[ Web Proxy ]
URL:
Viewing: https://raw.githubusercontent.com/allozaur/llama.cpp/master/tests/test-alloc.cpp [Back]  [Original]

#include "ggml-alloc.h"
#include "../ggml/src/ggml-backend-impl.h"
#include "ggml-cpp.h"
#include "../ggml/src/ggml-impl.h"
#include "ggml.h"

#include 
#include 
#include 
#include 

//
// dummy backend with configurable max_buffer_size, tracks allocations

uint8_t * const alloc_base = (uint8_t *) 16;

struct dummy_backend_context {
    size_t max_buffer_size = 64;
    size_t alignment       = 8;

    ggml_backend_buffer_i              buffer_interface;
    ggml_backend_device                device;
    ggml_backend                       backend;
    std::vector buffers;

    size_t allocated_total() const {
        size_t n = 0;
        for (ggml_backend_buffer_t buf : buffers) {
            n += ggml_backend_buffer_get_size(buf);
        }
        return n;
    }
};

// ggml_backend_buffer_type interface

static const char * dummy_backend_buffer_type_get_name(ggml_backend_buffer_type_t) {
    return "dummy_buffer_type";
}

static ggml_backend_buffer_t dummy_backend_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
    dummy_backend_context * ctx    = (dummy_backend_context *) buft->context;
    ggml_backend_buffer_t & buffer = ctx->buffers.emplace_back();
    buffer                         = ggml_backend_buffer_init(buft, ctx->buffer_interface, ctx, size);
    return buffer;
}

static size_t dummy_backend_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
    dummy_backend_context * ctx = (dummy_backend_context *) buft->context;
    return ctx->alignment;
}

static size_t dummy_backend_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
    dummy_backend_context * ctx = (dummy_backend_context *) buft->context;
    return ctx->max_buffer_size;
}

static bool dummy_backend_buffer_type_is_host(ggml_backend_buffer_type_t) {
    return true;
}

// ggml_backend_buffer interface

static void dummy_backend_buffer_free_buffer(ggml_backend_buffer_t buffer) {
    dummy_backend_context * ctx = (dummy_backend_context *) buffer->context;

    auto i = std::find(ctx->buffers.begin(), ctx->buffers.end(), buffer);
    GGML_ASSERT(i != ctx->buffers.end());
    ctx->buffers.erase(i);
}

static void * dummy_backend_buffer_get_base(ggml_backend_buffer_t) {
    return alloc_base;
}

static ggml_status dummy_backend_buffer_init_tensor(ggml_backend_buffer_t, ggml_tensor *) {
    return GGML_STATUS_SUCCESS;
}

static void dummy_backend_buffer_memset_tensor(ggml_backend_buffer_t, ggml_tensor *, uint8_t, size_t, size_t) {}

static void dummy_backend_buffer_set_tensor(ggml_backend_buffer_t, ggml_tensor *, const void *, size_t, size_t) {}

static void dummy_backend_buffer_get_tensor(ggml_backend_buffer_t, const ggml_tensor *, void *, size_t, size_t) {}

static void dummy_backend_buffer_clear(ggml_backend_buffer_t, uint8_t) {}

// ggml_backend_device interface

static enum ggml_backend_dev_type dummy_backend_device_get_type(ggml_backend_dev_t) {
    return GGML_BACKEND_DEVICE_TYPE_CPU;
}

static bool dummy_backend_device_supports_op(ggml_backend_dev_t, const ggml_tensor *) {
    return true;
}

static bool dummy_backend_device_supports_buft(ggml_backend_dev_t device, ggml_backend_buffer_type_t buft) {
    return device->context == buft->context;
}

// ggml_backend interface

static const char * dummy_backend_get_name(ggml_backend_t) {
    return "dummy_backend";
}

// dummy_backend

struct dummy_backend {
    std::unique_ptr context;
    ggml_backend_buffer_type               buffer_type;
};

static dummy_backend dummy_backend_init(size_t max_buffer_size, size_t alignment = 8) {
    dummy_backend b{};
    b.context                  = std::make_unique();
    b.context->alignment       = alignment;
    b.context->max_buffer_size = max_buffer_size;

    b.context->buffer_interface.free_buffer   = dummy_backend_buffer_free_buffer;
    b.context->buffer_interface.get_base      = dummy_backend_buffer_get_base;
    b.context->buffer_interface.init_tensor   = dummy_backend_buffer_init_tensor;
    b.context->buffer_interface.memset_tensor = dummy_backend_buffer_memset_tensor;
    b.context->buffer_interface.set_tensor    = dummy_backend_buffer_set_tensor;
    b.context->buffer_interface.get_tensor    = dummy_backend_buffer_get_tensor;
    b.context->buffer_interface.clear         = dummy_backend_buffer_clear;

