llama_cpp 0.3.2 → 0.3.3

data/ext/llama_cpp/src/ggml-metal.h

@@ -34,9 +34,13 @@ extern "C" {
 
 struct ggml_metal_context;
 
-struct ggml_metal_context * ggml_metal_init(void);
+// number of command buffers to use
+struct ggml_metal_context * ggml_metal_init(int n_cb);
 void ggml_metal_free(struct ggml_metal_context * ctx);
 
+// set the number of command buffers to use
+void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb);
+
 // creates a mapping between a host memory buffer and a device memory buffer
 // - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
 // - the mapping is used during computation to determine the arguments of the compute kernels

data/ext/llama_cpp/src/ggml-metal.m

@@ -25,6 +25,8 @@ struct ggml_metal_buffer {
 };
 
 struct ggml_metal_context {
+    int n_cb;
+
     float * logits;
 
     id<MTLDevice>       device;
@@ -86,11 +88,12 @@ static NSString * const msl_library_source = @"see metal.metal";
 @implementation GGMLMetalClass
 @end
 
-struct ggml_metal_context * ggml_metal_init(void) {
+struct ggml_metal_context * ggml_metal_init(int n_cb) {
     fprintf(stderr, "%s: allocating\n", __func__);
 
     struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
 
+    ctx->n_cb   = n_cb;
     ctx->device = MTLCreateSystemDefaultDevice();
     ctx->queue  = [ctx->device newCommandQueue];
     ctx->n_buffers = 0;
@@ -208,6 +211,10 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     free(ctx);
 }
 
+void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb) {
+    ctx->n_cb = n_cb;
+}
+
 // finds the Metal buffer that contains the tensor data on the GPU device
 // the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
 // Metal buffer based on the host memory pointer
@@ -354,7 +361,7 @@ void ggml_metal_graph_compute(
     // create multiple command buffers and enqueue them
     // then, we encode the graph into the command buffers in parallel
 
-    const int n_cb = gf->n_threads;
+    const int n_cb = ctx->n_cb;
 
     NSMutableArray * command_buffers = [NSMutableArray arrayWithCapacity:n_cb];
 
@@ -386,8 +393,8 @@ void ggml_metal_graph_compute(
             for (int i = node_start; i < node_end; ++i) {
                 metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
 
-                struct ggml_tensor * src0 = gf->nodes[i]->src0;
-                struct ggml_tensor * src1 = gf->nodes[i]->src1;
+                struct ggml_tensor * src0 = gf->nodes[i]->src[0];
+                struct ggml_tensor * src1 = gf->nodes[i]->src[1];
                 struct ggml_tensor * dst  = gf->nodes[i];
 
                 const int64_t  ne00 = src0 ? src0->ne[0] : 0;
@@ -443,6 +450,7 @@ void ggml_metal_graph_compute(
                 //}
 
                 switch (dst->op) {
+                    case GGML_OP_NONE:
                     case GGML_OP_RESHAPE:
                     case GGML_OP_VIEW:
                     case GGML_OP_TRANSPOSE:
@@ -731,7 +739,10 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:13];
                                 [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
 
-                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1) {
+                                if (src0t == GGML_TYPE_Q4_0) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01 / 8+((ne01 % 8) & 0x01), ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src0t == GGML_TYPE_Q4_1) {
                                     [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
                                     [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }

