(argv[0]);
     }
 
-    llama_init_backend(false);
+    llama_backend_init(false);
 
     // parse command line arguments
     const std::string fname_inp = argv[arg_idx];
@@ -257,5 +257,7 @@ int main(int argc, char ** argv) {
         printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0);
     }
 
+    llama_backend_free();
+
     return 0;
 }

examples/server/server.cpp

@@ -1079,7 +1079,7 @@ int main(int argc, char **argv)
         params.model_alias = params.model;
     }
 
-    llama_init_backend(params.numa);
+    llama_backend_init(params.numa);
 
     LOG_INFO("build info", {{"build", BUILD_NUMBER},
                             {"commit", BUILD_COMMIT}});
@@ -1309,5 +1309,7 @@ int main(int argc, char **argv)
         return 1;
     }
 
+    llama_backend_free();
+
     return 0;
 }

examples/simple/simple.cpp

@@ -66,7 +66,7 @@ int main(int argc, char ** argv)
     // Init LLM :
     //---------------------------------
 
-    llama_init_backend(params.numa);
+    llama_backend_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -173,6 +173,8 @@ int main(int argc, char ** argv)
     llama_free( ctx );
     llama_free_model( model );
 
+    llama_backend_free();
+
     return 0;
 }
 

ggml-mpi.c

@@ -0,0 +1,213 @@
+#include "ggml-mpi.h"
+
+#include "ggml.h"
+
+#include <mpi.h>
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+
+#define UNUSED GGML_UNUSED
+
+struct ggml_mpi_context {
+    int rank;
+    int size;
+};
+
+void ggml_mpi_backend_init(void) {
+    MPI_Init(NULL, NULL);
+}
+
+void ggml_mpi_backend_free(void) {
+    MPI_Finalize();
+}
+
+struct ggml_mpi_context * ggml_mpi_init(void) {
+    struct ggml_mpi_context * ctx = calloc(1, sizeof(struct ggml_mpi_context));
+
+    MPI_Comm_rank(MPI_COMM_WORLD, &ctx->rank);
+    MPI_Comm_size(MPI_COMM_WORLD, &ctx->size);
+
+    return ctx;
+}
+
+void ggml_mpi_free(struct ggml_mpi_context * ctx) {
+    free(ctx);
+}
+
+int ggml_mpi_rank(struct ggml_mpi_context * ctx) {
+    return ctx->rank;
+}
+
+void ggml_mpi_eval_init(
+        struct ggml_mpi_context * ctx_mpi,
+                            int * n_tokens,
+                            int * n_past,
+                            int * n_threads) {
+    UNUSED(ctx_mpi);
+
+    // synchronize the worker node parameters with the root node
+    MPI_Barrier(MPI_COMM_WORLD);
+
+    MPI_Bcast(n_tokens,  1, MPI_INT, 0, MPI_COMM_WORLD);
+    MPI_Bcast(n_past,    1, MPI_INT, 0, MPI_COMM_WORLD);
+    MPI_Bcast(n_threads, 1, MPI_INT, 0, MPI_COMM_WORLD);
+}
+
+int ggml_graph_get_node_idx(struct ggml_cgraph * gf, const char * name) {
+    struct ggml_tensor * t = ggml_graph_get_tensor(gf, name);
+    if (t == NULL) {
+        fprintf(stderr, "%s: tensor %s not found\n", __func__, name);
+        return -1;
+    }
+
+    for (int i = 0; i < gf->n_nodes; i++) {
+        if (gf->nodes[i] == t) {
+            return i;
+        }
+    }
+
+    fprintf(stderr, "%s: tensor %s not found in graph (should not happen)\n", __func__, name);
+    return -1;
+}
+
+// TODO: there are many improvements that can be done to this implementation
+void ggml_mpi_graph_compute(
+        struct ggml_mpi_context * ctx_mpi,
+        struct ggml_context     * ctx,
+             struct ggml_cgraph * gf,
+                            int   n_layers,
+                            int   n_threads) {
+    const int mpi_rank = ctx_mpi->rank;
+    const int mpi_size = ctx_mpi->size;
+
+    struct ggml_tensor * inp_tokens = ggml_graph_get_tensor(gf, "inp_tokens");
+    if (inp_tokens == NULL) {
+        fprintf(stderr, "%s: tensor 'inp_tokens' not found\n", __func__);
+        return;
+    }
+
+    struct ggml_tensor * inp0 = ggml_graph_get_tensor(gf, "layer_inp_0");
+    if (inp0 == NULL) {
+        fprintf(stderr, "%s: tensor 'inp0' not found\n", __func__);
+        return;
+    }
+
+    GGML_ASSERT(inp0 == gf->nodes[0]);
+
+    // distribute the compute graph into slices across the MPI nodes
+    //
+    // the main node (0) processes the last layers + the remainder of the compute graph
+    // and is responsible to pass the input tokens to the first node (1)
+    //
+    // node 1:   [(  0) * n_per_node, (  1) * n_per_node)
+    // node 2:   [(  1) * n_per_node, (  2) * n_per_node)
+    // ...
+    // node n-1: [(n-2) * n_per_node, (n-1) * n_per_node)
+    // node 0:   [(n-1) * n_per_node,            n_nodes)
+    //
+    if (mpi_rank > 0) {