    b.context->device.context             = b.context.get();
    b.context->device.iface.get_type      = dummy_backend_device_get_type;
    b.context->device.iface.supports_op   = dummy_backend_device_supports_op;
    b.context->device.iface.supports_buft = dummy_backend_device_supports_buft;

    b.context->backend.context        = b.context.get();
    b.context->backend.device         = &b.context->device;
    b.context->backend.iface.get_name = dummy_backend_get_name;

    b.buffer_type.device              = &b.context->device;
    b.buffer_type.context             = b.context.get();
    b.buffer_type.iface.get_name      = dummy_backend_buffer_type_get_name;
    b.buffer_type.iface.alloc_buffer  = dummy_backend_buffer_type_alloc_buffer;
    b.buffer_type.iface.get_alignment = dummy_backend_buffer_type_get_alignment;
    b.buffer_type.iface.get_max_size  = dummy_backend_buffer_type_get_max_size;
    b.buffer_type.iface.is_host       = dummy_backend_buffer_type_is_host;
    return b;
}

//
// test utilities

struct test_context_with_graph {
    ggml_context *   ctx;
    ggml_cgraph *    graph;
    ggml_context_ptr ctx_ptr;
};

static test_context_with_graph make_context() {
    ggml_init_params params{};
    params.mem_size = 48 * ggml_tensor_overhead() + ggml_graph_overhead();
    params.no_alloc = true;

    ggml_context *   ctx     = ggml_init(params);
    ggml_context_ptr ctx_ptr = ggml_context_ptr(ctx);
    ggml_cgraph *    graph   = ggml_new_graph(ctx);
    return { ctx, graph, std::move(ctx_ptr) };
}

static ggml_tensor * make_input_1d(ggml_context * ctx, int64_t n_elements) {
    ggml_tensor * t = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
    ggml_set_input(t);
    return t;
}

static ggml_tensor * make_input_with_size(ggml_context * ctx, size_t size_bytes) {
    GGML_ASSERT(size_bytes % 4 == 0);
    return make_input_1d(ctx, size_bytes / 4);
}

static void assign_names(ggml_context * ctx, const char * prefix = "x") {
    int i = 0;
    for (ggml_tensor * t = ggml_get_first_tensor(ctx); t; t = ggml_get_next_tensor(ctx, t)) {
        ggml_format_name(t, "%s%d", prefix, i++);
    }
}

static int get_leaf_id(ggml_cgraph * graph, const char * tensor_name) {
    for (int i = 0; i < graph->n_leafs; ++i) {
        if (strncmp(graph->leafs[i]->name, tensor_name, GGML_MAX_NAME) == 0) {
            return i;
        }
    }
    fprintf(stderr, "leaf not found: %s\n", tensor_name);
    return -1;
}

static int get_node_id(ggml_cgraph * graph, const char * tensor_name) {
    for (int i = 0; i < graph->n_nodes; ++i) {
        if (strncmp(graph->nodes[i]->name, tensor_name, GGML_MAX_NAME) == 0) {
            return i;
        }
    }
    fprintf(stderr, "node not found: %s", tensor_name);
    return -1;
}

static ggml_gallocr_ptr allocate_graph(ggml_cgraph * graph, ggml_tensor * out, ggml_backend_buffer_type_t buft) {
    ggml_set_output(out);
    ggml_build_forward_expand(graph, out);

    ggml_gallocr_ptr galloc = ggml_gallocr_ptr(ggml_gallocr_new(buft));
    bool             result = ggml_gallocr_alloc_graph(galloc.get(), graph);
    GGML_ASSERT(result);
    return galloc;
}

//
// correctness checks for result allocations

static void check_all_allocated(ggml_cgraph * graph) {
    for (int i = 0; i < ggml_graph_n_nodes(graph); ++i) {
        ggml_tensor * t = ggml_graph_node(graph, i);
        GGML_ASSERT(t->buffer != nullptr);
        GGML_ASSERT(t->data != nullptr);
    }
}

static void check_max_size(ggml_context * ctx) {
    for (ggml_tensor * t = ggml_get_first_tensor(ctx); t; t = ggml_get_next_tensor(ctx, t)) {
        auto   buft     = ggml_backend_buffer_get_type(t->buffer);
        size_t max_size = ggml_backend_buft_get_max_size(buft);
        size_t offset   = (char *) t->data - (char *) ggml_backend_buffer_get_base(t->buffer);
        GGML_ASSERT(t->data >= ggml_backend_buffer_get_base(t->buffer));
        GGML_ASSERT((size_t) offset + ggml_nbytes(t) flags & GGML_TENSOR_FLAG_OUTPUT) {
        return false;
    }
    // Check if `other` is still "alive", ie. an input to any node after the `current` op
    for (int i = current_i; i < ggml_graph_n_nodes(graph); ++i) {
        ggml_tensor * t = ggml_graph_node(graph, i);
        for (int s = 0; s < GGML_MAX_SRC; s++) {
            if (t == current && ggml_op_can_inplace(t->op)) {
                continue;
            }
            if (t->src[s] == other) {
                return false;
            }
            if (t->src[s] && t->src[s]->view_src == other) {
                return false;
            }
        }
    }
    return true;
}