data/ext/llama_cpp/src/ggml-metal.metal

@@ -365,6 +365,10 @@ kernel void kernel_rms_norm(
     }
 }
 
+// putting them in the kernel cause a significant performance penalty
+#define N_DST 4 // each SIMD group works on 4 rows
+#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
 kernel void kernel_mul_mat_q4_0_f32(
         device const  void * src0,
         device const float * src1,
@@ -372,64 +376,69 @@ kernel void kernel_mul_mat_q4_0_f32(
         constant   int64_t & ne00,
         constant   int64_t & ne10,
         constant   int64_t & ne0,
-        threadgroup float  * sum [[threadgroup(0)]],
+        constant   int64_t & ne01[[buffer(4)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2 tpitg[[thread_position_in_threadgroup]],
-        uint2  tptg[[threads_per_threadgroup]]) {
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
     const int nb = ne00/QK4_0;
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-
-    device const block_q4_0 * x = (device const block_q4_0 *) src0 + r0*nb;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    device const block_q4_0 * x = (device const block_q4_0 *) src0 + (r0 * N_SIMDGROUP + sgitg) * N_DST * nb;
     device const float      * y = (device const float      *) src1 + r1*ne10;
+    block_q4_0 qb_curr, qb_next;
+    float4 y_curr[8];       // src1 vector cache
+    float sumf[N_DST]={0.f}, all_sum;
+    thread float * yl=(thread float *)y_curr;
+
+    // bootstrap
+    qb_curr = x[tiisg];
+    // each thread in a SIMD group deals with 1 block.
+    for (int column = 0; column < nb / N_SIMDWIDTH; column++) {
+
+        for (int i = 0; i < QK4_0 / 4; i++) {
+            y_curr[i] = *((device float4  *)(y + N_SIMDWIDTH * (tiisg + column * QK4_0) + 4 * i));
+        }
 
-    const int nth = tptg.x*tptg.y;
-    const int ith = tptg.y*tpitg.x + tpitg.y;
-
-    const int ix = tpitg.y/4;           // 0 or 1
-    const int iy = tpitg.y - 4*ix;      // 0...3
-
-    const int first = 4 * iy;
-
-    float sumf = 0;
+        for (int row = 0; row < N_DST; row++) {
+            // prefetch next x block
+            qb_next = x[tiisg + ((row + 1) % N_DST) * nb + (column + ((row + 1) / N_DST)) * N_SIMDWIDTH];
 
-    for (int i = 2*tpitg.x + ix; i < nb; i += 2*tptg.x) {
+            // calculate
+            float d = qb_curr.d;
+            float2 acc = {0.0f, 0.0f};
+            for (int i = 0; i < 16; i++) {
+                acc[0] += yl[i] * (qb_curr.qs[i] & 0xF) + yl[i+16] * (qb_curr.qs[i] >> 4);
+                acc[1] += yl[i] + yl[i+16];
+            }
+            sumf[row] += d * (acc[0] - 8.f*acc[1]);
+            qb_curr = qb_next;
+        }
+    }
 
-        const float d = (float)x[i].d;
+    for (int i = 0; i < QK4_0 / 4; i++) {
+        y_curr[i] = *((device float4 *)(y + N_SIMDWIDTH * (tiisg + (nb / N_SIMDWIDTH) * QK4_0) + 4 * i));
+    }
 
-        device const uint8_t * xl = x[i].qs + first;
-        device const float   * yl = y + i * QK4_0 + first;
+    for (int row = 0; row < N_DST; row++) {
+        // prefetch next x block
+        qb_next = x[tiisg + ((row + 1) % N_DST) * nb + (nb / N_SIMDWIDTH + ((row + 1) / N_DST)) * N_SIMDWIDTH];
 
+        // calculate
+        float d = qb_curr.d;
         float2 acc = {0.0f, 0.0f};
-
-        for (int j = 0; j < 4; ++j) {
-
-            acc[0] += yl[j] * (xl[j] & 0xF) + yl[j+16] * (xl[j] >> 4);
-            acc[1] += yl[j] + yl[j+16];
-
+        for (int i = 0; i < 16; i++) {
+            acc[0] += yl[i] * (qb_curr.qs[i] & 0xF) + yl[i+16] * (qb_curr.qs[i] >> 4);
+            acc[1] += yl[i] + yl[i+16];
         }
+        if (tiisg < nb % N_SIMDWIDTH) {
+            sumf[row] += d * (acc[0] - 8.f*acc[1]);
+        }
+        qb_curr = qb_next;
 