+        if (mpi_rank == 1) { // the first node receives the input tokens from the main node
+            MPI_Status status; UNUSED(status);
+
+            const int mpi_rank_src = mpi_rank - 1;
+
+            const int retval = MPI_Recv(inp_tokens->data, ggml_nelements(inp_tokens), MPI_INT, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
+            GGML_ASSERT(retval == MPI_SUCCESS);
+        } else { // recv input data for each node into the "inp0" tensor (i.e. the first node in the compute graph)
+            MPI_Status status; UNUSED(status);
+
+            const int mpi_rank_src = mpi_rank - 1;
+
+            //printf("%s: node %d: waiting for %d elements from %d\n", __func__, mpi_rank, (int) ggml_nelements(inp0), mpi_rank_src);
+            const int retval = MPI_Recv(inp0->data, ggml_nelements(inp0), MPI_FLOAT, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
+            GGML_ASSERT(retval == MPI_SUCCESS);
+        }
+    } else if (mpi_size > 1) {
+        // node 0 sends the input tokens to node 1
+        {
+            const int mpi_rank_dst = mpi_rank + 1;
+
+            const int retval = MPI_Send(inp_tokens->data, ggml_nelements(inp_tokens), MPI_INT, mpi_rank_dst, 0, MPI_COMM_WORLD);
+            GGML_ASSERT(retval == MPI_SUCCESS);
+        }
+
+        // recv the output data from the last node
+        {
+            MPI_Status status; UNUSED(status);
+
+            const int mpi_rank_src = mpi_size - 1;
+
+            //fprintf(stderr, "%s: node %d: waiting for %d elements from %d\n", __func__, mpi_rank, (int) ggml_nelements(inp0), mpi_rank_src);
+            const int retval = MPI_Recv(inp0->data, ggml_nelements(inp0), MPI_FLOAT, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
+            GGML_ASSERT(retval == MPI_SUCCESS);
+        }
+    }
+
+    {
+        const int n_per_node = (n_layers + (mpi_size - 1)) / mpi_size;
+
+        const int mpi_idx = mpi_rank > 0 ? mpi_rank - 1 : mpi_size - 1;
+
+        const int il0 =               (mpi_idx + 0) * n_per_node;
+        const int il1 = MIN(n_layers, (mpi_idx + 1) * n_per_node);
+
+        char name_l0[GGML_MAX_NAME];
+        char name_l1[GGML_MAX_NAME];
+
+        snprintf(name_l0, sizeof(name_l0), "layer_inp_%d", il0);
+        snprintf(name_l1, sizeof(name_l1), "layer_inp_%d", il1);
+
+        const int idx_l0 =                ggml_graph_get_node_idx(gf, name_l0);
+        const int idx_l1 = mpi_rank > 0 ? ggml_graph_get_node_idx(gf, name_l1) + 1 : gf->n_nodes;
+
+        if (idx_l0 < 0 || idx_l1 < 0) {
+            fprintf(stderr, "%s: layer input nodes not found\n", __func__);
+            return;
+        }
+
+        // attach the input data to all nodes that need it
+        // TODO: not great - should be able to do this without modifying the compute graph (see next TODO below)
+        for (int i = idx_l0; i < idx_l1; i++) {
+            if (gf->nodes[i]->src0 == gf->nodes[idx_l0]) {
+                gf->nodes[i]->src0 =  inp0;
+            }
+            if (gf->nodes[i]->src1 == gf->nodes[idx_l0]) {
+                gf->nodes[i]->src1 =  inp0;
+            }
+        }
+
+        // TODO: instead of rearranging the nodes, we should be able to execute a subset of the compute graph
+        for (int i = 1; i < idx_l1 - idx_l0; i++) {
+            gf->nodes[i] = gf->nodes[idx_l0 + i];
+            gf->grads[i] = gf->grads[idx_l0 + i];
+        }
+
+        // the first node performs the "get_rows" operation, the rest of the nodes get the data from the previous node
+        if (mpi_idx != 0) {
+            gf->nodes[0]->op = GGML_OP_NONE;
+        }
+
+        gf->n_nodes = idx_l1 - idx_l0;
+
+        //fprintf(stderr, "%s: node %d: processing %d nodes [%d, %d)\n", __func__, mpi_rank, gf->n_nodes, il0, il1);
+    }
+
+    ggml_graph_compute_with_ctx(ctx, gf, n_threads);
+
+    //fprintf(stderr, "%s: node %d: done\n", __func__, mpi_rank);
+
+    // send the output data to the next node
+    if (mpi_rank > 0) {
+        struct ggml_tensor * output = gf->nodes[gf->n_nodes - 1];
+
+        const int mpi_rank_dst = (mpi_rank + 1) % mpi_size;
+
+        //fprintf(stderr, "%s: node %d: sending %d elements to node %d\n", __func__, mpi_rank, ggml_nelements(output), mpi_rank_dst);
+
+        const int retval = MPI_Send(output->data, ggml_nelements(output), MPI_FLOAT, mpi_rank_dst, 0, MPI_COMM_WORLD);
+        GGML_ASSERT(retval == MPI_SUCCESS);
+    }
+}