static bool memory_overlap(ggml_tensor * a, ggml_tensor * b) {
    if (a->buffer != b->buffer) {
        return false;
    }
    int64_t a0 = (int64_t) a->data;
    int64_t a1 = a0 + ggml_nbytes(a);
    int64_t b0 = (int64_t) b->data;
    int64_t b1 = b0 + ggml_nbytes(b);
    return a1 > b0 && b1 > a0;
}

static ggml_tensor * get_view_source(ggml_tensor * t) {
    while (t->view_src) {
        t = t->view_src;
    }
    return t;
}

static void check_no_overlap(ggml_cgraph * graph) {
    for (int i = 0; i < ggml_graph_n_nodes(graph); ++i) {
        for (int j = 0; j < i; ++j) {
            ggml_tensor * t = ggml_graph_node(graph, i);
            ggml_tensor * o = ggml_graph_node(graph, j);
            GGML_ASSERT(t != o);

            if (get_view_source(t) == get_view_source(o)) {
                continue;
            }
            if (memory_overlap(t, o)) {
                GGML_ASSERT(can_reuse_memory(graph, i, t, o));
            }
        }
    }
}

//
// test cases

// Scenario where the first backend buffer is completely exhausted and there are further
// tensors which require a second buffer
static void test_max_size_too_many_tensors() {
    dummy_backend backend      = dummy_backend_init(16);
    auto [ctx, graph, ctx_ptr] = make_context();

    ggml_tensor * x[7];
    x[0] = make_input_with_size(ctx, 8);
    x[1] = make_input_with_size(ctx, 8);
    x[2] = make_input_with_size(ctx, 8);
    x[3] = ggml_mul(ctx, x[0], x[1]);
    x[4] = ggml_add(ctx, x[1], x[2]);
    x[5] = ggml_add(ctx, x[3], x[0]);
    x[6] = ggml_add(ctx, x[4], x[5]);
    assign_names(ctx);

    ggml_gallocr_ptr galloc = allocate_graph(graph, x[6], &backend.buffer_type);
    check_all_allocated(graph);
    check_no_overlap(graph);
    check_max_size(ctx);
    GGML_ASSERT(backend.context->allocated_total() allocated_total() allocated_total() == 24);
}

// This test assumes a max of 16 buffer chunks, and tries to allocate tensors that would
// require more. Expectation is that the last buffer should grow to fit everything,
// leaving it to the backend to error out if it can't allocate that much.
static void test_not_enough_chunks() {
    const int max_chunks = 16;
    const int max_size   = 8;

    dummy_backend backend      = dummy_backend_init(max_size);
    auto [ctx, graph, ctx_ptr] = make_context();

    ggml_tensor * x[max_chunks + 1];
    for (int i = 0; i < max_chunks + 1; ++i) {
        x[i] = make_input_with_size(ctx, max_size);
    }
    ggml_tensor * acc = x[0];
    for (int i = 0; i < max_chunks; ++i) {
        acc = ggml_add(ctx, acc, x[i + 1]);
    }
    assign_names(ctx);

    ggml_gallocr_ptr galloc = allocate_graph(graph, acc, &backend.buffer_type);
    check_all_allocated(graph);
    check_no_overlap(graph);
    GGML_ASSERT(backend.context->allocated_total() > max_chunks * max_size);
}

// Fill up leftover unallocated space of a chunk after allocating a large tensor that
// requires a new chunk.
static void test_fill_leftover_space() {
    dummy_backend backend      = dummy_backend_init(16);
    auto [ctx, graph, ctx_ptr] = make_context();

    ggml_tensor * x[4];
    x[0] = make_input_with_size(ctx, 8);
    x[1] = ggml_pad(ctx, x[0], 2, 0, 0, 0);
    x[3] = ggml_mean(ctx, x[1]);
    assign_names(ctx);

    ggml_gallocr_ptr galloc = allocate_graph(graph, x[3], &backend.buffer_type);
    check_all_allocated(graph);
    check_no_overlap(graph);
    check_max_size(ctx);
    GGML_ASSERT(backend.context->allocated_total() 

Web Proxy Viewer  |  New URL  |  Original Page