-        sumf += d * (acc[0] - 8.f*acc[1]);
-    }
-
-    sum[ith] = sumf;
-
-    //
-    // Accumulate the sum from all threads in the threadgroup
-    //
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith%4 == 0) {
-        sum[ith] += sum[ith+1] + sum[ith+2] + sum[ith+3];
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith%16 == 0) {
-        sum[ith] += sum[ith+4] + sum[ith+8] + sum[ith+12];
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith == 0) {
-        for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
-        dst[r1*ne0 + r0] = sum[0];
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0 && ((r0 * N_SIMDGROUP + sgitg) * N_DST + row) < ne01) {
+            dst[r1*ne0 + (r0 * N_SIMDGROUP + sgitg) * N_DST + row] = all_sum;
+        }
     }
 }
 

data/ext/llama_cpp/src/ggml-mpi.c

@@ -0,0 +1,216 @@
+#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);
+}
+
+static 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;
+}
+
+static void ggml_mpi_tensor_send(struct ggml_tensor * t, int mpi_rank_dst) {
+    MPI_Datatype mpi_type;
+
+    switch (t->type) {
+        case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
+        case GGML_TYPE_F32: mpi_type = MPI_FLOAT;   break;
+        default: GGML_ASSERT(false && "not implemented");
+    }
+
+    const int retval = MPI_Send(t->data, ggml_nelements(t), mpi_type, mpi_rank_dst, 0, MPI_COMM_WORLD);
+    GGML_ASSERT(retval == MPI_SUCCESS);
+}
+
+static void ggml_mpi_tensor_recv(struct ggml_tensor * t, int mpi_rank_src) {
+    MPI_Datatype mpi_type;
+
+    switch (t->type) {
+        case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
+        case GGML_TYPE_F32: mpi_type = MPI_FLOAT;   break;
+        default: GGML_ASSERT(false && "not implemented");
+    }
+
+    MPI_Status status; UNUSED(status);
+
+    const int retval = MPI_Recv(t->data, ggml_nelements(t), mpi_type, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
+    GGML_ASSERT(retval == MPI_SUCCESS);
+}
+
+// TODO: there are many improvements that can be done to this implementation
+void ggml_mpi_graph_compute_pre(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers) {
+    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 (1) receives the input tokens from the main node (0)
+            ggml_mpi_tensor_recv(inp_tokens, 0);
+        } else {
+            // recv input data for each node into the "inp0" tensor (i.e. the first node in the compute graph)
+            ggml_mpi_tensor_recv(inp0, mpi_rank - 1);
+        }
+    } else if (mpi_size > 1) {
+        // node 0 sends the input tokens to node 1
+        ggml_mpi_tensor_send(inp_tokens, 1);
+
+        // recv the output data from the last node
+        ggml_mpi_tensor_recv(inp0, mpi_size - 1);
+    }
+
+    {
+        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]->src[0] == gf->nodes[idx_l0]) {
+                gf->nodes[i]->src[0] =  inp0;
+            }
+            if (gf->nodes[i]->src[1] == gf->nodes[idx_l0]) {
+                gf->nodes[i]->src[1] =  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);
+    }
+}
+
+void ggml_mpi_graph_compute_post(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers) {
+    UNUSED(n_layers);
+
+    const int mpi_rank = ctx_mpi->rank;
+    const int mpi_size = ctx_mpi->size;
+
+    // send the output data to the next node
+    if (mpi_rank > 0) {
+        ggml_mpi_tensor_send(gf->nodes[gf->n_nodes - 1], (mpi_rank + 1) % mpi_size);
+    }
+}

data/ext/llama_cpp/src/ggml-mpi.h

@@ -0,0 +1,39 @@
+#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_pre(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers);
+
+void ggml_mpi_graph_compute_post(
+        struct ggml_mpi_context * ctx_mpi,
+             struct ggml_cgraph * gf,
+                            int   n_layers);
+
+#ifdef __cplusplus
+}
+#endif