ggml-mpi.h

@@ -0,0 +1,36 @@
+#pragma once
+
+struct ggml_context;
+struct ggml_tensor;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_mpi_context;
+
+void ggml_mpi_backend_init(void);
+void ggml_mpi_backend_free(void);
+
+struct ggml_mpi_context * ggml_mpi_init(void);
+void ggml_mpi_free(struct ggml_mpi_context * ctx);
+
+int ggml_mpi_rank(struct ggml_mpi_context * ctx);
+
+void ggml_mpi_eval_init(
+        struct ggml_mpi_context * ctx_mpi,
+                            int * n_tokens,
+                            int * n_past,
+                            int * n_threads);
+
+void ggml_mpi_graph_compute(
+        struct ggml_mpi_context * ctx_mpi,
+        struct ggml_context     * ctx,
+             struct ggml_cgraph * gf,
+                            int   n_layers,
+                            int   n_threads);
+
+#ifdef __cplusplus
+}
+#endif

llama.cpp

@@ -19,6 +19,9 @@
 #ifdef GGML_USE_METAL
 #include "ggml-metal.h"
 #endif
+#ifdef GGML_USE_MPI
+#include "ggml-mpi.h"
+#endif
 #ifdef GGML_USE_K_QUANTS
 #ifndef QK_K
 #ifdef GGML_QKK_64
@@ -352,6 +355,10 @@ struct llama_context {
     ggml_metal_context * ctx_metal = NULL;
 #endif
 
+#ifdef GGML_USE_MPI
+    ggml_mpi_context * ctx_mpi = NULL;
+#endif
+
     int    buf_last = 0;
     size_t buf_max_size[LLAMA_MAX_SCRATCH_BUFFERS] = { 0 };
 
@@ -870,7 +877,7 @@ bool llama_mlock_supported() {
     return llama_mlock::SUPPORTED;
 }
 
-void llama_init_backend(bool numa) {
+void llama_backend_init(bool numa) {
     ggml_time_init();
 
     // needed to initialize f16 tables
@@ -883,6 +890,16 @@ void llama_init_backend(bool numa) {
     if (numa) {
         ggml_numa_init();
     }
+
+#ifdef GGML_USE_MPI
+    ggml_mpi_backend_init();
+#endif
+}
+
+void llama_backend_free() {
+#ifdef GGML_USE_MPI
+    ggml_mpi_backend_free();
+#endif
 }
 
 int64_t llama_time_us() {
@@ -1284,13 +1301,17 @@ static bool llama_eval_internal(
          llama_context & lctx,
      const llama_token * tokens,
            const float * embd,
-             const int   n_tokens,
-             const int   n_past,
+                   int   n_tokens,
+                   int   n_past,
                    int   n_threads,
             const char * cgraph_fname) {
 
     LLAMA_ASSERT((!tokens && embd) || (tokens && !embd));
 
+#ifdef GGML_USE_MPI
+    ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads);
+#endif
+
     const int64_t t_start_us = ggml_time_us();
 
     const int N = n_tokens;
@@ -1331,11 +1352,16 @@ static bool llama_eval_internal(
     struct ggml_tensor * inpL;
 
     if (tokens) {
-        struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
-        ggml_set_name(embd, "embd");
-        memcpy(embd->data, tokens, N*ggml_element_size(embd));
-        inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd);
+        struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+        memcpy(inp_tokens->data, tokens, N*ggml_element_size(inp_tokens));
+        ggml_set_name(inp_tokens, "inp_tokens");
+
+        inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens);
     } else {
+#ifdef GGML_USE_MPI
+        GGML_ASSERT(false && "not implemented");
+#endif
+
         inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N);
         memcpy(inpL->data, embd, N * n_embd * ggml_element_size(inpL));
     }
@@ -1353,18 +1379,20 @@ static bool llama_eval_internal(
     offload_func_t offload_func_v  = llama_nop;
 
 #ifdef GGML_USE_CUBLAS
-        if (n_gpu_layers > n_layer) {
-            offload_func_nr = ggml_cuda_assign_buffers;
-        }
-        if (n_gpu_layers > n_layer + 1) {
-            offload_func_v  = ggml_cuda_assign_buffers;
-        }
-        if (n_gpu_layers > n_layer + 2) {
-            offload_func_kq = ggml_cuda_assign_buffers;
-        }
+    if (n_gpu_layers > n_layer) {
+        offload_func_nr = ggml_cuda_assign_buffers;
+    }
+    if (n_gpu_layers > n_layer + 1) {
+        offload_func_v  = ggml_cuda_assign_buffers;
+    }
+    if (n_gpu_layers > n_layer + 2) {
+        offload_func_kq = ggml_cuda_assign_buffers;
+    }
 #endif // GGML_USE_CUBLAS
 
     for (int il = 0; il < n_layer; ++il) {
+        ggml_format_name(inpL, "layer_inp_%d", il);
+
         offload_func_t offload_func = llama_nop;
 
 #ifdef GGML_USE_CUBLAS
@@ -1571,7 +1599,6 @@ static bool llama_eval_internal(
 
         // input for next layer
         inpL = cur;
-
     }
 
     lctx.use_buf(ctx0, 0);
@@ -1579,7 +1606,6 @@ static bool llama_eval_internal(
     // used at the end to optionally extract the embeddings
     struct ggml_tensor * embeddings = NULL;
 
-
     // norm
     {
         cur = ggml_rms_norm(ctx0, inpL);
@@ -1594,7 +1620,6 @@ static bool llama_eval_internal(
         embeddings = cur;
     }
 
-
     // lm_head
     cur = ggml_mul_mat(ctx0, model.output, cur);
     ggml_set_name(cur, "result_output");
@@ -1631,6 +1656,10 @@ static bool llama_eval_internal(
 
         ggml_graph_compute_helper(lctx.work_buffer, &gf, n_threads);
     }
+#elif GGML_USE_MPI
+    ggml_mpi_graph_compute(lctx.ctx_mpi, ctx0, &gf, n_layer, n_threads);
+
+    cur = gf.nodes[gf.n_nodes - 1];
 #else
     ggml_graph_compute_helper(lctx.work_buffer, &gf, n_threads);
 #endif
@@ -1656,26 +1685,32 @@ static bool llama_eval_internal(
     // update kv token count
     lctx.kv_self.n = n_past + N;
 
-    // extract logits
+#ifdef GGML_USE_MPI
+    if (ggml_mpi_rank(lctx.ctx_mpi) == 0) {
+#else
     {
-        auto & logits_out = lctx.logits;
+#endif
+        // extract logits
+        {
+            auto & logits_out = lctx.logits;
 
-        if (lctx.logits_all) {
-            logits_out.resize(n_vocab * N);
-            memcpy(logits_out.data(), (float *) ggml_get_data(cur), sizeof(float)*n_vocab*N);
-        } else {
-            // return result for just the last token
-            logits_out.resize(n_vocab);
-            memcpy(logits_out.data(), (float *) ggml_get_data(cur) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+            if (lctx.logits_all) {
+                logits_out.resize(n_vocab * N);
+                memcpy(logits_out.data(), (float *) ggml_get_data(cur), sizeof(float)*n_vocab*N);
+            } else {
+                // return result for just the last token
+                logits_out.resize(n_vocab);
+                memcpy(logits_out.data(), (float *) ggml_get_data(cur) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+            }
         }
-    }
 
-    // extract embeddings
-    if (!lctx.embedding.empty()) {
-        auto & embedding_out = lctx.embedding;
+        // extract embeddings
+        if (!lctx.embedding.empty()) {
+            auto & embedding_out = lctx.embedding;
 
-        embedding_out.resize(n_embd);
-        memcpy(embedding_out.data(), (float *) ggml_get_data(embeddings) + (n_embd*(N - 1)), sizeof(float)*n_embd);
+            embedding_out.resize(n_embd);
+            memcpy(embedding_out.data(), (float *) ggml_get_data(embeddings) + (n_embd*(N - 1)), sizeof(float)*n_embd);
+        }
     }
 
     if (mem_per_token == 0) {
@@ -2697,6 +2732,18 @@ struct llama_context * llama_new_context_with_model(
     }
 #endif
 
+#ifdef GGML_USE_MPI
+    ctx->ctx_mpi = ggml_mpi_init();
+
+    if (ggml_mpi_rank(ctx->ctx_mpi) > 0) {
+        // Enter a blocking eval loop with dummy input, letting rank=0 drive the process
+        const std::vector<llama_token> tmp(ctx->model.hparams.n_ctx, llama_token_bos());
+        while (!llama_eval(ctx, tmp.data(), tmp.size(), 0, 0)) {};
+        llama_backend_free();
+        exit(1);
+    }
+#endif
+
     return ctx;
 }
 

llama.h

@@ -158,7 +158,9 @@ extern "C" {
     // Initialize the llama + ggml backend
     // If numa is true, use NUMA optimizations
     // Call once at the start of the program
-    LLAMA_API void llama_init_backend(bool numa);
+    LLAMA_API void llama_backend_init(bool numa);
+    // Call once at the end of the program - currently only used for MPI
+    LLAMA_API void llama_backend_free();
 
     LLAMA_API int64_t llama_time_us();
 

tests/test-tokenizer-0.cpp

@@ -31,6 +31,8 @@ int main(int argc, char **argv) {
     llama_model * model;
     llama_context * ctx;
 
+    llama_init_backend(false);
+
     // load the vocab
     {
         auto lparams = llama_context_default_params();
@@ -97,5 +99,7 @@ int main(int argc, char **argv) {
     llama_free_model(model);
     llama_free(ctx);
 
+    llama_finalize_backend();
+
     return 0;
 }